From e75ba4c0434eb759eb7ff74e034ebe729053e575 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= <sigbjorn.skjaeret@scala.com>
Date: Wed, 2 Jul 2025 21:02:35 +0200
Subject: [PATCH 01/10] gguf-py : add support for chat template jinja files
 (#14508)

* add support for chat template jinja files

* remove gemma3n hack
---
 convert_hf_to_gguf.py |  3 ---
 gguf-py/gguf/vocab.py | 15 ++++++++++++---
 2 files changed, 12 insertions(+), 6 deletions(-)

diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index bd9cd81449..dd80a4a05d 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -4408,9 +4408,6 @@ class Gemma3NModel(Gemma3Model):
         ]
 
     def set_vocab(self):
-        with open(self.dir_model / "chat_template.jinja") as f:
-            # quick hack to make sure chat template is added
-            self.gguf_writer.add_chat_template(f.read())
         super().set_vocab()
 
     def set_gguf_parameters(self):
diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py
index 3f541b0c02..635fcef35e 100644
--- a/gguf-py/gguf/vocab.py
+++ b/gguf-py/gguf/vocab.py
@@ -245,9 +245,18 @@ class SpecialVocab:
         if not tokenizer_config:
             return True
         chat_template_alt = None
-        chat_template_file = path / 'chat_template.json'
-        if chat_template_file.is_file():
-            with open(chat_template_file, encoding = 'utf-8') as f:
+        chat_template_json = path / 'chat_template.json'
+        chat_template_jinja = path / 'chat_template.jinja'
+        if chat_template_jinja.is_file():
+            with open(chat_template_jinja, encoding = 'utf-8') as f:
+                chat_template_alt = f.read()
+            if additional_templates := list((path / 'additional_chat_templates').glob('*.jinja')):
+                chat_template_alt = [{'name': 'default', 'template': chat_template_alt}]
+                for template_path in additional_templates:
+                    with open(template_path, encoding = 'utf-8') as fp:
+                        chat_template_alt.append({'name': template_path.stem, 'template': fp.read()})
+        elif chat_template_json.is_file():
+            with open(chat_template_json, encoding = 'utf-8') as f:
                 chat_template_alt = json.load(f).get('chat_template')
         chat_template = tokenizer_config.get('chat_template', chat_template_alt)
         if chat_template is None or isinstance(chat_template, (str, list)):

From 55c2646b452419be7552157b9b11373ccc1bf2f2 Mon Sep 17 00:00:00 2001
From: Aman Gupta <amangupta052@gmail.com>
Date: Thu, 3 Jul 2025 07:45:11 +0800
Subject: [PATCH 02/10] CUDA: add dynamic shared mem to softmax, refactor
 general usage (#14497)

---
 ggml/src/ggml-cuda/common.cuh            | 14 +++++
 ggml/src/ggml-cuda/cross-entropy-loss.cu | 16 +----
 ggml/src/ggml-cuda/mmq.cuh               | 10 +--
 ggml/src/ggml-cuda/softmax.cu            | 78 ++++++++++++------------
 tests/test-backend-ops.cpp               |  1 +
 5 files changed, 57 insertions(+), 62 deletions(-)

diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index ea20355023..954f74d408 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -175,6 +175,20 @@ static const char * cu_get_error_str(CUresult err) {
 #define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str)
 #endif
 
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#define CUDA_SET_SHARED_MEMORY_LIMIT(kernel, nbytes) \
+    do { \
+        static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false}; \
+        const int id = ggml_cuda_get_device(); \
+        if (!shared_memory_limit_raised[id]) { \
+            CUDA_CHECK(cudaFuncSetAttribute(kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes)); \
+            shared_memory_limit_raised[id] = true; \
+        } \
+    } while (0)
+#else
+#define CUDA_SET_SHARED_MEMORY_LIMIT(kernel, nbytes) do {} while (0)
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+
 #if CUDART_VERSION >= 11010 || defined(GGML_USE_MUSA)
 #define GGML_CUDA_ASSUME(x) __builtin_assume(x)
 #else
diff --git a/ggml/src/ggml-cuda/cross-entropy-loss.cu b/ggml/src/ggml-cuda/cross-entropy-loss.cu
index 0ce4afbb22..0c8b081972 100644
--- a/ggml/src/ggml-cuda/cross-entropy-loss.cu
+++ b/ggml/src/ggml-cuda/cross-entropy-loss.cu
@@ -123,13 +123,7 @@ void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor *
     ggml_cuda_pool_alloc<float> dst_tmp(pool, blocks_num.x);
 
     if (nbytes_shared <= smpbo) {
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
-        static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
-        if (!shared_memory_limit_raised[id]) {
-            CUDA_CHECK(cudaFuncSetAttribute(cross_entropy_loss_f32<true>, cudaFuncAttributeMaxDynamicSharedMemorySize, smpbo));
-            shared_memory_limit_raised[id] = true;
-        }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+        CUDA_SET_SHARED_MEMORY_LIMIT((cross_entropy_loss_f32<true>), smpbo);
         cross_entropy_loss_f32<true><<<blocks_num, blocks_dim, nbytes_shared, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
     } else {
         cross_entropy_loss_f32<false><<<blocks_num, blocks_dim, 0, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
@@ -175,13 +169,7 @@ void ggml_cuda_cross_entropy_loss_back(ggml_backend_cuda_context & ctx, ggml_ten
     const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
 
     if (nbytes_shared <= smpbo) {
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
-        static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
-        if (!shared_memory_limit_raised[id]) {
-            CUDA_CHECK(cudaFuncSetAttribute(cross_entropy_loss_back_f32<true>, cudaFuncAttributeMaxDynamicSharedMemorySize, smpbo));
-            shared_memory_limit_raised[id] = true;
-        }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+        CUDA_SET_SHARED_MEMORY_LIMIT((cross_entropy_loss_back_f32<true>), smpbo);
         cross_entropy_loss_back_f32<true><<<blocks_num, blocks_dim, nbytes_shared, stream>>>(grad_d, src0f_d, src1f_d, dst_d, ne00);
     } else {
         cross_entropy_loss_back_f32<false><<<blocks_num, blocks_dim, 0, stream>>>(grad_d, src0f_d, src1f_d, dst_d, ne00);
diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
index 80baf459c1..9696a32046 100644
--- a/ggml/src/ggml-cuda/mmq.cuh
+++ b/ggml/src/ggml-cuda/mmq.cuh
@@ -3016,14 +3016,8 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 
     const int nbytes_shared = mmq_get_nbytes_shared<type>(mmq_x, mmq_y, cc);
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
-    static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
-    if (!shared_memory_limit_raised[id]) {
-        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared));
-        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>,  cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared));
-        shared_memory_limit_raised[id] = true;
-    }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+    CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q<type, mmq_x, MMQ_NWARPS, false>), nbytes_shared);
+    CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q<type, mmq_x, MMQ_NWARPS, true>),  nbytes_shared);
 
     const int nty  = (args.nrows_x   + mmq_y - 1) / mmq_y;
     const int ntx  = (args.ncols_dst + mmq_x - 1) / mmq_x;
diff --git a/ggml/src/ggml-cuda/softmax.cu b/ggml/src/ggml-cuda/softmax.cu
index e0eb921d85..14543e978c 100644
--- a/ggml/src/ggml-cuda/softmax.cu
+++ b/ggml/src/ggml-cuda/softmax.cu
@@ -2,6 +2,7 @@
 #include "ggml.h"
 #include "softmax.cuh"
 #include <cstdint>
+#include <utility>
 
 template <typename T>
 static __device__ __forceinline__ float t2f32(T val) {
@@ -181,6 +182,37 @@ static __global__ void soft_max_back_f32(
     }
 }
 
+template<int... Ns, typename T>
+static void launch_soft_max_kernels(const float * x, const T * mask, float * dst,
+                             const soft_max_params & p, cudaStream_t stream, dim3 block_dims, dim3 block_nums, size_t nbytes_shared)
+{
+    const int id       = ggml_cuda_get_device();
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
+
+    auto launch_kernel = [=](auto I) -> bool {
+        constexpr int ncols = decltype(I)::value;
+        constexpr int block = (ncols > 1024 ? 1024 : ncols);
+
+        if (p.ncols == ncols) {
+            CUDA_SET_SHARED_MEMORY_LIMIT((soft_max_f32<true, ncols, block, T>), smpbo);
+            soft_max_f32<true, ncols, block><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, mask, dst, p);
+            return true;
+        }
+        return false;
+    };
+
+    // unary fold over launch_kernel
+    if ((launch_kernel(std::integral_constant<int, Ns>{}) || ...)) {
+        return;
+    }
+
+    //default case
+    CUDA_SET_SHARED_MEMORY_LIMIT((soft_max_f32<true, 0, 0, T>), smpbo);
+    soft_max_f32<true, 0, 0><<<block_nums, block_dims, nbytes_shared, stream>>>(x, mask, dst, p);
+}
+
+
 template<typename T>
 static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const soft_max_params & params, cudaStream_t stream) {
     int nth = WARP_SIZE;
@@ -193,46 +225,12 @@ static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, cons
     static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
 
 
-    // FIXME: this limit could be raised by ~2-4x on Ampere or newer
-    if (nbytes_shared < ggml_cuda_info().devices[ggml_cuda_get_device()].smpb) {
-        switch (ncols_x) {
-            case 32:
-                soft_max_f32<true,   32,   32><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, params);
-                break;
-            case 64:
-                soft_max_f32<true,   64,   64><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, params);
-                break;
-            case 128:
-                soft_max_f32<true,  128,  128><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, params);
-                break;
-            case 256:
-                soft_max_f32<true,  256,  256><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, params);
-                break;
-            case 512:
-                soft_max_f32<true,  512,  512><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, params);
-                break;
-            case 1024:
-                soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, params);
-                break;
-            case 2048:
-                soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, params);
-                break;
-            case 4096:
-                soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, params);
-                break;
-            default:
-                soft_max_f32<true,    0,    0><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, params);
-                break;
-        }
+    const int id       = ggml_cuda_get_device();
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
+
+
+    if (nbytes_shared <= smpbo) {
+        launch_soft_max_kernels<32, 64, 128, 256, 512, 1024, 2048, 4096>(x, mask, dst, params, stream, block_dims, block_nums, nbytes_shared);
     } else {
         const size_t nbytes_shared_low = WARP_SIZE*sizeof(float);
         soft_max_f32<false, 0, 0><<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, dst, params);
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 51bbcf92c9..5c3db854cb 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -4932,6 +4932,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {3072, 512, 2, 1}, {0, 2, 1, 3}));
 
     test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {4096, 4096, 5, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {12888, 256, 5, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
     test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 4096, 5, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
     test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {1024, 1024, 10, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
     test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 1024, 10, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));

From d4cdd9c1c3cebdafca735958597de4ff7b7c0f54 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Thu, 3 Jul 2025 07:48:32 +0300
Subject: [PATCH 03/10] ggml : remove kompute backend (#14501)

ggml-ci
---
 .../ISSUE_TEMPLATE/010-bug-compilation.yml    |    2 +-
 .github/ISSUE_TEMPLATE/011-bug-results.yml    |    2 +-
 .github/labeler.yml                           |    6 -
 .github/workflows/build.yml                   |   11 +-
 .gitmodules                                   |    3 -
 CMakeLists.txt                                |    1 -
 ggml/CMakeLists.txt                           |    2 -
 ggml/include/ggml-kompute.h                   |   50 -
 ggml/src/CMakeLists.txt                       |    1 -
 ggml/src/ggml-backend-reg.cpp                 |    8 -
 ggml/src/ggml-kompute/CMakeLists.txt          |  166 --
 ggml/src/ggml-kompute/ggml-kompute.cpp        | 2251 -----------------
 ggml/src/ggml-kompute/kompute                 |    1 -
 .../ggml-kompute/kompute-shaders/common.comp  |  112 -
 .../ggml-kompute/kompute-shaders/op_add.comp  |   58 -
 .../kompute-shaders/op_addrow.comp            |   25 -
 .../kompute-shaders/op_cpy_f16_f16.comp       |   52 -
 .../kompute-shaders/op_cpy_f16_f32.comp       |   52 -
 .../kompute-shaders/op_cpy_f32_f16.comp       |   52 -
 .../kompute-shaders/op_cpy_f32_f32.comp       |   52 -
 .../kompute-shaders/op_diagmask.comp          |   30 -
 .../ggml-kompute/kompute-shaders/op_gelu.comp |   22 -
 .../kompute-shaders/op_getrows.comp           |   17 -
 .../kompute-shaders/op_getrows_f16.comp       |   31 -
 .../kompute-shaders/op_getrows_f32.comp       |   31 -
 .../kompute-shaders/op_getrows_q4_0.comp      |   38 -
 .../kompute-shaders/op_getrows_q4_1.comp      |   39 -
 .../kompute-shaders/op_getrows_q6_k.comp      |   44 -
 .../ggml-kompute/kompute-shaders/op_mul.comp  |   52 -
 .../kompute-shaders/op_mul_mat_f16.comp       |   69 -
 .../kompute-shaders/op_mul_mat_mat_f32.comp   |   51 -
 .../kompute-shaders/op_mul_mat_q4_0.comp      |   33 -
 .../kompute-shaders/op_mul_mat_q4_1.comp      |   35 -
 .../kompute-shaders/op_mul_mat_q4_k.comp      |  140 -
 .../kompute-shaders/op_mul_mat_q6_k.comp      |  106 -
 .../kompute-shaders/op_mul_mat_q8_0.comp      |   73 -
 .../kompute-shaders/op_mul_mv_q_n.comp        |   52 -
 .../kompute-shaders/op_mul_mv_q_n_pre.comp    |   28 -
 .../ggml-kompute/kompute-shaders/op_norm.comp |   84 -
 .../ggml-kompute/kompute-shaders/op_relu.comp |   21 -
 .../kompute-shaders/op_rmsnorm.comp           |   53 -
 .../kompute-shaders/op_rope_neox_f16.comp     |   52 -
 .../kompute-shaders/op_rope_neox_f32.comp     |   52 -
 .../kompute-shaders/op_rope_norm_f16.comp     |   52 -
 .../kompute-shaders/op_rope_norm_f32.comp     |   52 -
 .../kompute-shaders/op_scale.comp             |   19 -
 .../kompute-shaders/op_scale_8.comp           |   23 -
 .../ggml-kompute/kompute-shaders/op_silu.comp |   22 -
 .../kompute-shaders/op_softmax.comp           |   72 -
 .../kompute-shaders/rope_common.comp          |   71 -
 scripts/sync-ggml-am.sh                       |    3 -
 scripts/sync-ggml.sh                          |    1 -
 tests/test-c.c                                |    4 -
 53 files changed, 3 insertions(+), 4376 deletions(-)
 delete mode 100644 ggml/include/ggml-kompute.h
 delete mode 100644 ggml/src/ggml-kompute/CMakeLists.txt
 delete mode 100644 ggml/src/ggml-kompute/ggml-kompute.cpp
 delete mode 160000 ggml/src/ggml-kompute/kompute
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/common.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_add.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_mul.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_norm.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_relu.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_scale.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_silu.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp
 delete mode 100644 ggml/src/ggml-kompute/kompute-shaders/rope_common.comp

diff --git a/.github/ISSUE_TEMPLATE/010-bug-compilation.yml b/.github/ISSUE_TEMPLATE/010-bug-compilation.yml
index d538fdff10..95a0b5cc75 100644
--- a/.github/ISSUE_TEMPLATE/010-bug-compilation.yml
+++ b/.github/ISSUE_TEMPLATE/010-bug-compilation.yml
@@ -40,7 +40,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
         multiple: true
     validations:
       required: true
diff --git a/.github/ISSUE_TEMPLATE/011-bug-results.yml b/.github/ISSUE_TEMPLATE/011-bug-results.yml
index 2acec7114f..d1034bbb69 100644
--- a/.github/ISSUE_TEMPLATE/011-bug-results.yml
+++ b/.github/ISSUE_TEMPLATE/011-bug-results.yml
@@ -42,7 +42,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
         multiple: true
     validations:
       required: true
diff --git a/.github/labeler.yml b/.github/labeler.yml
index a4c6e3d33a..df6a7a40ed 100644
--- a/.github/labeler.yml
+++ b/.github/labeler.yml
@@ -1,10 +1,4 @@
 # https://github.com/actions/labeler
-Kompute:
-    - changed-files:
-        - any-glob-to-any-file:
-            - ggml/include/ggml-kompute.h
-            - ggml/src/ggml-kompute/**
-            - README-kompute.md
 Apple Metal:
     - changed-files:
         - any-glob-to-any-file:
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 5d4fb5272e..42d63b7c54 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -740,9 +740,6 @@ jobs:
           - build: 'llvm-arm64-opencl-adreno'
             arch: 'arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
-         # - build: 'kompute-x64'
-         #   arch: 'x64'
-         #   defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'
 
     steps:
       - name: Clone
@@ -756,12 +753,6 @@ jobs:
           variant: ccache
           evict-old-files: 1d
 
-      - name: Clone Kompute submodule
-        id: clone_kompute
-        if: ${{ matrix.build == 'kompute-x64' }}
-        run: |
-          git submodule update --init ggml/src/ggml-kompute/kompute
-
       - name: Download OpenBLAS
         id: get_openblas
         if: ${{ matrix.build == 'openblas-x64' }}
@@ -777,7 +768,7 @@ jobs:
 
       - name: Install Vulkan SDK
         id: get_vulkan
-        if: ${{ matrix.build == 'kompute-x64' || matrix.build == 'vulkan-x64' }}
+        if: ${{ matrix.build == 'vulkan-x64' }}
         run: |
           curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
           & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
diff --git a/.gitmodules b/.gitmodules
index 23ce5ff059..e69de29bb2 100644
--- a/.gitmodules
+++ b/.gitmodules
@@ -1,3 +0,0 @@
-[submodule "kompute"]
-	path = ggml/src/ggml-kompute/kompute
-	url = https://github.com/nomic-ai/kompute.git
diff --git a/CMakeLists.txt b/CMakeLists.txt
index d2becb04c6..c79ccd09e0 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -120,7 +120,6 @@ endfunction()
 
 llama_option_depr(FATAL_ERROR LLAMA_CUBLAS              GGML_CUDA)
 llama_option_depr(WARNING     LLAMA_CUDA                GGML_CUDA)
-llama_option_depr(WARNING     LLAMA_KOMPUTE             GGML_KOMPUTE)
 llama_option_depr(WARNING     LLAMA_METAL               GGML_METAL)
 llama_option_depr(WARNING     LLAMA_METAL_EMBED_LIBRARY GGML_METAL_EMBED_LIBRARY)
 llama_option_depr(WARNING     LLAMA_NATIVE              GGML_NATIVE)
diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index 675f3bf756..eaba9c7046 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -181,7 +181,6 @@ option(GGML_VULKAN_MEMORY_DEBUG             "ggml: enable Vulkan memory debug ou
 option(GGML_VULKAN_SHADER_DEBUG_INFO        "ggml: enable Vulkan shader debug info"           OFF)
 option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"                  OFF)
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
-option(GGML_KOMPUTE                         "ggml: use Kompute"                               OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
 option(GGML_METAL_USE_BF16                  "ggml: use bfloat if available"                   OFF)
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
@@ -266,7 +265,6 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-cann.h
     include/ggml-cpp.h
     include/ggml-cuda.h
-    include/ggml-kompute.h
     include/ggml-opt.h
     include/ggml-metal.h
     include/ggml-rpc.h
diff --git a/ggml/include/ggml-kompute.h b/ggml/include/ggml-kompute.h
deleted file mode 100644
index 154aa56a74..0000000000
--- a/ggml/include/ggml-kompute.h
+++ /dev/null
@@ -1,50 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#include <stdbool.h>
-#include <stddef.h>
-#include <stdint.h>
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#define GGML_KOMPUTE_MAX_DEVICES 16
-
-struct ggml_vk_device {
-    int index;
-    int type; // same as VkPhysicalDeviceType
-    size_t heapSize;
-    const char * name;
-    const char * vendor;
-    int subgroupSize;
-    uint64_t bufferAlignment;
-    uint64_t maxAlloc;
-};
-
-struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count);
-bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name);
-bool ggml_vk_has_vulkan(void);
-bool ggml_vk_has_device(void);
-struct ggml_vk_device ggml_vk_current_device(void);
-
-//
-// backend API
-//
-
-// forward declaration
-typedef struct ggml_backend * ggml_backend_t;
-
-GGML_BACKEND_API ggml_backend_t ggml_backend_kompute_init(int device);
-
-GGML_BACKEND_API bool ggml_backend_is_kompute(ggml_backend_t backend);
-
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
-
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
-
-#ifdef __cplusplus
-}
-#endif
diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
index 9cb2c228dc..8760c2d35e 100644
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -365,7 +365,6 @@ ggml_add_backend(BLAS)
 ggml_add_backend(CANN)
 ggml_add_backend(CUDA)
 ggml_add_backend(HIP)
-ggml_add_backend(Kompute)
 ggml_add_backend(METAL)
 ggml_add_backend(MUSA)
 ggml_add_backend(RPC)
diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
index 2d93771fd1..042ea77aca 100644
--- a/ggml/src/ggml-backend-reg.cpp
+++ b/ggml/src/ggml-backend-reg.cpp
@@ -61,10 +61,6 @@
 #include "ggml-cann.h"
 #endif
 
