llama : add gpt-oss (#15091)

* oai moe

* compat with new checkpoint

* add attn sink impl

* add rope scaling yarn

* logits match with latest transformers code

* wip chat template

* rm trailing space

* use ggml_scale_bias

* rm redundant is_swa_all

* convert interleaved gate_up

* graph : fix activation function to match reference (#7)

* vocab : handle o200k_harmony special tokens

* ggml : add attention sinks support (#1)

* llama : add attn sinks

* ggml : add attn sinks

* cuda : add attn sinks

* vulkan : add support for sinks in softmax

remove unnecessary return

* ggml : add fused swiglu_oai op (#11)

* ggml : add fused swiglu_oai op

* Update ggml/src/ggml-cpu/ops.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* update CUDA impl

* cont : metal impl

* add vulkan impl

* test-backend-ops : more test cases, clean up

* llama : remove unfused impl

* remove extra lines

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

---------

Co-authored-by: slaren <slarengh@gmail.com>

* repack mxfp4 upon conversion

* clean up a bit

* enable thinking

* add quick hack to render only some special tokens

* fix bf16 conversion

* remove vocab hack

* webui ok

* support chat parsing for gpt-oss

* fix webui

* direct mapping mxfp4, FINALLY

* force using mxfp4

* properly use lazy tensor

* ggml : add mxfp4

ggml : use e8m0 conversion instead of powf

Co-authored-by: Diego Devesa <slarengh@gmail.com>

change kvalues_mxfp4 table to match e2m1 (#6)

metal : remove quantization for now (not used)

cuda : fix disabled CUDA graphs due to ffn moe bias

vulkan : add support for mxfp4

cont : add cm2 dequant

* ggml : add ggml_add_id (#13)

* ggml : add ggml_add_id

* add cuda impl

* llama : add weight support check for add_id

* perf opt

* add vulkan impl

* rename cuda files

* add metal impl

* allow in-place ggml_add_id

* llama : keep biases on CPU with --cpu-moe

* llama : fix compile error

ggml-ci

* cuda : add fallback for __nv_cvt_e8m0_to_bf16raw

ggml-ci

* cleanup

ggml-ci

* sycl : fix supports_op for MXFP4

ggml-ci

* fix Unknown reasoning format

* ggml-cpu : fix AVX build

ggml-ci

* fix hip build

ggml-ci

* cuda : add mxfp4 dequantization support for cuBLAS

ggml-ci

* ggml-cpu : fix mxfp4 fallback definitions for some architectures

ggml-ci

* cuda : fix version required for __nv_cvt_e8m0_to_bf16raw

---------

Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: slaren <slarengh@gmail.com>

ggml/src/ggml-vulkan/vulkan-shaders/add_id.comp

@@ -0,0 +1,42 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : require
+
+#include "types.comp"
+
+layout (push_constant) uniform parameter
+{
+    uint ne0;
+    uint ne1;
+    uint s01;
+    uint s02;
+    uint s11;
+    uint s21;
+} p;
+
+#define BLOCK_SIZE 512
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
+layout (binding = 2) readonly buffer Z {int32_t data_c[];};
+layout (binding = 3) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i1 = gl_WorkGroupID.x;
+    const uint i2 = gl_WorkGroupID.y;
+
+    const uint i11 = data_c[i1 + i2 * p.s21];
+
+    const uint s1 = p.ne0;
+    const uint s2 = p.ne0 * p.ne1;
+
+    const uint d0 = i1 * s1 + i2 * s2;
+    const uint a0 = i1 * p.s01 + i2 * p.s02;
+    const uint b0 = i11 * p.s11;
+
+    for (uint i0 = gl_LocalInvocationID.x; i0 < p.ne0; i0 += BLOCK_SIZE) {
+        data_d[d0 + i0] = data_a[a0 + i0] + data_b[b0 + i0];
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp

@@ -4,8 +4,8 @@
 #include "generic_unary_head.comp"
 #include "dequant_funcs.comp"
 
-#if defined(DATA_A_IQ4_NL)
-// 16 invocations needed for init_iq4nl_shmem
+#if defined(DATA_A_IQ4_NL) || defined(DATA_A_MXFP4)
+// 16 invocations needed for init_iq_shmem
 layout(local_size_x = 16, local_size_y = 1, local_size_z = 1) in;
 #else
 layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp

@@ -434,6 +434,18 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
 }
 #endif
 
