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-quants.c

@@ -21,6 +21,17 @@
 
 #define UNUSED GGML_UNUSED
 
+static inline int best_index_int8(int n, const int8_t * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
 // reference implementation for deterministic creation of model files
 void quantize_row_q4_0_ref(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k) {
     static const int qk = QK4_0;
@@ -246,6 +257,53 @@ void quantize_row_q8_1_ref(const float * GGML_RESTRICT x, block_q8_1 * GGML_REST
     }
 }
 
+static inline int best_index_mxfp4(float x, float e) {
+    int best_index = 0;
+    float best_err = fabsf(kvalues_mxfp4[0]*e - x);
+    for (int i = 1; i < 16; i++) {
+        float err = fabsf(kvalues_mxfp4[i]*e - x);
+        if (err < best_err) {
+            best_index = i;
+            best_err = err;
+        }
+    }
+    return best_index;
+}
+
+void quantize_row_mxfp4_ref(const float * GGML_RESTRICT x, block_mxfp4 * GGML_RESTRICT y, int64_t k) {
+    static const int qk = QK_MXFP4;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        float amax = 0.0f; // absolute max
+
+        for (int j = 0; j < qk; j++) {
+            const float v = x[i*qk + j];
+
+            if (amax < fabsf(v)) {
+                amax = fabsf(v);
+            }
+        }
+
+        const uint8_t e = (uint8_t) (floorf(log2f(amax)) - 2 + 127);
+
+        const float d = GGML_E8M0_TO_FP32_HALF(e);
+
+        y[i].e = e;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t x0 = best_index_mxfp4(x[i*qk + 0    + j], d);
+            const uint8_t x1 = best_index_mxfp4(x[i*qk + qk/2 + j], d);
+
+            y[i].qs[j]  = x0;
+            y[i].qs[j] |= x1 << 4;
+        }
+    }
+}
+
 void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     static const int qk = QK4_0;
 
@@ -356,6 +414,26 @@ void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRI
     }
 }
 
+void dequantize_row_mxfp4(const block_mxfp4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+    static const int qk = QK_MXFP4;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        const float d = GGML_E8M0_TO_FP32_HALF(x[i].e);
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int8_t x0 = kvalues_mxfp4[x[i].qs[j] & 0x0F];
+            const int8_t x1 = kvalues_mxfp4[x[i].qs[j] >>   4];
+
+            y[i*qk + j + 0   ] = x0*d;
+            y[i*qk + j + qk/2] = x1*d;
+        }
+    }
+}
+
 //
 // 2-6 bit quantization in super-blocks
 //
@@ -2014,6 +2092,12 @@ size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst,
     return nrow * row_size;
 }
 
+size_t quantize_mxfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    GGML_UNUSED(quant_weights);
+    quantize_row_mxfp4_ref(src, dst, (int64_t)nrow*n_per_row);
+    return nrow * ggml_row_size(GGML_TYPE_MXFP4, n_per_row);
+}
+
 // ====================== Ternary (de)-quantization (BitNet b1.58 and TriLMs)
 
 void quantize_row_tq1_0_ref(const float * GGML_RESTRICT x, block_tq1_0 * GGML_RESTRICT y, int64_t k) {
@@ -4551,17 +4635,6 @@ size_t quantize_iq1_m(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst,
 
 // ============================ 4-bit non-linear quants
 
-static inline int best_index_int8(int n, const int8_t * val, float x) {
-    if (x <= val[0]) return 0;
-    if (x >= val[n-1]) return n-1;
-    int ml = 0, mu = n-1;
-    while (mu-ml > 1) {
-        int mav = (ml+mu)/2;
-        if (x < val[mav]) mu = mav; else ml = mav;
-    }
-    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
-}
-
 static void quantize_row_iq4_nl_impl(const int super_block_size, const int block_size, const float * GGML_RESTRICT x,
         ggml_fp16_t * dh, uint8_t * q4, uint16_t * scales_h, uint8_t * scales_l,
         float * scales, float * weight, uint8_t * L,
@@ -4961,6 +5034,15 @@ static bool validate_fp16(ggml_fp16_t f, size_t i) {
     return true;
 }
 
+static bool validate_e_e8m0(uint8_t e, size_t i) {
+    if (e == 0xff) {
+        fprintf(stderr, "ggml_validate_row_data: found invalid e value %d at block %zu\n", e, i);
+        return false;
+    }
+
+    return true;
+}
+
 #define VALIDATE_ROW_DATA_D_F16_IMPL(type, data, nb) \
     const type * q = (const type *) (data); \
     for (size_t i = 0; i < (nb); ++i) { \
@@ -4977,6 +5059,14 @@ static bool validate_fp16(ggml_fp16_t f, size_t i) {
         } \
     }
 
+#define VALIDATE_ROW_DATA_E_E8M0_IMPL(type, data, nb) \
+    const type * q = (const type *) (data); \
+    for (size_t i = 0; i < (nb); ++i) { \
+        if (!validate_e_e8m0(q[i].e, i)) { \
+            return false; \
+        } \
+    }
+
 #define VALIDATE_ROW_DATA_DVEC_F16_IMPL(type, data, nb, nr) \
     const type * q = (const type *) (data); \
     for (size_t i = 0; i < (nb); ++i) { \
@@ -5130,6 +5220,10 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte
             {
                 VALIDATE_ROW_DATA_D_F16_IMPL(block_q8_0, data, nb);
             } break;
+        case GGML_TYPE_MXFP4:
+            {
+                VALIDATE_ROW_DATA_E_E8M0_IMPL(block_mxfp4, data, nb);
+            } break;
         case GGML_TYPE_Q2_K:
             {
                 VALIDATE_ROW_DATA_DM_F16_IMPL(block_q2_K, data, nb, d, dmin);