-#ifdef GGML_USE_KOMPUTE
-#include "ggml-kompute.h"
-#endif
-
 // disable C++17 deprecation warning for std::codecvt_utf8
 #if defined(__clang__)
 #    pragma clang diagnostic push
@@ -189,9 +185,6 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_RPC
         register_backend(ggml_backend_rpc_reg());
 #endif
-#ifdef GGML_USE_KOMPUTE
-        register_backend(ggml_backend_kompute_reg());
-#endif
 #ifdef GGML_USE_CPU
         register_backend(ggml_backend_cpu_reg());
 #endif
@@ -575,7 +568,6 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
     ggml_backend_load_best("cann", silent, dir_path);
     ggml_backend_load_best("cuda", silent, dir_path);
     ggml_backend_load_best("hip", silent, dir_path);
-    ggml_backend_load_best("kompute", silent, dir_path);
     ggml_backend_load_best("metal", silent, dir_path);
     ggml_backend_load_best("rpc", silent, dir_path);
     ggml_backend_load_best("sycl", silent, dir_path);
diff --git a/ggml/src/ggml-kompute/CMakeLists.txt b/ggml/src/ggml-kompute/CMakeLists.txt
deleted file mode 100644
index c9109d5e8e..0000000000
--- a/ggml/src/ggml-kompute/CMakeLists.txt
+++ /dev/null
@@ -1,166 +0,0 @@
-
-find_package(Vulkan COMPONENTS glslc REQUIRED)
-find_program(glslc_executable NAMES glslc HINTS Vulkan::glslc)
-
-if (NOT glslc_executable)
-    message(FATAL_ERROR "glslc not found")
-endif()
-
-ggml_add_backend_library(ggml-kompute
-                         ggml-kompute.cpp
-                         ../../include/ggml-kompute.h
-                        )
-
-target_link_libraries(ggml-kompute PRIVATE ggml-base kompute)
-target_include_directories(ggml-kompute PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
-
-add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
-
-function(compile_shader)
-    set(options)
-    set(oneValueArgs)
-    set(multiValueArgs SOURCES)
-    cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-    foreach(source ${compile_shader_SOURCES})
-        get_filename_component(filename ${source} NAME)
-        set(spv_file ${filename}.spv)
-        add_custom_command(
-            OUTPUT ${spv_file}
-            DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
-            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
-            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
-            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
-            COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-            COMMENT "Compiling ${source} to ${spv_file}"
-            )
-
-        get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
-        set(FILE_NAME "shader${RAW_FILE_NAME}")
-        string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
-        string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
-        string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
-        set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
-        message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
-        if(CMAKE_GENERATOR MATCHES "Visual Studio")
-            add_custom_command(
-                OUTPUT ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                DEPENDS ${spv_file} xxd
-                COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
-                )
-        else()
-            add_custom_command(
-                OUTPUT ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                DEPENDS ${spv_file} xxd
-                COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
-                )
-        endif()
-    endforeach()
-endfunction()
-
-if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
-    message(STATUS "Kompute found")
-    set(KOMPUTE_OPT_LOG_LEVEL Error CACHE STRING "Kompute log level")
-    add_subdirectory(kompute)
-
-    # Compile our shaders
-    compile_shader(SOURCES
-        kompute-shaders/op_scale.comp
-        kompute-shaders/op_scale_8.comp
-        kompute-shaders/op_add.comp
-        kompute-shaders/op_addrow.comp
-        kompute-shaders/op_mul.comp
-        kompute-shaders/op_silu.comp
-        kompute-shaders/op_relu.comp
-        kompute-shaders/op_gelu.comp
-        kompute-shaders/op_softmax.comp
-        kompute-shaders/op_norm.comp
-        kompute-shaders/op_rmsnorm.comp
-        kompute-shaders/op_diagmask.comp
-        kompute-shaders/op_mul_mat_mat_f32.comp
-        kompute-shaders/op_mul_mat_f16.comp
-        kompute-shaders/op_mul_mat_q8_0.comp
-        kompute-shaders/op_mul_mat_q4_0.comp
-        kompute-shaders/op_mul_mat_q4_1.comp
-        kompute-shaders/op_mul_mat_q4_k.comp
-        kompute-shaders/op_mul_mat_q6_k.comp
-        kompute-shaders/op_getrows_f32.comp
-        kompute-shaders/op_getrows_f16.comp
-        kompute-shaders/op_getrows_q4_0.comp
-        kompute-shaders/op_getrows_q4_1.comp
-        kompute-shaders/op_getrows_q6_k.comp
-        kompute-shaders/op_rope_norm_f16.comp
-        kompute-shaders/op_rope_norm_f32.comp
-        kompute-shaders/op_rope_neox_f16.comp
-        kompute-shaders/op_rope_neox_f32.comp
-        kompute-shaders/op_cpy_f16_f16.comp
-        kompute-shaders/op_cpy_f16_f32.comp
-        kompute-shaders/op_cpy_f32_f16.comp
-        kompute-shaders/op_cpy_f32_f32.comp
-    )
-
-    # Create a custom target for our generated shaders
-    add_custom_target(generated_shaders DEPENDS
-        shaderop_scale.h
-        shaderop_scale_8.h
-        shaderop_add.h
-        shaderop_addrow.h
-        shaderop_mul.h
-        shaderop_silu.h
-        shaderop_relu.h
-        shaderop_gelu.h
-        shaderop_softmax.h
-        shaderop_norm.h
-        shaderop_rmsnorm.h
-        shaderop_diagmask.h
-        shaderop_mul_mat_mat_f32.h
-        shaderop_mul_mat_f16.h
-        shaderop_mul_mat_q8_0.h
-        shaderop_mul_mat_q4_0.h
-        shaderop_mul_mat_q4_1.h
-        shaderop_mul_mat_q4_k.h
-        shaderop_mul_mat_q6_k.h
-        shaderop_getrows_f32.h
-        shaderop_getrows_f16.h
-        shaderop_getrows_q4_0.h
-        shaderop_getrows_q4_1.h
-        shaderop_getrows_q6_k.h
-        shaderop_rope_norm_f16.h
-        shaderop_rope_norm_f32.h
-        shaderop_rope_neox_f16.h
-        shaderop_rope_neox_f32.h
-        shaderop_cpy_f16_f16.h
-        shaderop_cpy_f16_f32.h
-        shaderop_cpy_f32_f16.h
-        shaderop_cpy_f32_f32.h
-    )
-
-    # Create a custom command that depends on the generated_shaders
-    add_custom_command(
-        OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
-        COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
-        DEPENDS generated_shaders
-        COMMENT "Ensuring shaders are generated before compiling ggml-kompute.cpp"
-    )
-
-    # Add the stamp to the main sources to ensure dependency tracking
-    target_sources(ggml-kompute PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
-else()
-    message(WARNING "Kompute not found")
-endif()
diff --git a/ggml/src/ggml-kompute/ggml-kompute.cpp b/ggml/src/ggml-kompute/ggml-kompute.cpp
deleted file mode 100644
index 5057922718..0000000000
--- a/ggml/src/ggml-kompute/ggml-kompute.cpp
+++ /dev/null
@@ -1,2251 +0,0 @@
-#include "ggml-impl.h"
-#include "ggml-backend.h"
-#include "ggml-backend-impl.h"
-#include "ggml-kompute.h"
-
-// These are generated at build time by cmake custom command
-#include "shaderop_scale.h"
-#include "shaderop_scale_8.h"
-#include "shaderop_add.h"
-#include "shaderop_addrow.h"
-#include "shaderop_mul.h"
-#include "shaderop_silu.h"
-#include "shaderop_relu.h"
-#include "shaderop_gelu.h"
-#include "shaderop_softmax.h"
-#include "shaderop_norm.h"
-#include "shaderop_rmsnorm.h"
-#include "shaderop_diagmask.h"
-#include "shaderop_mul_mat_f16.h"
-#include "shaderop_mul_mat_q8_0.h"
-#include "shaderop_mul_mat_q4_0.h"
-#include "shaderop_mul_mat_q4_1.h"
-#include "shaderop_mul_mat_q4_k.h"
-#include "shaderop_mul_mat_q6_k.h"
-#include "shaderop_mul_mat_mat_f32.h"
-#include "shaderop_getrows_f32.h"
-#include "shaderop_getrows_f16.h"
-#include "shaderop_getrows_q4_0.h"
-#include "shaderop_getrows_q4_1.h"
-#include "shaderop_getrows_q6_k.h"
-#include "shaderop_rope_norm_f16.h"
-#include "shaderop_rope_norm_f32.h"
-#include "shaderop_rope_neox_f16.h"
-#include "shaderop_rope_neox_f32.h"
-#include "shaderop_cpy_f16_f16.h"
-#include "shaderop_cpy_f16_f32.h"
-#include "shaderop_cpy_f32_f16.h"
-#include "shaderop_cpy_f32_f32.h"
-
-#include <algorithm>
-#include <array>
-#include <cassert>
-#include <cstdint>
-#include <cstdio>
-#include <cstring>
-#include <iostream>
-#include <memory>
-#include <mutex>
-#include <stdexcept>
-#include <string>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <kompute/Kompute.hpp>
-#include <vulkan/vulkan.hpp>
-
-#ifdef __linux__
-#include <cstdlib> // for setenv
-#endif
-
-#define QK4_0 32
-#define QR4_0 2
-#define QK4_1 32
-#define QK_NL 16
-
-typedef ggml_fp16_t half;
-
-static std::string ggml_kompute_format_name(int device) {
-    return "Kompute" + std::to_string(device);
-}
-
-struct ggml_kompute_context {
-    int device;
-    std::string name;
-    std::shared_ptr<vk::DescriptorPool> pool;
-
-    ggml_kompute_context(int device)
-        : device(device), name(ggml_kompute_format_name(device)) {}
-};
-
-// FIXME: It would be good to consolidate the kompute manager and the kompute context into one object
-// and consolidate the init functions and simplify object lifetime management. As it currently stands,
-// we *have* to have the kompute manager no matter what for device discovery, but the kompute context
-// is only created when a device is set and vulkan is explicitly turned on.
-static ggml_kompute_context *s_kompute_context = nullptr;
-
-class kompute_manager {
-    kp::Manager *s_mgr = nullptr;
-
-public:
-    kp::Manager *operator()() {
-        if (s_mgr && !s_mgr->hasInstance()) {
-            destroy();
-        }
-        if (!s_mgr) {
-            s_mgr = new kp::Manager;
-        }
-        return s_mgr;
-    }
-
-    void destroy() {
-        delete s_mgr;
-        s_mgr = nullptr;
-    }
-};
-
-static kompute_manager komputeManager;
-
-struct ggml_vk_memory {
-    void *data = nullptr;
-    size_t size = 0;
-    vk::DeviceMemory *primaryMemory = nullptr;
-    vk::Buffer *primaryBuffer = nullptr;
-    vk::DeviceMemory *stagingMemory = nullptr;
-    vk::Buffer *stagingBuffer = nullptr;
-};
-
-#ifdef __linux__
-__attribute__((constructor))
-static void enable_sam() {
-    setenv("RADV_PERFTEST", "sam", false);
-}
-#endif
-
-static bool ggml_vk_checkPhysicalDeviceFeatures(vk::PhysicalDevice physical_device) {
-    vk::PhysicalDeviceFeatures availableFeatures;
-    physical_device.getFeatures(&availableFeatures);
-
-    if (!availableFeatures.shaderInt16)
-        return false;
-
-    vk::PhysicalDeviceVulkan11Features availableFeatures11;
-    vk::PhysicalDeviceVulkan12Features availableFeatures12;
-
-    availableFeatures11.pNext = &availableFeatures12;
-    availableFeatures12.pNext = nullptr;
-
-    vk::PhysicalDeviceFeatures2 features2;
-    features2.pNext = &availableFeatures11;
-
-    physical_device.getFeatures2(&features2);
-
-    if (!availableFeatures11.uniformAndStorageBuffer16BitAccess ||
-        !availableFeatures11.storageBuffer16BitAccess) {
-        return false;
-    }
-
-    if (!availableFeatures12.storageBuffer8BitAccess ||
-        !availableFeatures12.uniformAndStorageBuffer8BitAccess ||
-        !availableFeatures12.shaderFloat16 ||
-        !availableFeatures12.shaderInt8) {
-        return false;
-    }
-
-    return true;
-}
-
-static const char * ggml_vk_getVendorName(uint32_t vendorID) {
-    switch (vendorID) {
-        case 0x10DE:
-            return "nvidia";
-        case 0x1002:
-            return "amd";
-        case 0x8086:
-            return "intel";
-        default:
-            return "unknown";
-    }
-}
-
-static std::vector<ggml_vk_device> ggml_vk_available_devices_internal(size_t memoryRequired) {
-    std::vector<ggml_vk_device> results;
-    if (!komputeManager()->hasVulkan() || !komputeManager()->hasInstance())
-        return results;
-
-    std::vector<vk::PhysicalDevice> physical_devices;
-    try {
-        physical_devices = komputeManager()->listDevices();
-    } catch (vk::SystemError & err) {
-        std::cerr << __func__ << ": ignoring Vulkan exception: " << err.what() << "\n";
-        return results;
-    }
-
-    uint32_t deviceCount = physical_devices.size();
-    if (deviceCount == 0)
-        return results;
-
-    std::unordered_map<std::string, size_t> count_by_name;
-
-    for (uint32_t i = 0; i < deviceCount; i++) {
-        const auto & physical_device = physical_devices[i];
-
-        VkPhysicalDeviceProperties dev_props = physical_device.getProperties();
-        VkPhysicalDeviceMemoryProperties memoryProperties = physical_device.getMemoryProperties();
-        const uint32_t major = VK_VERSION_MAJOR(dev_props.apiVersion);
-        const uint32_t minor = VK_VERSION_MINOR(dev_props.apiVersion);
-        if (major < 1 || minor < 2)
-            continue;
-
-        if (!ggml_vk_checkPhysicalDeviceFeatures(physical_device))
-            continue;
-
-        size_t heapSize = 0;
-        for (uint32_t j = 0; j < memoryProperties.memoryHeapCount; ++j) {
-            VkMemoryHeap heap = memoryProperties.memoryHeaps[j];
-            if (heap.flags & VK_MEMORY_HEAP_DEVICE_LOCAL_BIT) {
-                heapSize = heap.size;
-                break;
-            }
-        }
-
-        if (heapSize < memoryRequired)
-            continue;
-
-        auto ext_props = physical_device.enumerateDeviceExtensionProperties();
-        bool has_maintenance4 = false;
-
-        // Check if maintenance4 is supported
-        for (const auto & properties : ext_props) {
-            if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
-                has_maintenance4 = true;
-            }
-        }
-
-        vk::PhysicalDeviceSubgroupProperties subgroup_props;
-        vk::PhysicalDeviceProperties2 dev_props2;
-        vk::PhysicalDeviceMaintenance3Properties dev_props3;
-        vk::PhysicalDeviceMaintenance4Properties dev_props4;
-        dev_props2.pNext = &dev_props3;
-        dev_props3.pNext = &subgroup_props;
-        if (has_maintenance4) {
-            subgroup_props.pNext = &dev_props4;
-        }
-        physical_device.getProperties2(&dev_props2);
-
-        if (subgroup_props.subgroupSize < 32)
-            continue;
-
-        ggml_vk_device d;
-        d.index = i;
-        d.type = dev_props.deviceType;
-        d.heapSize = heapSize;
-        d.vendor = strdup(ggml_vk_getVendorName(dev_props.vendorID));
-        d.subgroupSize = subgroup_props.subgroupSize;
-        d.bufferAlignment = dev_props.limits.minStorageBufferOffsetAlignment;
-
-        if (has_maintenance4) {
-            d.maxAlloc = std::min(dev_props3.maxMemoryAllocationSize, dev_props4.maxBufferSize);
-        } else {
-            d.maxAlloc = dev_props3.maxMemoryAllocationSize;
-        }
-
-        std::string name(dev_props.deviceName);
-        size_t n_idx = ++count_by_name[name];
-        if (n_idx > 1) {
-            name += " (" + std::to_string(n_idx) + ")";
-        }
-        d.name = strdup(name.c_str());
-
-        results.push_back(d);
-    }
-
-    std::stable_sort(results.begin(), results.end(),
-        [](const ggml_vk_device& lhs, const ggml_vk_device& rhs) -> bool {
-            if (lhs.type != rhs.type) {
-                if (lhs.type == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU) return true;
-                if (rhs.type == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU) return false;
-
-                if (lhs.type == VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU) return true;
-                if (rhs.type == VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU) return false;
-            }
-            return lhs.heapSize < rhs.heapSize;
-        }
-    );
-
-    return results;
-}
-
-static std::vector<ggml_vk_device>& ggml_vk_available_devices() {
-    static std::vector<ggml_vk_device> devices = ggml_vk_available_devices_internal(0);
-    return devices;
-}
-
-static void ggml_vk_filterByVendor(std::vector<ggml_vk_device>& devices, const std::string& targetVendor) {
-    devices.erase(
-        std::remove_if(devices.begin(), devices.end(),
-            [&targetVendor](const ggml_vk_device& device) {
-                return device.vendor != targetVendor;
-            }),
-        devices.end()
-    );
-}
-
-static void ggml_vk_filterByName(std::vector<ggml_vk_device>& devices, const std::string& targetName) {
-    devices.erase(
-        std::remove_if(devices.begin(), devices.end(),
-            [&targetName](const ggml_vk_device& device) {
-                return device.name != targetName;
-            }),
-        devices.end()
-    );
-}
-
-static bool ggml_vk_get_device(ggml_vk_device * device, size_t memoryRequired, const std::string & name) {
-    if (name.empty())
-        return false;
-
-    auto devices = ggml_vk_available_devices_internal(memoryRequired);
-    if (name == "amd" || name == "nvidia" || name == "intel") {
-        ggml_vk_filterByVendor(devices, name);
-    } else if (name != "gpu") {
-        ggml_vk_filterByName(devices, name);
-    }
-
-    if (devices.empty())
-        return false;
-
-    *device = devices.front();
-    return true;
-}
-
-bool ggml_vk_get_device(ggml_vk_device * device, size_t memoryRequired, const char * name) {
-    return ggml_vk_get_device(device, memoryRequired, std::string(name));
-}
-
-bool ggml_vk_has_vulkan() {
-    return komputeManager()->hasVulkan();
-}
-
-bool ggml_vk_has_device() {
-    return komputeManager()->hasDevice();
-}
-
-ggml_vk_device ggml_vk_current_device() {
-    if (!komputeManager()->hasDevice())
-        return ggml_vk_device();
-
-    auto devices = ggml_vk_available_devices();
-    ggml_vk_filterByName(devices, komputeManager()->physicalDevice()->getProperties().deviceName.data());
-    GGML_ASSERT(!devices.empty());
-    return devices.front();
-}
-
-static
-void ggml_vk_allocate_descriptor_pool(struct ggml_kompute_context * ctx, size_t size) {
-    std::vector<vk::DescriptorPoolSize> descriptorPoolSizes = {
-        vk::DescriptorPoolSize(
-          vk::DescriptorType::eStorageBuffer,
-          4 * size // Descriptor count is number of possible tensors to pass into an algorithm
-          )
-    };
-
-    vk::DescriptorPoolCreateInfo descriptorPoolInfo(
-      vk::DescriptorPoolCreateFlags(),
-      size, // Max sets
-      static_cast<uint32_t>(descriptorPoolSizes.size()),
-      descriptorPoolSizes.data());
-
-    ctx->pool = std::make_shared<vk::DescriptorPool>();
-    vk::Result r = komputeManager()->device()->createDescriptorPool(
-      &descriptorPoolInfo, nullptr, ctx->pool.get());
-    if (r != vk::Result::eSuccess)
-        std::cerr << "Error allocating descriptor pool" << vk::to_string(r);
-}
-
-static
-void ggml_vk_free_descriptor_pool(struct ggml_kompute_context * ctx) {
-    if (ctx->pool) {
-        komputeManager()->device()->destroy(
-          *ctx->pool,
-          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
-        ctx->pool = nullptr;
-    }
-}
-
-static
-vk::Buffer *ggml_vk_allocate_buffer(size_t size) {
-    vk::BufferCreateInfo bufferCreateInfo;
-    bufferCreateInfo.size = size;
-    bufferCreateInfo.usage = vk::BufferUsageFlagBits::eStorageBuffer |
-                             vk::BufferUsageFlagBits::eTransferSrc |
-                             vk::BufferUsageFlagBits::eTransferDst;
-    bufferCreateInfo.sharingMode = vk::SharingMode::eExclusive;
-
-    vk::Buffer *vkBuffer = new vk::Buffer;
-    vk::Result r = komputeManager()->device()->createBuffer(&bufferCreateInfo, nullptr, vkBuffer);
-    if (r != vk::Result::eSuccess)
-        std::cerr << "Error allocating buffer " << vk::to_string(r) << std::endl;
-    return vkBuffer;
-}
-
-static
-vk::DeviceMemory *ggml_vk_allocate(size_t size, vk::MemoryPropertyFlags flags, vk::MemoryRequirements requirements, bool *isHostVisible) {
-
-    uint32_t memoryTypeIndex = -1;
-    bool memoryTypeIndexFound = false;
-    vk::PhysicalDeviceMemoryProperties memoryProperties = komputeManager()->physicalDevice()->getMemoryProperties();
-    for (uint32_t i = 0; i < memoryProperties.memoryTypeCount; i++) {
-        const vk::MemoryType &memoryType = memoryProperties.memoryTypes[i];
-        const vk::MemoryHeap &memoryHeap = memoryProperties.memoryHeaps[memoryType.heapIndex];
-        if (memoryHeap.size < size) {
-            continue;
-        }
-
-        if (requirements.memoryTypeBits & (1 << i)) {
-            if (((memoryProperties.memoryTypes[i]).propertyFlags &
-                 flags) == flags) {
-                memoryTypeIndex = i;
-                memoryTypeIndexFound = true;
-                if (isHostVisible && (memoryProperties.memoryTypes[i].propertyFlags & vk::MemoryPropertyFlagBits::eHostVisible)) {
-                    *isHostVisible = true;
-                }
-                break;
-            }
-        }
-    }
-    if (!memoryTypeIndexFound) {
-        throw std::runtime_error(
-          "Memory type index for buffer creation not found");
-    }
-
-    vk::MemoryAllocateInfo allocInfo;
-    allocInfo.allocationSize = size;
-    allocInfo.memoryTypeIndex = memoryTypeIndex;
-    vk::DeviceMemory *vkDeviceMemory =  new vk::DeviceMemory;
-    vk::Result r = komputeManager()->device()->allocateMemory(&allocInfo, nullptr, vkDeviceMemory);
-    if (r != vk::Result::eSuccess) {
-        std::cerr << "Error allocating memory " << vk::to_string(r) << std::endl;
-        throw std::runtime_error("Error allocating vulkan memory.");
-    }
-    return vkDeviceMemory;
-}
-
-static size_t ggml_vk_aligned_offset(ggml_backend_buffer_t buffer, size_t offset) {
-    size_t minStorageBufferOffsetAlignment = ggml_backend_buffer_get_alignment(buffer);
-
-    // If offset is already aligned, return it directly
-    if (offset % minStorageBufferOffsetAlignment == 0) {
-        return offset;
-    }
-
-    // Otherwise, return the largest multiple of minStorageBufferOffsetAlignment less than offset
-    return (offset / minStorageBufferOffsetAlignment) * minStorageBufferOffsetAlignment;
-}
-
-static ggml_vk_memory ggml_vk_allocate(size_t size) {
-    ggml_vk_memory memory;
-    bool isHostVisible = false;
-    {
-        memory.primaryBuffer = ggml_vk_allocate_buffer(size);
-        vk::MemoryRequirements memoryRequirements = komputeManager()->device()->getBufferMemoryRequirements(*memory.primaryBuffer);
-        vk::MemoryPropertyFlags memoryPropertyFlags = vk::MemoryPropertyFlagBits::eDeviceLocal;
-        memory.primaryMemory = ggml_vk_allocate(size, memoryPropertyFlags, memoryRequirements, &isHostVisible);
-        komputeManager()->device()->bindBufferMemory(*memory.primaryBuffer, *memory.primaryMemory, 0);
-        if (isHostVisible) {
-            vk::Result r = komputeManager()->device()->mapMemory(*memory.primaryMemory, 0, size, vk::MemoryMapFlags(), &memory.data);
-            if (r != vk::Result::eSuccess)
-                std::cerr << "Error mapping memory" << vk::to_string(r);
-        }
-    }
-
-    if (!isHostVisible) {
-        memory.stagingBuffer = ggml_vk_allocate_buffer(size);
-        vk::MemoryRequirements memoryRequirements = komputeManager()->device()->getBufferMemoryRequirements(*memory.stagingBuffer);
-        vk::MemoryPropertyFlags memoryPropertyFlags = vk::MemoryPropertyFlagBits::eHostVisible |
-                                                      vk::MemoryPropertyFlagBits::eHostCoherent |
-                                                      vk::MemoryPropertyFlagBits::eHostCached;
-        memory.stagingMemory = ggml_vk_allocate(size, memoryPropertyFlags, memoryRequirements, &isHostVisible);
-        komputeManager()->device()->bindBufferMemory(*memory.stagingBuffer, *memory.stagingMemory, 0);
-        vk::Result r = komputeManager()->device()->mapMemory(*memory.stagingMemory, 0, size, vk::MemoryMapFlags(), &memory.data);
-        if (r != vk::Result::eSuccess)
-            std::cerr << "Error mapping memory" << vk::to_string(r);
-    }
-
-    memory.size = size;
-    return memory;
-}
-
-static void ggml_vk_free_memory(ggml_vk_memory &memory)
-{
-    komputeManager()->device()->destroy(
-      *memory.primaryBuffer,
-      (vk::Optional<const vk::AllocationCallbacks>)nullptr);
-    if (memory.stagingBuffer) {
-        komputeManager()->device()->destroy(
-          *memory.stagingBuffer,
-          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
-    }
-    komputeManager()->device()->freeMemory(
-      *memory.primaryMemory,
-      (vk::Optional<const vk::AllocationCallbacks>)nullptr);
-    if (memory.stagingMemory) {
-        komputeManager()->device()->freeMemory(
-          *memory.stagingMemory,
-          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
-    }
-}
-
-static const char * ggml_backend_kompute_buffer_type_get_name(ggml_backend_buffer_type_t buft);
-
-static
-ggml_vk_memory * ggml_vk_find_tensor(const struct ggml_tensor * t, uint64_t & offset) {
-    ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
-
-    // compatibility with ggml-backend
-    GGML_ASSERT(buffer && buffer->buft->iface.get_name == ggml_backend_kompute_buffer_type_get_name);
-
-    ggml_vk_memory * buf_ctx = static_cast<ggml_vk_memory *>(buffer->context);
-
-    const intptr_t ioffs = intptr_t(t->data) - intptr_t(buf_ctx->data);
-
-    GGML_ASSERT(ioffs >= 0 && ioffs + int64_t(ggml_nbytes(t)) <= int64_t(buffer->size));
-
-    offset = uint64_t(ioffs);
-    return buf_ctx;
-}
-
-static
-const std::shared_ptr<kp::Tensor> ggml_vk_get_tensor(const struct ggml_tensor * t, uint32_t * alignedOffset = nullptr) {
-    uint64_t originalOffset = 0;
-    auto * res = ggml_vk_find_tensor(t, originalOffset);
-    if (!res) {
-        static std::shared_ptr<kp::Tensor> nullTensor = nullptr;
-        return nullTensor;
-    }
-
-    // Create a tensor whose memory will be composed of our buffers at the correct offset
-    const size_t nelements = ggml_nelements(t);
-    size_t nbytes = ggml_nbytes(t);
-
-    size_t vulkanOffset = ggml_vk_aligned_offset(t->buffer, originalOffset);
-    if (alignedOffset) {
-        *alignedOffset = originalOffset - vulkanOffset;
-        nbytes += *alignedOffset;
-    }
-
-    return komputeManager()->tensor(
-        t->data,
-        nelements,
-        nbytes, kp::Tensor::TensorDataTypes::eFloat,
-        res->primaryMemory, res->primaryBuffer,
-        res->stagingMemory, res->stagingBuffer,
-        vulkanOffset);
-}
-
-static std::vector<uint32_t> getSpirvShader(const unsigned char* rawData, size_t size) {
-    if (size % sizeof(uint32_t) != 0) {
-        throw std::runtime_error("Invalid size: must be divisible by sizeof(uint32_t)");
-    }
-
-    const uint32_t* data_ptr = reinterpret_cast<const uint32_t*>(rawData);
-    size_t count = size / sizeof(uint32_t);
-    return std::vector<uint32_t>(data_ptr, data_ptr + count);
-}
-
-inline static
-uint32_t safe_divide(uint32_t a, uint32_t b) {
-    if (b <= 1) {
-        return a;
-    }
-    if ((a % b) != 0) {
-        fprintf(stderr, "((%u %% %u) == %u) != 0\n", a, b, a % b);
-        GGML_ABORT("safe_divide result would've had remainder");
-    }
-    return a / b;
-}
-
-static void ggml_vk_add(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
-    int32_t nb00, int32_t nb01, int32_t nb02, int32_t nb03,
-    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
-    int32_t nb10, int32_t nb11, int32_t nb12, int32_t nb13,
-    int32_t ne0,
-    int32_t nb0,  int32_t nb1,  int32_t nb2,  int32_t nb3
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_add_comp_spv,
-        kp::shader_data::op_add_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00;
-        int32_t nb00, nb01, nb02, nb03;
-        int32_t ne10, ne11, ne12, ne13;
-        int32_t nb10, nb11, nb12, nb13;
-        int32_t ne0;
-        int32_t nb0, nb1, nb2, nb3;
-    } const pushConsts {
-        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00,
-        nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12, ne13,
-        nb10, nb11, nb12, nb13,
-        ne0,
-        nb0, nb1, nb2, nb3
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_addrow(kp::Sequence& seq,
-                 const std::shared_ptr<kp::Tensor>& inA,
-                 const std::shared_ptr<kp::Tensor>& inB,
-                 const std::shared_ptr<kp::Tensor>& out,
-                 uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-                 uint32_t size, uint32_t row = 0) {
-
-    const static auto spirv = getSpirvShader(kp::shader_data::op_addrow_comp_spv,
-        kp::shader_data::op_addrow_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        uint32_t row;
-    } const pushConsts {
-        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        row
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__))
-        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {size}, {}, {pushConsts});
-    else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({size});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_mul(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
-    int32_t nb00, int32_t nb01, int32_t nb02, int32_t nb03,
-    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
-    int32_t nb10, int32_t nb11, int32_t nb12, int32_t nb13,
-    int32_t ne0,
-    int32_t nb0,  int32_t nb1,  int32_t nb2,  int32_t nb3
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_comp_spv,
-        kp::shader_data::op_mul_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00;
-        int32_t nb00, nb01, nb02, nb03;
-        int32_t ne10, ne11, ne12, ne13;
-        int32_t nb10, nb11, nb12, nb13;
-        int32_t ne0;
-        int32_t nb0, nb1, nb2, nb3;
-    } const pushConsts {
-        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00,
-        nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12, ne13,
-        nb10, nb11, nb12, nb13,
-        ne0,
-        nb0, nb1, nb2, nb3
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_scale(kp::Sequence& seq,
-                   const std::shared_ptr<kp::Tensor>& in,
-                   const std::shared_ptr<kp::Tensor>& out,
-                   uint32_t inOff, uint32_t outOff,
-                   uint32_t size, float scale) {
-    const static auto spirv_1 = getSpirvShader(
-        kp::shader_data::op_scale_comp_spv, kp::shader_data::op_scale_comp_spv_len
-    );
-    const static auto spirv_8 = getSpirvShader(
-        kp::shader_data::op_scale_8_comp_spv, kp::shader_data::op_scale_8_comp_spv_len
-    );
-
-    struct PushConstants {
-        uint32_t inOff, outOff;
-        float scale;
-    } const pushConsts {
-        safe_divide(inOff, 4), safe_divide(outOff, 4),
-        scale
-    };
-
-    const auto * spirv = &spirv_1;
-    std::string name(__func__);
-    if (size % 8 == 0) {
-        size /= 8;
-        name += "_8";
-        spirv = &spirv_8;
-    }
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, *spirv, {size}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({in, out});
-        s_algo->setWorkgroup({size});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_xxlu(
-    const std::vector<uint32_t>& spirv, const char * suffix, kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& in,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inOff, uint32_t outOff,
-    uint32_t size
-) {
-    struct PushConstants {
-        uint32_t inOff, outOff;
-    } const pushConsts {
-        safe_divide(inOff, 4), safe_divide(outOff, 4),
-    };
-
-    auto name = std::string(__func__) + "_" + suffix;
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {size}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({in, out});
-        s_algo->setWorkgroup({size});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-template <typename... Args>
-static void ggml_vk_silu(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_silu_comp_spv,
-        kp::shader_data::op_silu_comp_spv_len);
-
-    ggml_vk_xxlu(spirv, "silu", std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_relu(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_relu_comp_spv,
-        kp::shader_data::op_relu_comp_spv_len);
-
-    ggml_vk_xxlu(spirv, "relu", std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_gelu(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_gelu_comp_spv,
-        kp::shader_data::op_gelu_comp_spv_len);
-
-    ggml_vk_xxlu(spirv, "gelu", std::forward<Args>(args)...);
-}
-
-static void ggml_vk_soft_max(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02, uint32_t ne03,
-    float scale, float max_bias, float m0, float m1,
-    uint32_t n_head_log2
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_softmax_comp_spv,
-        kp::shader_data::op_softmax_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne01, ne02;
-        float scale, max_bias, m0, m1;
-        uint32_t n_head_log2;
-        int32_t mask;
-    } pushConsts {
-        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne01, ne02,
-        scale, max_bias, m0, m1,
-        n_head_log2,
-        bool(inB)
-    };
-
-    auto & inB_ = inB ? inB : inA;
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        // FIXME: The softmax kernel needs to be fixed to use the subgroupsize which can vary by device
-        const uint32_t local_x = 32;
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB_, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {local_x}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB_, out});
-        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_norm_(
-    const std::vector<uint32_t>& spirv, const char * suffix, kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& in,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inOff, uint32_t outOff,
-    int32_t ne00, int32_t nb01,
-    int32_t nrows, float epsilon
-) {
-    GGML_ASSERT(nb01%sizeof(float) == 0);
-    GGML_ASSERT(ne00%sizeof(float) == 0);
-
-    struct PushConstants {
-        uint32_t inOff, outOff;
-        uint32_t ne00, nb01;
-        float eps;
-    } pushConsts {
-        safe_divide(inOff, 4), safe_divide(outOff, 4),
-        (uint32_t)ne00, (uint32_t)nb01, epsilon
-    };
-
-    auto name = std::string(__func__) + "_" + suffix;
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {(uint32_t)nrows}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({in, out});
-        s_algo->setWorkgroup({(uint32_t)nrows});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-template <typename... Args>
-static void ggml_vk_norm(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_norm_comp_spv,
-        kp::shader_data::op_norm_comp_spv_len);
-
-    ggml_vk_norm_(spirv, "norm", std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_rms_norm(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_rmsnorm_comp_spv,
-        kp::shader_data::op_rmsnorm_comp_spv_len);
-
-    ggml_vk_norm_(spirv, "rms", std::forward<Args>(args)...);
-}
-
-static void ggml_vk_diag_mask_inf(kp::Sequence& seq,
-                           const std::shared_ptr<kp::Tensor>& in,
-                           const std::shared_ptr<kp::Tensor>& out,
-                           uint32_t inOff, uint32_t outOff,
-                           uint32_t n_past,
-                           int32_t ne00, int32_t ne01, int32_t ne02) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_diagmask_comp_spv,
-        kp::shader_data::op_diagmask_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inOff, outOff;
-        uint32_t n_past;
-        int32_t ne00, ne01;
-    } pushConsts {
-        safe_divide(inOff, 4), safe_divide(outOff, 4),
-        n_past,
-        ne00, ne01
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__))
-        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {in, out}, spirv, {unsigned(ne00), unsigned(ne01), unsigned(ne02)}, {}, {pushConsts});
-    else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({in, out});
-        s_algo->setWorkgroup({unsigned(ne00), unsigned(ne01), unsigned(ne02)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_mul_mat_f16(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02,
-    uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03,
-    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
-    uint32_t nb10, uint32_t nb11, uint32_t nb12, uint32_t nb13,
-    int32_t ne0, int32_t ne1,
-    uint32_t r2, uint32_t r3
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_f16_comp_spv,
-        kp::shader_data::op_mul_mat_f16_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne01, ne02;
-        uint32_t nb00, nb01, nb02, nb03;
-        int32_t ne10, ne11, ne12;
-        uint32_t nb10, nb11, nb12, nb13;
-        int32_t ne0, ne1;
-        uint32_t r2, r3;
-    } pushConsts {
-        safe_divide(inAOff, 2), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne01, ne02,
-        nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12,
-        nb10, nb11, nb12, nb13,
-        ne0, ne1,
-        r2, r3
-    };
-
-    const unsigned ny = unsigned((ne11 + 4 - 1)/4);
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2;
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), ny, unsigned(ne12*ne13)}, {local_x}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned(ne01), ny, unsigned(ne12*ne13)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_mul_mat_mat_f32(kp::Sequence& seq,
-                         const std::shared_ptr<kp::Tensor>& inA,
-                         const std::shared_ptr<kp::Tensor>& inB,
-                         const std::shared_ptr<kp::Tensor>& out,
-                         uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-                         int32_t ne00, int32_t ne01, int32_t ne02,
-                         uint32_t nb01, uint32_t nb02,
-                         int32_t ne11, int32_t ne12,
-                         uint32_t nb11, uint32_t nb12,
-                         uint32_t nb1, uint32_t nb2) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_mat_f32_comp_spv,
-        kp::shader_data::op_mul_mat_mat_f32_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne01, ne02, ne11, ne12;
-        uint32_t nb01, nb02;
-        uint32_t nb11, nb12;
-        uint32_t nb1, nb2;
-    } pushConsts {
-        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne01, ne02, ne11, ne12,
-        nb01, nb02, nb11, nb12,
-        nb1, nb2
-    };
-
-    const uint32_t local_x = ggml_vk_current_device().subgroupSize;
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(),
-        {inA, inB, out}, spirv,
-        {unsigned(ne01),
-         unsigned(ne11),
-         unsigned(std::max(ne12, ne02))
-         },
-        {local_x},
-        {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned(ne01),
-                              unsigned(ne11),
-                              unsigned(std::max(ne12, ne02)),
-                              });
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_mul_mat_impl(
-    const std::vector<uint32_t>& spirv, const char * suffix, uint32_t block_size, kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02,
-    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
-    int32_t ne0, int32_t ne1,
-    uint32_t nb01, uint32_t nb02, uint32_t nb03,
-    uint32_t nb11, uint32_t nb12, uint32_t nb13,
-    uint32_t r2, uint32_t r3
-) {
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne01, ne02;
-        int32_t ne10, ne12;
-        int32_t ne0, ne1;
-        uint32_t nb01, nb02, nb03;
-        uint32_t nb11, nb12, nb13;
-        uint32_t r2, r3;
-    } pushConsts {
-        safe_divide(inAOff, block_size), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne01, ne02,
-        ne10, ne12,
-        ne0, ne1,
-        nb01, nb02, nb03,
-        nb11, nb12, nb13,
-        r2, r3
-    };
-
-    auto name = std::string(__func__) + "_" + suffix;
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        const uint32_t local_x = (ggml_vk_current_device().subgroupSize * 2) / 8;
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(name, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 7)/8), unsigned(ne11), unsigned(ne12*ne13)}, {local_x}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned((ne01 + 7)/8), unsigned(ne11), unsigned(ne12*ne13)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-template <typename... Args>
-static void ggml_vk_mul_mat_q4_0(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_0_comp_spv,
-        kp::shader_data::op_mul_mat_q4_0_comp_spv_len);
-
-    ggml_vk_mul_mat_impl(spirv, "q4_0", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_mul_mat_q4_1(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_1_comp_spv,
-        kp::shader_data::op_mul_mat_q4_1_comp_spv_len);
-
-    ggml_vk_mul_mat_impl(spirv, "q4_1", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_mul_mat_q8_0(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q8_0_comp_spv,
-        kp::shader_data::op_mul_mat_q8_0_comp_spv_len);
-
-    ggml_vk_mul_mat_impl(spirv, "q8_0", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
-}
-
-static void ggml_vk_mul_mat_q4_k(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02,
-    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
-    int32_t ne0, int32_t ne1,
-    uint32_t nb01, uint32_t nb02, uint32_t nb03,
-    uint32_t nb11, uint32_t nb12, uint32_t nb13,
-    uint32_t r2, uint32_t r3
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_k_comp_spv,
-        kp::shader_data::op_mul_mat_q4_k_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne10, ne0, ne1, ne01, ne02, ne12;
-        uint32_t nb01, nb02, nb03, nb11, nb12, nb13;
-        uint32_t r2, r3;
-    } pushConsts {
-        inAOff, safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne10, ne0, ne1, ne01, ne02, ne12,
-        nb01, nb02, nb03, nb11, nb12, nb13,
-        r2, r3
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 3)/4), unsigned(ne11), unsigned(ne12) * unsigned(ne13)}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned((ne01 + 3)/4), unsigned(ne11), unsigned(ne12) * unsigned(ne13)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_mul_mat_q6_k(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02,
-    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
-    int32_t ne0, int32_t ne1,
-    uint32_t nb01, uint32_t nb02, uint32_t nb03,
-    uint32_t nb11, uint32_t nb12, uint32_t nb13,
-    uint32_t r2, uint32_t r3
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q6_k_comp_spv,
-        kp::shader_data::op_mul_mat_q6_k_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne10, ne0, ne1, ne01, ne02, ne12;
-        uint32_t nb01, nb02, nb03, nb11, nb12, nb13;
-        uint32_t r2, r3;
-    } pushConsts {
-        inAOff, safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne10, ne0, ne1, ne01, ne02, ne12,
-        nb01, nb02, nb03, nb11, nb12, nb13,
-        r2, r3
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        const uint32_t local_x = 2;
-        const uint32_t local_y = ggml_vk_current_device().subgroupSize;
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)*unsigned(ne13)}, {local_x, local_y}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)*unsigned(ne13)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_get_rows(
-    const std::vector<uint32_t>& spirv,
-    const char * suffix,
-    unsigned element_size, unsigned qk,
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t nb01, int32_t nb1,
-    uint32_t size
-) {
-    GGML_ASSERT(nb01%element_size == 0);
-    GGML_ASSERT(nb1%sizeof(float) == 0);
-    if (qk) GGML_ASSERT(ne00%qk == 0);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, nb01, nb1;
-    } pushConsts {
-        safe_divide(inAOff, element_size), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, nb01, nb1
-    };
-
-    auto name = std::string(__func__) + "_" + suffix;
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {size}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({size});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-template <typename... Args>
-static void ggml_vk_get_rows_f32(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_f32_comp_spv,
-        kp::shader_data::op_getrows_f32_comp_spv_len);
-
-    ggml_vk_get_rows(spirv, "f32", sizeof(float), 0, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_get_rows_f16(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_f16_comp_spv,
-        kp::shader_data::op_getrows_f16_comp_spv_len);
-
-    ggml_vk_get_rows(spirv, "f16", sizeof(half), 0, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_get_rows_q4_0(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q4_0_comp_spv,
-        kp::shader_data::op_getrows_q4_0_comp_spv_len);
-
-    ggml_vk_get_rows(spirv, "q4_0", 1/*We access blocks unaligned*/, QK4_0, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_get_rows_q4_1(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q4_1_comp_spv,
-        kp::shader_data::op_getrows_q4_1_comp_spv_len);
-
-    ggml_vk_get_rows(spirv, "q4_1", 1/*We access blocks unaligned*/, QK4_1, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_get_rows_q6_k(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q6_k_comp_spv,
-        kp::shader_data::op_getrows_q6_k_comp_spv_len);
-    ggml_vk_get_rows(spirv, "q6_k", 1/*We access blocks unaligned*/, QK_NL, std::forward<Args>(args)...);
-}
-
-static void ggml_vk_rope(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& inC,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t inCOff, uint32_t outOff,
-    ggml_type src0t, int32_t n_dims, int32_t mode, int32_t n_ctx_orig,
-    float freq_base, float freq_scale, bool has_freq_factors, float ext_factor, float attn_factor, float beta_fast, float beta_slow,
-    int32_t ne01, int32_t ne02, int32_t ne03,
-    uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03,
-    int32_t ne0,
-    uint32_t nb0, uint32_t nb1, uint32_t nb2, uint32_t nb3
-) {
-    GGML_ASSERT(src0t == GGML_TYPE_F16 || src0t == GGML_TYPE_F32);
-
-    static const auto spirv_norm_f16 = getSpirvShader(
-        kp::shader_data::op_rope_norm_f16_comp_spv, kp::shader_data::op_rope_norm_f16_comp_spv_len
-    );
-    static const auto spirv_norm_f32 = getSpirvShader(
-        kp::shader_data::op_rope_norm_f32_comp_spv, kp::shader_data::op_rope_norm_f32_comp_spv_len
-    );
-    static const auto spirv_neox_f16 = getSpirvShader(
-        kp::shader_data::op_rope_neox_f16_comp_spv, kp::shader_data::op_rope_neox_f16_comp_spv_len
-    );
-    static const auto spirv_neox_f32 = getSpirvShader(
-        kp::shader_data::op_rope_neox_f32_comp_spv, kp::shader_data::op_rope_neox_f32_comp_spv_len
-    );
-
-    int type_size = src0t == GGML_TYPE_F16 ? 2 : 4;
-
-    GGML_ASSERT(nb03 % type_size == 0);
-    GGML_ASSERT(nb02 % type_size == 0);
-    GGML_ASSERT(nb01 % type_size == 0);
-    GGML_ASSERT(nb00 % type_size == 0);
-    GGML_ASSERT(nb3  % type_size == 0);
-    GGML_ASSERT(nb2  % type_size == 0);
-    GGML_ASSERT(nb1  % type_size == 0);