+#if defined(DATA_A_MXFP4)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return vec2(kvalues_mxfp4[vui & 0xF], kvalues_mxfp4[vui >> 4]);
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    vec2 v0 = dequantize(ib, iqs, a_offset);
+    vec2 v1 = dequantize(ib, iqs + 1, a_offset);
+    return vec4(v0.x, v0.y, v1.x, v1.y);
+}
+#endif
+
 #if defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16)
 vec2 get_dm(uint ib, uint a_offset) {
     return vec2(0, 0);
@@ -455,6 +467,12 @@ vec2 get_dm(uint ib, uint a_offset) {
 }
 #endif
 
+#if defined(DATA_A_MXFP4)
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(e8m0_to_fp32(data_a[a_offset + ib].e), 0);
+}
+#endif
+
 #if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1)
 vec2 get_dm(uint ib, uint a_offset) {
     return vec2(float(data_a[a_offset + ib].d), float(data_a[a_offset + ib].m));

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp

@@ -654,6 +654,25 @@ float16_t dequantFuncIQ4_NL(const in decodeBufIQ4_NL bl, const in uint blockCoor
 }
 #endif
 
+#if defined(DATA_A_MXFP4)
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufMXFP4 {
+   block_mxfp4 block;
+};
+
+float16_t dequantFuncMXFP4(const in decodeBufMXFP4 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const float d = e8m0_to_fp32(bl.block.e);
+    const uint idx = coordInBlock[1];
+    const uint iqs = idx & 0xF;
+    const uint shift = (idx & 0x10) >> 2;
+    uint32_t qs = bl.block.qs[iqs];
+    qs >>= shift;
+    qs &= 0xF;
+    float16_t ret = float16_t(kvalues_mxfp4[qs] * d);
+    return ret;
+}
+#endif
+
 #if defined(DATA_A_Q4_0)
 #define dequantFuncA dequantFuncQ4_0
 #elif defined(DATA_A_Q4_1)
@@ -696,4 +715,6 @@ float16_t dequantFuncIQ4_NL(const in decodeBufIQ4_NL bl, const in uint blockCoor
 #define dequantFuncA dequantFuncIQ4_XS
 #elif defined(DATA_A_IQ4_NL)
 #define dequantFuncA dequantFuncIQ4_NL
+#elif defined(DATA_A_MXFP4)
+#define dequantFuncA dequantFuncMXFP4
 #endif

ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp

@@ -0,0 +1,32 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_mxfp4 data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    const uint tid = gl_LocalInvocationID.x % 64;
+    const uint il  = tid/32;
+    const uint ir  = tid%32;
+    const uint ib = 32*i + ir;
+    if (ib >= p.nel / 32) {
+        return;
+    }
+
+    const uint q_idx = 8*il;
+    const uint b_idx = 1024*i + 32*ir + q_idx;
+
+    const float d = e8m0_to_fp32(data_a[ib].e);
+
+    [[unroll]] for (uint l = 0; l < 8; ++l) {
+        data_b[b_idx + l +  0] = D_TYPE(d * kvalues_mxfp4[data_a[ib].qs[q_idx + l] & 0xF]);
+        data_b[b_idx + l + 16] = D_TYPE(d * kvalues_mxfp4[data_a[ib].qs[q_idx + l] >>  4]);
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/glu_head.comp

@@ -14,4 +14,6 @@ layout (push_constant) uniform parameter
     uint ne00;
     uint ne20;
     uint mode;
+    float alpha;
+    float limit;
 } p;

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp

@@ -747,6 +747,21 @@ void main() {
             buf_a[buf_idx + 1 ] = FLOAT_TYPE(kvalues_iq4nl[bitfieldExtract(vui, 8, 4)]) * d;
             buf_a[buf_idx + 16] = FLOAT_TYPE(kvalues_iq4nl[bitfieldExtract(vui, 4, 4)]) * d;
             buf_a[buf_idx + 17] = FLOAT_TYPE(kvalues_iq4nl[vui >> 12]) * d;
+#elif defined(DATA_A_MXFP4)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + 2 * loadr_a;
+
+            const uint ib = idx / 8;
+            const uint iqs = (idx & 0x07) * 2;
+
+            const float d = e8m0_to_fp32(data_a[ib].e);
+            const uint vui = uint(data_a[ib].qs[iqs]);
+            const uint vui2 = uint(data_a[ib].qs[iqs+1]);
+
+            buf_a[buf_idx     ] = FLOAT_TYPE(kvalues_mxfp4[vui & 0xF] * d);
+            buf_a[buf_idx + 16] = FLOAT_TYPE(kvalues_mxfp4[vui >>  4] * d);
+            buf_a[buf_idx +  1] = FLOAT_TYPE(kvalues_mxfp4[vui2 & 0xF] * d);
+            buf_a[buf_idx + 17] = FLOAT_TYPE(kvalues_mxfp4[vui2 >>  4] * d);
 #endif
         }
         [[unroll]] for (uint l = 0; l < BN; l += loadstride_b) {

ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp

@@ -92,6 +92,12 @@ FLOAT_TYPE get_d(uint ib) {
 }
 #endif
 
+#if defined(DATA_A_MXFP4)
+FLOAT_TYPE get_d(uint ib) {
+    return FLOAT_TYPE(e8m0_to_fp32(data_a[ib].e));
+}
+#endif
+
 #if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1)
 FLOAT_TYPE_VEC2 get_dm(uint ib) {
     return FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);

ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp

@@ -20,6 +20,7 @@ layout (push_constant) uniform parameter
     float m1;
     uint n_head_log2;
     uint nrows_x;
+    uint has_sinks;
 } p;
 
 #include "types.comp"
@@ -29,7 +30,8 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
-layout (binding = 2) buffer D {D_TYPE data_d[];};
+layout (binding = 2) readonly buffer Z {float data_c[];};
+layout (binding = 3) buffer D {D_TYPE data_d[];};
 
 shared FLOAT_TYPE vals[BLOCK_SIZE];
 
@@ -60,13 +62,13 @@ void soft_max(uint num_iters) {
         const uint h = (rowx / p.ne01) % p.ne02; // head index
 
         const float base = h < p.n_head_log2 ? p.m0 : p.m1;
-        const uint   exp  = h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1;
+        const uint   exp = h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1;
 
         slope = pow(base, exp);
     }
 
     // Find max
-    FLOAT_TYPE max_val = uintBitsToFloat(0xFF800000);
+    FLOAT_TYPE max_val = p.has_sinks == 0 ? uintBitsToFloat(0xFF800000) : data_c[i02];
 
     // Cache values while we compute the max, so we don't need to read them
     // again when we're ready to compute exp(x-max).
@@ -148,6 +150,10 @@ void soft_max(uint num_iters) {
     }
     sum = vals[0];
 
+    if (p.has_sinks != 0) {
+        sum += FLOAT_TYPE(exp(FLOAT_TYPE(data_c[i02]) - max_val));
+    }
+
     FLOAT_TYPE rcpdivisor = 1.0/sum;
 
     [[unroll]] for (uint col0 = 0, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) {

ggml/src/ggml-vulkan/vulkan-shaders/swiglu_oai.comp

@@ -0,0 +1,14 @@
+#version 450
+
+#include "glu_head.comp"
+
+float op(float a, float b) {
+    float xi = min(a, p.limit);
+    float gi = max(min(b, p.limit), -p.limit);
+
+    float out_glu = xi / (1.0f + exp(-xi * p.alpha));
+    out_glu = out_glu * (1.0f + gi);
+    return out_glu;
+}
+
+#include "glu_main.comp"

ggml/src/ggml-vulkan/vulkan-shaders/types.comp

@@ -1337,6 +1337,29 @@ struct block_iq4_nl_packed16
 #define A_TYPE_PACKED16 block_iq4_nl_packed16
 #endif
 
+#define QUANT_K_MXFP4 32
+#define QUANT_R_MXFP4 2
+
+struct block_mxfp4
+{
+    uint8_t e;
+    uint8_t qs[QUANT_K_MXFP4/2];
+};
+
+//struct block_mxfp4_packed16
+//{
+//    uint8_t e;
+//    uint16_t qs[QUANT_K_MXFP4/2/2];
+//};
+
+#if defined(DATA_A_MXFP4)
+#define QUANT_K QUANT_K_MXFP4
+#define QUANT_R QUANT_R_MXFP4
+#define QUANT_AUXF 1
+#define A_TYPE block_mxfp4
+//#define A_TYPE_PACKED16 block_mxfp4_packed16
+#endif
+
 #if defined(DATA_A_IQ4_NL) || defined(DATA_A_IQ4_XS)
 const int8_t kvalues_iq4nl_const[16] = {
     int8_t(-127), int8_t(-104), int8_t(-83), int8_t(-65), int8_t(-49), int8_t(-35), int8_t(-22), int8_t(-10),
@@ -1356,6 +1379,25 @@ void init_iq_shmem(uvec3 wgsize)
 }
 #endif
 
