MXFP4 support

This implements the Open Compute Microscaling (MX) FP4 format
as a tensor type with backend implementations focusing
on mulmat and mulmatid on CPU, CUDA, and Metal.

ml/backend/ggml/ggml/src/ggml-quants.h

@@ -37,6 +37,8 @@ GGML_API void quantize_row_iq4_xs_ref (const float * GGML_RESTRICT x, block_iq4_
 GGML_API void quantize_row_iq3_s_ref  (const float * GGML_RESTRICT x, block_iq3_s   * GGML_RESTRICT y, int64_t k);
 GGML_API void quantize_row_iq2_s_ref  (const float * GGML_RESTRICT x, block_iq2_s   * GGML_RESTRICT y, int64_t k);
 
+GGML_API void quantize_row_mxfp4_ref(const float * GGML_RESTRICT x, block_mxfp4 * GGML_RESTRICT y, int64_t k);
+
 // Dequantization
 GGML_API void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 GGML_API void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
@@ -65,6 +67,8 @@ GGML_API void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, floa
 GGML_API void dequantize_row_iq4_xs (const block_iq4_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 GGML_API void dequantize_row_iq3_s  (const block_iq3_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 
+GGML_API void dequantize_row_mxfp4(const block_mxfp4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
 // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
 GGML_API size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 GGML_API size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
@@ -90,6 +94,8 @@ GGML_API size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTR
 GGML_API size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 GGML_API size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 
+GGML_API size_t quantize_mxfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
 GGML_API void iq2xs_init_impl(enum ggml_type type);
 GGML_API void iq2xs_free_impl(enum ggml_type type);
 GGML_API void iq3xs_init_impl(int grid_size);

ml/backend/ggml/ggml/src/ggml.c

@@ -589,11 +589,13 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .to_float                 = (ggml_to_float_t) dequantize_row_q4_1,
         .from_float_ref           = (ggml_from_float_t) quantize_row_q4_1_ref,
     },
-    [4] = { // GGML_TYPE_Q4_2
-        .type_name                = "DEPRECATED",
-        .blck_size                = 0,
-        .type_size                = 0,
-        .is_quantized             = false,
+    [GGML_TYPE_MXFP4] = { // formerly deprecated GGML_TYPE_Q4_2
+        .type_name                = "mxfp4",
+        .blck_size                = MXFP4,
+        .type_size                = sizeof(block_mxfp4),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_mxfp4,
+        .from_float_ref           = (ggml_from_float_t) quantize_row_mxfp4_ref,
     },
     [5] = { // GGML_TYPE_Q4_3
         .type_name                = "DEPRECATED",
@@ -6446,6 +6448,7 @@ size_t ggml_quantize_chunk(
         case GGML_TYPE_IQ1_M:   result = quantize_iq1_m  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ4_NL:  result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ4_XS:  result = quantize_iq4_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_MXFP4:   result = quantize_mxfp4  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_F16:
             {
                 size_t elemsize = sizeof(ggml_fp16_t);

ml/backend/ggml/quantization.go

@@ -44,6 +44,8 @@ func ConvertToF32(data []byte, dtype uint32, nelements uint64) []float32 {
 		C.dequantize_row_q6_K((*C.block_q6_K)(unsafe.Pointer(&data[0])), (*C.float)(&f32s[0]), elems)
 	case C.GGML_TYPE_BF16:
 		C.ggml_bf16_to_fp32_row((*C.ggml_bf16_t)(unsafe.Pointer(&data[0])), (*C.float)(&f32s[0]), elems)
+	case C.GGML_TYPE_MXFP4:
+		C.dequantize_row_mxfp4((*C.block_mxfp4)(unsafe.Pointer(&data[0])), (*C.float)(&f32s[0]), elems)
 	default:
 		panic("unsupported quantization format")
 	}