-    GGML_ASSERT(nb0  % type_size == 0);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, inCOff, outOff;
-        int32_t n_dims, mode, n_ctx_orig;
-        float freq_base, freq_scale;
-        bool has_freq_factors;
-        float ext_factor, attn_factor, beta_fast, beta_slow;
-        uint32_t nb00, nb01, nb02, nb03;
-        int32_t ne0;
-        uint32_t nb0, nb1, nb2, nb3;
-    } pushConsts {
-        safe_divide(inAOff, type_size), safe_divide(inBOff, 4), safe_divide(inCOff, type_size), safe_divide(outOff, type_size),
-        n_dims, mode, n_ctx_orig,
-        freq_base, freq_scale,
-        has_freq_factors,
-        ext_factor, attn_factor, beta_fast, beta_slow,
-        nb00, nb01, nb02, nb03,
-        ne0,
-        nb0, nb1, nb2, nb3
-    };
-
-    auto & inC_ = inC ? inC : inA;
-    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
-    const bool is_f16 = src0t == GGML_TYPE_F16;
-
-    auto name = std::string(__func__) + (is_neox ? "_neox" : "_norm") + (src0t == GGML_TYPE_F16 ? "_f16" : "_f32");
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        auto & spirv = is_neox ? is_f16 ? spirv_neox_f16 : spirv_neox_f32 : is_f16 ? spirv_norm_f16 : spirv_norm_f32;
-        s_algo = komputeManager()->algorithm<float, PushConstants>(
-            name, s_kompute_context->pool.get(), {inA, inB, inC_, out}, spirv,
-            {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts}
-        );
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({inA, inB, inC_, out});
-        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_cpy(
-    const std::vector<uint32_t>& spirv,
-    uint32_t in_element_size, uint32_t out_element_size,
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& in,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
-    uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03,
-    int32_t ne0, int32_t ne1, int32_t ne2,
-    uint32_t nb0, uint32_t nb1, uint32_t nb2, uint32_t nb3
-) {
-    struct PushConstants {
-        uint32_t inOff, outOff;
-        int32_t ne00, ne01, ne02;
-        uint32_t nb00, nb01, nb02, nb03;
-        int32_t ne0, ne1, ne2;
-        uint32_t nb0, nb1, nb2, nb3;
-    } pushConsts {
-        safe_divide(inOff, in_element_size), safe_divide(outOff, out_element_size),
-        ne00, ne01, ne02,
-        nb00, nb01, nb02, nb03,
-        ne0, ne1, ne2,
-        nb0, nb1, nb2, nb3
-    };
-
-    std::string name = std::string(__func__)
-                       + "_i_" + std::to_string(in_element_size)
-                       + "_o_" + std::to_string(out_element_size);
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name))
-        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
-    else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({in, out});
-        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-template <typename... Args>
-static void ggml_vk_cpy_f32_f16(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f32_f16_comp_spv,
-        kp::shader_data::op_cpy_f32_f16_comp_spv_len);
-    ggml_vk_cpy(spirv, 4, 2, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_cpy_f32_f32(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f32_f32_comp_spv,
-        kp::shader_data::op_cpy_f32_f32_comp_spv_len);
-    ggml_vk_cpy(spirv, 4, 4, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_cpy_f16_f16(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f16_f16_comp_spv,
-        kp::shader_data::op_cpy_f16_f16_comp_spv_len);
-    ggml_vk_cpy(spirv, 2, 2, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_cpy_f16_f32(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f16_f32_comp_spv,
-        kp::shader_data::op_cpy_f16_f32_comp_spv_len);
-    ggml_vk_cpy(spirv, 2, 4, std::forward<Args>(args)...);
-}
-
-static bool ggml_backend_kompute_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
-    int64_t n = ggml_nelements(op);
-    switch (op->op) {
-        case GGML_OP_UNARY:
-            if (n % 4 != 0) return false;
-            switch (ggml_get_unary_op(op)) {
-                case GGML_UNARY_OP_GELU:
-                    if (n % 8 != 0) return false;
-                    // fall through
-                case GGML_UNARY_OP_RELU:
-                case GGML_UNARY_OP_SILU:
-                    return ggml_is_contiguous(op->src[0]);
-                default:
-                    ;
-            }
-            break;
-        case GGML_OP_NONE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_VIEW:
-        case GGML_OP_TRANSPOSE:
-        case GGML_OP_PERMUTE:
-        case GGML_OP_ADD:
-        case GGML_OP_MUL:
-        case GGML_OP_SCALE:
-        case GGML_OP_SOFT_MAX:
-        case GGML_OP_RMS_NORM:
-        case GGML_OP_NORM:
-            return true;
-        case GGML_OP_ROPE:
-            {
-                const int mode = ((const int32_t *) op->op_params)[2];
-                if (mode & GGML_ROPE_TYPE_MROPE) {
-                    return false;
-                }
-                if (mode & GGML_ROPE_TYPE_VISION) {
-                    return false;
-                }
-                return true;
-            }
-        case GGML_OP_DUP:
-        case GGML_OP_CPY:
-        case GGML_OP_CONT:
-            switch (op->src[0]->type) {
-                case GGML_TYPE_F32:
-                case GGML_TYPE_F16:
-                    break;
-                default:
-                    return false;
-            }
-            switch (op->type) {
-                case GGML_TYPE_F32:
-                case GGML_TYPE_F16:
-                    break;
-                default:
-                    return false;
-            }
-            return true;
-        case GGML_OP_DIAG_MASK_INF:
-            return op->ne[3] == 1;
-        case GGML_OP_GET_ROWS:
-            switch (op->src[0]->type) {
-                case GGML_TYPE_F32:
-                case GGML_TYPE_F16:
-                case GGML_TYPE_Q4_0:
-                case GGML_TYPE_Q4_1:
-                case GGML_TYPE_Q6_K:
-                    return op->ne[2] == 1 && op->ne[3] == 1;
-                default:
-                    ;
-            }
-            return false;
-        case GGML_OP_MUL_MAT:
-            if (op->src[1]->type != GGML_TYPE_F32 || ggml_is_transposed(op->src[0]) || ggml_is_transposed(op->src[1]))
-                return false;
-
-            switch (op->src[0]->type) {
-                case GGML_TYPE_F32:
-                    return op->ne[3] == 1;
-                case GGML_TYPE_Q6_K:
-                case GGML_TYPE_F16:
-                case GGML_TYPE_Q8_0:
-                case GGML_TYPE_Q4_0:
-                case GGML_TYPE_Q4_1:
-                case GGML_TYPE_Q4_K:
-                    return true;
-                default:
-                    ;
-            }
-        default:
-            ;
-    }
-    return false;
-
-    GGML_UNUSED(dev);
-}
-
-static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph * gf) {
-    const int n_seq = 8;
-
-    // FIXME: Figure out if we can somehow optimize the size of the pool... right now we're setting
-    // it to the size of the graph, but I think it can be made smaller?
-    ggml_vk_allocate_descriptor_pool(ctx, gf->n_nodes);
-
-    std::vector<std::shared_ptr<kp::Sequence>> sequences(n_seq);
-
-    for (auto& sequence : sequences) {
-        sequence = komputeManager()->sequence();
-    }
-    for (int seq_idx = 0; seq_idx < n_seq; ++seq_idx) {
-        const int n_nodes_per_seq = (gf->n_nodes + n_seq - 1) / n_seq;
-
-        auto& seq = *sequences[seq_idx];
-
-        const int node_start = (seq_idx + 0) * n_nodes_per_seq;
-        const int node_end   = std::min((seq_idx == n_seq - 1) ? gf->n_nodes : (seq_idx + 1) * n_nodes_per_seq, gf->n_nodes);
-
-        bool any_commands_recorded = false;
-
-        for (int i = node_start; i < node_end; ++i) {
-            struct ggml_tensor * src0 = gf->nodes[i]->src[0];
-            struct ggml_tensor * src1 = gf->nodes[i]->src[1];
-            struct ggml_tensor * src2 = gf->nodes[i]->src[2]; GGML_UNUSED(src2);
-            struct ggml_tensor * dst = gf->nodes[i];
-            GGML_ASSERT(dst->data != nullptr);
-
-            if (ggml_is_empty(dst)) {
-                continue;
-            }
-
-            switch (dst->op) {
-                case GGML_OP_NONE:
-                case GGML_OP_RESHAPE:
-                case GGML_OP_VIEW:
-                case GGML_OP_TRANSPOSE:
-                case GGML_OP_PERMUTE:
-                    continue; // noop -> next node
-                default:
-                    break;
-            }
-
-            any_commands_recorded = true;
-
-            const int32_t ne00 = src0 ? src0->ne[0] : 0;
-            const int32_t ne01 = src0 ? src0->ne[1] : 0;
-            const int32_t ne02 = src0 ? src0->ne[2] : 0;
-            const int32_t ne03 = src0 ? src0->ne[3] : 0;
-
-            const uint32_t nb00 = src0 ? src0->nb[0] : 0;
-            const uint32_t nb01 = src0 ? src0->nb[1] : 0;
-            const uint32_t nb02 = src0 ? src0->nb[2] : 0;
-            const uint32_t nb03 = src0 ? src0->nb[3] : 0;
-
-            const int32_t ne10 = src1 ? src1->ne[0] : 0;
-            const int32_t ne11 = src1 ? src1->ne[1] : 0;
-            const int32_t ne12 = src1 ? src1->ne[2] : 0;
-            const int32_t ne13 = src1 ? src1->ne[3] : 0;
-
-            const uint32_t nb10 = src1 ? src1->nb[0] : 0;
-            const uint32_t nb11 = src1 ? src1->nb[1] : 0;
-            const uint32_t nb12 = src1 ? src1->nb[2] : 0;
-            const uint32_t nb13 = src1 ? src1->nb[3] : 0;
-
-            const int32_t ne0 = dst ? dst->ne[0] : 0;
-            const int32_t ne1 = dst ? dst->ne[1] : 0;
-            const int32_t ne2 = dst ? dst->ne[2] : 0;
-//            const int32_t ne3 = dst ? dst->ne[3] : 0;
-
-            const uint32_t nb0 = dst ? dst->nb[0] : 0;
-            const uint32_t nb1 = dst ? dst->nb[1] : 0;
-            const uint32_t nb2 = dst ? dst->nb[2] : 0;
-            const uint32_t nb3 = dst ? dst->nb[3] : 0;
-
-            const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
-            const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
-            const enum ggml_type dstt = dst ? dst->type : GGML_TYPE_COUNT;
-
-            const static std::shared_ptr<kp::Tensor> nullTensor = nullptr;
-            uint32_t off_src0 = 0;
-            uint32_t off_src1 = 0;
-            uint32_t off_src2 = 0;
-            uint32_t off_dst  = 0;
-            const std::shared_ptr<kp::Tensor>& id_src0 = src0 ? ggml_vk_get_tensor(src0, &off_src0) : nullTensor;
-            const std::shared_ptr<kp::Tensor>& id_src1 = src1 ? ggml_vk_get_tensor(src1, &off_src1) : nullTensor;
-            const std::shared_ptr<kp::Tensor>& id_src2 = src2 ? ggml_vk_get_tensor(src2, &off_src2) : nullTensor;
-            const std::shared_ptr<kp::Tensor>& id_dst  = dst  ? ggml_vk_get_tensor(dst,  &off_dst)  : nullTensor;
-
-            switch (dst->op) {
-                case GGML_OP_ADD:
-                    {
-                        if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
-                            // src1 is a row
-                            ggml_vk_addrow(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ggml_nelements(dst)/4, ne00);
-                        } else {
-                            ggml_vk_add(
-                                seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                ne00, ne01, ne02, ne03,
-                                nb00, nb01, nb02, nb03,
-                                ne10, ne11, ne12, ne13,
-                                nb10, nb11, nb12, nb13,
-                                ne0,
-                                nb0, nb1, nb2, nb3
-                            );
-                        }
-                    } break;
-                case GGML_OP_MUL:
-                    {
-                        ggml_vk_mul(
-                            seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                            ne00, ne01, ne02, ne03,
-                            nb00, nb01, nb02, nb03,
-                            ne10, ne11, ne12, ne13,
-                            nb10, nb11, nb12, nb13,
-                            ne0,
-                            nb0, nb1, nb2, nb3
-                        );
-                    } break;
-                case GGML_OP_SCALE:
-                    {
-                        float scale; memcpy(&scale, dst->op_params, sizeof(float));
-
-                        ggml_vk_scale(seq, id_src0, id_dst, off_src0, off_dst, ggml_nelements(dst), scale);
-                    } break;
-                case GGML_OP_UNARY:
-                    {
-                        int64_t n = ggml_nelements(dst);
-                        GGML_ASSERT(n % 4 == 0);
-                        switch (ggml_get_unary_op(gf->nodes[i])) {
-                            case GGML_UNARY_OP_SILU:
-                                {
-                                    ggml_vk_silu(seq, id_src0, id_dst, off_src0, off_dst, n/4);
-                                } break;
-                            case GGML_UNARY_OP_RELU:
-                                {
-                                    ggml_vk_relu(seq, id_src0, id_dst, off_src0, off_dst, n/4);
-                                } break;
-                            case GGML_UNARY_OP_GELU:
-                                {
-                                    GGML_ASSERT(n % 8 == 0);
-                                    ggml_vk_gelu(seq, id_src0, id_dst, off_src0, off_dst, n/8);
-                                } break;
-                            default:
-                                {
-                                    fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
-                                    GGML_ABORT("fatal error");
-                                }
-                        }
-                    } break;
-                case GGML_OP_SOFT_MAX:
-                    {
-                        float scale;
-                        float max_bias;
-
-                        memcpy(&scale,    (float *)dst->op_params + 0, sizeof(float));
-                        memcpy(&max_bias, (float *)dst->op_params + 1, sizeof(float));
-
-#pragma message("TODO: add ggml_vk_soft_max() F16 src1 support")
-#pragma message("ref:  https://github.com/ggerganov/llama.cpp/pull/5021")
-                        GGML_ASSERT(!src1 || src1t == GGML_TYPE_F32);
-
-                        const int64_t nrows_x = ggml_nrows(src0);
-                        const int64_t nrows_y = src0->ne[1];
-
-                        const uint32_t n_head      = nrows_x/nrows_y;
-                        const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
-
-                        const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
-                        const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
-
-                        ggml_vk_soft_max(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne03, scale, max_bias, m0, m1, n_head_log2);
-                    } break;
-                case GGML_OP_DIAG_MASK_INF:
-                    {
-                        const int n_past = ((int32_t *)(dst->op_params))[0];
-                        ggml_vk_diag_mask_inf(seq, id_src0, id_dst, off_src0, off_dst, n_past, ne00, ne01, ne02);
-                    } break;
-                case GGML_OP_NORM:
-                    {
-                        float eps;
-                        memcpy(&eps, dst->op_params, sizeof(float));
-                        ggml_vk_norm(seq, id_src0, id_dst, off_src0, off_dst, ne00, nb01, ggml_nrows(src0), eps);
-                    } break;
-                case GGML_OP_RMS_NORM:
-                    {
-                        GGML_ASSERT(ne00 % 4 == 0);
-
-                        float eps;
-                        memcpy(&eps, dst->op_params, sizeof(float));
-                        ggml_vk_rms_norm(seq, id_src0, id_dst, off_src0, off_dst, ne00, nb01, ggml_nrows(src0), eps);
-                    } break;
-                case GGML_OP_MUL_MAT:
-                    {
-                        GGML_ASSERT(ne00 == ne10);
-
-                        GGML_ASSERT(ne12 % ne02 == 0);
-                        GGML_ASSERT(ne13 % ne03 == 0);
-
-                        const uint32_t r2 = ne12/ne02;
-                        const uint32_t r3 = ne13/ne03;
-
-                        if (src1t != GGML_TYPE_F32) {
-                            fprintf(stderr, "%s: %s: Unsupported src1 type: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
-                            goto not_implemented;
-                        }
-
-                        if (ggml_is_transposed(src0) ||
-                            ggml_is_transposed(src1)) {
-                            fprintf(stderr, "%s: %s: matmul on tranposed tensor not supported: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
-                            goto not_implemented;
-                        }
-
-                        switch (src0t) {
-                            case GGML_TYPE_F32:
-                                ggml_vk_mul_mat_mat_f32(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, nb01, nb02, ne11, ne12, nb11, nb12, nb1, nb2
-                                );
-                                break;
-                            case GGML_TYPE_F16:
-                                ggml_vk_mul_mat_f16(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                    ne10, ne11, ne12, ne13, nb10, nb11, nb12, nb13,
-                                    ne0, ne1, r2, r3
-                                );
-                                break;
-                            case GGML_TYPE_Q8_0:
-                                ggml_vk_mul_mat_q8_0(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1,
-                                    nb01, nb02, nb03, nb11, nb12, nb13, r2, r3
-                                );
-                                break;
-                            case GGML_TYPE_Q4_0:
-                                ggml_vk_mul_mat_q4_0(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1,
-                                    nb01, nb02, nb03, nb11, nb12, nb13, r2, r3
-                                );
-                                break;
-                            case GGML_TYPE_Q4_1:
-                                ggml_vk_mul_mat_q4_1(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1,
-                                    nb01, nb02, nb03, nb11, nb12, nb13, r2, r3
-                                );
-                                break;
-                            case GGML_TYPE_Q4_K:
-                                ggml_vk_mul_mat_q4_k(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1,
-                                    nb01, nb02, nb03, nb11, nb12, nb13, r2, r3
-                                );
-                                break;
-                            case GGML_TYPE_Q6_K:
-                                ggml_vk_mul_mat_q6_k(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1,
-                                    nb01, nb02, nb03, nb11, nb12, nb13, r2, r3
-                                );
-                                break;
-                            default: {
-                                fprintf(stderr, "%s: %s: Unsupported quantization: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
-                                goto not_implemented;
-                            }
-                        }
-
-                    } break;
-                case GGML_OP_GET_ROWS:
-                    {
-                        if (src0t == GGML_TYPE_F32) {
-                            ggml_vk_get_rows_f32(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
-                        } else if (src0t == GGML_TYPE_F16) {
-                            ggml_vk_get_rows_f16(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
-                        } else if (src0t == GGML_TYPE_Q4_0) {
-                            ggml_vk_get_rows_q4_0(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
-                        } else if (src0t == GGML_TYPE_Q4_1) {
-                            ggml_vk_get_rows_q4_1(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
-                        } else if (src0t == GGML_TYPE_Q6_K) {
-                            ggml_vk_get_rows_q6_k(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
-                        } else {
-                            fprintf(stderr, "%s: %s: Unsupported quantization: %u\n", __func__, ggml_op_name(dst->op), src0t);
-                            goto not_implemented;
-                        }
-                    } break;
-                case GGML_OP_ROPE:
-                    {
-                        GGML_ASSERT(ne10 == ne02);
-                        GGML_ASSERT(src0t == dstt);
-                        // const int n_past = ((int32_t *) dst->op_params)[0];
-                        const int n_dims     = ((int32_t *) dst->op_params)[1];
-                        const int mode       = ((int32_t *) dst->op_params)[2];
-                        // skip 3, n_ctx used in GLM RoPE, unimplemented in Vulkan
-                        const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
-
-                        const bool has_freq_factors = dst->src[2] != nullptr;
-
-                        float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
-                        memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
-                        memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
-                        memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
-                        memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
-                        memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
-                        memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
-                        ggml_vk_rope(
-                            seq, id_src0, id_src1, id_src2, id_dst, off_src0, off_src1, off_src2, off_dst, src0t, n_dims, mode, n_ctx_orig,
-                            freq_base, freq_scale, has_freq_factors, ext_factor, attn_factor, beta_fast, beta_slow,
-                            ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, nb0, nb1, nb2, nb3
-                        );
-                    } break;
-                case GGML_OP_DUP:
-                case GGML_OP_CPY:
-                case GGML_OP_CONT:
-                    {
-                        switch (src0t) {
-                            case GGML_TYPE_F32:
-                                {
-                                    switch (dstt) {
-                                        case GGML_TYPE_F16: ggml_vk_cpy_f32_f16(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
-                                        case GGML_TYPE_F32: ggml_vk_cpy_f32_f32(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
-                                        default: goto not_implemented;
-                                    }
-                                } break;
-                            case GGML_TYPE_F16:
-                                {
-                                    switch (dstt) {
-                                        case GGML_TYPE_F16: ggml_vk_cpy_f16_f16(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
-                                        case GGML_TYPE_F32: ggml_vk_cpy_f16_f32(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
-                                    default: goto not_implemented;
-                                } break;
-                            default: goto not_implemented;
-                            }
-                        }
-                    } break;
-                default: goto not_implemented;
-            }
-            continue;
-            not_implemented: {}
-            fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
-            //GGML_ABORT("fatal error");
-        }
-
-        // Evaluate sequence
-        if (any_commands_recorded) {
-            seq.evalAsync();
-        }
-    }
-
-    // Wait for all sequences to finish
-    for (auto& sequence : sequences) {
-        if (sequence->isRunning())
-            sequence->evalAwait();
-    }
-
-    ggml_vk_free_descriptor_pool(ctx);
-}
-
-template<>
-kp::Tensor::TensorDataTypes
-kp::TensorT<half>::dataType()
-{
-    return TensorDataTypes::eFloat;
-}
-
-template<>
-kp::Tensor::TensorDataTypes
-kp::TensorT<uint8_t>::dataType()
-{
-    return TensorDataTypes::eUnsignedInt;
-}
-
-////////////////////////////////////////////////////////////////////////////////
-
-// backend interface
-
-struct ggml_backend_kompute_buffer_type_context {
-    int         device;
-    int         device_ref = 0;
-    uint64_t    buffer_alignment;
-    uint64_t    max_alloc;
-    std::string name;
-
-    ggml_backend_kompute_buffer_type_context(int device, uint64_t buffer_alignment, uint64_t max_alloc)
-        : device(device), buffer_alignment(buffer_alignment), max_alloc(max_alloc), name(ggml_kompute_format_name(device)) {}
-};
-
-static void ggml_backend_kompute_device_ref(ggml_backend_buffer_type_t buft) {
-    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
-
-    if (!ctx->device_ref) {
-        komputeManager()->initializeDevice(
-            ctx->device, {}, {
-                "VK_KHR_shader_float16_int8", "VK_KHR_8bit_storage",
-                "VK_KHR_16bit_storage", "VK_KHR_shader_non_semantic_info"
-            }
-        );
-    }
-
-    assert(ggml_vk_has_device());
-    ctx->device_ref++;
-}
-
-static void ggml_backend_kompute_device_unref(ggml_backend_buffer_type_t buft) {
-    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
-
-    assert(ctx->device_ref > 0);
-
-    ctx->device_ref--;
-
-    if (!ctx->device_ref) {
-        komputeManager.destroy();
-    }
-}
-
-static void ggml_backend_kompute_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    auto * memory = (ggml_vk_memory *)buffer->context;
-    if (ggml_vk_has_device()) {
-        ggml_vk_free_memory(*memory);
-    }
-    delete memory;
-}
-
-static void * ggml_backend_kompute_buffer_get_base(ggml_backend_buffer_t buffer) {
-    return ((ggml_vk_memory *)buffer->context)->data;
-}
-
-static void ggml_backend_kompute_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    GGML_UNUSED(buffer);
-
-    const auto res = ggml_vk_get_tensor(tensor);
-    GGML_ASSERT(res);
-
-    memcpy((char *)tensor->data + offset, data, size);
-
-    komputeManager()->sequence()->eval<kp::OpTensorSyncDevice>({res});
-}
-
-static void ggml_backend_kompute_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_UNUSED(buffer);
-
-    const auto res = ggml_vk_get_tensor(tensor);
-    GGML_ASSERT(res);
-
-    komputeManager()->sequence()->eval<kp::OpTensorSyncLocal>({res});
-
-    memcpy(data, (const char *)tensor->data + offset, size);
-}
-
-static void ggml_backend_kompute_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    auto * memory = (ggml_vk_memory *)buffer->context;
-    memset(memory->data, value, buffer->size);
-
-    if (memory->stagingBuffer)
-        komputeManager()->sequence()->eval<kp::OpBufferSyncDevice>(memory->primaryBuffer, memory->stagingBuffer, memory->size);
-}
-
-static ggml_backend_buffer_i ggml_backend_kompute_buffer_i = {
-    /* .free_buffer     = */ ggml_backend_kompute_buffer_free_buffer,
-    /* .get_base        = */ ggml_backend_kompute_buffer_get_base,
-    /* .init_tensor     = */ NULL,
-    /* .memset_tensor   = */ NULL,
-    /* .set_tensor      = */ ggml_backend_kompute_buffer_set_tensor,
-    /* .get_tensor      = */ ggml_backend_kompute_buffer_get_tensor,
-    /* .cpy_tensor      = */ NULL,
-    /* .clear           = */ ggml_backend_kompute_buffer_clear,
-    /* .reset           = */ NULL,
-};
-
-// default buffer type
-
-static const char * ggml_backend_kompute_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
-    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
-    return ctx->name.c_str();
-}
-
-static ggml_backend_buffer_t ggml_backend_kompute_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    ggml_backend_kompute_device_ref(buft);
-    auto * ctx = new ggml_vk_memory(ggml_vk_allocate(size));
-    return ggml_backend_buffer_init(buft, ggml_backend_kompute_buffer_i, ctx, size);
-}
-
-static size_t ggml_backend_kompute_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
-    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
-    return ctx->buffer_alignment;
-}
-
-static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
-    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
-    return ctx->max_alloc;
-}
-
-static ggml_backend_buffer_type_i ggml_backend_kompute_buffer_type_interface = {
-    /* .get_name         = */ ggml_backend_kompute_buffer_type_get_name,
-    /* .alloc_buffer     = */ ggml_backend_kompute_buffer_type_alloc_buffer,
-    /* .get_alignment    = */ ggml_backend_kompute_buffer_type_get_alignment,
-    /* .get_max_size     = */ ggml_backend_vk_buffer_type_get_max_size,
-    /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
-    /* .is_host          = */ NULL,
-};
-
-ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device) {
-    static std::mutex mutex;
-    std::lock_guard<std::mutex> lock(mutex);
-
-    auto devices = ggml_vk_available_devices();
-    int32_t device_count = (int32_t) devices.size();
-    GGML_ASSERT(device < device_count);
-    GGML_ASSERT(devices.size() <= GGML_KOMPUTE_MAX_DEVICES);
-
-    static ggml_backend_buffer_type
-        ggml_backend_kompute_buffer_types[GGML_KOMPUTE_MAX_DEVICES];
-
-    static bool ggml_backend_kompute_buffer_type_initialized = false;
-
-    if (!ggml_backend_kompute_buffer_type_initialized) {
-        for (int32_t i = 0; i < device_count; i++) {
-            ggml_backend_kompute_buffer_types[i] = {
-                /* .iface    = */ ggml_backend_kompute_buffer_type_interface,
-                /* .device   = */ ggml_backend_reg_dev_get(ggml_backend_kompute_reg(), i),
-                /* .context  = */ new ggml_backend_kompute_buffer_type_context{ i, devices[i].bufferAlignment, devices[i].maxAlloc },
-            };
-        }
-        ggml_backend_kompute_buffer_type_initialized = true;
-    }
-
-    return &ggml_backend_kompute_buffer_types[device];
-}
-
-// backend
-
-static const char * ggml_backend_kompute_name(ggml_backend_t backend) {
-    auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
-    return ctx->name.c_str();
-}
-
-static void ggml_backend_kompute_free(ggml_backend_t backend) {
-    auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
-
-    assert(ctx == s_kompute_context);
-    s_kompute_context = nullptr;
-    if (ctx != nullptr) {
-        delete ctx;
-    }
-
-    delete backend;
-}
-
-static ggml_status ggml_backend_kompute_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
-    ggml_vk_graph_compute(ctx, cgraph);
-    return GGML_STATUS_SUCCESS;
-}
-
-static struct ggml_backend_i kompute_backend_i = {
-    /* .get_name                = */ ggml_backend_kompute_name,
-    /* .free                    = */ ggml_backend_kompute_free,
-    /* .set_tensor_async        = */ NULL,
-    /* .get_tensor_async        = */ NULL,
-    /* .cpy_tensor_async        = */ NULL,
-    /* .synchronize             = */ NULL,
-    /* .graph_plan_create       = */ NULL,
-    /* .graph_plan_free         = */ NULL,
-    /* .graph_plan_update       = */ NULL,
-    /* .graph_plan_compute      = */ NULL,
-    /* .graph_compute           = */ ggml_backend_kompute_graph_compute,
-    /* .event_record            = */ NULL,
-    /* .event_wait              = */ NULL,
-};
-
-static ggml_guid_t ggml_backend_kompute_guid() {
-    static ggml_guid guid = { 0x7b, 0x57, 0xdc, 0xaf, 0xde, 0x12, 0x1d, 0x49, 0xfb, 0x35, 0xfa, 0x9b, 0x18, 0x31, 0x1d, 0xca };
-    return &guid;
-}
-
-ggml_backend_t ggml_backend_kompute_init(int device) {
-    GGML_ASSERT(s_kompute_context == nullptr);
-    s_kompute_context = new ggml_kompute_context(device);
-
-    ggml_backend_t kompute_backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_kompute_guid(),
-        /* .interface = */ kompute_backend_i,
-        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_kompute_reg(), device),
-        /* .context   = */ s_kompute_context,
-    };
-
-    return kompute_backend;
-}
-
-bool ggml_backend_is_kompute(ggml_backend_t backend) {
-    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_kompute_guid());
-}
-
-static size_t ggml_backend_kompute_get_device_count() {
-    auto devices = ggml_vk_available_devices();
-    return devices.size();
-}
-
-static void ggml_backend_kompute_get_device_description(int device, char * description, size_t description_size) {
-    auto devices = ggml_vk_available_devices();
-    GGML_ASSERT((size_t) device < devices.size());
-    snprintf(description, description_size, "%s", devices[device].name);
-}
-
-static void ggml_backend_kompute_get_device_memory(int device, size_t * free, size_t * total) {
-    auto devices = ggml_vk_available_devices();
-    GGML_ASSERT((size_t) device < devices.size());
-    *total = devices[device].heapSize;
-    *free = devices[device].heapSize;
-}
-
-//////////////////////////
-
-struct ggml_backend_kompute_device_context {
-    int device;
-    std::string name;
-    std::string description;
-};
-
-static const char * ggml_backend_kompute_device_get_name(ggml_backend_dev_t dev) {
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    return ctx->name.c_str();
-}
-
-static const char * ggml_backend_kompute_device_get_description(ggml_backend_dev_t dev) {
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    return ctx->description.c_str();
-}
-
-static void ggml_backend_kompute_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    ggml_backend_kompute_get_device_memory(ctx->device, free, total);
-}
-
-static ggml_backend_buffer_type_t ggml_backend_kompute_device_get_buffer_type(ggml_backend_dev_t dev) {
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    return ggml_backend_kompute_buffer_type(ctx->device);
-}
-
-static bool ggml_backend_kompute_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    if (buft->iface.get_name != ggml_backend_kompute_buffer_type_get_name) {
-        return false;
-    }
-
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    ggml_backend_kompute_buffer_type_context * buft_ctx = (ggml_backend_kompute_buffer_type_context *)buft->context;
-
-    return buft_ctx->device == ctx->device;
-}
-
-static enum ggml_backend_dev_type ggml_backend_kompute_device_get_type(ggml_backend_dev_t dev) {
-    GGML_UNUSED(dev);
-    return GGML_BACKEND_DEVICE_TYPE_GPU;
-}
-
-static void ggml_backend_kompute_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
-    props->name        = ggml_backend_kompute_device_get_name(dev);
-    props->description = ggml_backend_kompute_device_get_description(dev);
-    props->type        = ggml_backend_kompute_device_get_type(dev);
-    ggml_backend_kompute_device_get_memory(dev, &props->memory_free, &props->memory_total);
-    props->caps = {
-        /* async                  = */ false,
-        /* host_buffer            = */ false,
-        /* .buffer_from_host_ptr  = */ false,
-        /* events                 = */ false,
-    };
-}
-
-static ggml_backend_t ggml_backend_kompute_device_init(ggml_backend_dev_t dev, const char * params) {
-    GGML_UNUSED(params);
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    return ggml_backend_kompute_init(ctx->device);
-}
-
-static bool ggml_backend_kompute_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
-    const int min_batch_size = 32;
-
-    return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
-           (op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
-
-    GGML_UNUSED(dev);
-}
-
-static const struct ggml_backend_device_i ggml_backend_kompute_device_i = {
-    /* .get_name             = */ ggml_backend_kompute_device_get_name,
-    /* .get_description      = */ ggml_backend_kompute_device_get_description,
-    /* .get_memory           = */ ggml_backend_kompute_device_get_memory,
-    /* .get_type             = */ ggml_backend_kompute_device_get_type,
-    /* .get_props            = */ ggml_backend_kompute_device_get_props,
-    /* .init_backend         = */ ggml_backend_kompute_device_init,
-    /* .get_buffer_type      = */ ggml_backend_kompute_device_get_buffer_type,
-    /* .get_host_buffer_type = */ NULL,
-    /* .buffer_from_host_ptr = */ NULL,
-    /* .supports_op          = */ ggml_backend_kompute_device_supports_op,
-    /* .supports_buft        = */ ggml_backend_kompute_device_supports_buft,
-    /* .offload_op           = */ ggml_backend_kompute_device_offload_op,
-    /* .event_new            = */ NULL,
-    /* .event_free           = */ NULL,
-    /* .event_synchronize    = */ NULL,
-};
-
-static const char * ggml_backend_kompute_reg_get_name(ggml_backend_reg_t reg) {
-    GGML_UNUSED(reg);
-    return "Kompute";
-}
-
-static size_t ggml_backend_kompute_reg_get_device_count(ggml_backend_reg_t reg) {
-    GGML_UNUSED(reg);
-    return ggml_backend_kompute_get_device_count();
-}
-
-static ggml_backend_dev_t ggml_backend_kompute_reg_get_device(ggml_backend_reg_t reg, size_t device) {
-    static std::vector<ggml_backend_dev_t> devices;
-
-    static bool initialized = false;
-
-    {
-        static std::mutex mutex;
-        std::lock_guard<std::mutex> lock(mutex);
-        if (!initialized) {
-            for (size_t i = 0; i < ggml_backend_kompute_get_device_count(); i++) {
-                ggml_backend_kompute_device_context * ctx = new ggml_backend_kompute_device_context;
-                char desc[256];
-                ggml_backend_kompute_get_device_description(i, desc, sizeof(desc));
-                ctx->device = i;
-                ctx->name = "Kompute" + std::to_string(i);
-                ctx->description = desc;
-                devices.push_back(new ggml_backend_device {
-                    /* .iface   = */ ggml_backend_kompute_device_i,
-                    /* .reg     = */ reg,
-                    /* .context = */ ctx,
-                });
-            }
-            initialized = true;
-        }
-    }
-
-    GGML_ASSERT(device < devices.size());
-    return devices[device];
-}
-
-static const struct ggml_backend_reg_i ggml_backend_kompute_reg_i = {
-    /* .get_name         = */ ggml_backend_kompute_reg_get_name,
-    /* .get_device_count = */ ggml_backend_kompute_reg_get_device_count,
-    /* .get_device       = */ ggml_backend_kompute_reg_get_device,
-    /* .get_proc_address = */ NULL,
-};
-
-ggml_backend_reg_t ggml_backend_kompute_reg() {
-    static ggml_backend_reg reg = {
-        /* .api_version = */ GGML_BACKEND_API_VERSION,
-        /* .iface       = */ ggml_backend_kompute_reg_i,
-        /* .context     = */ nullptr,
-    };
-
-    return &reg;
-}
-
-GGML_BACKEND_DL_IMPL(ggml_backend_kompute_reg)
diff --git a/ggml/src/ggml-kompute/kompute b/ggml/src/ggml-kompute/kompute
deleted file mode 160000
index 4565194ed7..0000000000
--- a/ggml/src/ggml-kompute/kompute
+++ /dev/null
@@ -1 +0,0 @@
-Subproject commit 4565194ed7c32d1d2efa32ceab4d3c6cae006306
diff --git a/ggml/src/ggml-kompute/kompute-shaders/common.comp b/ggml/src/ggml-kompute/kompute-shaders/common.comp
deleted file mode 100644
index dbe4cf804e..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/common.comp
+++ /dev/null
@@ -1,112 +0,0 @@
-#extension GL_EXT_shader_16bit_storage: require
-#extension GL_EXT_shader_8bit_storage: require
-#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
-#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
-#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
-#extension GL_EXT_shader_explicit_arithmetic_types_int64: require
-#extension GL_EXT_control_flow_attributes: enable
-#extension GL_KHR_shader_subgroup_arithmetic : require
-#extension GL_EXT_debug_printf : enable
-
-#define QK4_0 32
-#define QK4_1 32
-
-#define GELU_COEF_A 0.044715
-#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
-#define TWOPI_F 6.283185307179586f
-
-#define QK_K 256
-#define K_SCALE_SIZE 12
-
-#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
-#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
-#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
-#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
-
-#define sizeof_block_q4_0 0x12
-struct block_q4_0 {
-    float16_t d;
-    uint8_t qs[QK4_0 / 2];
-};
-mat4 dequantize_q4_0(const block_q4_0 xb, uint il) {
-    const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
-    const float d2 = d1 / 256.f;
-    const float md = -8.f * xb.d;
-    const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
-    const uint16_t mask1 = mask0 << 8;
-
-    mat4 reg;
-    for (int i=0;i<8;i++) {
-        uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
-        reg[i/2][2*(i%2)+0] = d1 * (b & mask0) + md;
-        reg[i/2][2*(i%2)+1] = d2 * (b & mask1) + md;
-    }
-    return reg;
-}
-
-#define sizeof_block_q4_1 0x14
-struct block_q4_1 {
-    float16_t d;
-    float16_t m;
-    uint8_t qs[QK4_1 / 2];
-};
-mat4 dequantize_q4_1(const block_q4_1 xb, uint il) {
-    const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
-    const float d2 = d1 / 256.f;
-    const float  m = xb.m;
-    const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
-    const uint16_t mask1 = mask0 << 8;
-
-    mat4 reg;
-    for (int i=0;i<8;i++) {
-        uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
-        reg[i/2][2*(i%2)+0] = ((b & mask0) * d1) + m;
-        reg[i/2][2*(i%2)+1] = ((b & mask1) * d2) + m;
-    }
-    return reg;
-}
-
-#define sizeof_block_q4_k 144
-struct block_q4_k {
-    float16_t d;
-    float16_t dmin;
-    uint8_t scales[K_SCALE_SIZE];
-    uint8_t qs[QK_K/2];
-};
-
-#define sizeof_block_q6_k 210
-struct block_q6_k {
-    uint8_t ql[QK_K/2];      // quants, lower 4 bits
-    uint8_t qh[QK_K/4];      // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
-    float16_t d;             // super-block scale
-};
-mat4 dequantize_q6_k(const block_q6_k xb, uint il) {
-    const float16_t d_all = xb.d;
-
-    const uint qlIndex = 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
-    const uint qhIndex = 32*(il/8) + 16*(il&1);
-    float16_t sc = xb.scales[(il%2) + 2 * ((il/2))];
-    il = (il/2) & 3;
-
-    const uint16_t  kmask1 = il>1 ? uint16_t(il>2 ? 192 : 48) : uint16_t(il>0 ? 12 : 3);
-    const uint16_t  kmask2 = il>1 ? uint8_t(0xF0)             : uint8_t(0x0F);
-    const float16_t coef   = il>1 ? float16_t(1.f/16.f)       : float16_t(1.f);
-    const float16_t ml = float16_t(d_all * sc * 32.f);
-    const float16_t dl = float16_t(d_all * sc * coef);
-    mat4 reg;
-    for (int i = 0; i < 16; ++i) {
-        const float16_t q = (il&1) != 0 ? ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 2))
-                                        : ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 4));
-        reg[i/4][i%4] = dl * q - ml;
-    }
-    return reg;
-}
-
-
-#define QK8_0 32
-// struct block_q8_0 {
-//     float16_t d;         // delta
-//     int8_t    qs[QK8_0]; // quants
-// };
-#define sizeof_block_q8_0 34
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_add.comp b/ggml/src/ggml-kompute/kompute-shaders/op_add.comp
deleted file mode 100644
index b7b76a79db..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_add.comp
+++ /dev/null
@@ -1,58 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1024) in;
-
-layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb00;
-    int nb01;
-    int nb02;
-    int nb03;
-    int ne10;
-    int ne11;
-    int ne12;
-    int ne13;
-    int nb10;
-    int nb11;
-    int nb12;
-    int nb13;
-    int ne0;
-    int nb0;
-    int nb1;
-    int nb2;
-    int nb3;
-  //int offs; // TODO: needed for GGML_OP_ACC, see metal code
-} pcs;
-
-// general-purpose kernel for addition of two tensors
-// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
-// cons: not very efficient
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const uint i13 = i03 % pcs.ne13;
-    const uint i12 = i02 % pcs.ne12;
-    const uint i11 = i01 % pcs.ne11;
-
-    int offs = 0; // TMP (see above)
-
-    uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + offs) / 4);
-    uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11       ) / 4);
-    uint dst_off  = uint((i03*pcs.nb3  + i02*pcs.nb2  + i01*pcs.nb1  + offs) / 4);
-
-    for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
-        const uint i10 = i0 % pcs.ne10;
-        out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] + inB[pcs.inBOff + src1_off + i10];
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp b/ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp
deleted file mode 100644
index 2376a6b8f0..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp
+++ /dev/null
@@ -1,25 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    uint row;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 4;
-
-    for (uint x = 0; x < 4; x++) {
-        const uint i = baseIndex + x;
-        out_[i + pcs.outOff] = inA[i + pcs.inAOff] + inB[(i % pcs.row) + pcs.inBOff];
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp
deleted file mode 100644
index d57247d2dc..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define IN_TYPE float16_t
-#define IN_TYPE_SIZE 2
-#define OUT_TYPE float16_t
-#define OUT_TYPE_SIZE 2
-
-layout(local_size_x = 1024) in;
-
-layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
-layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    int ne1;
-    int ne2;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
-
-    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
-    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
-    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
-    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
-
-    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
-
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
-        out_[dst_data+i00] = OUT_TYPE(in_[src]);
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp
deleted file mode 100644
index b568bcd7b2..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define IN_TYPE float16_t
-#define IN_TYPE_SIZE 2
-#define OUT_TYPE float
-#define OUT_TYPE_SIZE 4
-
-layout(local_size_x = 1024) in;
-
-layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
-layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    int ne1;
-    int ne2;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
-
-    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
-    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
-    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
-    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
-
-    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
-
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
-        out_[dst_data+i00] = OUT_TYPE(in_[src]);