+#if defined(DATA_A_MXFP4)
+const FLOAT_TYPE kvalues_mxfp4_const[16] = {
+    FLOAT_TYPE(0.0f), FLOAT_TYPE(0.5f), FLOAT_TYPE(1.0f), FLOAT_TYPE(1.5f), FLOAT_TYPE(2.0f), FLOAT_TYPE(3.0f), FLOAT_TYPE(4.0f), FLOAT_TYPE(6.0f),
+    FLOAT_TYPE(-0.0f), FLOAT_TYPE(-0.5f), FLOAT_TYPE(-1.0f), FLOAT_TYPE(-1.5f), FLOAT_TYPE(-2.0f), FLOAT_TYPE(-3.0f), FLOAT_TYPE(-4.0f), FLOAT_TYPE(-6.0f)
+};
+
+shared FLOAT_TYPE kvalues_mxfp4[16];
+
+#define NEEDS_INIT_IQ_SHMEM
+void init_iq_shmem(uvec3 wgsize)
+{
+    // copy the table into shared memory and sync
+    for (uint i = gl_LocalInvocationIndex.x; i < kvalues_mxfp4.length(); i += wgsize.x) {
+        kvalues_mxfp4[i] = kvalues_mxfp4_const[i];
+    }
+    barrier();
+}
+#endif
+
 // returns the bfloat value in the low 16b.
 // See ggml_compute_fp32_to_bf16
 uint32_t fp32_to_bf16(float f)
@@ -1370,4 +1412,17 @@ float bf16_to_fp32(uint32_t u)
     return uintBitsToFloat(u << 16);
 }
 
+float e8m0_to_fp32(uint8_t x) {
+    uint32_t bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = x;
+        bits = bits << 23;
+    }
+
+    return uintBitsToFloat(bits);
+}
+
 #endif // !defined(GGML_TYPES_COMP)

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -64,6 +64,7 @@ const std::vector<std::string> type_names = {
     "iq3_s",
     "iq4_xs",
     "iq4_nl",
+    "mxfp4",
     "bf16",
 };
 
@@ -118,7 +119,7 @@ void execute_command(const std::string& command, std::string& stdout_str, std::s
     CloseHandle(pi.hProcess);
     CloseHandle(pi.hThread);
 #else
-int stdout_pipe[2];
+    int stdout_pipe[2];
     int stderr_pipe[2];
 
     if (pipe(stdout_pipe) != 0 || pipe(stderr_pipe) != 0) {
@@ -362,7 +363,7 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
         std::string load_vec_quant = "2";
         if ((tname == "q4_0") || (tname == "q4_1") || (tname == "iq1_s") || (tname == "iq1_m") || (tname == "iq2_xxs") || (tname == "iq2_xs") || (tname == "iq2_s"))
             load_vec_quant = "8";
-        else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "iq3_xxs") || (tname == "iq3_s") || (tname == "iq4_nl"))
+        else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "iq3_xxs") || (tname == "iq3_s") || (tname == "iq4_nl") || (tname == "mxfp4"))
             load_vec_quant = "4";
 
         if (tname == "bf16") {
@@ -602,6 +603,8 @@ void process_shaders() {
         string_to_spv("reglu_f32" + suffix,      "reglu.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
         string_to_spv("swiglu_f16" + suffix,     "swiglu.comp",      {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
         string_to_spv("swiglu_f32" + suffix,     "swiglu.comp",      {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("swiglu_oai_f16" + suffix, "swiglu_oai.comp",  {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("swiglu_oai_f32" + suffix, "swiglu_oai.comp",  {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
         string_to_spv("geglu_erf_f16" + suffix,  "geglu_erf.comp",   {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
         string_to_spv("geglu_erf_f32" + suffix,  "geglu_erf.comp",   {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
         string_to_spv("geglu_quick_f16" + suffix,"geglu_quick.comp", {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
@@ -671,6 +674,8 @@ void process_shaders() {
 
     string_to_spv("roll_f32", "roll.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
+    string_to_spv("add_id_f32", "add_id.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+
     for (auto &c : compiles) {
         c.wait();
     }