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp
deleted file mode 100644
index 99b2283430..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define IN_TYPE float
-#define IN_TYPE_SIZE 4
-#define OUT_TYPE float16_t
-#define OUT_TYPE_SIZE 2
-
-layout(local_size_x = 1024) in;
-
-layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
-layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    int ne1;
-    int ne2;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
-
-    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
-    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
-    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
-    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
-
-    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
-
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
-        out_[dst_data+i00] = OUT_TYPE(in_[src]);
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp
deleted file mode 100644
index 2fc998492b..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define IN_TYPE float
-#define IN_TYPE_SIZE 4
-#define OUT_TYPE float
-#define OUT_TYPE_SIZE 4
-
-layout(local_size_x = 1024) in;
-
-layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
-layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    int ne1;
-    int ne2;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
-
-    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
-    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
-    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
-    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
-
-    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
-
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
-        out_[dst_data+i00] = OUT_TYPE(in_[src]);
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp b/ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp
deleted file mode 100644
index 291c3fc189..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp
+++ /dev/null
@@ -1,30 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-    uint n_past;
-    int ne00;
-    int ne01;
-} pcs;
-
-void main() {
-    const uint i02 = gl_WorkGroupID.z;
-    const uint i01 = gl_WorkGroupID.y;
-    const uint i00 = gl_WorkGroupID.x;
-
-    const uint index = i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00 + i00;
-
-    if (i00 > pcs.n_past + i01) {
-        out_[index + pcs.outOff] = uintBitsToFloat(0xFF800000);
-    } else {
-        out_[index + pcs.outOff] = in_[index + pcs.inOff];
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp b/ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp
deleted file mode 100644
index 9d8c53710a..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp
+++ /dev/null
@@ -1,22 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 8;
-
-    for (uint x = 0; x < 8; x++) {
-        const uint i = baseIndex + x;
-        const float y = in_[i + pcs.inOff];
-        out_[i + pcs.outOff] = 0.5*y*(1.0 + tanh(clamp(SQRT_2_OVER_PI*y*(1.0 + GELU_COEF_A*y*y), -15.0, 15.0)));
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp
deleted file mode 100644
index 1a5581b23a..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp
+++ /dev/null
@@ -1,17 +0,0 @@
-void main() {
-    const uint i = gl_WorkGroupID.x;
-    const int r = inB[i + pcs.inBOff];
-
-    int z = 0;
-    for (uint ind = gl_LocalInvocationID.x; ind < pcs.ne00/16; ind += gl_WorkGroupSize.x) {
-        const uint inIndex = (r * pcs.nb01 + pcs.inAOff) + ind/NL * SIZE_OF_BLOCK;
-        const mat4 result = dequantize_block(inIndex, ind%NL);
-        for (uint j = 0; j < 4; ++j) {
-            for (uint k = 0; k < 4; ++k) {
-                const uint outIndex = i * pcs.nb1/BYTES_FOR_TYPE + pcs.outOff + z;
-                out_[outIndex] = result[j][k];
-                ++z;
-            }
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp
deleted file mode 100644
index 48c9361081..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp
+++ /dev/null
@@ -1,31 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-void dequantize_row_f16(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
-    for (int j = 0; j < k; j++) {
-        out_[y + j] = inA[x + j];
-    }
-}
-
-void main() {
-    const uint i = gl_WorkGroupID.x;
-    const int r = inB[i + pcs.inBOff];
-
-    dequantize_row_f16(r*pcs.nb01/2/*bytes for float16*/ + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp
deleted file mode 100644
index 9d7acdaf8a..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp
+++ /dev/null
@@ -1,31 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { float inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-void dequantize_row_f32(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
-    for (int j = 0; j < k; j++) {
-        out_[y + j] = inA[x + j];
-    }
-}
-
-void main() {
-    const uint i = gl_WorkGroupID.x;
-    const int r = inB[i + pcs.inBOff];
-
-    dequantize_row_f32(r*pcs.nb01/4 + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp
deleted file mode 100644
index 32b2e891e8..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp
+++ /dev/null
@@ -1,38 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define NL 2
-#define BYTES_FOR_TYPE 4 /*bytes for float*/
-#define SIZE_OF_BLOCK sizeof_block_q4_0
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-block_q4_0 get_unaligned_block_q4_0(uint index) {
-    block_q4_0 fres;
-    fres.d = u8BufToFloat16(inA, index);
-    [[unroll]] for (uint it = 0; it != QK4_0 / 2; it++) {
-        fres.qs[it] = inA[index+2+it];
-    }
-    return fres;
-}
-
-mat4 dequantize_block(uint index, uint il) {
-    const block_q4_0 block = get_unaligned_block_q4_0(index);
-    return dequantize_q4_0(block, il);
-}
-
-#include "op_getrows.comp"
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp
deleted file mode 100644
index 87f2fbe17b..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp
+++ /dev/null
@@ -1,39 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define NL 2
-#define BYTES_FOR_TYPE 4 /*bytes for float*/
-#define SIZE_OF_BLOCK sizeof_block_q4_1
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-block_q4_1 get_unaligned_block_q4_1(uint index) {
-    block_q4_1 fres;
-    fres.d = u8BufToFloat16(inA, index);
-    fres.m = u8BufToFloat16(inA, index+2);
-    [[unroll]] for (uint it = 0; it != QK4_1 / 2; it++) {
-        fres.qs[it] = inA[index+4+it];
-    }
-    return fres;
-}
-
-mat4 dequantize_block(uint index, uint il) {
-    const block_q4_1 block = get_unaligned_block_q4_1(index);
-    return dequantize_q4_1(block, il);
-}
-
-#include "op_getrows.comp"
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp
deleted file mode 100644
index 9ce3545d1e..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp
+++ /dev/null
@@ -1,44 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define NL 16
-#define BYTES_FOR_TYPE 4 /*bytes for float*/
-#define SIZE_OF_BLOCK sizeof_block_q6_k
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-block_q6_k get_unaligned_block_q6_k(uint index) {
-    block_q6_k fres;
-    [[unroll]] for (uint it = 0; it != QK_K / 2; it++) {
-        fres.ql[it] = inA[index + it];
-    }
-    [[unroll]] for (uint it = 0; it != QK_K / 4; it++) {
-        fres.qh[it] = inA[index + QK_K/2 + it];
-    }
-    [[unroll]] for (uint it = 0; it != QK_K / 16; it++) {
-        fres.scales[it] = int8_t(inA[index + QK_K/2 + QK_K/4 + it]);
-    }
-    fres.d = u8BufToFloat16(inA, index + QK_K/2 + QK_K/4 + QK_K/16);
-    return fres;
-}
-
-mat4 dequantize_block(uint index, uint il) {
-    const block_q6_k block = get_unaligned_block_q6_k(index);
-    return dequantize_q6_k(block, il);
-}
-
-#include "op_getrows.comp"
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul.comp
deleted file mode 100644
index c92647c4db..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1024) in;
-
-layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb00;
-    int nb01;
-    int nb02;
-    int nb03;
-    int ne10;
-    int ne11;
-    int ne12;
-    int ne13;
-    int nb10;
-    int nb11;
-    int nb12;
-    int nb13;
-    int ne0;
-    int nb0;
-    int nb1;
-    int nb2;
-    int nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const uint i13 = i03 % pcs.ne13;
-    const uint i12 = i02 % pcs.ne12;
-    const uint i11 = i01 % pcs.ne11;
-
-    uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01) / 4);
-    uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11) / 4);
-    uint dst_off  = uint((i03*pcs.nb3  + i02*pcs.nb2  + i01*pcs.nb1)  / 4);
-
-    for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
-        const uint i10 = i0 % pcs.ne10;
-        out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] * inB[pcs.inBOff + src1_off + i10];
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp
deleted file mode 100644
index 0ab1b2fc20..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp
+++ /dev/null
@@ -1,69 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#extension GL_KHR_shader_subgroup_arithmetic : require
-
-layout(local_size_x_id = 0) in;
-
-layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { float inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne10;
-    int ne11;
-    int ne12;
-    uint nb10;
-    uint nb11;
-    uint nb12;
-    uint nb13;
-    int ne0;
-    int ne1;
-    uint r2;
-    uint r3;
-} pcs;
-
-#define N_F16_F32 4
-
-void main() {
-    const uint r0 = gl_WorkGroupID.x;
-    const uint rb = gl_WorkGroupID.y*N_F16_F32;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    const uint offset0 = r0*pcs.nb01 + (i12/pcs.r2)*pcs.nb02 + (i13/pcs.r3)*pcs.nb03;
-
-    const uint x = offset0 / 2 + pcs.inAOff; // Based from inA
-
-    for (uint row = 0; row < N_F16_F32; ++row) {
-        uint r1 = rb + row;
-        if (r1 >= pcs.ne11) {
-            break;
-        }
-
-        const uint y = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff;
-
-        float sumf = 0;
-        for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
-            sumf += float(inA[x+i]) * float(inB[y+i]);
-        }
-
-        const float all_sum = subgroupAdd(sumf);
-        if (subgroupElect()) {
-            out_[im*pcs.ne1*pcs.ne0 + r1*pcs.ne0 + r0 + pcs.outOff] = all_sum;
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp
deleted file mode 100644
index d1ca4ad6c2..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp
+++ /dev/null
@@ -1,51 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#extension GL_KHR_shader_subgroup_arithmetic : require
-#extension GL_EXT_debug_printf : enable
-
-// device subgroup size
-layout (local_size_x_id = 0) in;
-
-layout(binding = 0) readonly buffer tensorInA { float inA[]; };
-layout(binding = 1) readonly buffer tensorInB { float inB[]; };
-layout(binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout(push_constant) uniform parameter {
-  uint inAOff;
-  uint inBOff;
-  uint outOff;
-  int ne00;
-  int ne01;
-  int ne02;
-  int ne11;
-  int ne12;
-  uint nb01;
-  uint nb02;
-  uint nb11;
-  uint nb12;
-  uint nb1;
-  uint nb2;
-}
-pcs;
-
-
-void main() {
-  uvec3 gid = gl_WorkGroupID;
-
-  uint bc_ab = pcs.ne12 > pcs.ne02 ? gid.z / (pcs.ne12 / pcs.ne02) : gid.z;
-  uint bc_ba = pcs.ne02 > pcs.ne12 ? gid.z / (pcs.ne02 / pcs.ne12) : gid.z;
-
-  const uint x = (gid.x*pcs.nb01 + bc_ab*pcs.nb02) / 4 + pcs.inAOff; // Based from inA
-  const uint y = (gid.y*pcs.nb11 + bc_ba*pcs.nb12) / 4 + pcs.inBOff; // based from inB
-  float sum = 0.0f;
-  for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
-      sum += float(inA[x+i]) * float(inB[y+i]);
-  }
-
-  const float all_sum = subgroupAdd(sum);
-  if (subgroupElect()) {
-    out_[gid.z*(pcs.nb2/4) + gid.y*(pcs.nb1/4) + gid.x + pcs.outOff] = all_sum;
-  }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp
deleted file mode 100644
index b0cea8bbe6..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp
+++ /dev/null
@@ -1,33 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define BLOCKS_IN_QUANT QK4_0
-#define SIZE_OF_BLOCK sizeof_block_q4_0
-#define N_ROWS 4
-
-#include "op_mul_mv_q_n_pre.comp"
-
-// The q4_0 version of this function
-float block_q_n_dot_y(uint block_index, uint yb, uint il) {
-    vec2 acc = vec2(0.0, 0.0);
-    const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
-    float d = float(u8BufToFloat16(inA, index));
-    float sumy = 0.0f;
-    for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
-        const uint16_t b = u8BufToU16(inA, index + 2 + il + i);
-
-        const float yl0 = inB[yb + i];
-        const float yl1 = inB[yb + i + 1];
-        const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
-        const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
-
-        sumy += yl0 + yl1 + yl8 + yl9;
-
-        acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
-        acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
-    }
-    return d * (sumy * -8.f + acc[0] + acc[1]);
-}
-
-#include "op_mul_mv_q_n.comp"
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp
deleted file mode 100644
index 8582c61a3b..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp
+++ /dev/null
@@ -1,35 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define BLOCKS_IN_QUANT QK4_1
-#define SIZE_OF_BLOCK sizeof_block_q4_1
-#define N_ROWS 4
-
-#include "op_mul_mv_q_n_pre.comp"
-
-// The q4_1 version of this function
-float block_q_n_dot_y(uint block_index, uint yb, uint il) {
-    vec2 acc = vec2(0.0, 0.0);
-    const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
-    float d = float(u8BufToFloat16(inA, index));
-    float m = float(u8BufToFloat16(inA, index+2));
-
-    float sumy = 0.0f;
-    for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
-        const uint16_t b = u8BufToU16(inA, index + 4 + il + i);
-
-        const float yl0 = inB[yb + i];
-        const float yl1 = inB[yb + i + 1];
-        const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
-        const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
-
-        sumy += yl0 + yl1 + yl8 + yl9;
-
-        acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
-        acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
-    }
-    return d * (acc[0] + acc[1]) + sumy * m;
-}
-
-#include "op_mul_mv_q_n.comp"
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp
deleted file mode 100644
index a5752a3a00..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp
+++ /dev/null
@@ -1,140 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define N_DST 4
-#define SIZE_OF_BLOCK sizeof_block_q4_k
-
-layout(local_size_x = 4) in;
-layout(local_size_y = 8) in;
-layout(local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { block_q4_k inA[]; };
-layout (binding = 1) readonly buffer tensorInB { float inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int ne10;
-    int ne0;
-    int ne1;
-    int ne01;
-    int ne02;
-    int ne12;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    uint nb11;
-    uint nb12;
-    uint nb13;
-    uint r2;
-    uint r3;
-} pcs;
-
-void main() {
-    const uint16_t kmask1 = uint16_t(0x3f3f);
-    const uint16_t kmask2 = uint16_t(0x0f0f);
-    const uint16_t kmask3 = uint16_t(0xc0c0);
-
-    const uint ix = gl_SubgroupInvocationID/8;  // 0...3
-    const uint it = gl_SubgroupInvocationID%8;  // 0...7
-    const uint iq = it/4;     // 0 or 1
-    const uint ir = it%4;     // 0...3
-
-    const uint nb = pcs.ne00/QK_K;
-
-    const uint r0 = gl_WorkGroupID.x;
-    const uint r1 = gl_WorkGroupID.y;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint first_row = r0 * N_DST;
-    const uint ib_row = first_row * nb;
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    const uint offset0 = first_row*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
-    const uint offset1 =        r1*pcs.nb11 + (i12       )*pcs.nb12 + (i13       )*pcs.nb13;
-
-    const uint xblk = offset0 + pcs.inAOff;
-    const uint y = (offset1 / 4) + pcs.inBOff;
-
-    float yl[16];
-    float yh[16];
-    float sumf[N_DST] = {0.f, 0.f, 0.f, 0.f};
-    float all_sum = 0.f;
-
-    uint y4 = y + ix * QK_K + 64 * iq + 8 * ir;
-
-    for (uint ib = ix; ib < nb; ib += 4) {
-        const uint blk_idx = ib + xblk;
-
-        float sumy[4] = {0.f, 0.f, 0.f, 0.f};
-        for (int i = 0; i < 8; ++i) {
-            yl[i+0] = inB[y4+i+  0]; sumy[0] += yl[i+0];
-            yl[i+8] = inB[y4+i+ 32]; sumy[1] += yl[i+8];
-            yh[i+0] = inB[y4+i+128]; sumy[2] += yh[i+0];
-            yh[i+8] = inB[y4+i+160]; sumy[3] += yh[i+8];
-        }
-
-        for (int row = 0; row < N_DST; row++) {
-            uint row_idx = row * (pcs.nb01 / SIZE_OF_BLOCK);
-
-            uint16_t sc_0 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 0);
-            uint16_t sc_1 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 2);
-            uint16_t sc_2 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 4);
-            uint16_t sc_3 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 6);
-            uint16_t sc_4 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 8);
-
-            uint16_t sc16[4];
-            sc16[0] = sc_0 & kmask1;
-            sc16[1] = sc_2 & kmask1;
-            sc16[2] = ((sc_4 >> 0) & kmask2) | ((sc_0 & kmask3) >> 2);
-            sc16[3] = ((sc_4 >> 4) & kmask2) | ((sc_2 & kmask3) >> 2);
-
-            float acc1[4] = {0.f, 0.f, 0.f, 0.f};
-            float acc2[4] = {0.f, 0.f, 0.f, 0.f};
-            for (int i = 0; i < 8; i += 2) {
-                uint16_t q1 = u8BufToU16(inA[blk_idx + row_idx].qs, 32 * iq + 8 * ir + i);
-                uint16_t q2 = u8BufToU16(inA[blk_idx + row_idx].qs, 64 + 32 * iq + 8 * ir + i);
-                acc1[0] += yl[i+0] * (q1 & 0x000F);
-                acc1[1] += yl[i+1] * (q1 & 0x0F00);
-                acc1[2] += yl[i+8] * (q1 & 0x00F0);
-                acc1[3] += yl[i+9] * (q1 & 0xF000);
-                acc2[0] += yh[i+0] * (q2 & 0x000F);
-                acc2[1] += yh[i+1] * (q2 & 0x0F00);
-                acc2[2] += yh[i+8] * (q2 & 0x00F0);
-                acc2[3] += yh[i+9] * (q2 & 0xF000);
-            }
-
-            uint8_t sc8_0 = uint8_t(sc16[0] & 0xFF);
-            uint8_t sc8_1 = uint8_t(sc16[0] >> 8 );
-            uint8_t sc8_2 = uint8_t(sc16[1] & 0xFF);
-            uint8_t sc8_3 = uint8_t(sc16[1] >> 8 );
-            uint8_t sc8_4 = uint8_t(sc16[2] & 0xFF);
-            uint8_t sc8_5 = uint8_t(sc16[2] >> 8 );
-            uint8_t sc8_6 = uint8_t(sc16[3] & 0xFF);
-            uint8_t sc8_7 = uint8_t(sc16[3] >> 8 );
-
-            float dall = float(inA[blk_idx + row_idx].d);
-            float dmin = float(inA[blk_idx + row_idx].dmin);
-            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc8_0 +
-                               (acc1[2] + 1.f/256.f * acc1[3]) * sc8_1 * 1.f/16.f +
-                               (acc2[0] + 1.f/256.f * acc2[1]) * sc8_4 +
-                               (acc2[2] + 1.f/256.f * acc2[3]) * sc8_5 * 1.f/16.f) -
-                dmin * (sumy[0] * sc8_2 + sumy[1] * sc8_3 + sumy[2] * sc8_6 + sumy[3] * sc8_7);
-        }
-
-        y4 += 4 * QK_K;
-    }
-
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = subgroupAdd(sumf[row]);
-        if (subgroupElect()) {
-            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = all_sum;
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp
deleted file mode 100644
index d331d1a705..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp
+++ /dev/null
@@ -1,106 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define SIZE_OF_BLOCK sizeof_block_q6_k
-
-layout(local_size_x_id = 0) in;
-layout(local_size_y_id = 1) in;
-layout(local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { float inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int ne10;
-    int ne0;
-    int ne1;
-    int ne01;
-    int ne02;
-    int ne12;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    uint nb11;
-    uint nb12;
-    uint nb13;
-    uint r2;
-    uint r3;
-} pcs;
-
-void main() {
-    const uint8_t kmask1 = uint8_t(0x03);
-    const uint8_t kmask2 = uint8_t(0x0C);
-    const uint8_t kmask3 = uint8_t(0x30);
-    const uint8_t kmask4 = uint8_t(0xC0);
-
-    const uint nb = pcs.ne00/QK_K;
-
-    const uint r0 = gl_WorkGroupID.x;
-    const uint r1 = gl_WorkGroupID.y;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint row = (r0 * gl_NumSubgroups + gl_SubgroupID);
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    const uint x = row*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
-    const uint yy = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff;
-
-    float sumf = 0;
-
-    // bits of invocation ID for gl_SubgroupSize=32:
-    //  x   x   x   x   x
-    //  4   3   2   1   0
-    // (     tid     ) ix
-    //  ip (   il    )
-
-    const uint block_stride = gl_SubgroupSize / 16;         // number of blocks each subgroup processes
-    const uint tid  = gl_SubgroupInvocationID/block_stride; // first block_stride groups have tid=0
-    const uint ix   = gl_SubgroupInvocationID%block_stride; // first block is 0..block_stride-1
-    const uint ip   = tid/8;        // first or second half of block (0 or 1)
-    const uint il   = tid%8;        // each half has 8 parts, one per scale
-    const uint n    = 4;            // 4 scales at a time (and 4 sums)
-    const uint l0   = n*il;         // offset into half-block, 0..28
-    const uint is   = 8*ip + l0/16; // 0, 1, 8, 9
-
-    const uint y_offset = 128*ip + l0;
-    const uint q_offset_l = 64*ip + l0;
-    const uint q_offset_h = 32*ip + l0;
-
-    for (uint i = ix; i < nb; i += block_stride) {
-
-        const uint baseIndex = (x + i) * SIZE_OF_BLOCK + pcs.inAOff;
-
-        const uint qlIndex = q_offset_l;
-        const uint q2Index = qlIndex + QK_K/8;
-        const uint qhIndex = q_offset_h;
-        const uint y = yy + i * QK_K + y_offset;
-
-        float sums[4] = {0.0f, 0.0f, 0.0f, 0.0f};
-        for (uint l = 0; l < n; ++l) {
-            const uint8_t currentQ1 = inA[baseIndex + qlIndex + l];
-            const uint8_t currentQ2 = inA[baseIndex + q2Index + l];
-            const uint8_t currentQh = inA[baseIndex + QK_K/2 + qhIndex + l];
-
-            sums[0] += inB[y+l+ 0] * (int8_t((currentQ1 & 0xF) | ((currentQh & kmask1) << 4)) - 32);
-            sums[1] += inB[y+l+32] * (int8_t((currentQ2 & 0xF) | ((currentQh & kmask2) << 2)) - 32);
-            sums[2] += inB[y+l+64] * (int8_t((currentQ1  >> 4) | ((currentQh & kmask3) << 0)) - 32);
-            sums[3] += inB[y+l+96] * (int8_t((currentQ2  >> 4) | ((currentQh & kmask4) >> 2)) - 32);
-        }
-
-        float d = u8BufToFloat16(inA, baseIndex + QK_K/2 + QK_K/4 + QK_K/16);
-        sumf += d * (sums[0] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + is]) + sums[1] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 2 + is]) + sums[2] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 4 + is]) + sums[3] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 6 + is]));
-    }
-
-    const float tot = subgroupAdd(sumf);
-    if (subgroupElect()) {
-        out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + row + pcs.outOff] = tot;
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp
deleted file mode 100644
index 34d015e90b..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp
+++ /dev/null
@@ -1,73 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#include "op_mul_mv_q_n_pre.comp"
-
-#define SIZE_OF_D 2
-
-#define N_DST 4 // each SIMD group works on 4 rows
-#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
-#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
-
-#define NB_Q8_0 8
-
-void main() {
-    // NB: hack to make compatible with AMD GPUs that have a subgroup size of 64
-    if (gl_SubgroupInvocationID > 31)
-        return;
-
-    const int nr  = N_DST;
-    const int nsg = N_SIMDGROUP;
-    const int nw  = N_SIMDWIDTH;
-
-    const int nb = pcs.ne00/QK8_0;
-    const uint r0 = gl_WorkGroupID.x;
-    const uint r1 = gl_WorkGroupID.y;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint first_row = (r0 * nsg + gl_SubgroupID) * nr;
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    const uint offset0 = first_row * nb + (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
-
-    const uint x = offset0*sizeof_block_q8_0 + pcs.inAOff; // Based from inA
-    const uint y = r1*pcs.ne10 + im*pcs.ne00*pcs.ne1 + pcs.inBOff; // based from inB
-
-    float yl[NB_Q8_0];
-    float sumf[N_DST]={0.f, 0.f, 0.f, 0.f};
-
-    const uint ix = gl_SubgroupInvocationID.x/4;
-    const uint il = gl_SubgroupInvocationID.x%4;
-
-    uint yb = y + ix * QK8_0 + NB_Q8_0*il;
-
-    // each thread in a SIMD group deals with NB_Q8_0 quants at a time
-    for (uint ib = ix; ib < nb; ib += nw/4) {
-        for (int i = 0; i < NB_Q8_0; ++i) {
-            yl[i] = inB[yb + i];
-        }
-
-        for (int row = 0; row < nr; row++) {
-            const uint block_offset = (ib+row*nb) * sizeof_block_q8_0;
-            float sumq = 0.f;
-            for (int iq = 0; iq < NB_Q8_0; ++iq) {
-                const int8_t qs_iq = int8_t(inA[x + block_offset + SIZE_OF_D + NB_Q8_0*il + iq]);
-                sumq += qs_iq * yl[iq];
-            }
-            const float16_t d = u8BufToFloat16(inA, x + block_offset);
-            sumf[row] += sumq*d;
-        }
-
-        yb += NB_Q8_0 * nw;
-    }
-
-    for (int row = 0; row < nr; ++row) {
-        const float tot = subgroupAdd(sumf[row]);
-        if (subgroupElect() && first_row + row < pcs.ne01) {
-            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row] = tot;
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp
deleted file mode 100644
index a6517cc1f1..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-void main() {
-    // NB: hack to make compatible with AMD GPUs that have a subgroup size of 64
-    if (gl_SubgroupInvocationID > 31)
-        return;
-
-    const uint nb = uint(pcs.ne00/BLOCKS_IN_QUANT);
-
-    const uint r0 = gl_WorkGroupID.x;
-    const uint r1 = gl_WorkGroupID.y;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint first_row = (r0 * gl_NumSubgroups + gl_SubgroupID) * N_ROWS;
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    // pointers to src0 rows
-    uint ax[N_ROWS];
-    for (int row = 0; row < N_ROWS; ++row) {
-        const uint offset0 = (first_row + row)*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
-
-        ax[row] = offset0 + pcs.inAOff;
-    }
-
-    const uint y = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff;
-
-    float sumf[N_ROWS] = {0.0f, 0.0f, 0.0f, 0.0f};
-
-    const uint ix = gl_SubgroupInvocationID/2;
-    const uint il = (BLOCKS_IN_QUANT/4)*(gl_SubgroupInvocationID%2);
-
-    uint yb = y + ix * BLOCKS_IN_QUANT + il;
-
-    //debugPrintfEXT("gl_NumSubgroups=%d, gl_SubgroupID=%d, gl_SubgroupInvocationID=%d, glSubgroupSize=%d, gl_WorkGroupSize.x=%d, gl_WorkGroupSize.y=%d, gl_WorkGroupSize.z=%d\n",
-    //    gl_NumSubgroups, gl_SubgroupID, gl_SubgroupInvocationID, gl_SubgroupSize,
-    //    gl_WorkGroupSize.x, gl_WorkGroupSize.y, gl_WorkGroupSize.z);
-
-    for (uint ib = ix; ib < nb; ib += 16) {
-        for (int row = 0; row < N_ROWS; row++) {
-            sumf[row] += block_q_n_dot_y(ax[row] + ib, yb, il);
-        }
-
-        yb += BLOCKS_IN_QUANT * 16;
-    }
-
-    for (int row = 0; row < N_ROWS; ++row) {
-        const float tot = subgroupAdd(sumf[row]);
-        if (first_row + row < pcs.ne01 && subgroupElect()) {
-            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = tot;
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp
deleted file mode 100644
index a9a2f22180..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp
+++ /dev/null
@@ -1,28 +0,0 @@
-layout(local_size_x_id = 0) in;
-layout(local_size_y = 8) in;
-layout(local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { float inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int  ne00;
-    int  ne01;
-    int  ne02;
-    int  ne10;
-    int  ne12;
-    int  ne0;
-    int  ne1;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    uint nb11;
-    uint nb12;
-    uint nb13;
-    uint r2;
-    uint r3;
-} pcs;
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_norm.comp b/ggml/src/ggml-kompute/kompute-shaders/op_norm.comp
deleted file mode 100644
index ad0c3c01b9..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_norm.comp
+++ /dev/null
@@ -1,84 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 256) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-    uint ne00;
-    uint nb01;
-    float eps;
-} pcs;
-
-shared float sum[gl_WorkGroupSize.x];
-
-void main() {
-    const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
-    // MEAN
-    // parallel sum
-    sum[gl_LocalInvocationID.x] = 0.0;
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        sum[gl_LocalInvocationID.x] += in_[x+i00];
-    }
-
-    // reduce
-    barrier();
-    memoryBarrierShared();
-    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
-        if (gl_LocalInvocationID.x < i) {
-            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
-        }
-        barrier();
-        memoryBarrierShared();
-    }
-
-    // broadcast
-    if (gl_LocalInvocationID.x == 0) {
-        sum[0] /= float(pcs.ne00);
-    }
-    barrier();
-    memoryBarrierShared();
-    const float mean = sum[0];
-
-    // recenter
-    const uint y = (gl_WorkGroupID.x*pcs.ne00) + pcs.outOff; // Based from out_
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        out_[y+i00] = in_[x+i00] - mean;
-    }
-
-    // VARIANCE
-    // parallel sum
-    sum[gl_LocalInvocationID.x] = 0.0;
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        sum[gl_LocalInvocationID.x] += out_[y+i00] * out_[y+i00];
-    }
-
-    // reduce
-    barrier();
-    memoryBarrierShared();
-    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
-        if (gl_LocalInvocationID.x < i) {
-            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
-        }
-        barrier();
-        memoryBarrierShared();
-    }
-
-    // broadcast
-    if (gl_LocalInvocationID.x == 0) {
-        sum[0] /= float(pcs.ne00);
-    }
-    barrier();
-    memoryBarrierShared();
-    const float variance = sum[0];
-
-    const float scale = 1.0f/sqrt(variance + pcs.eps);
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        out_[y+i00] *= scale;
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_relu.comp b/ggml/src/ggml-kompute/kompute-shaders/op_relu.comp
deleted file mode 100644
index 52a601fe6d..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_relu.comp
+++ /dev/null
@@ -1,21 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 4;
-
-    for (uint x = 0; x < 4; x++) {
-        const uint i = baseIndex + x;
-        out_[i + pcs.outOff] = max(0.0, in_[i + pcs.inOff]);
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp
deleted file mode 100644
index da658c1601..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp
+++ /dev/null
@@ -1,53 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 512) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-    uint ne00;
-    uint nb01;
-    float eps;
-} pcs;
-
-shared float sum[gl_WorkGroupSize.x];
-
-void main() {
-    const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
-
-    // parallel sum
-    sum[gl_LocalInvocationID.x] = 0.0;
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        sum[gl_LocalInvocationID.x] += in_[x+i00] * in_[x+i00];
-    }
-
-    // reduce
-    barrier();
-    memoryBarrierShared();
-    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
-        if (gl_LocalInvocationID.x < i) {
-            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
-        }
-        barrier();
-        memoryBarrierShared();
-    }
-
-    // broadcast
-    if (gl_LocalInvocationID.x == 0) {
-        sum[0] /= float(pcs.ne00);
-    }
-    barrier();
-    memoryBarrierShared();
-
-    const float scale = 1.0f/sqrt(sum[0] + pcs.eps);
-
-    const uint y = (gl_WorkGroupID.x*pcs.ne00) + pcs.outOff; // Based from out_
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        out_[y+i00] = in_[x+i00] * scale;
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp
deleted file mode 100644
index 63659cbfe5..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "rope_common.comp"
-
-layout(binding = 0) buffer restrict readonly  tensorInA { float16_t inA[]; };
-layout(binding = 1) buffer restrict readonly  tensorInB { int       inB[]; };
-layout(binding = 2) buffer restrict readonly  tensorInC { float     inC[]; };
-layout(binding = 3) buffer restrict writeonly tensorOut { float16_t out_[]; };
-
-void main() {
-    const uint i3 = gl_WorkGroupID.z;
-    const uint i2 = gl_WorkGroupID.y;
-    const uint i1 = gl_WorkGroupID.x;
-
-    float corr_dims[2];
-    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
-
-    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
-
-    float theta_base = float(inB[pcs.inBOff + i2]);
-    float inv_ndims = -1.f/pcs.n_dims;
-
-    float cos_theta;
-    float sin_theta;
-
-    for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
-        if (i0 < pcs.n_dims) {
-            uint ic = i0/2;
-
-            float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);
-
-            const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;
-
-            rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + ic*pcs.nb00) / 2) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + ic*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
-
-            const float x0 = float(inA[src]);
-            const float x1 = float(inA[src+pcs.n_dims/2]);
-
-            out_[dst_data]              = float16_t(x0*cos_theta - x1*sin_theta);
-            out_[dst_data+pcs.n_dims/2] = float16_t(x0*sin_theta + x1*cos_theta);
-        } else {
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
-
-            out_[dst_data]   = inA[src];
-            out_[dst_data+1] = inA[src+1];
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp
deleted file mode 100644
index 4df56204d7..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "rope_common.comp"
-
-layout(binding = 0) buffer restrict readonly  tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly  tensorInB { int       inB[]; };
-layout(binding = 2) buffer restrict readonly  tensorInC { float inC[]; };
-layout(binding = 3) buffer restrict writeonly tensorOut { float out_[]; };
-
-void main() {
-    const uint i3 = gl_WorkGroupID.z;
-    const uint i2 = gl_WorkGroupID.y;
-    const uint i1 = gl_WorkGroupID.x;
-
-    float corr_dims[2];
-    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
-
-    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
-
-    float theta_base = float(inB[pcs.inBOff + i2]);
-    float inv_ndims = -1.f/pcs.n_dims;
-
-    float cos_theta;
-    float sin_theta;
-
-    for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
-        if (i0 < pcs.n_dims) {
-            uint ic = i0/2;
-
-            float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);
-
-            const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;
-
-            rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + ic*pcs.nb00) / 4) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + ic*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
-
-            const float x0 = inA[src];
-            const float x1 = inA[src+pcs.n_dims/2];
-
-            out_[dst_data]              = x0*cos_theta - x1*sin_theta;
-            out_[dst_data+pcs.n_dims/2] = x0*sin_theta + x1*cos_theta;
-        } else {
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
-
-            out_[dst_data]   = inA[src];
-            out_[dst_data+1] = inA[src+1];
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp
deleted file mode 100644
index a3c0eda8bd..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "rope_common.comp"
-
-layout(binding = 0) buffer restrict readonly  tensorInA { float16_t inA[]; };
-layout(binding = 1) buffer restrict readonly  tensorInB { int       inB[]; };
-layout(binding = 2) buffer restrict readonly  tensorInC { float     inC[]; };
-layout(binding = 3) buffer restrict writeonly tensorOut { float16_t out_[]; };
-
-void main() {
-    const uint i3 = gl_WorkGroupID.z;
-    const uint i2 = gl_WorkGroupID.y;
-    const uint i1 = gl_WorkGroupID.x;
-
-    float corr_dims[2];
-    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
-
-    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
-
-    float theta_base = float(inB[pcs.inBOff + i2]);
-    float inv_ndims = -1.f/pcs.n_dims;
-
-    float cos_theta;
-    float sin_theta;
-
-    for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
-        if (i0 < pcs.n_dims) {
-            uint ic = i0/2;
-
-            float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);
-
-            const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;
-
-            rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
-
-            const float x0 = float(inA[src]);
-            const float x1 = float(inA[src+1]);
-
-            out_[dst_data]   = float16_t(x0*cos_theta - x1*sin_theta);
-            out_[dst_data+1] = float16_t(x0*sin_theta + x1*cos_theta);
-        } else {
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
-
-            out_[dst_data]   = inA[src];
-            out_[dst_data+1] = inA[src+1];
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp
deleted file mode 100644
index b7963ae725..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "rope_common.comp"
-
-layout(binding = 0) buffer restrict readonly  tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly  tensorInB { int   inB[]; };
-layout(binding = 2) buffer restrict readonly  tensorInC { float inC[]; };
-layout(binding = 3) buffer restrict writeonly tensorOut { float out_[]; };
-
-void main() {
-    const uint i3 = gl_WorkGroupID.z;
-    const uint i2 = gl_WorkGroupID.y;
-    const uint i1 = gl_WorkGroupID.x;
-
-    float corr_dims[2];
-    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
-
-    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
-
-    float theta_base = float(inB[pcs.inBOff + i2]);
-    float inv_ndims = -1.f/pcs.n_dims;
-
-    float cos_theta;
-    float sin_theta;
-
-    for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
-        if (i0 < pcs.n_dims) {
-            uint ic = i0/2;
-
-            float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);
-
-            const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;
-
-            rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
-
-            const float x0 = inA[src];
-            const float x1 = inA[src+1];
-
-            out_[dst_data]   = x0*cos_theta - x1*sin_theta;
-            out_[dst_data+1] = x0*sin_theta + x1*cos_theta;
-        } else {
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
-
-            out_[dst_data]   = inA[src];
-            out_[dst_data+1] = inA[src+1];
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_scale.comp b/ggml/src/ggml-kompute/kompute-shaders/op_scale.comp
deleted file mode 100644
index bdae267382..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_scale.comp
+++ /dev/null
@@ -1,19 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-    float scale;
-} pcs;
-
-void main() {
-    const uint i = gl_WorkGroupID.x;
-    out_[i + pcs.outOff] = in_[i + pcs.inOff] * pcs.scale;
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp b/ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp
deleted file mode 100644
index ada69754b2..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp
+++ /dev/null
@@ -1,23 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-    float scale;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 8;
-
-    for (uint x = 0; x < 8; x++) {
-        const uint i = baseIndex + x;
-        out_[i + pcs.outOff] = in_[i + pcs.inOff] * pcs.scale;
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_silu.comp b/ggml/src/ggml-kompute/kompute-shaders/op_silu.comp
deleted file mode 100644
index 0fb8e4b740..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_silu.comp
+++ /dev/null
@@ -1,22 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 4;
-
-    for (uint x = 0; x < 4; x++) {
-        const uint i = baseIndex + x;
-        const float y = in_[i + pcs.inOff];
-        out_[i + pcs.outOff] = y / (1.0 + exp(-y));
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp b/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp
deleted file mode 100644
index 4165295bf4..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp
+++ /dev/null
@@ -1,72 +0,0 @@
-// TODO: implement multi-simd softmax (llama.cpp commit e16b9fa4)
-
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x_id = 0) in;
-
-layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    float scale;
-    float max_bias;
-    float m0;
-    float m1;
-    uint n_head_log2;
-    int mask;
-} pcs;
-
-void main() {
-    if (gl_SubgroupInvocationID > 31)
-        return;
-
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const uint extra_off = i03*pcs.ne02*pcs.ne01*pcs.ne00 + i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00;
-    const uint psrc0 = extra_off + pcs.inAOff; // Based from inA
-    const uint pmask = i01*pcs.ne00 + pcs.inBOff; // Based from inB
-    const uint pdst = extra_off + pcs.outOff; // Based from out_
-
-    float slope = 1.0f;
-
-    // ALiBi
-    if (pcs.max_bias > 0.0f) {
-        int64_t h = i02;
-
-        float base = h < pcs.n_head_log2 ? pcs.m0 : pcs.m1;
-        int64_t exp = h < pcs.n_head_log2 ? h + 1 : 2*(h - pcs.n_head_log2) + 1;
-
-        slope = pow(base, float(exp));
-    }
-
-    // parallel max
-    float localMax = uintBitsToFloat(0xFF800000);
-    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
-        localMax = max(localMax, inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? slope*inB[pmask + i00] : 0.0f));
-    }
-    float max_ = subgroupMax(localMax);
-
-    // parallel sum
-    float localSum = 0.0f;
-    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
-        const float exp_psrc0 = exp(inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? slope*inB[pmask + i00] : 0.0f) - max_);
-        localSum += exp_psrc0;
-        out_[pdst + i00] = exp_psrc0;
-    }
-
-    const float sum = subgroupAdd(localSum);
-    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
-        out_[pdst + i00] /= sum;
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp b/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp
deleted file mode 100644
index 0fca640dcc..0000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp
+++ /dev/null
@@ -1,71 +0,0 @@
-#include "common.comp"
-
-#define GGML_ROPE_TYPE_NEOX 2
-
-// TODO: use a local size of 32 or more (Metal uses 1024)
-layout(local_size_x = 1) in;
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint inCOff;
-    uint outOff;
-    int n_dims;
-    int mode;
-    int n_ctx_orig;
-    float freq_base;
-    float freq_scale;
-    bool has_freq_factors;
-    float ext_factor;
-    float attn_factor;
-    float beta_fast;
-    float beta_slow;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-float rope_yarn_ramp(const float low, const float high, const float i0) {
-    const float y = (i0 / 2 - low) / max(0.001f, high - low);
-    return 1.0f - min(1.0f, max(0.0f, y));
-}
-
-// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
-// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
-void rope_yarn(
-    float theta_extrap, float freq_scale, float corr_dims[2], float i0, float ext_factor, float mscale,
-    out float cos_theta, out float sin_theta
-) {
-    // Get n-d rotational scaling corrected for extrapolation
-    float theta_interp = freq_scale * theta_extrap;
-    float theta = theta_interp;
-    if (ext_factor != 0.0f) {
-        float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
-        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
-
-        // Get n-d magnitude scaling corrected for interpolation
-        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
-    }
-    cos_theta = cos(theta) * mscale;
-    sin_theta = sin(theta) * mscale;
-}
-
-// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
-// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
-float rope_yarn_corr_factor(int n_dims, int n_ctx_orig, float n_rot, float base) {
-    return n_dims * log(n_ctx_orig / (n_rot * TWOPI_F)) / (2 * log(base));
-}
-
-void rope_yarn_corr_dims(
-    int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, out float dims[2]
-) {
-    // start and end correction dims
-    dims[0] = max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_fast, freq_base)));
-    dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_slow, freq_base)));
-}
diff --git a/scripts/sync-ggml-am.sh b/scripts/sync-ggml-am.sh
index 73d4eec0b5..29d30e0a18 100755
--- a/scripts/sync-ggml-am.sh
+++ b/scripts/sync-ggml-am.sh
@@ -83,7 +83,6 @@ while read c; do
         src/ggml-cpu/* \
         src/ggml-cuda/* \
         src/ggml-hip/* \
-        src/ggml-kompute/* \
         src/ggml-metal/* \
         src/ggml-musa/* \
         src/ggml-opencl/* \
@@ -141,7 +140,6 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     # src/ggml-cpu/*       -> ggml/src/ggml-cpu/*
     # src/ggml-cuda/*      -> ggml/src/ggml-cuda/*
     # src/ggml-hip/*       -> ggml/src/ggml-hip/*
-    # src/ggml-kompute/*   -> ggml/src/ggml-kompute/*
     # src/ggml-metal/*     -> ggml/src/ggml-metal/*
     # src/ggml-musa/*      -> ggml/src/ggml-musa/*
     # src/ggml-opencl/*    -> ggml/src/ggml-opencl/*
@@ -174,7 +172,6 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/([[:space:]]| [ab]\/)src\/ggml-cpu\//\1ggml\/src\/ggml-cpu\//g' \
         -e 's/([[:space:]]| [ab]\/)src\/ggml-cuda\//\1ggml\/src\/ggml-cuda\//g' \
         -e 's/([[:space:]]| [ab]\/)src\/ggml-hip\//\1ggml\/src\/ggml-hip\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-kompute\//\1ggml\/src\/ggml-kompute\//g' \
         -e 's/([[:space:]]| [ab]\/)src\/ggml-metal\//\1ggml\/src\/ggml-metal\//g' \
         -e 's/([[:space:]]| [ab]\/)src\/ggml-opencl\//\1ggml\/src\/ggml-opencl\//g' \
         -e 's/([[:space:]]| [ab]\/)src\/ggml-rpc\//\1ggml\/src\/ggml-rpc\//g' \
diff --git a/scripts/sync-ggml.sh b/scripts/sync-ggml.sh
index 6460a77f1c..9b98329e09 100755
--- a/scripts/sync-ggml.sh
+++ b/scripts/sync-ggml.sh
@@ -15,7 +15,6 @@ cp -rpv ../ggml/src/ggml-cann/*    ./ggml/src/ggml-cann/
 cp -rpv ../ggml/src/ggml-cpu/*     ./ggml/src/ggml-cpu/
 cp -rpv ../ggml/src/ggml-cuda/*    ./ggml/src/ggml-cuda/
 cp -rpv ../ggml/src/ggml-hip/*     ./ggml/src/ggml-hip/
-cp -rpv ../ggml/src/ggml-kompute/* ./ggml/src/ggml-kompute/
 cp -rpv ../ggml/src/ggml-metal/*   ./ggml/src/ggml-metal/
 cp -rpv ../ggml/src/ggml-musa/*    ./ggml/src/ggml-musa/
 cp -rpv ../ggml/src/ggml-opencl/*  ./ggml/src/ggml-opencl/
diff --git a/tests/test-c.c b/tests/test-c.c
index 95ba73df39..a05071080a 100644
--- a/tests/test-c.c
+++ b/tests/test-c.c
@@ -1,7 +1,3 @@
 #include "llama.h"
 
-#ifdef GGML_USE_KOMPUTE
-#include "ggml-kompute.h"
-#endif
-
 int main(void) {}

From 9067487c4411efb20400103fcccfdd389c80d428 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Thu, 3 Jul 2025 10:46:57 +0300
Subject: [PATCH 04/10] ggml : fix FA mask dim 2 and 3 (#14505)

* ggml : fix FA mask dim 2 and 3

ggml-ci

* backends : unsupport batched FA in CUDA and Vulkan

ggml-ci

* vulkan : disable FA for mask->ne[2] != 1
---
 ggml/include/ggml.h                   | 13 +++++++------
 ggml/src/ggml-cpu/ops.cpp             |  2 +-
 ggml/src/ggml-cuda/ggml-cuda.cu       |  3 ++-
 ggml/src/ggml-metal/ggml-metal-impl.h |  2 ++
 ggml/src/ggml-metal/ggml-metal.m      |  2 ++
 ggml/src/ggml-metal/ggml-metal.metal  |  4 ++--
 ggml/src/ggml-vulkan/ggml-vulkan.cpp  |  6 ++++++
 ggml/src/ggml.c                       |  5 ++---
 tests/test-backend-ops.cpp            |  4 ++--
 9 files changed, 26 insertions(+), 15 deletions(-)

diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index 78983bcc56..b2ad13f194 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -1983,15 +1983,16 @@ extern "C" {
 
 #define GGML_KQ_MASK_PAD 64
 
-    // q:    [n_embd_k, n_batch,     n_head,    ne3]
-    // k:    [n_embd_k, n_kv,        n_head_kv, ne3]
-    // v:    [n_embd_v, n_kv,        n_head_kv, ne3] !! not transposed !!
-    // mask: [n_kv,     n_batch_pad, ne32,      1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
-    // res:  [n_embd_v, n_head,      n_batch,   ne3] !! permuted !!
+    // q:    [n_embd_k, n_batch,     n_head,    ne3 ]
+    // k:    [n_embd_k, n_kv,        n_head_kv, ne3 ]
+    // v:    [n_embd_v, n_kv,        n_head_kv, ne3 ] !! not transposed !!
+    // mask: [n_kv,     n_batch_pad, ne32,      ne33] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
+    // res:  [n_embd_v, n_head,      n_batch,   ne3 ] !! permuted !!
     //
     // broadcast:
     //   n_head % n_head_kv == 0
-    //   ne3    % ne32      == 0
+    //   n_head % ne32      == 0
+    //   ne3    % ne33      == 0
     //
     GGML_API struct ggml_tensor * ggml_flash_attn_ext(
             struct ggml_context * ctx,
diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
index 2f15341954..0fb2c08b5b 100644
--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
@@ -7799,7 +7799,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
             memset(VKQ32, 0, DV*sizeof(float));
         }
 
-        const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1] + (iq3%mask->ne[2])*mask->nb[2]) : NULL;
+        const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1] + (iq2%mask->ne[2])*mask->nb[2] + (iq3%mask->ne[3])*mask->nb[3]) : NULL;
 
         // k indices
         const int ik3 = iq3 / rk3;
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 121e83641b..1c04bba52e 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -3390,7 +3390,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 return false;
             }
             // TODO: support broadcast
-            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
+            // note: this was initially implemented in https://github.com/ggml-org/llama.cpp/pull/14500, but
+            //       the interface of ggml_flash_attn_ext() changed in https://github.com/ggml-org/llama.cpp/pull/14505
             if (op->src[0]->ne[3] != 1) {
                 return false;
             }
diff --git a/ggml/src/ggml-metal/ggml-metal-impl.h b/ggml/src/ggml-metal/ggml-metal-impl.h
index adb65c0e7b..752d55c216 100644
--- a/ggml/src/ggml-metal/ggml-metal-impl.h
+++ b/ggml/src/ggml-metal/ggml-metal-impl.h
@@ -230,8 +230,10 @@ typedef struct {
     uint64_t nb22;
     uint64_t nb23;
     int32_t  ne32;
+    int32_t  ne33;
     uint64_t nb31;
     uint64_t nb32;
+    uint64_t nb33;
     int32_t  ne1;
     int32_t  ne2;
     float    scale;
diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
index de40430ef3..5e5467e88a 100644
--- a/ggml/src/ggml-metal/ggml-metal.m
+++ b/ggml/src/ggml-metal/ggml-metal.m
@@ -5018,8 +5018,10 @@ static bool ggml_metal_encode_node(
                     /*.nb22          =*/ nb22,
                     /*.nb23          =*/ nb23,
                     /*.ne32          =*/ ne32,
+                    /*.ne33          =*/ ne33,
                     /*.nb31          =*/ nb31,
                     /*.nb32          =*/ nb32,
+                    /*.nb33          =*/ nb33,
                     /*.ne1           =*/ ne1,
                     /*.ne2           =*/ ne2,
                     /*.scale         =*/ scale,
diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
index a7657e0fc3..ebde005f33 100644
--- a/ggml/src/ggml-metal/ggml-metal.metal
+++ b/ggml/src/ggml-metal/ggml-metal.metal
@@ -3857,7 +3857,7 @@ kernel void kernel_flash_attn_ext(
                 // load the mask in shared memory
                 #pragma unroll(Q)
                 for (short j = 0; j < Q; ++j) {
-                    device const half * pm = (device const half *) ((device const char *) mask + (iq1 + j)*args.nb31 + (iq3%args.ne32)*args.nb32);
+                    device const half * pm = (device const half *) ((device const char *) mask + (iq1 + j)*args.nb31 + (iq2%args.ne32)*args.nb32 + (iq3%args.ne33)*args.nb33);
 
                     const float m = pm[ic + tiisg];
 
@@ -4343,7 +4343,7 @@ kernel void kernel_flash_attn_ext_vec(
         const bool has_mask = mask != q;
 
         // pointer to the mask
-        device const half * pm = (device const half *) (mask + iq1*args.nb31 + (iq3%args.ne32)*args.nb32);
+        device const half * pm = (device const half *) (mask + iq1*args.nb31 + (iq2%args.ne32)*args.nb32 + (iq3%args.ne33)*args.nb33);
 
         float slope = 1.0f;
 
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 25f70127a6..ccb0d47b67 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -10265,6 +10265,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 if (op->src[3] && op->src[3]->type != GGML_TYPE_F16) {
                     return false;
                 }
+                // TODO: support broadcast
+                // note: this was initially implemented in https://github.com/ggml-org/llama.cpp/pull/14449, but
+                //       the interface of ggml_flash_attn_ext() changed in https://github.com/ggml-org/llama.cpp/pull/14505
+                if (op->src[0]->ne[3] != 1 || (op->src[3] && op->src[3]->ne[2] != 1)) {
+                    return false;
+                }
                 // It's straightforward to support different K/V dequant, but would
                 // significantly increase the number of pipelines
                 if (op->src[1]->type != op->src[2]->type) {
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 97da26b378..fdb57e1785 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -3674,7 +3674,6 @@ static struct ggml_tensor * ggml_soft_max_impl(
     if (mask) {
         GGML_ASSERT(mask->type == GGML_TYPE_F16 || mask->type == GGML_TYPE_F32);
         GGML_ASSERT(ggml_is_contiguous(mask));
-        GGML_ASSERT(ggml_is_3d(mask));
         GGML_ASSERT(mask->ne[0] == a->ne[0]);
         GGML_ASSERT(mask->ne[1] >= a->ne[1]);
         GGML_ASSERT(a->ne[2]%mask->ne[2] == 0);
@@ -4704,12 +4703,12 @@ struct ggml_tensor * ggml_flash_attn_ext(
 
     if (mask) {
         GGML_ASSERT(ggml_is_contiguous(mask));
-        GGML_ASSERT(mask->ne[2] == q->ne[3]);
         GGML_ASSERT(mask->ne[1] >= GGML_PAD(q->ne[1], GGML_KQ_MASK_PAD) &&
                 "the Flash-Attention kernel requires the mask to be padded to GGML_KQ_MASK_PAD and at least n_queries big");
         //GGML_ASSERT(ggml_can_repeat_rows(mask, qk));
 
-        GGML_ASSERT(q->ne[3] % mask->ne[2] == 0);
+        GGML_ASSERT(q->ne[2] % mask->ne[2] == 0);
+        GGML_ASSERT(q->ne[3] % mask->ne[3] == 0);
     }
 
     if (max_bias > 0.0f) {
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 5c3db854cb..2ab6dd06fc 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -3637,7 +3637,7 @@ struct test_flash_attn_ext : public test_case {
 
         ggml_tensor * m = nullptr;
         if (mask) {
-            m = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, GGML_PAD(nb, GGML_KQ_MASK_PAD), nr23[1], 1);
+            m = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, GGML_PAD(nb, GGML_KQ_MASK_PAD), nr23[0], nr23[1]);
             ggml_set_name(m, "m");
         }
 
@@ -4751,7 +4751,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                                 test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, m_prec, {1, 1}, scale, max_bias));
 
                                 if (ne0 <= 32 && ne1 <= 32) {
-                                    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0,   ne1,   1, 1}, mask, m_prec, {3, 1}, scale, max_bias));
+                                    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0,   ne1,   1, 3}, mask, m_prec, {3, 1}, scale, max_bias));
                                     test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, m_prec, {2, 3}, scale, max_bias));
                                 }
                             }

From a70c8a0c4b4c1606cd9a0ba889ce61aa88610095 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Thu, 3 Jul 2025 10:53:35 +0300
Subject: [PATCH 05/10] kv-cache : use ggml_set_rows (#14285)

* kv-cache : use ggml_set_rows

ggml-ci

* graph : separate k and v indices

ggml-ci

* cont : remove redundant ifs

ggml-ci

* kv-cache : improve find_slot impl

* kv-cache : bounds-check when accessing slot_info indices

* kv-cache : add comments

ggml-ci

* ggml : add TODOs for adding GGML_OP_SET_ROWS support in the backends

ggml-ci
---
 ggml/src/ggml-cann/ggml-cann.cpp     |   6 +
 ggml/src/ggml-opencl/ggml-opencl.cpp |   6 +
 ggml/src/ggml-sycl/ggml-sycl.cpp     |   6 +
 ggml/src/ggml-vulkan/ggml-vulkan.cpp |   6 +
 src/llama-graph.cpp                  |  68 ++++---
 src/llama-graph.h                    |  24 ++-
 src/llama-kv-cache-unified-iswa.cpp  |  32 ++--
 src/llama-kv-cache-unified-iswa.h    |   6 +-
 src/llama-kv-cache-unified.cpp       | 274 ++++++++++++++++++++-------
 src/llama-kv-cache-unified.h         |  86 ++++++---
 src/llama-kv-cells.h                 |  72 ++++++-
 src/llama-memory-hybrid.cpp          |   4 +-
 src/llama-memory-hybrid.h            |   4 +-
 13 files changed, 451 insertions(+), 143 deletions(-)

diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
index 8a3d2026ed..eae575cc04 100755
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -2086,6 +2086,12 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                     return false;
             }
         } break;
+        case GGML_OP_SET_ROWS:
+            {
+                // TODO: add support
+                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
+                return false;
+            } break;
         case GGML_OP_CPY: {
             ggml_tensor *src = op->src[0];
             if ((op->type != GGML_TYPE_F32 && op->type != GGML_TYPE_F16) ||
diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index 71829c05a5..9436e6ea9a 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -2222,6 +2222,12 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 default:
                     return false;
             }
+        case GGML_OP_SET_ROWS:
+            {
+                // TODO: add support
+                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
+                return false;
+            } break;
         case GGML_OP_CPY:
         case GGML_OP_DUP:
         case GGML_OP_CONT:
diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
index 1d41d7a4be..ee4c8f7b2f 100644
--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -4285,6 +4285,12 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                         return false;
                 }
             }
+        case GGML_OP_SET_ROWS:
+            {
+                // TODO: add support
+                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
+                return false;
+            } break;
         case GGML_OP_CPY:
             {
                 ggml_type src0_type = op->src[0]->type;
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index ccb0d47b67..8114cb4204 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -10339,6 +10339,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                         return false;
                 }
             } break;
+        case GGML_OP_SET_ROWS:
+            {
+                // TODO: add support
+                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
+                return false;
+            } break;
         case GGML_OP_CONT:
         case GGML_OP_CPY:
         case GGML_OP_DUP:
diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index f2fae6d1b7..4443420132 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -281,19 +281,22 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
 }
 
 void llm_graph_input_attn_kv_unified::set_input(const llama_ubatch * ubatch) {
-    if (self_kq_mask) {
-        mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
-    }
+    mctx->set_input_k_idxs(self_k_idxs, ubatch);
+    mctx->set_input_v_idxs(self_v_idxs, ubatch);
+
+    mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
 }
 
 void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch) {
-    if (self_kq_mask) {
-        mctx->get_base()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
-    }
+    mctx->get_base()->set_input_k_idxs(self_k_idxs, ubatch);
+    mctx->get_base()->set_input_v_idxs(self_v_idxs, ubatch);
 
-    if (self_kq_mask_swa) {
-        mctx->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn);
-    }
+    mctx->get_base()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
+
+    mctx->get_swa()->set_input_k_idxs(self_k_idxs_swa, ubatch);
+    mctx->get_swa()->set_input_v_idxs(self_v_idxs_swa, ubatch);
+
+    mctx->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn);
 }
 
 void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
@@ -333,9 +336,10 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
 }
 
 void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) {
-    if (self_kq_mask) {
-        mctx->get_attn()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
-    }
+    mctx->get_attn()->set_input_k_idxs(self_k_idxs, ubatch);
+    mctx->get_attn()->set_input_v_idxs(self_v_idxs, ubatch);
+
+    mctx->get_attn()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
 
     const int64_t n_rs = mctx->get_recr()->get_n_rs();
 
@@ -350,7 +354,8 @@ void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) {
     }
 }
 
-void llm_graph_input_one::set_input(const llama_ubatch *) {
+void llm_graph_input_one::set_input(const llama_ubatch * ubatch) {
+    GGML_UNUSED(ubatch);
     GGML_ASSERT(one && ggml_nelements(one) == 1);
     float f_one = 1.0f;
     ggml_backend_tensor_set(one, &f_one, 0, sizeof(float));
@@ -997,6 +1002,9 @@ llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
 
         const auto n_kv = inp->mctx->get_attn()->get_n_kv();
 
+        inp->self_k_idxs = mctx_cur->get_attn()->build_input_k_idxs(ctx0, ubatch);
+        inp->self_v_idxs = mctx_cur->get_attn()->build_input_v_idxs(ctx0, ubatch);
+
         inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
         //cb(inp->self_kq_mask, "KQ_mask", -1);
         ggml_set_input(inp->self_kq_mask);
@@ -1198,8 +1206,10 @@ llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified()
 
         const auto n_kv = mctx_cur->get_n_kv();
 
+        inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);
+        inp->self_v_idxs = mctx_cur->build_input_v_idxs(ctx0, ubatch);
+
         inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        //cb(inp->self_kq_mask, "KQ_mask", -1);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1230,8 +1240,11 @@ ggml_tensor * llm_graph_context::build_attn(
 
     // store to KV cache
     {
-        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, il));
-        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, il));
+        const auto & k_idxs = inp->get_k_idxs();
+        const auto & v_idxs = inp->get_v_idxs();
+
+        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
+        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, v_idxs, il));
     }
 
     const auto & kq_mask = inp->get_kq_mask();
@@ -1290,11 +1303,15 @@ ggml_tensor * llm_graph_context::build_attn(
 
     // optionally store to KV cache
     if (k_cur) {
-        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, il));
+        const auto & k_idxs = is_swa ? inp->get_k_idxs_swa() : inp->get_k_idxs();
+
+        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
     }
 
     if (v_cur) {
-        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, il));
+        const auto & v_idxs = is_swa ? inp->get_v_idxs_swa() : inp->get_v_idxs();
+
+        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, v_idxs, il));
     }
 
     const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask();
@@ -1398,8 +1415,11 @@ ggml_tensor * llm_graph_context::build_attn(
 
     // store to KV cache
     {
-        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, il));
-        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, il));
+        const auto & k_idxs = inp->get_k_idxs();
+        const auto & v_idxs = inp->get_v_idxs();
+
+        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
+        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, v_idxs, il));
     }
 
     const auto & kq_mask = inp->get_kq_mask();
@@ -1434,8 +1454,10 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
     {
         const auto n_kv = mctx_cur->get_base()->get_n_kv();
 
+        inp->self_k_idxs = mctx_cur->get_base()->build_input_k_idxs(ctx0, ubatch);
+        inp->self_v_idxs = mctx_cur->get_base()->build_input_v_idxs(ctx0, ubatch);
+
         inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        //cb(inp->self_kq_mask, "KQ_mask", -1);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1446,8 +1468,10 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
 
         const auto n_kv = mctx_cur->get_swa()->get_n_kv();
 
+        inp->self_k_idxs_swa = mctx_cur->get_swa()->build_input_k_idxs(ctx0, ubatch);
+        inp->self_v_idxs_swa = mctx_cur->get_swa()->build_input_v_idxs(ctx0, ubatch);
+
         inp->self_kq_mask_swa = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        //cb(inp->self_kq_mask_swa, "KQ_mask_swa", -1);
         ggml_set_input(inp->self_kq_mask_swa);
 
         inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
diff --git a/src/llama-graph.h b/src/llama-graph.h
index db4e14805c..c8b74a1474 100644
--- a/src/llama-graph.h
+++ b/src/llama-graph.h
@@ -249,8 +249,14 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    ggml_tensor * get_k_idxs() const { return self_k_idxs; }
+    ggml_tensor * get_v_idxs() const { return self_v_idxs; }
+
     ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
 
+    ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch]
+
     ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch]
     ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch]
 
@@ -274,9 +280,19 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    ggml_tensor * get_k_idxs()     const { return self_k_idxs; }
+    ggml_tensor * get_v_idxs()     const { return self_v_idxs; }
+    ggml_tensor * get_k_idxs_swa() const { return self_k_idxs_swa; }
+    ggml_tensor * get_v_idxs_swa() const { return self_v_idxs_swa; }
+
     ggml_tensor * get_kq_mask()     const { return self_kq_mask_cnv; }
     ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }
 
+    ggml_tensor * self_k_idxs     = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs     = nullptr; // I64 [n_batch]
+    ggml_tensor * self_k_idxs_swa = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs_swa = nullptr; // I64 [n_batch]
+
     ggml_tensor * self_kq_mask         = nullptr; // F32 [n_kv, n_batch]
     ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_kv, n_batch]
     ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch]
@@ -319,8 +335,14 @@ public:
 
     ggml_tensor * s_copy; // I32 [kv_size]
 
+    ggml_tensor * get_k_idxs() const { return self_k_idxs; }
+    ggml_tensor * get_v_idxs() const { return self_v_idxs; }
+
     ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
 
+    ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch]
+
     ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch]
     ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch]
 
@@ -336,7 +358,7 @@ public:
     llm_graph_input_one() {}
     virtual ~llm_graph_input_one() = default;
 
-    void set_input(const llama_ubatch *) override;
+    void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * one = nullptr; // F32
 };
diff --git a/src/llama-kv-cache-unified-iswa.cpp b/src/llama-kv-cache-unified-iswa.cpp
index d1f839b63a..ee202cc710 100644
--- a/src/llama-kv-cache-unified-iswa.cpp
+++ b/src/llama-kv-cache-unified-iswa.cpp
@@ -113,20 +113,20 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
             ubatches.push_back(std::move(ubatch)); // NOLINT
         }
 
-        auto heads_base = kv_base->prepare(ubatches);
-        if (heads_base.empty()) {
+        auto sinfos_base = kv_base->prepare(ubatches);
+        if (sinfos_base.empty()) {
             break;
         }
 
-        auto heads_swa = kv_swa->prepare(ubatches);
-        if (heads_swa.empty()) {
+        auto sinfos_swa = kv_swa->prepare(ubatches);
+        if (sinfos_swa.empty()) {
             break;
         }
 
-        assert(heads_base.size() == heads_swa.size());
+        assert(sinfos_base.size() == sinfos_swa.size());
 
         return std::make_unique<llama_kv_cache_unified_iswa_context>(
-                this, std::move(heads_base), std::move(heads_swa), std::move(ubatches));
+                this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
     } while (false);
 
     // if it fails, try equal split
@@ -144,20 +144,20 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
             ubatches.push_back(std::move(ubatch)); // NOLINT
         }
 
-        auto heads_base = kv_base->prepare(ubatches);
-        if (heads_base.empty()) {
+        auto sinfos_base = kv_base->prepare(ubatches);
+        if (sinfos_base.empty()) {
             break;
         }
 
-        auto heads_swa = kv_swa->prepare(ubatches);
-        if (heads_swa.empty()) {
+        auto sinfos_swa = kv_swa->prepare(ubatches);
+        if (sinfos_swa.empty()) {
             break;
         }
 
-        assert(heads_base.size() == heads_swa.size());
+        assert(sinfos_base.size() == sinfos_swa.size());
 
         return std::make_unique<llama_kv_cache_unified_iswa_context>(
-                this, std::move(heads_base), std::move(heads_swa), std::move(ubatches));
+                this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
     } while (false);
 
     // TODO: if we fail again, we should attempt different splitting strategies
@@ -220,13 +220,13 @@ llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
 
 llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
         llama_kv_cache_unified_iswa * kv,
-        std::vector<uint32_t> heads_base,
-        std::vector<uint32_t> heads_swa,
+        slot_info_vec_t sinfos_base,
+        slot_info_vec_t sinfos_swa,
         std::vector<llama_ubatch> ubatches) :
     ubatches(std::move(ubatches)),
     // note: here we copy the ubatches. not sure if this is ideal
-    ctx_base(new llama_kv_cache_unified_context(kv->get_base(), std::move(heads_base), this->ubatches)),
-    ctx_swa (new llama_kv_cache_unified_context(kv->get_swa (), std::move(heads_swa),  this->ubatches)),
+    ctx_base(new llama_kv_cache_unified_context(kv->get_base(), std::move(sinfos_base), this->ubatches)),
+    ctx_swa (new llama_kv_cache_unified_context(kv->get_swa (), std::move(sinfos_swa),  this->ubatches)),
     status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
 }
 
diff --git a/src/llama-kv-cache-unified-iswa.h b/src/llama-kv-cache-unified-iswa.h
index 46c1ed614f..23205d826b 100644
--- a/src/llama-kv-cache-unified-iswa.h
+++ b/src/llama-kv-cache-unified-iswa.h
@@ -74,6 +74,8 @@ private:
 
 class llama_kv_cache_unified_iswa_context : public llama_memory_context_i {
 public:
+    using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t;
+
     // used for errors
     llama_kv_cache_unified_iswa_context(llama_memory_status status);
 
@@ -90,8 +92,8 @@ public:
     // used to create a batch processing context from a batch
     llama_kv_cache_unified_iswa_context(
             llama_kv_cache_unified_iswa * kv,
-            std::vector<uint32_t> heads_base,
-            std::vector<uint32_t> heads_swa,
+            slot_info_vec_t sinfos_base,
+            slot_info_vec_t sinfos_swa,
             std::vector<llama_ubatch> ubatches);
 
     virtual ~llama_kv_cache_unified_iswa_context();
diff --git a/src/llama-kv-cache-unified.cpp b/src/llama-kv-cache-unified.cpp
index 7f7b162ffd..ff22079851 100644
--- a/src/llama-kv-cache-unified.cpp
+++ b/src/llama-kv-cache-unified.cpp
@@ -156,6 +156,13 @@ llama_kv_cache_unified::llama_kv_cache_unified(
 
     const char * LLAMA_KV_CACHE_DEBUG = getenv("LLAMA_KV_CACHE_DEBUG");
     debug = LLAMA_KV_CACHE_DEBUG ? atoi(LLAMA_KV_CACHE_DEBUG) : 0;
+
+    const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS");
+    supports_set_rows = LLAMA_SET_ROWS ? atoi(LLAMA_SET_ROWS) : 0;
+
+    if (!supports_set_rows) {
+        LLAMA_LOG_WARN("%s: LLAMA_SET_ROWS=0, using old ggml_cpy() method for backwards compatibility\n", __func__);
+    }
 }
 
 void llama_kv_cache_unified::clear(bool data) {
@@ -353,13 +360,13 @@ llama_memory_context_ptr llama_kv_cache_unified::init_batch(
             ubatches.push_back(std::move(ubatch)); // NOLINT
         }
 
-        auto heads = prepare(ubatches);
-        if (heads.empty()) {
+        auto sinfos = prepare(ubatches);
+        if (sinfos.empty()) {
             break;
         }
 
         return std::make_unique<llama_kv_cache_unified_context>(
-                this, std::move(heads), std::move(ubatches));
+                this, std::move(sinfos), std::move(ubatches));
     } while (false);
 
     return std::make_unique<llama_kv_cache_unified_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
@@ -402,12 +409,13 @@ llama_memory_context_ptr llama_kv_cache_unified::init_update(llama_context * lct
     return std::make_unique<llama_kv_cache_unified_context>(this, lctx, do_shift, std::move(dinfo));
 }
 
-llama_kv_cache_unified::ubatch_heads llama_kv_cache_unified::prepare(const std::vector<llama_ubatch> & ubatches) {
-    llama_kv_cache_unified::ubatch_heads res;
+llama_kv_cache_unified::slot_info_vec_t llama_kv_cache_unified::prepare(const std::vector<llama_ubatch> & ubatches) {
+    llama_kv_cache_unified::slot_info_vec_t res;
 
     struct state {
         uint32_t head_old; // old position of the head, before placing the ubatch
-        uint32_t head_new; // new position of the head, after placing the ubatch
+
+        slot_info sinfo; // slot info for the ubatch
 
         llama_kv_cells_unified cells; // copy of the old cells, before placing the ubatch
     };
@@ -418,26 +426,29 @@ llama_kv_cache_unified::ubatch_heads llama_kv_cache_unified::prepare(const std::
     bool success = true;
 
     for (const auto & ubatch : ubatches) {
+        // non-continuous slots require support for ggml_set_rows()
+        const bool cont = supports_set_rows ? false : true;
+
         // only find a suitable slot for the ubatch. don't modify the cells yet
-        const int32_t head_new = find_slot(ubatch);
-        if (head_new < 0) {
+        const auto sinfo_new = find_slot(ubatch, cont);
+        if (sinfo_new.empty()) {
             success = false;
             break;
         }
 
         // remeber the position that we found
-        res.push_back(head_new);
+        res.push_back(sinfo_new);
 
         // store the old state of the cells in the recovery stack
-        states.push_back({head, (uint32_t) head_new, cells.cp(head_new, ubatch.n_tokens)});
+        states.push_back({head, sinfo_new, cells.cp(sinfo_new.idxs)});
 
         // now emplace the ubatch
-        apply_ubatch(head_new, ubatch);
+        apply_ubatch(sinfo_new, ubatch);
     }
 
     // iterate backwards and restore the cells to their original state
     for (auto it = states.rbegin(); it != states.rend(); ++it) {
-        cells.set(it->head_new, it->cells);
+        cells.set(it->sinfo.idxs, it->cells);
         head = it->head_old;
     }
 
@@ -539,7 +550,7 @@ bool llama_kv_cache_unified::update(llama_context * lctx, bool do_shift, const d
     return updated;
 }
 
-int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
+llama_kv_cache_unified::slot_info llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch, bool cont) const {
     const uint32_t n_tokens = ubatch.n_tokens;
 
     uint32_t head_cur = this->head;
@@ -552,7 +563,7 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
 
     if (n_tokens > cells.size()) {
         LLAMA_LOG_ERROR("%s: n_tokens = %d > size = %u\n", __func__, n_tokens, cells.size());
-        return -1;
+        return { };
     }
 
     if (debug > 0) {
@@ -615,15 +626,26 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
 
     uint32_t n_tested = 0;
 
+    // for continuous slots, we test that all tokens in the ubatch fit, starting from the current head
+    // for non-continuous slots, we test the tokens one by one
+    const uint32_t n_test = cont ? n_tokens : 1;
+
+    slot_info res;
+
+    auto & idxs = res.idxs;
+
+    idxs.reserve(n_tokens);
+
     while (true) {
-        if (head_cur + n_tokens > cells.size()) {
+        if (head_cur + n_test > cells.size()) {
             n_tested += cells.size() - head_cur;
             head_cur = 0;
             continue;
         }
 
-        bool found = true;
-        for (uint32_t i = 0; i < n_tokens; i++) {
+        for (uint32_t i = 0; i < n_test; i++) {
+            const auto idx = head_cur;
+
             //const llama_pos    pos    = ubatch.pos[i];
             //const llama_seq_id seq_id = ubatch.seq_id[i][0];
 
@@ -633,19 +655,19 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
             //    - (disabled) mask causally, if the sequence is the same as the one we are inserting
             //    - mask SWA, using current max pos for that sequence in the cache
             //                always insert in the cell with minimum pos
-            bool can_use = cells.is_empty(head_cur + i);
+            bool can_use = cells.is_empty(idx);
 
-            if (!can_use && cells.seq_count(head_cur + i) == 1) {
-                const llama_pos pos_cell = cells.pos_get(head_cur + i);
+            if (!can_use && cells.seq_count(idx) == 1) {
+                const llama_pos pos_cell = cells.pos_get(idx);
 
                 // (disabled) causal mask
                 // note: it's better to purge any "future" tokens beforehand
-                //if (cells.seq_has(head_cur + i, seq_id)) {
+                //if (cells.seq_has(idx, seq_id)) {
                 //    can_use = pos_cell >= pos;
                 //}
 
                 if (!can_use) {
-                    const llama_seq_id seq_id_cell = cells.seq_get(head_cur + i);
+                    const llama_seq_id seq_id_cell = cells.seq_get(idx);
 
                     // SWA mask
                     if (is_masked_swa(pos_cell, cells.seq_pos_max(seq_id_cell) + 1)) {
@@ -654,28 +676,39 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
                 }
             }
 
-            if (!can_use) {
-                found = false;
-                head_cur += i + 1;
-                n_tested += i + 1;
+            head_cur++;
+            n_tested++;
+
+            if (can_use) {
+                idxs.push_back(idx);
+            } else {
                 break;
             }
         }
 
-        if (found) {
+        if (idxs.size() == n_tokens) {
             break;
         }
 
+        if (cont) {
+            idxs.clear();
+        }
+
         if (n_tested >= cells.size()) {
             //LLAMA_LOG_ERROR("%s: failed to find a slot for %d tokens\n", __func__, n_tokens);
-            return -1;
+            return { };
         }
     }
 
-    return head_cur;
+    // we didn't find a suitable slot - return empty result
+    if (idxs.size() < n_tokens) {
+        res.clear();
+    }
+
+    return res;
 }
 
-void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch & ubatch) {
+void llama_kv_cache_unified::apply_ubatch(const slot_info & sinfo, const llama_ubatch & ubatch) {
     // keep track of the max sequence position that we would overwrite with this ubatch
     // for non-SWA cache, this would be always empty
     llama_seq_id seq_pos_max_rm[LLAMA_MAX_SEQ];
@@ -683,22 +716,26 @@ void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch
         seq_pos_max_rm[s] = -1;
     }
 
-    for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
-        if (!cells.is_empty(head_cur + i)) {
-            assert(cells.seq_count(head_cur + i) == 1);
+    assert(ubatch.n_tokens == sinfo.idxs.size());
 
-            const llama_seq_id seq_id = cells.seq_get(head_cur + i);
-            const llama_pos    pos    = cells.pos_get(head_cur + i);
+    for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
+        const auto idx = sinfo.idxs.at(i);
+
+        if (!cells.is_empty(idx)) {
+            assert(cells.seq_count(idx) == 1);
+
+            const llama_seq_id seq_id = cells.seq_get(idx);
+            const llama_pos    pos    = cells.pos_get(idx);
 
             seq_pos_max_rm[seq_id] = std::max(seq_pos_max_rm[seq_id], pos);
 
-            cells.rm(head_cur + i);
+            cells.rm(idx);
         }
 
-        cells.pos_set(head_cur + i, ubatch.pos[i]);
+        cells.pos_set(idx, ubatch.pos[i]);
 
         for (int32_t s = 0; s < ubatch.n_seq_id[i]; s++) {
-            cells.seq_add(head_cur + i, ubatch.seq_id[i][s]);
+            cells.seq_add(idx, ubatch.seq_id[i][s]);
         }
     }
 
@@ -719,7 +756,7 @@ void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch
     }
 
     // move the head at the end of the slot
-    head = head_cur + ubatch.n_tokens;
+    head = sinfo.idxs.back() + 1;
 }
 
 bool llama_kv_cache_unified::get_can_shift() const {
@@ -772,47 +809,133 @@ ggml_tensor * llama_kv_cache_unified::get_v(ggml_context * ctx, int32_t il, uint
             0);
 }
 
-ggml_tensor * llama_kv_cache_unified::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il, uint32_t head_cur) const {
+ggml_tensor * llama_kv_cache_unified::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const {
     const int32_t ikv = map_layer_ids.at(il);
 
     auto * k = layers[ikv].k;
 
+    const int64_t n_embd_k_gqa = k->ne[0];
     const int64_t n_tokens = k_cur->ne[2];
 
+    k_cur = ggml_reshape_2d(ctx, k_cur, k->ne[0], n_tokens);
+
+    if (k_idxs && supports_set_rows) {
+        return ggml_set_rows(ctx, k, k_cur, k_idxs);
+    }
+
+    // TODO: fallback to old ggml_cpy() method for backwards compatibility
+    //       will be removed when ggml_set_rows() is adopted by all backends
+
     ggml_tensor * k_view = ggml_view_1d(ctx, k,
-            n_tokens*hparams.n_embd_k_gqa(il),
-            ggml_row_size(k->type, hparams.n_embd_k_gqa(il))*head_cur);
+            n_tokens*n_embd_k_gqa,
+            ggml_row_size(k->type, n_embd_k_gqa)*sinfo.head());
 
     return ggml_cpy(ctx, k_cur, k_view);
 }
 
-ggml_tensor * llama_kv_cache_unified::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il, uint32_t head_cur) const {
+ggml_tensor * llama_kv_cache_unified::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il, const slot_info & sinfo) const {
     const int32_t ikv = map_layer_ids.at(il);
 
     auto * v = layers[ikv].v;
 
+    const int64_t n_embd_v_gqa = v->ne[0];
     const int64_t n_tokens = v_cur->ne[2];
 
-    v_cur = ggml_reshape_2d(ctx, v_cur, hparams.n_embd_v_gqa(il), n_tokens);
+    v_cur = ggml_reshape_2d(ctx, v_cur, n_embd_v_gqa, n_tokens);
+
+    if (v_idxs && supports_set_rows) {
+        if (!v_trans) {
+            return ggml_set_rows(ctx, v, v_cur, v_idxs);
+        }
+
+        // the row becomes a single element
+        ggml_tensor * v_view = ggml_reshape_3d(ctx, v, 1, v->ne[1], v->ne[0]);
+
+        // note: the V cache is transposed when not using flash attention
+        v_cur = ggml_permute(ctx, ggml_reshape_3d(ctx, v_cur, v_cur->ne[0], 1, v_cur->ne[1]), 2, 0, 1, 3);
+
+        // note: we can be more explicit here at the cost of extra cont
+        //       however, above we take advantage that a row of single element is always continuous regardless of the row stride
+        //v_cur = ggml_transpose(ctx, v_cur);
+        //v_cur = ggml_cont_3d(ctx, v_cur, 1, v_cur->ne[0], v_cur->ne[1]);
+
+        // we broadcast the KV indices n_embd_v_gqa times
+        // v      [1,        n_kv,     n_embd_v_gqa]
+        // v_cur  [1,        n_tokens, n_embd_v_gqa]
+        // v_idxs [n_tokens, 1,        1]
+        return ggml_set_rows(ctx, v_view, v_cur, v_idxs);
+    }
+
+    // TODO: fallback to old ggml_cpy() method for backwards compatibility
+    //       will be removed when ggml_set_rows() is adopted by all backends
 
     ggml_tensor * v_view = nullptr;
 
     if (!v_trans) {
         v_view = ggml_view_1d(ctx, v,
-                n_tokens*hparams.n_embd_v_gqa(il),
-                ggml_row_size(v->type, hparams.n_embd_v_gqa(il))*head_cur);
+                n_tokens*n_embd_v_gqa,
+                ggml_row_size(v->type, n_embd_v_gqa)*sinfo.head());
     } else {
-        // note: the V cache is transposed when not using flash attention
-        v_view = ggml_view_2d(ctx, v, n_tokens, hparams.n_embd_v_gqa(il),
-                (v->ne[1])*ggml_element_size(v),
-                (head_cur)*ggml_element_size(v));
-
         v_cur = ggml_transpose(ctx, v_cur);
+
+        v_view = ggml_view_2d(ctx, v, n_tokens, n_embd_v_gqa,
+                (v->ne[1]    )*ggml_element_size(v),
+                (sinfo.head())*ggml_element_size(v));
     }
 
     return ggml_cpy(ctx, v_cur, v_view);
 }
 
+ggml_tensor * llama_kv_cache_unified::build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
+    const uint32_t n_tokens = ubatch.n_tokens;
+
+    ggml_tensor * k_idxs = ggml_new_tensor_1d(ctx, GGML_TYPE_I64, n_tokens);
+
+    ggml_set_input(k_idxs);
+
+    return k_idxs;
+}
+
+ggml_tensor * llama_kv_cache_unified::build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
+    const uint32_t n_tokens = ubatch.n_tokens;
+
+    ggml_tensor * v_idxs = ggml_new_tensor_1d(ctx, GGML_TYPE_I64, n_tokens);
+
+    ggml_set_input(v_idxs);
+
+    return v_idxs;
+}
+
+void llama_kv_cache_unified::set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const {
+    if (!supports_set_rows) {
+        return;
+    }
+
+    const uint32_t n_tokens = ubatch->n_tokens;
+
+    GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
+    int64_t * data = (int64_t *) dst->data;
+
+    for (int64_t i = 0; i < n_tokens; ++i) {
+        data[i] = sinfo.idxs.at(i);
+    }
+}
+
+void llama_kv_cache_unified::set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const {
+    if (!supports_set_rows) {
+        return;
+    }
+
+    const uint32_t n_tokens = ubatch->n_tokens;
+
+    GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
+    int64_t * data = (int64_t *) dst->data;
+
+    for (int64_t i = 0; i < n_tokens; ++i) {
+        data[i] = sinfo.idxs.at(i);
+    }
+}
+
 void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const {
     const uint32_t n_tokens = ubatch->n_tokens;
 
@@ -1552,13 +1675,15 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell
             ubatch.seq_id[i]   = &dest_seq_id;
         }
 
-        const auto head_cur = find_slot(ubatch);
-        if (head_cur < 0) {
+        const auto sinfo = find_slot(ubatch, true);
+        if (sinfo.empty()) {
             LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__);
             return false;
         }
 
-        apply_ubatch(head_cur, ubatch);
+        apply_ubatch(sinfo, ubatch);
+
+        const auto head_cur = sinfo.head();
 
         // keep the head at the old position because we will read the KV data into it in state_read_data()
         head = head_cur;
@@ -1744,7 +1869,11 @@ llama_kv_cache_unified_context::llama_kv_cache_unified_context(llama_memory_stat
 llama_kv_cache_unified_context::llama_kv_cache_unified_context(
         llama_kv_cache_unified * kv) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv) {
     n_kv = kv->get_size();
-    head = 0;
+
+    // create a dummy slot info - the actual data is irrelevant. we just need to build the graph
+    sinfos.resize(1);
+    sinfos[0].idxs.resize(1);
+    sinfos[0].idxs[0] = 0;
 }
 
 llama_kv_cache_unified_context::llama_kv_cache_unified_context(
@@ -1759,8 +1888,8 @@ llama_kv_cache_unified_context::llama_kv_cache_unified_context(
 
 llama_kv_cache_unified_context::llama_kv_cache_unified_context(
         llama_kv_cache_unified * kv,
-        llama_kv_cache_unified::ubatch_heads heads,
-        std::vector<llama_ubatch> ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), heads(std::move(heads)), ubatches(std::move(ubatches)) {
+        llama_kv_cache_unified::slot_info_vec_t sinfos,
+        std::vector<llama_ubatch> ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), sinfos(std::move(sinfos)), ubatches(std::move(ubatches)) {
 }
 
 llama_kv_cache_unified_context::~llama_kv_cache_unified_context() = default;
@@ -1768,7 +1897,7 @@ llama_kv_cache_unified_context::~llama_kv_cache_unified_context() = default;
 bool llama_kv_cache_unified_context::next() {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    if (++i_next >= ubatches.size()) {
+    if (++i_cur >= ubatches.size()) {
         return false;
     }
 
@@ -1785,10 +1914,9 @@ bool llama_kv_cache_unified_context::apply() {
         return true;
     }
 
-    kv->apply_ubatch(heads[i_next], ubatches[i_next]);
+    kv->apply_ubatch(sinfos[i_cur], ubatches[i_cur]);
 
     n_kv = kv->get_n_kv();
-    head = heads[i_next];
 
     return true;
 }
@@ -1800,7 +1928,7 @@ llama_memory_status llama_kv_cache_unified_context::get_status() const {
 const llama_ubatch & llama_kv_cache_unified_context::get_ubatch() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    return ubatches[i_next];
+    return ubatches[i_cur];
 }
 
 uint32_t llama_kv_cache_unified_context::get_n_kv() const {
@@ -1815,18 +1943,34 @@ ggml_tensor * llama_kv_cache_unified_context::get_v(ggml_context * ctx, int32_t
     return kv->get_v(ctx, il, n_kv);
 }
 
-ggml_tensor * llama_kv_cache_unified_context::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il) const {
-    return kv->cpy_k(ctx, k_cur, il, head);
+ggml_tensor * llama_kv_cache_unified_context::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il) const {
+    return kv->cpy_k(ctx, k_cur, k_idxs, il, sinfos[i_cur]);
 }
 
-ggml_tensor * llama_kv_cache_unified_context::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il) const {
-    return kv->cpy_v(ctx, v_cur, il, head);
+ggml_tensor * llama_kv_cache_unified_context::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il) const {
+    return kv->cpy_v(ctx, v_cur, v_idxs, il, sinfos[i_cur]);
+}
+
+ggml_tensor * llama_kv_cache_unified_context::build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
+    return kv->build_input_k_idxs(ctx, ubatch);
+}
+
+ggml_tensor * llama_kv_cache_unified_context::build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
+    return kv->build_input_v_idxs(ctx, ubatch);
 }
 
 void llama_kv_cache_unified_context::set_input_k_shift(ggml_tensor * dst) const {
     kv->set_input_k_shift(dst);
 }
 
+void llama_kv_cache_unified_context::set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const {
+    kv->set_input_k_idxs(dst, ubatch, sinfos[i_cur]);
+}
+
+void llama_kv_cache_unified_context::set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const {
+    kv->set_input_v_idxs(dst, ubatch, sinfos[i_cur]);
+}
+
 void llama_kv_cache_unified_context::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const {
     kv->set_input_kq_mask(dst, ubatch, causal_attn);
 }
diff --git a/src/llama-kv-cache-unified.h b/src/llama-kv-cache-unified.h
index 4c53f1273a..b8b0356e83 100644
--- a/src/llama-kv-cache-unified.h
+++ b/src/llama-kv-cache-unified.h
@@ -24,8 +24,6 @@ public:
     // this callback is used to filter out layers that should not be included in the cache
     using layer_filter_cb = std::function<bool(int32_t il)>;
 
-    using ubatch_heads = std::vector<uint32_t>;
-
     struct defrag_info {
         bool empty() const {
             return ids.empty();
@@ -37,6 +35,32 @@ public:
         std::vector<uint32_t> ids;
     };
 
+    // for each ubatch, create a slot_info that contains information about where the ubatch should be inserted in the
+    //   KV cells. for example, cell indices for each token, such that: token[i] -> goes to cells[idxs[i]]
+    struct slot_info {
+        // data for ggml_set_rows
+        using idx_vec_t = std::vector<uint32_t>;
+
+        idx_vec_t idxs;
+
+        uint32_t head() const {
+            return idxs.at(0);
+        }
+
+        bool empty() const {
+            return idxs.empty();
+        }
+
+        void clear() {
+            idxs.clear();
+        }
+
+        // TODO: implement
+        //std::vector<idx_vec_t> seq_idxs;
+    };
+
+    using slot_info_vec_t = std::vector<slot_info>;
+
     llama_kv_cache_unified(
             const llama_model &  model,
               layer_filter_cb && filter,
@@ -102,30 +126,37 @@ public:
     ggml_tensor * get_v(ggml_context * ctx, int32_t il, uint32_t n_kv) const;
 
     // store k_cur and v_cur in the cache based on the provided head location
-    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il, uint32_t head_cur) const;
-    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il, uint32_t head_cur) const;
+    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const;
+    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il, const slot_info & sinfo) const;
 
     //
     // preparation API
     //
 
-    // find places for the provided ubatches in the cache, returns the head locations
+    // find places for the provided ubatches in the cache, returns the slot infos
     // return empty vector on failure
-    ubatch_heads prepare(const std::vector<llama_ubatch> & ubatches);
+    slot_info_vec_t prepare(const std::vector<llama_ubatch> & ubatches);
 
     bool update(llama_context * lctx, bool do_shift, const defrag_info & dinfo);
 
-    // return the cell position where we can insert the ubatch
-    // return -1 on failure to find a contiguous slot of kv cells
-    int32_t find_slot(const llama_ubatch & ubatch) const;
+    // find a slot of kv cells that can hold the ubatch
+    // if cont == true, then the slot must be continuous
+    // return empty slot_info on failure
+    slot_info find_slot(const llama_ubatch & ubatch, bool cont) const;
 
-    // emplace the ubatch context into slot: [head_cur, head_cur + ubatch.n_tokens)
-    void apply_ubatch(uint32_t head_cur, const llama_ubatch & ubatch);
+    // emplace the ubatch context into slot: [sinfo.idxs[0...ubatch.n_tokens - 1]]
+    void apply_ubatch(const slot_info & sinfo, const llama_ubatch & ubatch);
 
     //
-    // set_input API
+    // input API
     //
 
+    ggml_tensor * build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
+    ggml_tensor * build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
+
+    void set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const;
+    void set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const;
+
     void set_input_kq_mask   (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
     void set_input_k_shift   (ggml_tensor * dst) const;
     void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;
@@ -157,8 +188,13 @@ private:
     // SWA
     const uint32_t n_swa = 0;
 
+    // env: LLAMA_KV_CACHE_DEBUG
     int debug = 0;
 
+    // env: LLAMA_SET_ROWS (temporary)
+    // ref: https://github.com/ggml-org/llama.cpp/pull/14285
+    int supports_set_rows = false;
+
     const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 
     std::vector<ggml_context_ptr>        ctxs;
@@ -211,8 +247,8 @@ private:
 class llama_kv_cache_unified_context : public llama_memory_context_i {
 public:
     // some shorthands
-    using ubatch_heads = llama_kv_cache_unified::ubatch_heads;
-    using defrag_info  = llama_kv_cache_unified::defrag_info;
+    using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t;
+    using defrag_info     = llama_kv_cache_unified::defrag_info;
 
     // used for errors
     llama_kv_cache_unified_context(llama_memory_status status);
@@ -231,7 +267,7 @@ public:
     // used to create a batch procesing context from a batch
     llama_kv_cache_unified_context(
             llama_kv_cache_unified * kv,
-            ubatch_heads heads,
+            slot_info_vec_t sinfos,
             std::vector<llama_ubatch> ubatches);
 
     virtual ~llama_kv_cache_unified_context();
@@ -257,11 +293,16 @@ public:
     ggml_tensor * get_v(ggml_context * ctx, int32_t il) const;
 
     // store k_cur and v_cur in the cache based on the provided head location
-    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il) const;
-    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il) const;
+    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il) const;
+    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il) const;
 
-    void set_input_k_shift(ggml_tensor * dst) const;
+    ggml_tensor * build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
+    ggml_tensor * build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
 
+    void set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const;
+    void set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const;
+
+    void set_input_k_shift   (ggml_tensor * dst) const;
     void set_input_kq_mask   (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
     void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;
 
@@ -283,10 +324,10 @@ private:
     // batch processing context
     //
 
-    // the index of the next ubatch to process
-    size_t i_next = 0;
+    // the index of the cur ubatch to process
+    size_t i_cur = 0;
 
-    ubatch_heads heads;
+    slot_info_vec_t sinfos;
 
     std::vector<llama_ubatch> ubatches;
 
@@ -297,7 +338,4 @@ private:
     // a heuristic, to avoid attending the full cache if it is not yet utilized
     // as the cache gets filled, the benefit from this heuristic disappears
     int32_t n_kv;
-
-    // the beginning of the current slot in which the ubatch will be inserted
-    int32_t head;
 };
diff --git a/src/llama-kv-cells.h b/src/llama-kv-cells.h
index c95d635948..0d0dd316fd 100644
--- a/src/llama-kv-cells.h
+++ b/src/llama-kv-cells.h
@@ -105,10 +105,30 @@ public:
         res.resize(n);
 
         for (uint32_t j = 0; j < n; ++j) {
-            res.pos[j] = pos[i + j];
-            res.seq[j] = seq[i + j];
+            const auto idx = i + j;
 
-            assert(shift[i + j] == 0);
+            res.pos[j] = pos[idx];
+            res.seq[j] = seq[idx];
+
+            assert(shift[idx] == 0);
+        }
+
+        return res;
+    }
+
+    // copy the state of cells [idxs[0], idxs[1], ..., idxs[idxs.size() - 1])
+    llama_kv_cells_unified cp(const std::vector<uint32_t> & idxs) const {
+        llama_kv_cells_unified res;
+
+        res.resize(idxs.size());
+
+        for (uint32_t j = 0; j < idxs.size(); ++j) {
+            const auto idx = idxs[j];
+
+            res.pos[j] = pos[idx];
+            res.seq[j] = seq[idx];
+
+            assert(shift[idx] == 0);
         }
 
         return res;
@@ -119,26 +139,58 @@ public:
         assert(i + other.pos.size() <= pos.size());
 
         for (uint32_t j = 0; j < other.pos.size(); ++j) {
-            if (pos[i + j] == -1 && other.pos[j] != -1) {
+            const auto idx = i + j;
+
+            if (pos[idx] == -1 && other.pos[j] != -1) {
                 used.insert(i + j);
             }
 
-            if (pos[i + j] != -1 && other.pos[j] == -1) {
+            if (pos[idx] != -1 && other.pos[j] == -1) {
                 used.erase(i + j);
             }
 
-            if (pos[i + j] != -1) {
+            if (pos[idx] != -1) {
                 seq_pos_rm(i + j);
             }
 
-            pos[i + j] = other.pos[j];
-            seq[i + j] = other.seq[j];
+            pos[idx] = other.pos[j];
+            seq[idx] = other.seq[j];
 
-            if (pos[i + j] != -1) {
+            if (pos[idx] != -1) {
                 seq_pos_add(i + j);
             }
 
-            assert(shift[i + j] == 0);
+            assert(shift[idx] == 0);
+        }
+    }
+
+    // set the state of cells [idxs[0], idxs[1], ..., idxs[idxs.size() - 1])
+    void set(const std::vector<uint32_t> & idxs, const llama_kv_cells_unified & other) {
+        assert(idxs.size() == other.pos.size());
+
+        for (uint32_t j = 0; j < other.pos.size(); ++j) {
+            const auto idx = idxs[j];
+
+            if (pos[idx] == -1 && other.pos[j] != -1) {
+                used.insert(idx);
+            }
+
+            if (pos[idx] != -1 && other.pos[j] == -1) {
+                used.erase(idx);
+            }
+
+            if (pos[idx] != -1) {
+                seq_pos_rm(idx);
+            }
+
+            pos[idx] = other.pos[j];
+            seq[idx] = other.seq[j];
+
+            if (pos[idx] != -1) {
+                seq_pos_add(idx);
+            }
+
+            assert(shift[idx] == 0);
         }
     }
 
diff --git a/src/llama-memory-hybrid.cpp b/src/llama-memory-hybrid.cpp
index 67cbf95548..03d974d852 100644
--- a/src/llama-memory-hybrid.cpp
+++ b/src/llama-memory-hybrid.cpp
@@ -195,11 +195,11 @@ llama_memory_hybrid_context::llama_memory_hybrid_context(
 
 llama_memory_hybrid_context::llama_memory_hybrid_context(
               llama_memory_hybrid * mem,
-            std::vector<uint32_t>   heads_attn,
+                  slot_info_vec_t   sinfos_attn,
         std::vector<llama_ubatch>   ubatches) :
     ubatches(std::move(ubatches)),
     // note: here we copy the ubatches. not sure if this is ideal
-    ctx_attn(new llama_kv_cache_unified_context(mem->get_mem_attn(), std::move(heads_attn), this->ubatches)),
+    ctx_attn(new llama_kv_cache_unified_context(mem->get_mem_attn(), std::move(sinfos_attn), this->ubatches)),
     ctx_recr(new llama_memory_recurrent_context(mem->get_mem_recr(),                        this->ubatches)),
     status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
 }
diff --git a/src/llama-memory-hybrid.h b/src/llama-memory-hybrid.h
index f0c2420e9a..4ac3181757 100644
--- a/src/llama-memory-hybrid.h
+++ b/src/llama-memory-hybrid.h
@@ -92,6 +92,8 @@ private:
 
 class llama_memory_hybrid_context : public llama_memory_context_i {
 public:
+    using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t;
+
     // init failure
     explicit llama_memory_hybrid_context(llama_memory_status status);
 
@@ -107,7 +109,7 @@ public:
     // init success
     llama_memory_hybrid_context(
               llama_memory_hybrid * mem,
-            std::vector<uint32_t>   heads_attn,
+                  slot_info_vec_t   sinfos_attn,
         std::vector<llama_ubatch>   ubatches);
 
     ~llama_memory_hybrid_context() = default;

From 0c2ee38ab729f3f6a7c28dc4c5dcd8d5f0cbf8fc Mon Sep 17 00:00:00 2001
From: Xuan-Son Nguyen <son@huggingface.co>
Date: Thu, 3 Jul 2025 10:03:06 +0200
Subject: [PATCH 06/10] convert : correct gemma 3n conversion (#14450)

* convert : correct gemma 3n conversion

* rm redundant code
---
 gguf-py/gguf/gguf_writer.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py
index 697e057c09..a7ecf3d312 100644
--- a/gguf-py/gguf/gguf_writer.py
+++ b/gguf-py/gguf/gguf_writer.py
@@ -714,8 +714,8 @@ class GGUFWriter:
     def add_clamp_kqv(self, value: float) -> None:
         self.add_float32(Keys.Attention.CLAMP_KQV.format(arch=self.arch), value)
 
-    def add_shared_kv_layers(self, value: float) -> None:
-        self.add_float32(Keys.Attention.SHARED_KV_LAYERS.format(arch=self.arch), value)
+    def add_shared_kv_layers(self, value: int) -> None:
+        self.add_uint32(Keys.Attention.SHARED_KV_LAYERS.format(arch=self.arch), value)
 
     def add_sliding_window_pattern(self, value: Sequence[bool]) -> None:
         self.add_array(Keys.Attention.SLIDING_WINDOW_PATTERN.format(arch=self.arch), value)

From 7b63a71a6b0f54effe9b94073d4d0519dcf53676 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nicol=C3=B2=20Scipione?= <nicolo.scipione@codeplay.com>
Date: Thu, 3 Jul 2025 11:00:03 +0200
Subject: [PATCH 07/10] Fix conditional enabling following arch checks for
 ggml-sycl (#14504)

Signed-off-by: nscipione <nicolo.scipione@codeplay.com>
---
 ggml/src/ggml-sycl/ggml-sycl.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
index ee4c8f7b2f..b063a698de 100644
--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -83,7 +83,7 @@ static ggml_sycl_device_info ggml_sycl_init() {
 
         info.devices[i].cc =
             100 * prop.get_major_version() + 10 * prop.get_minor_version();
-        info.devices[i].opt_feature.reorder = !device.ext_oneapi_architecture_is(syclex::arch_category::intel_gpu);
+        info.devices[i].opt_feature.reorder = device.ext_oneapi_architecture_is(syclex::arch_category::intel_gpu);
         info.max_work_group_sizes[i] = prop.get_max_work_group_size();
     }
 

From c8c4495b8d3a8799e2d46778f993965b0ac1ae43 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Thu, 3 Jul 2025 17:05:18 +0200
Subject: [PATCH 08/10] ggml: backward pass for split swiglu (#14483)

---
 ggml/src/ggml.c            | 19 +++++++++++++++++--
 tests/test-backend-ops.cpp |  4 ++++
 2 files changed, 21 insertions(+), 2 deletions(-)

diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index fdb57e1785..b481193da3 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -6050,13 +6050,28 @@ static void ggml_compute_backward(
             }
             GGML_ASSERT(!src1_needs_grads && "backward pass for labels not implemented");
         } break;
+        case GGML_OP_GLU: {
+            switch (ggml_get_glu_op(tensor)) {
+                case GGML_GLU_OP_SWIGLU: {
+                    if (src0_needs_grads) {
+                        GGML_ASSERT(src1 && "backward pass only implemented for split swiglu");
+                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_silu_back(ctx, ggml_mul(ctx, grad, src1), src0));
+                    }
+                    if (src1_needs_grads) {
+                        ggml_add_or_set(ctx, cgraph, isrc1, ggml_mul(ctx, ggml_silu(ctx, src0), grad));
+                    }
+                } break;
+                default: {
+                    GGML_ABORT("unsupported glu op for backward pass: %s", ggml_glu_op_name(ggml_get_glu_op(tensor)));
+                } //break;
+            }
+        } break;
         case GGML_OP_NONE: {
             // noop
         } break;
         case GGML_OP_COUNT:
         default: {
-            fprintf(stderr, "%s: unsupported ggml op for backward pass: %s\n", __func__, ggml_op_name(tensor->op));
-            GGML_ABORT("fatal error");
+            GGML_ABORT("%s: unsupported ggml op for backward pass: %s\n", __func__, ggml_op_name(tensor->op));
         } //break;
     }
 
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 2ab6dd06fc..c76635793f 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -1175,21 +1175,25 @@ struct test_glu_split : public test_case {
         if (v & 1) {
             auto ne = ne_a; ne[0] *= 3;
             a = ggml_new_tensor(ctx, type, 4, ne.data());
+            ggml_set_param(a);
             ggml_set_name(a, "a");
 
             a = ggml_view_4d(ctx, a, ne_a[0], ne_a[1], ne_a[2], ne_a[3], a->nb[1], a->nb[2], a->nb[3], 0);
             ggml_set_name(a, "view_of_a");
 
             b = ggml_new_tensor(ctx, type, 4, ne.data());
+            ggml_set_param(b);
             ggml_set_name(b, "b");
 
             b = ggml_view_4d(ctx, b, ne_a[0], ne_a[1], ne_a[2], ne_a[3], b->nb[1], b->nb[2], b->nb[3], 0);
             ggml_set_name(a, "view_of_b");
         } else {
             a = ggml_new_tensor(ctx, type, 4, ne_a.data());
+            ggml_set_param(a);
             ggml_set_name(a, "a");
 
             b = ggml_new_tensor(ctx, type, 4, ne_a.data());
+            ggml_set_param(b);
             ggml_set_name(b, "b");
         }
 

From 2b72bedec198a90bb5b0cceaf1d0aff9e34ffbc2 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Thu, 3 Jul 2025 13:21:14 -0500
Subject: [PATCH 09/10] vulkan: support mixed/deepseekR1 FA head sizes (#14509)

* vulkan: better parameterize FA by head sizes

* vulkan: support mixed/deepseekR1 FA head sizes
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp          | 228 ++++++++++--------
 .../vulkan-shaders/flash_attn.comp            |  43 ++--
 .../vulkan-shaders/flash_attn_base.comp       |   8 +-
 .../vulkan-shaders/flash_attn_cm1.comp        |  62 ++---
 .../vulkan-shaders/flash_attn_cm2.comp        |  50 ++--
 5 files changed, 208 insertions(+), 183 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 8114cb4204..c0032cba21 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -224,6 +224,21 @@ enum vk_device_architecture {
     INTEL_XE2,
 };
 
+// HSK x HSV
+enum FaHeadSizes {
+    FA_HEAD_SIZE_64,
+    FA_HEAD_SIZE_80,
+    FA_HEAD_SIZE_96,
+    FA_HEAD_SIZE_112,
+    FA_HEAD_SIZE_128,
+    FA_HEAD_SIZE_192,
+    FA_HEAD_SIZE_192_128,
+    FA_HEAD_SIZE_256,
+    FA_HEAD_SIZE_576_512,
+    FA_HEAD_SIZE_UNSUPPORTED,
+    FA_HEAD_SIZE_COUNT = FA_HEAD_SIZE_UNSUPPORTED,
+};
+
 static vk_device_architecture get_device_architecture(const vk::PhysicalDevice& device) {
     vk::PhysicalDeviceProperties props = device.getProperties();
 
@@ -467,26 +482,11 @@ struct vk_device_struct {
     vk_pipeline pipeline_conv2d_dw_cwhn_f32;
 
     // [2][2][2] is for {f16acc,f32acc}x{large,small_rows}x{unaligned, aligned}
-    vk_pipeline pipeline_flash_attn_f32_f16_D64_cm2[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D80_cm2[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D96_cm2[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D112_cm2[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D128_cm2[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D256_cm2[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_cm2[GGML_TYPE_COUNT][FA_HEAD_SIZE_COUNT][2][2][2];
 
-    vk_pipeline pipeline_flash_attn_f32_f16_D64_cm1[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D80_cm1[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D96_cm1[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D112_cm1[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D128_cm1[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D256_cm1[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_cm1[GGML_TYPE_COUNT][FA_HEAD_SIZE_COUNT][2][2][2];
 
-    vk_pipeline pipeline_flash_attn_f32_f16_D64[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D80[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D96[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D112[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D128[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D256[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16[GGML_TYPE_COUNT][FA_HEAD_SIZE_COUNT][2][2][2];
 
     vk_pipeline pipeline_flash_attn_split_k_reduce;
 
@@ -1003,7 +1003,7 @@ struct ggml_backend_vk_context {
 
     // number of additional consecutive nodes that are being fused with the
     // node currently being processed
-    uint32_t num_additional_fused_ops {};
+    int num_additional_fused_ops {};
 };
 
 static void * const vk_ptr_base = (void *)(uintptr_t) 0x1000;  // NOLINT
@@ -1699,6 +1699,35 @@ enum FaCodePath {
     FA_COOPMAT2,
 };
 
+static FaHeadSizes fa_get_head_sizes(uint32_t hsk, uint32_t hsv) {
+    if (hsk != 192 && hsk != 576 && hsk != hsv) {
+        return FA_HEAD_SIZE_UNSUPPORTED;
+    }
+    switch (hsk) {
+    case 64: return FA_HEAD_SIZE_64;
+    case 80: return FA_HEAD_SIZE_80;
+    case 96: return FA_HEAD_SIZE_96;
+    case 112: return FA_HEAD_SIZE_112;
+    case 128: return FA_HEAD_SIZE_128;
+    case 192:
+        if (hsv == 192) {
+            return FA_HEAD_SIZE_192;
+        } else if (hsv == 128) {
+            return FA_HEAD_SIZE_192_128;
+        } else {
+            return FA_HEAD_SIZE_UNSUPPORTED;
+        }
+    case 256: return FA_HEAD_SIZE_256;
+    case 576:
+        if (hsv == 512) {
+            return FA_HEAD_SIZE_576_512;
+        } else {
+            return FA_HEAD_SIZE_UNSUPPORTED;
+        }
+    default: return FA_HEAD_SIZE_UNSUPPORTED;
+    }
+}
+
 // number of rows/cols for flash attention shader
 static constexpr uint32_t flash_attention_num_small_rows = 32;
 static constexpr uint32_t scalar_flash_attention_num_small_rows = 1;
@@ -1719,8 +1748,9 @@ static uint32_t get_fa_num_small_rows(FaCodePath path) {
     }
 }
 
-static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) {
+static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows) {
     GGML_UNUSED(clamp);
+    GGML_UNUSED(hsv);
 
     if (path == FA_SCALAR) {
         if (small_rows) {
@@ -1744,7 +1774,7 @@ static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t D, uint32_
     }
 
     // small cols to reduce register count
-    if (ggml_is_quantized(type) || D == 256) {
+    if (ggml_is_quantized(type) || hsk >= 256) {
         return {64, 32};
     }
     return {64, 64};
@@ -2037,19 +2067,21 @@ static void ggml_vk_load_shaders(vk_device& device) {
                                       parameter_count, wg_denoms, specialization_constants, disable_robustness, require_full_subgroups, required_subgroup_size));
     };
 
-    auto const &fa_wg_denoms = [&](FaCodePath path, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::array<uint32_t, 3> {
-        return {fa_rows_cols(path, D, clamp, type, small_rows)[0], 1, 1};
+    auto const &fa_wg_denoms = [&](FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows) -> std::array<uint32_t, 3> {
+        return {fa_rows_cols(path, hsk, hsv, clamp, type, small_rows)[0], 1, 1};
     };
 
-    auto const &fa_spec_constants = [&](FaCodePath path, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::vector<uint32_t> {
+    auto const &fa_spec_constants = [&](FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows) -> std::vector<uint32_t> {
         // For large number of rows, 128 invocations seems to work best.
         // For small number of rows (e.g. N==1), 256 works better. But matrix granularity for 256 is 32, so we
         // can't use 256 for D==80.
         // For scalar, use 128 (arbitrary)
+        // The same D_split value is used for both HSK and HSV, so just base it on the union of the LSBs.
+        const uint32_t D = (hsk|hsv);
         uint32_t wg_size = (path == FA_SCALAR || path == FA_COOPMAT1)
                             ? scalar_flash_attention_workgroup_size
                             : ((small_rows && (D % 32) == 0) ? 256 : 128);
-        auto rows_cols = fa_rows_cols(path, D, clamp, type, small_rows);
+        auto rows_cols = fa_rows_cols(path, hsk, hsv, clamp, type, small_rows);
 
         // D_split can't be larger than a subgroup because we use subgroupShuffle to reduce it.
         // D_split can't be larger than the LSB of D divided by 4 due to vectorization in the shader.
@@ -2058,26 +2090,29 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
         // mask dim1 is padded to 64, we rely on this to avoid clamping mask loads
         GGML_ASSERT((GGML_KQ_MASK_PAD % rows_cols[0]) == 0);
-        return {wg_size, rows_cols[0], rows_cols[1], (D), clamp, D_split};
+        return {wg_size, rows_cols[0], rows_cols[1], hsk, hsv, clamp, D_split};
     };
 
-#define CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, D) \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][0], "flash_attn_f32_f16_D" #D "_f16acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,false), fa_spec_constants(FAPATH, D,1,TYPE,false), 1,                                      true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,false), fa_spec_constants(FAPATH, D,0,TYPE,false), fa_rows_cols(FAPATH,D,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][0], "flash_attn_f32_f16_D" #D "_f32acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,false), fa_spec_constants(FAPATH, D,1,TYPE,false), 1,                                      true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,false), fa_spec_constants(FAPATH, D,0,TYPE,false), fa_rows_cols(FAPATH,D,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][0], "flash_attn_f32_f16_D" #D "_f16acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,true), fa_spec_constants(FAPATH, D,1,TYPE,true),   1,                                      true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,true), fa_spec_constants(FAPATH, D,0,TYPE,true),   fa_rows_cols(FAPATH,D,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][0], "flash_attn_f32_f16_D" #D "_f32acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,true), fa_spec_constants(FAPATH, D,1,TYPE,true),   1,                                      true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,true), fa_spec_constants(FAPATH, D,0,TYPE,true),   fa_rows_cols(FAPATH,D,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+#define CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, HSK, HSV, HEAD_SIZES) \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][0][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f16acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,false), 1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][0][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f16acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,false), fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][0][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f32acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,false), 1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][0][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f32acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,false), fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][1][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f16acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,true),   1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][1][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,true),   fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][1][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f32acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,true),   1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][1][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,true),   fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
 
 #define CREATE_FA(TYPE, NAMELC, FAPATH, SUFFIX) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 64) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 80) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 96) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 112) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 128) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 256)
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 64, 64, 64) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 80, 80, 80) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 96, 96, 96) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 112, 112, 112) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 128, 128, 128) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 192, 192, 192) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 192, 128, 192_128) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 256, 256, 256) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 576, 512, 576_512)
 
     CREATE_FA(GGML_TYPE_F16, f16, FA_SCALAR, )
     CREATE_FA(GGML_TYPE_Q4_0, q4_0, FA_SCALAR, )
@@ -3688,7 +3723,6 @@ static void ggml_vk_instance_init() {
 
     }
 
-    size_t num_available_devices = vk_instance.instance.enumeratePhysicalDevices().size();
     vk_perf_logger_enabled = getenv("GGML_VK_PERF_LOGGER") != nullptr;
 
     // Emulate behavior of CUDA_VISIBLE_DEVICES for Vulkan
@@ -6002,24 +6036,47 @@ static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx
     }
 }
 
-static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const uint32_t D, bool f32acc) {
+static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, const uint32_t hsk, uint32_t hsv) {
     // Needs to be kept up to date on shader changes
+    GGML_UNUSED(hsv);
     const uint32_t wg_size = scalar_flash_attention_workgroup_size;
     const uint32_t Br = scalar_flash_attention_num_large_rows;
     const uint32_t Bc = scalar_flash_attention_Bc;
 
+    const uint32_t tmpsh = wg_size * sizeof(float);
+    const uint32_t tmpshv4 = wg_size * 4 * sizeof(float);
+
+    const uint32_t masksh = Bc * Br * sizeof(float);
+
+    const uint32_t Qf = Br * (hsk / 4 + 2) * 4 * sizeof(float);
+
+    const uint32_t total_size = tmpsh + tmpshv4 + masksh + Qf;
+    const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize;
+
+    VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", total_size=" << total_size << ", supported=" << supported);
+
+    return supported;
+}
+
+static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const uint32_t hsk, uint32_t hsv, bool f32acc) {
+    // Needs to be kept up to date on shader changes
+    GGML_UNUSED(hsv);
+    const uint32_t wg_size = scalar_flash_attention_workgroup_size;
+    const uint32_t Br = coopmat1_flash_attention_num_large_rows;
+    const uint32_t Bc = scalar_flash_attention_Bc;
+
     const uint32_t acctype = f32acc ? 4 : 2;
     const uint32_t f16vec4 = 8;
 
     const uint32_t tmpsh = wg_size * sizeof(float);
     const uint32_t tmpshv4 = wg_size * 4 * acctype;
 
-    const uint32_t Qf = Br * (D / 4 + 2) * f16vec4;
+    const uint32_t Qf = Br * (hsk / 4 + 2) * f16vec4;
 
-    const uint32_t sfshstride = (D <= 128) ? (Br + 8) : Br;
+    const uint32_t sfshstride = (hsk <= 128) ? (Br + 8) : Br;
     const uint32_t sfsh = Bc * sfshstride * acctype;
 
-    const uint32_t kshstride = D / 4 + 2;
+    const uint32_t kshstride = hsk / 4 + 2;
     const uint32_t ksh = Bc * kshstride * f16vec4;
 
     const uint32_t slope = Br * sizeof(float);
@@ -6027,7 +6084,7 @@ static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, co
     const uint32_t total_size = tmpsh + tmpshv4 + Qf + sfsh + ksh + slope;
     const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize;
 
-    VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(D=" << D << ", f32acc=" << f32acc << ", total_size=" << total_size << ", supported=" << supported);
+    VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", f32acc=" << f32acc << ", total_size=" << total_size << ", supported=" << supported);
 
     return supported;
 }
@@ -6051,11 +6108,12 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     const uint32_t nem1 = mask ? mask->ne[1] : 0;
     const uint32_t nem2 = mask ? mask->ne[2] : 0;
 
-    const uint32_t D = neq0;
+    const uint32_t HSK = nek0;
+    const uint32_t HSV = nev0;
     uint32_t N = neq1;
     const uint32_t KV = nek1;
 
-    GGML_ASSERT(ne0 == D);
+    GGML_ASSERT(ne0 == HSV);
     GGML_ASSERT(ne2 == N);
 
     // input tensor rows must be contiguous
@@ -6063,12 +6121,9 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     GGML_ASSERT(nbk0 == ggml_type_size(k->type));
     GGML_ASSERT(nbv0 == ggml_type_size(v->type));
 
-    GGML_ASSERT(neq0 == D);
-    GGML_ASSERT(nek0 == D);
-    GGML_ASSERT(nev0 == D);
+    GGML_ASSERT(neq0 == HSK);
 
     GGML_ASSERT(neq1 == N);
-    GGML_ASSERT(nev0 == D);
 
     GGML_ASSERT(nev1 == nek1);
 
@@ -6089,7 +6144,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         const bool coopmat_shape_supported = (dst->op_params[3] == GGML_PREC_F32 && ctx->device->coopmat_support_16x16x16_f32acc) ||
                                              (dst->op_params[3] != GGML_PREC_F32 && ctx->device->coopmat_support_16x16x16_f16acc);
 
-        const bool coopmat_shmem_supported = ggml_vk_flash_attn_coopmat_shmem_support(ctx->device, D, dst->op_params[3] == GGML_PREC_F32);
+        const bool coopmat_shmem_supported = ggml_vk_flash_attn_coopmat_shmem_support(ctx->device, HSK, HSV, dst->op_params[3] == GGML_PREC_F32);
 
         if (!coopmat_shape_supported || !coopmat_shmem_supported) {
             path = FA_SCALAR;
@@ -6142,47 +6197,25 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         path = FA_SCALAR;
     }
 
+    // with large hsk/hsv, scalar path may need to use small_rows to fit in shared memory
+    if (path == FA_SCALAR &&
+        !ggml_vk_flash_attn_scalar_shmem_support(ctx->device, HSK, HSV)) {
+        small_rows = true;
+    }
+
     bool f32acc = path == FA_SCALAR || dst->op_params[3] == GGML_PREC_F32;
 
+    FaHeadSizes head_sizes = fa_get_head_sizes(k->ne[0], v->ne[0]);
+
     switch (path) {
     case FA_SCALAR:
-        switch (D) {
-        case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64[k->type][f32acc][small_rows][0]; break;
-        case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80[k->type][f32acc][small_rows][0]; break;
-        case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96[k->type][f32acc][small_rows][0]; break;
-        case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112[k->type][f32acc][small_rows][0]; break;
-        case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128[k->type][f32acc][small_rows][0]; break;
-        case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256[k->type][f32acc][small_rows][0]; break;
-        default:
-            GGML_ASSERT(!"unsupported D value");
-            return;
-        }
+        pipelines = &ctx->device->pipeline_flash_attn_f32_f16[k->type][head_sizes][f32acc][small_rows][0];
         break;
     case FA_COOPMAT1:
-        switch (D) {
-        case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64_cm1[k->type][f32acc][small_rows][0]; break;
-        case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80_cm1[k->type][f32acc][small_rows][0]; break;
-        case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96_cm1[k->type][f32acc][small_rows][0]; break;
-        case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112_cm1[k->type][f32acc][small_rows][0]; break;
-        case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128_cm1[k->type][f32acc][small_rows][0]; break;
-        case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256_cm1[k->type][f32acc][small_rows][0]; break;
-        default:
-            GGML_ASSERT(!"unsupported D value");
-            return;
-        }
+        pipelines = &ctx->device->pipeline_flash_attn_f32_f16_cm1[k->type][head_sizes][f32acc][small_rows][0];
         break;
     case FA_COOPMAT2:
-        switch (D) {
-        case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64_cm2[k->type][f32acc][small_rows][0]; break;
-        case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80_cm2[k->type][f32acc][small_rows][0]; break;
-        case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96_cm2[k->type][f32acc][small_rows][0]; break;
-        case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112_cm2[k->type][f32acc][small_rows][0]; break;
-        case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128_cm2[k->type][f32acc][small_rows][0]; break;
-        case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256_cm2[k->type][f32acc][small_rows][0]; break;
-        default:
-            GGML_ASSERT(!"unsupported D value");
-            return;
-        }
+        pipelines = &ctx->device->pipeline_flash_attn_f32_f16_cm2[k->type][head_sizes][f32acc][small_rows][0];
         break;
     default:
         GGML_ASSERT(0);
@@ -6212,7 +6245,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     // Try to use split_k when KV is large enough to be worth the overhead
     if (workgroups_x == 1 && shader_core_count > 0 && KV >= 512) {
         // Try to run two workgroups per SM.
-        split_k = ctx->device->shader_core_count * 2 / (workgroups_y * workgroups_z);
+        split_k = shader_core_count * 2 / (workgroups_y * workgroups_z);
         if (split_k > 1) {
             // Try to evenly split KV into split_k chunks, but it needs to be a multiple
             // of "align", so recompute split_k based on that.
@@ -6224,7 +6257,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
 
     // Reserve space for split_k temporaries. For each split x batch, we need to store the O matrix (D x ne1)
     // and the per-row m and L values (ne1 rows). We store all the matrices first, followed by the rows.
-    const uint64_t split_k_size = split_k > 1 ? (D * ne1 * sizeof(float) + ne1 * sizeof(float) * 2) * split_k * ne3 : 0;
+    const uint64_t split_k_size = split_k > 1 ? (HSV * ne1 * sizeof(float) + ne1 * sizeof(float) * 2) * split_k * ne3 : 0;
     if (split_k_size > ctx->device->max_memory_allocation_size) {
         GGML_ABORT("Requested preallocation size is too large");
     }
@@ -6342,7 +6375,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                     pc, { workgroups_x * pipeline->wg_denoms[0], workgroups_y, workgroups_z });
 
         ggml_vk_sync_buffers(subctx);
-        const std::array<uint32_t, 4> pc2 = { D, (uint32_t)ne1, (uint32_t)ne3, split_k };
+        const std::array<uint32_t, 4> pc2 = { HSV, (uint32_t)ne1, (uint32_t)ne3, split_k };
         ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_flash_attn_split_k_reduce,
                                     {
                                         vk_subbuffer{ctx->prealloc_split_k, 0, VK_WHOLE_SIZE},
@@ -10241,19 +10274,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
                 auto device = ggml_vk_get_device(ctx->device);
                 bool coopmat2 = device->coopmat2;
-                switch (op->src[0]->ne[0]) {
-                case 64:
-                case 80:
-                case 96:
-                case 112:
-                case 128:
-                case 256:
-                    break;
-                default:
-                    return false;
-                }
-                if (op->src[1]->ne[0] != op->src[2]->ne[0]) {
-                    // different head sizes of K and V are not supported yet
+                FaHeadSizes head_sizes = fa_get_head_sizes(op->src[1]->ne[0], op->src[2]->ne[0]);
+                if (head_sizes == FA_HEAD_SIZE_UNSUPPORTED) {
                     return false;
                 }
                 if (op->src[0]->type != GGML_TYPE_F32) {
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
index 6f80101d1c..788a5e065d 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
@@ -11,7 +11,8 @@
 #include "types.comp"
 #include "flash_attn_base.comp"
 
-const uint32_t D_per_thread = D / D_split;
+const uint32_t HSK_per_thread = HSK / D_split;
+const uint32_t HSV_per_thread = HSV / D_split;
 
 const uint32_t cols_per_iter = WorkGroupSize / D_split;
 const uint32_t cols_per_thread = Bc / cols_per_iter;
@@ -29,7 +30,7 @@ layout (binding = 3) readonly buffer M {float16_t data_m[];};
 // Rows index by Q's dimension 2, and the first N rows are valid.
 D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
 {
-    uint32_t offset = (iq2 + r) * D + c;
+    uint32_t offset = (iq2 + r) * HSV + c;
     data_o[o_offset + offset] = D_TYPE(elem);
     return elem;
 }
@@ -38,7 +39,7 @@ shared FLOAT_TYPE tmpsh[WorkGroupSize];
 shared vec4 tmpshv4[WorkGroupSize];
 
 shared float masksh[Bc][Br];
-shared vec4 Qf[Br][D / 4];
+shared vec4 Qf[Br][HSK / 4];
 
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
@@ -53,18 +54,18 @@ void main() {
 
     uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
 
-    [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) {
-        uint32_t d = (idx + tid) % (D / 4);
-        uint32_t r = (idx + tid) / (D / 4);
-        if (r < Br && d < D / 4 &&
+    [[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
+        uint32_t d = (idx + tid) % (HSK / 4);
+        uint32_t r = (idx + tid) / (HSK / 4);
+        if (r < Br && d < HSK / 4 &&
             i * Br + r < N) {
             Qf[r][d] = vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d]) * p.scale;
         }
     }
     barrier();
 
-    vec4 Of[Br][D_per_thread / 4];
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    vec4 Of[Br][HSV_per_thread / 4];
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             Of[r][d] = vec4(0.0);
         }
@@ -116,7 +117,7 @@ void main() {
 
 
         [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+            [[unroll]] for (uint32_t d = 0; d < HSK_per_thread / 4; ++d) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
                 uint ib = coord / BLOCK_SIZE;
@@ -195,14 +196,14 @@ void main() {
             Lf[r] = eMf[r]*Lf[r] + rowsumf[r];
         }
 
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
             [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
                 Of[r][d] = eMf[r] * Of[r][d];
             }
         }
 
         [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+            [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
                 uint ib = coord / BLOCK_SIZE;
@@ -259,7 +260,7 @@ void main() {
         Lf[r] = tmpsh[d_tid];
         barrier();
 
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 
             Of[r][d] = eMf * Of[r][d];
             tmpshv4[tid] = Of[r][d];
@@ -281,11 +282,11 @@ void main() {
     // If there is split_k, then the split_k resolve shader does the final
     // division by L. Store the intermediate O value and per-row m and L values.
     if (p.k_num > 1) {
-        uint32_t o_offset = D * p.ne1 * (split_k_index + iq3 * p.k_num);
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
 
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
                     }
@@ -293,7 +294,7 @@ void main() {
             }
         }
 
-        o_offset = D * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
+        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
                 perElemOpStoreCol0(r, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
@@ -309,18 +310,18 @@ void main() {
         Lfrcp[r] = 1.0 / Lf[r];
     }
 
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             Of[r][d] *= Lfrcp[r];
         }
     }
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1*D;
+    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
 
     if (p.gqa_ratio > 1) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
                     }
@@ -330,9 +331,9 @@ void main() {
     } else {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (i * Br + r < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        data_o[o_offset + iq2 * D + (i * Br + r) * p.ne1 * D + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
+                        data_o[o_offset + iq2 * HSV + (i * Br + r) * p.ne1 * HSV + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
                     }
                 }
             }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp
index 67b935cf57..6609f0bad3 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp
@@ -4,10 +4,10 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 layout (constant_id = 0) const uint32_t WorkGroupSize = 128;
 layout (constant_id = 1) const uint32_t Br = 1;
 layout (constant_id = 2) const uint32_t Bc = 32;
-layout (constant_id = 3) const uint32_t D = 32;
-layout (constant_id = 4) const uint32_t Clamp = 0;
-layout (constant_id = 5) const uint32_t D_split = 16;
-
+layout (constant_id = 3) const uint32_t HSK = 32;
+layout (constant_id = 4) const uint32_t HSV = 32;
+layout (constant_id = 5) const uint32_t Clamp = 0;
+layout (constant_id = 6) const uint32_t D_split = 16;
 
 layout (push_constant) uniform parameter {
     uint32_t N;
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
index 26fe50c7a8..e74e2fa934 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
@@ -13,7 +13,9 @@
 #include "types.comp"
 #include "flash_attn_base.comp"
 
-const uint32_t D_per_thread = D / D_split;
+const uint32_t HSK_per_thread = HSK / D_split;
+const uint32_t HSV_per_thread = HSV / D_split;
+
 const uint32_t row_split = 4;
 const uint32_t rows_per_thread = Br / row_split;
 const uint32_t cols_per_iter = gl_WorkGroupSize.x / D_split / row_split;
@@ -32,7 +34,7 @@ layout (binding = 3) readonly buffer M {float16_t data_m[];};
 // Rows index by Q's dimension 2, and the first N rows are valid.
 D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
 {
-    uint32_t offset = (iq2 + r) * D + c;
+    uint32_t offset = (iq2 + r) * HSV + c;
     data_o[o_offset + offset] = D_TYPE(elem);
     return elem;
 }
@@ -44,14 +46,14 @@ const uint32_t MatBc = 16;
 shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x];
 shared ACC_TYPEV4 tmpshv4[gl_WorkGroupSize.x];
 
-const uint32_t qstride = D / 4 + 2; // in units of f16vec4
+const uint32_t qstride = HSK / 4 + 2; // in units of f16vec4
 shared f16vec4 Qf[Br * qstride];
 
-// Avoid padding for D==256 to make it fit in 48KB shmem.
-const uint32_t sfshstride = (D <= 128) ? (Br + 8) : Br;
+// Avoid padding for hsk==256 to make it fit in 48KB shmem.
+const uint32_t sfshstride = (HSK <= 128) ? (Br + 8) : Br;
 shared ACC_TYPE sfsh[Bc * sfshstride];
 
-const uint32_t kshstride = D / 4 + 2; // in units of f16vec4
+const uint32_t kshstride = HSK / 4 + 2; // in units of f16vec4
 shared f16vec4 ksh[Bc * kshstride];
 
 shared float slope[Br];
@@ -74,18 +76,18 @@ void main() {
 
     uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
 
-    [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) {
-        uint32_t d = (idx + tid) % (D / 4);
-        uint32_t r = (idx + tid) / (D / 4);
-        if (r < Br && d < D / 4 &&
+    [[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
+        uint32_t d = (idx + tid) % (HSK / 4);
+        uint32_t r = (idx + tid) / (HSK / 4);
+        if (r < Br && d < HSK / 4 &&
             i * Br + r < N) {
             Qf[r * qstride + d] = f16vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d] * p.scale);
         }
     }
     barrier();
 
-    ACC_TYPEV4 Of[rows_per_thread][D_per_thread / 4];
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    ACC_TYPEV4 Of[rows_per_thread][HSV_per_thread / 4];
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             Of[r][d] = ACC_TYPEV4(0.0);
         }
@@ -131,10 +133,10 @@ void main() {
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
 
-        [[unroll]] for (uint32_t idx = 0; idx < Bc * D / 4; idx += gl_WorkGroupSize.x) {
-            uint32_t d = (idx + tid) % (D / 4);
-            uint32_t c = (idx + tid) / (D / 4);
-            if (c < Bc && d < D / 4) {
+        [[unroll]] for (uint32_t idx = 0; idx < Bc * HSK / 4; idx += gl_WorkGroupSize.x) {
+            uint32_t d = (idx + tid) % (HSK / 4);
+            uint32_t c = (idx + tid) / (HSK / 4);
+            if (c < Bc && d < HSK / 4) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d;
                 uint ib = coord / BLOCK_SIZE;
@@ -149,14 +151,14 @@ void main() {
         }
         barrier();
 
-        // K * Q^T -> S^T: Bc x D * D x Br -> Bc x Br
-        // Bc split across workgroup (four subgroups), loop over D in chunks of 16: 16 x 16 * 16 x 16 -> 16 x 16
+        // K * Q^T -> S^T: Bc x HSK * HSK x Br -> Bc x Br
+        // Bc split across workgroup (four subgroups), loop over HSK in chunks of 16: 16 x 16 * 16 x 16 -> 16 x 16
         // This is written transposed in order to allow for N being 8 if implementations need it
         coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator> SfMat = coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator>(0);
         coopmat<float16_t, gl_ScopeSubgroup, MatBc, 16, gl_MatrixUseA> KMat;
         coopmat<float16_t, gl_ScopeSubgroup, 16, MatBr, gl_MatrixUseB> QMat;
 
-        for (uint32_t d = 0; d < D / 16; ++d) {
+        for (uint32_t d = 0; d < HSK / 16; ++d) {
             coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor);
 
             uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4;
@@ -206,7 +208,7 @@ void main() {
             eMf[r] = exp(Moldf - Mf[r]);
         }
 
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
             [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
                 Of[r][d] = float16_t(eMf[r]) * Of[r][d];
             }
@@ -221,7 +223,7 @@ void main() {
                 Pf[r] = exp(sfsh[tile_row(r) + (c * cols_per_iter + col_tid) * sfshstride] - Mf[r]);
                 Lf[r] += Pf[r];
             }
-            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+            [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
                 uint ib = coord / BLOCK_SIZE;
@@ -284,7 +286,7 @@ void main() {
     }
 
     [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 
             Of[r][d] = float16_t(eMf[r]) * Of[r][d];
             tmpshv4[tid] = Of[r][d];
@@ -304,11 +306,11 @@ void main() {
     // If there is split_k, then the split_k resolve shader does the final
     // division by L. Store the intermediate O value and per-row m and L values.
     if (p.k_num > 1) {
-        uint32_t o_offset = D * p.ne1 * (split_k_index + iq3 * p.k_num);
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
 
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (tile_row(r) < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N);
                     }
@@ -316,7 +318,7 @@ void main() {
             }
         }
 
-        o_offset = D * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
+        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (tile_row(r) < N) {
                 perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
@@ -332,18 +334,18 @@ void main() {
         Lfrcp[r] = 1.0 / Lf[r];
     }
 
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             Of[r][d] *= float16_t(Lfrcp[r]);
         }
     }
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1*D;
+    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
 
     if (p.gqa_ratio > 1) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (tile_row(r) < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N);
                     }
@@ -353,9 +355,9 @@ void main() {
     } else {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (i * Br + tile_row(r) < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        data_o[o_offset + iq2 * D + (i * Br + tile_row(r)) * p.ne1 * D + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
+                        data_o[o_offset + iq2 * HSV + (i * Br + tile_row(r)) * p.ne1 * HSV + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
                     }
                 }
             }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
index cf47bd53e2..8792d5195e 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
@@ -61,8 +61,8 @@ ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE ele
 // Rows index by Q's dimension 2, and the first N rows are valid.
 D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
 {
-    if (r < N && c < D) {
-        uint32_t offset = (iq2 + r) * D + c;
+    if (r < N && c < HSV) {
+        uint32_t offset = (iq2 + r) * HSV + c;
         data_o[o_offset + offset] = D_TYPE(elem);
     }
     return elem;
@@ -86,9 +86,9 @@ void main() {
     tensorLayoutV = setTensorLayoutBlockSizeNV(tensorLayoutV, 1, BLOCK_SIZE);
 #endif
 
-    tensorLayoutQ = setTensorLayoutDimensionNV(tensorLayoutQ, N, D);
-    tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, D);
-    tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, D);
+    tensorLayoutQ = setTensorLayoutDimensionNV(tensorLayoutQ, N, HSK);
+    tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, HSK);
+    tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, HSV);
 
     // hint to the compiler that strides are aligned for the aligned variant of the shader
     if (Clamp != gl_CooperativeMatrixClampModeConstantNV)
@@ -104,16 +104,16 @@ void main() {
     tensorLayoutK = setTensorLayoutStrideNV(tensorLayoutK, k_stride, 1);
     tensorLayoutV = setTensorLayoutStrideNV(tensorLayoutV, v_stride, 1);
 
-    coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> Q;
-    coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA> Qf16;
+    coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseAccumulator> Q;
+    coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseA> Qf16;
 
     uint32_t q_offset = iq2*p.nb02+iq3*p.nb03;
-    coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, D));
+    coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, HSK));
 
-    Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA>(Q);
+    Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseA>(Q);
     Qf16 *= float16_t(p.scale);
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(0);
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(0);
 
     coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> L, M;
 
@@ -140,10 +140,10 @@ void main() {
 
         coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> S = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
 
-        coopmat<float16_t, gl_ScopeWorkgroup, D, Bc, gl_MatrixUseB> K_T;
+        coopmat<float16_t, gl_ScopeWorkgroup, HSK, Bc, gl_MatrixUseB> K_T;
 
         uint32_t k_offset = ik2*p.nb12 + ik3*p.nb13;
-        coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, D), tensorViewTranspose DECODEFUNC);
+        coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, HSK), tensorViewTranspose DECODEFUNC);
         S = coopMatMulAdd(Qf16, K_T, S);
 
         if (p.logit_softcap != 0.0f) {
@@ -208,42 +208,42 @@ void main() {
         rowsum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0.0);
         rowsum = coopMatMulAdd(P_A, One, rowsum);
 
-        coopmat<float16_t, gl_ScopeWorkgroup, Bc, D, gl_MatrixUseB> V;
+        coopmat<float16_t, gl_ScopeWorkgroup, Bc, HSV, gl_MatrixUseB> V;
         uint32_t v_offset = iv2*p.nb22 + iv3*p.nb23;
-        coopMatLoadTensorNV(V,  data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, D) DECODEFUNC);
+        coopMatLoadTensorNV(V,  data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, HSV) DECODEFUNC);
 
         L = eM*L + rowsum;
 
         // This is the "diagonal" matrix in the paper, but since we do componentwise
         // multiply rather than matrix multiply it has the diagonal element smeared
         // across the row
-        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> eMdiag;
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> eMdiag;
 
         // resize eM by using smear/reduce
         coopMatReduceNV(eMdiag, eM, gl_CooperativeMatrixReduceRowNV, smearReduce);
 
         // multiply with fp16 accumulation, then add to O.
-        coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(0);
+        coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(0);
         PV = coopMatMulAdd(P_A, V, PV);
 
-        O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(PV);
+        O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(PV);
     }
 
     // If there is split_k, then the split_k resolve shader does the final
     // division by L. Store the intermediate O value and per-row m and L values.
     if (p.k_num > 1) {
-        coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O);
+        coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(O);
 
-        uint32_t o_offset = D * p.ne1 * (split_k_index + iq3 * p.k_num);
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
         coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
 
-        o_offset = D * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
+        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
         coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N);
         coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N);
         return;
     }
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> Ldiag;
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> Ldiag;
 
     // resize L by using smear/reduce
     coopMatReduceNV(Ldiag, L, gl_CooperativeMatrixReduceRowNV, smearReduce);
@@ -255,18 +255,18 @@ void main() {
 
     O = Ldiag*O;
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1*D;
+    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
 
-    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O);
+    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(O);
     if (p.gqa_ratio > 1) {
         coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
     } else {
         tensorLayoutNV<3, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutD = createTensorLayoutNV(3, gl_CooperativeMatrixClampModeConstantNV);
-        tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.ne2, p.ne1, D);
+        tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.ne2, p.ne1, HSV);
 
         // permute dimensions
         tensorViewNV<3, false, 1, 0, 2> tensorViewPermute = createTensorViewNV(3, false, 1, 0, 2);
 
-        coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, D), tensorViewPermute);
+        coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, HSV), tensorViewPermute);
     }
 }

From bee28421be25fd447f61cb6db64d556cbfce32ec Mon Sep 17 00:00:00 2001
From: lhez <quic_lih@quicinc.com>
Date: Thu, 3 Jul 2025 11:22:24 -0700
Subject: [PATCH 10/10] opencl : broadcast for soft_max (#14510)

---
 ggml/src/ggml-opencl/ggml-opencl.cpp          | 51 +++++++++++++------
 ggml/src/ggml-opencl/kernels/softmax_4_f16.cl | 35 +++++++++----
 ggml/src/ggml-opencl/kernels/softmax_4_f32.cl | 35 +++++++++----
 ggml/src/ggml-opencl/kernels/softmax_f16.cl   | 35 +++++++++----
 ggml/src/ggml-opencl/kernels/softmax_f32.cl   | 35 +++++++++----
 5 files changed, 132 insertions(+), 59 deletions(-)

diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index 9436e6ea9a..2450100b43 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -5763,19 +5763,31 @@ static void ggml_cl_soft_max(ggml_backend_t backend, const ggml_tensor * src0, c
 
     cl_ulong offset1 = extra1 ? extra1->offset + src1->view_offs : offset0;
 
-    const int  ne00 = src0 ? src0->ne[0] : 0;
-    const int  ne01 = src0 ? src0->ne[1] : 0;
-    const int  ne02 = src0 ? src0->ne[2] : 0;
-    const int  ne03 = src0 ? src0->ne[3] : 0;
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_long nb01 = src0->nb[1];
+    const cl_long nb02 = src0->nb[2];
+    const cl_long nb03 = src0->nb[3];
+
+    const int ne12 = src1 ? src1->ne[2] : 0;
+    const int ne13 = src1 ? src1->ne[3] : 0;
+
+    const cl_long nb11 = src1 ? src1->nb[1] : 0;
+    const cl_long nb12 = src1 ? src1->nb[2] : 0;
+    const cl_long nb13 = src1 ? src1->nb[3] : 0;
+
+    const cl_long nb1 = dst->nb[1];
+    const cl_long nb2 = dst->nb[2];
+    const cl_long nb3 = dst->nb[3];
 
     float scale, max_bias;
     memcpy(&scale,    dst->op_params + 0, sizeof(float));
     memcpy(&max_bias, dst->op_params + 1, sizeof(float));
 
-    const int nrows_x = ggml_nrows(src0);
-    const int nrows_y = src0->ne[1];
-
-    const int n_head      = nrows_x/nrows_y;
+    const int n_head      = src0->ne[2];
     const int n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
 
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
@@ -5820,13 +5832,22 @@ static void ggml_cl_soft_max(ggml_backend_t backend, const ggml_tensor * src0, c
     CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
     CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
     CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
-    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
-    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
-    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(float),    &scale));
-    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(float),    &max_bias));
-    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(float),    &m0));
-    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float),    &m1));
-    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &n_head_log2));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb13));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb3));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(float),    &scale));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(float),    &max_bias));
+    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(float),    &m0));
+    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(float),    &m1));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &n_head_log2));
 
     size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
     size_t local_work_size[] = {(size_t)nth, 1, 1};
diff --git a/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl b/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl
index 62c05369a8..a6d8ede670 100644
--- a/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl
+++ b/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl
@@ -22,32 +22,45 @@
 REQD_SUBGROUP_SIZE_64
 #endif
 kernel void kernel_soft_max_4_f16(
-        global float * src0,
+        global char * src0,
         ulong offset0,
-        global half * src1,
+        global char * src1,
         ulong offset1,
-        global float * dst,
+        global char * dst,
         ulong offsetd,
         int ne00,
-        int ne01,
-        int ne02,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
         float scale,
         float max_bias,
         float m0,
         float m1,
         int n_head_log2
 ) {
-    src0 = (global float *)((global char *)src0 + offset0);
-    src1 = (global half *)((global char *)src1 + offset1);
-    dst = (global float *)((global char *)dst + offsetd);
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
 
     int i03 = get_group_id(2);
     int i02 = get_group_id(1);
     int i01 = get_group_id(0);
 
-    global float4 * psrc4 = (global float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
-    global half4  * pmask = (global char *)src1 != (global char *)src0 ? (global half4 *)(src1 + i01*ne00) : 0;
-    global float4 * pdst4 = (global float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+    int i13 = i03%ne13;
+    int i12 = i02%ne12;
+    int i11 = i01;
+
+    global float4 * psrc4 = (global float4 *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+    global half4  * pmask = src1 != src0 ? (global half4 *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float4 * pdst4 = (global float4 *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
 
     float slope = 1.0f;
 
diff --git a/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl b/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl
index d562774eab..35b5573b46 100644
--- a/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl
+++ b/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl
@@ -22,32 +22,45 @@
 REQD_SUBGROUP_SIZE_64
 #endif
 kernel void kernel_soft_max_4(
-        global float * src0,
+        global char * src0,
         ulong offset0,
-        global float * src1,
+        global char * src1,
         ulong offset1,
-        global float * dst,
+        global char * dst,
         ulong offsetd,
         int ne00,
-        int ne01,
-        int ne02,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
         float scale,
         float max_bias,
         float m0,
         float m1,
         int n_head_log2
 ) {
-    src0 = (global float*)((global char*)src0 + offset0);
-    src1 = (global float*)((global char*)src1 + offset1);
-    dst = (global float*)((global char*)dst + offsetd);
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
 
     int i03 = get_group_id(2);
     int i02 = get_group_id(1);
     int i01 = get_group_id(0);
 
-    global float4 * psrc4 = (global float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
-    global float4 * pmask = src1 != src0 ? (global float4 *)(src1 + i01*ne00) : 0;
-    global float4 * pdst4 = (global float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+    int i13 = i03%ne13;
+    int i12 = i02%ne12;
+    int i11 = i01;
+
+    global float4 * psrc4 = (global float4 *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+    global float4 * pmask = src1 != src0 ? (global float4 *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float4 * pdst4 = (global float4 *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
 
     float slope = 1.0f;
 
diff --git a/ggml/src/ggml-opencl/kernels/softmax_f16.cl b/ggml/src/ggml-opencl/kernels/softmax_f16.cl
index d38d099671..9d292b5746 100644
--- a/ggml/src/ggml-opencl/kernels/softmax_f16.cl
+++ b/ggml/src/ggml-opencl/kernels/softmax_f16.cl
@@ -22,32 +22,45 @@
 REQD_SUBGROUP_SIZE_64
 #endif
 kernel void kernel_soft_max_f16(
-        global float * src0,
+        global char * src0,
         ulong offset0,
-        global half * src1,
+        global char * src1,
         ulong offset1,
-        global float * dst,
+        global char * dst,
         ulong offsetd,
         int ne00,
-        int ne01,
-        int ne02,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
         float scale,
         float max_bias,
         float m0,
         float m1,
         int n_head_log2
 ) {
-    src0 = (global float *)((global char *)src0 + offset0);
-    src1 = (global half *)((global char *)src1 + offset1);
-    dst = (global float *)((global char *)dst + offsetd);
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
 
     int i03 = get_group_id(2);
     int i02 = get_group_id(1);
     int i01 = get_group_id(0);
 
-    global float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
-    global half  * pmask = (global char *)src1 != (global char *)src0 ? src1 + i01*ne00 : 0;
-    global float * pdst  = dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    int i13 = i03%ne13;
+    int i12 = i02%ne12;
+    int i11 = i01;
+
+    global float * psrc0 = (global float *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+    global half  * pmask = src1 != src0 ? (global half *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float * pdst  = (global float *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
 
     float slope = 1.0f;
 
diff --git a/ggml/src/ggml-opencl/kernels/softmax_f32.cl b/ggml/src/ggml-opencl/kernels/softmax_f32.cl
index 001b587abe..7c53dfbe5a 100644
--- a/ggml/src/ggml-opencl/kernels/softmax_f32.cl
+++ b/ggml/src/ggml-opencl/kernels/softmax_f32.cl
@@ -22,32 +22,45 @@
 REQD_SUBGROUP_SIZE_64
 #endif
 kernel void kernel_soft_max(
-        global float * src0,
+        global char * src0,
         ulong offset0,
-        global float * src1,
+        global char * src1,
         ulong offset1,
-        global float * dst,
+        global char * dst,
         ulong offsetd,
         int ne00,
-        int ne01,
-        int ne02,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
         float scale,
         float max_bias,
         float m0,
         float m1,
         int n_head_log2
 ) {
-    src0 = (global float*)((global char*)src0 + offset0);
-    src1 = (global float*)((global char*)src1 + offset1);
-    dst = (global float*)((global char*)dst + offsetd);
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
 
     int i03 = get_group_id(2);
     int i02 = get_group_id(1);
     int i01 = get_group_id(0);
 
-    global float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
-    global float * pmask = src1 != src0 ? src1 + i01*ne00 : 0;
-    global float * pdst  = dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    int i13 = i03%ne13;
+    int i12 = i02%ne12;
+    int i11 = i01;
+
+    global float * psrc0 = (global float *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+    global float * pmask = src1 != src0 ? (global float *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float * pdst  = (global float *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
 
     float slope = 1.0f;