From d903f30e25f3024c37d1eedd4b46ed0f5b13ff88 Mon Sep 17 00:00:00 2001 From: Aman Gupta Date: Fri, 27 Feb 2026 18:15:09 +0800 Subject: [PATCH 1/4] ggml-cpu: add repack for mxfp4 (#19738) --- ggml/src/ggml-cpu/arch-fallback.h | 28 +++ ggml/src/ggml-cpu/arch/arm/repack.cpp | 156 +++++++++++++ ggml/src/ggml-cpu/arch/x86/repack.cpp | 104 ++++++++- ggml/src/ggml-cpu/repack.cpp | 318 ++++++++++++++++++++++++++ ggml/src/ggml-cpu/repack.h | 21 ++ 5 files changed, 625 insertions(+), 2 deletions(-) diff --git a/ggml/src/ggml-cpu/arch-fallback.h b/ggml/src/ggml-cpu/arch-fallback.h index 4dfe28e1d..ebbd4b47e 100644 --- a/ggml/src/ggml-cpu/arch-fallback.h +++ b/ggml/src/ggml-cpu/arch-fallback.h @@ -48,6 +48,8 @@ #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 +#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0 +#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 @@ -62,6 +64,8 @@ #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 +#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0 +#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 #elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64) @@ -69,8 +73,10 @@ #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 +#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 +#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64) // repack.cpp @@ -84,6 +90,7 @@ #define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 +#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 @@ -94,6 +101,7 @@ #define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 +#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 #elif defined(__POWERPC__) || defined(__powerpc__) @@ -120,6 +128,8 @@ #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 +#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0 +#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 @@ -134,6 +144,8 @@ #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 +#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0 +#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 #elif defined(__loongarch64) @@ -160,6 +172,8 @@ #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 +#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0 +#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 @@ -174,6 +188,8 @@ #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 +#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0 +#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 #elif defined(__riscv) @@ -201,6 +217,8 @@ #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 +#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0 +#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 @@ -214,6 +232,8 @@ #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 +#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0 +#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 #elif defined(__s390x__) @@ -246,6 +266,8 @@ #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 +#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0 +#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 @@ -260,6 +282,8 @@ #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 +#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0 +#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 #elif defined(__wasm__) @@ -294,6 +318,8 @@ #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 +#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0 +#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 @@ -308,6 +334,8 @@ #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 +#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0 +#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 #endif diff --git a/ggml/src/ggml-cpu/arch/arm/repack.cpp b/ggml/src/ggml-cpu/arch/arm/repack.cpp index c2e4623f3..3eed0105b 100644 --- a/ggml/src/ggml-cpu/arch/arm/repack.cpp +++ b/ggml/src/ggml-cpu/arch/arm/repack.cpp @@ -498,6 +498,81 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const ggml_gemv_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc); } +void ggml_gemv_mxfp4_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 4; + + assert (n % qk == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) + const int8x16_t kvalues = vld1q_s8(kvalues_mxfp4); + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + float * res_ptr = s; + + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_mxfp4x4 * b_ptr = (const block_mxfp4x4 *) vx + (x * nb); + + float32x4_t sumf = vdupq_n_f32(0); + for (int l = 0; l < nb; l++) { + uint8x16_t b_0 = vld1q_u8(b_ptr[l].qs + 0); + uint8x16_t b_1 = vld1q_u8(b_ptr[l].qs + 16); + uint8x16_t b_2 = vld1q_u8(b_ptr[l].qs + 32); + uint8x16_t b_3 = vld1q_u8(b_ptr[l].qs + 48); + + int8x16_t b_0_hi = vqtbl1q_s8(kvalues, b_0 >> 4); + int8x16_t b_0_lo = vqtbl1q_s8(kvalues, b_0 & 0x0F); + int8x16_t b_1_hi = vqtbl1q_s8(kvalues, b_1 >> 4); + int8x16_t b_1_lo = vqtbl1q_s8(kvalues, b_1 & 0x0F); + int8x16_t b_2_hi = vqtbl1q_s8(kvalues, b_2 >> 4); + int8x16_t b_2_lo = vqtbl1q_s8(kvalues, b_2 & 0x0F); + int8x16_t b_3_hi = vqtbl1q_s8(kvalues, b_3 >> 4); + int8x16_t b_3_lo = vqtbl1q_s8(kvalues, b_3 & 0x0F); + + int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 0); + int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16); + + int32x4_t sumi = vdupq_n_s32(0); + sumi = vdotq_laneq_s32(sumi, b_0_lo, a_0, 0); + sumi = vdotq_laneq_s32(sumi, b_0_hi, a_1, 0); + sumi = vdotq_laneq_s32(sumi, b_1_lo, a_0, 1); + sumi = vdotq_laneq_s32(sumi, b_1_hi, a_1, 1); + sumi = vdotq_laneq_s32(sumi, b_2_lo, a_0, 2); + sumi = vdotq_laneq_s32(sumi, b_2_hi, a_1, 2); + sumi = vdotq_laneq_s32(sumi, b_3_lo, a_0, 3); + sumi = vdotq_laneq_s32(sumi, b_3_hi, a_1, 3); + + float32x4_t a_d = vcvt_f32_f16(vld1_dup_f16((const float16_t *)&a_ptr[l].d)); + float32x4_t b_d = { + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[0]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[1]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[2]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[3]), + }; + float32x4_t d = a_d * b_d; + + sumf = vmlaq_f32(sumf, d, vcvtq_f32_s32(sumi)); + } + + vst1q_f32(res_ptr + x * 4, sumf); + } + return; +#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) + ggml_gemv_mxfp4_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc); +} + void ggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { constexpr int qk = QK_K; const int nb = n / qk; @@ -3164,6 +3239,87 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const ggml_gemm_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc); } +void ggml_gemm_mxfp4_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 4; + + assert (n % qk == 0); + assert (nr % 4 == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) + const int8x16_t kvalues = vld1q_s8(kvalues_mxfp4); + + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_mxfp4x4 * b_ptr = (const block_mxfp4x4 *) vx + (x * nb); + + float32x4_t sumf[4]; + for (int m = 0; m < 4; m++) { + sumf[m] = vdupq_n_f32(0); + } + + for (int l = 0; l < nb; l++) { + float32x4_t a_d = vcvt_f32_f16(vld1_f16((const float16_t *)a_ptr[l].d)); + float32x4_t b_d = { + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[0]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[1]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[2]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[3]), + }; + + int32x4_t sumi_0 = vdupq_n_s32(0); + int32x4_t sumi_1 = vdupq_n_s32(0); + int32x4_t sumi_2 = vdupq_n_s32(0); + int32x4_t sumi_3 = vdupq_n_s32(0); + + for (int k = 0; k < 4; k++) { + int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 16 * k + 0); + int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16 * k + 64); + + uint8x16_t b = vld1q_u8(b_ptr[l].qs + 16 * k); + int8x16_t b_hi = vqtbl1q_s8(kvalues, b >> 4); + int8x16_t b_lo = vqtbl1q_s8(kvalues, b & 0xF); + + sumi_0 = vdotq_laneq_s32(sumi_0, b_lo, a_0, 0); + sumi_1 = vdotq_laneq_s32(sumi_1, b_lo, a_0, 1); + sumi_2 = vdotq_laneq_s32(sumi_2, b_lo, a_0, 2); + sumi_3 = vdotq_laneq_s32(sumi_3, b_lo, a_0, 3); + sumi_0 = vdotq_laneq_s32(sumi_0, b_hi, a_1, 0); + sumi_1 = vdotq_laneq_s32(sumi_1, b_hi, a_1, 1); + sumi_2 = vdotq_laneq_s32(sumi_2, b_hi, a_1, 2); + sumi_3 = vdotq_laneq_s32(sumi_3, b_hi, a_1, 3); + } + + sumf[0] = vmlaq_f32(sumf[0], vmulq_laneq_f32(b_d, a_d, 0), vcvtq_f32_s32(sumi_0)); + sumf[1] = vmlaq_f32(sumf[1], vmulq_laneq_f32(b_d, a_d, 1), vcvtq_f32_s32(sumi_1)); + sumf[2] = vmlaq_f32(sumf[2], vmulq_laneq_f32(b_d, a_d, 2), vcvtq_f32_s32(sumi_2)); + sumf[3] = vmlaq_f32(sumf[3], vmulq_laneq_f32(b_d, a_d, 3), vcvtq_f32_s32(sumi_3)); + } + + for (int m = 0; m < 4; m++) { + vst1q_f32(s + (y * 4 + m) * bs + x * 4, sumf[m]); + } + } + } + return; +#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) + ggml_gemm_mxfp4_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc); +} + void ggml_gemm_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { constexpr int qk = QK_K; const int nb = n / qk; diff --git a/ggml/src/ggml-cpu/arch/x86/repack.cpp b/ggml/src/ggml-cpu/arch/x86/repack.cpp index 7dda9eea0..bd6906c41 100644 --- a/ggml/src/ggml-cpu/arch/x86/repack.cpp +++ b/ggml/src/ggml-cpu/arch/x86/repack.cpp @@ -522,7 +522,8 @@ template static void gemv_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc, __m256i signextendlut) { static_assert( std::is_same_v || - std::is_same_v, + std::is_same_v || + std::is_same_v, "Unsupported block type"); const int qk = QK8_0; @@ -580,6 +581,18 @@ static void gemv_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t std::is_same_v || std::is_same_v) { col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask); + } else if constexpr (std::is_same_v) { + // Load 8 E8M0 exponents and convert to float via LUT + // Rearranged to match changemask order: 0,4,1,5,2,6,3,7 + col_scale_f32 = _mm256_set_ps( + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[7]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[3]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[6]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[2]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[5]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[1]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[4]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[0])); } // Load and convert to FP32 scale from block_q8_0 @@ -628,7 +641,8 @@ template static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc, __m256i signextendlut) { static_assert( std::is_same_v || - std::is_same_v, + std::is_same_v || + std::is_same_v, "Unsupported block type"); const int qk = QK8_0; @@ -749,6 +763,25 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t std::is_same_v || std::is_same_v) { col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d); + } else if constexpr (std::is_same_v) { + //TODO: simd-ify + col_scale_f32 = _mm512_set_ps( + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[7]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[6]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[5]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[4]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[3]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[2]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[1]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[0]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[7]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[6]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[5]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[4]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[3]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[2]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[1]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[0])); } // Process LHS in pairs of rows @@ -941,6 +974,25 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t std::is_same_v || std::is_same_v) { col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d); + } else if constexpr (std::is_same_v) { + //TODO: simd-ify + col_scale_f32 = _mm512_set_ps( + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[7]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[6]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[5]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[4]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[3]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[2]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[1]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[0]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[7]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[6]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[5]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[4]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[3]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[2]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[1]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[0])); } // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3 @@ -1123,6 +1175,16 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t std::is_same_v || std::is_same_v) { col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d); + } else if constexpr (std::is_same_v) { + col_scale_f32 = _mm256_set_ps( + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[7]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[6]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[5]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[4]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[3]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[2]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[1]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[0])); } // Process LHS in groups of four @@ -1283,6 +1345,16 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t std::is_same_v || std::is_same_v) { col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d); + } else if constexpr (std::is_same_v) { + col_scale_f32 = _mm256_set_ps( + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[7]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[6]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[5]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[4]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[3]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[2]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[1]), + GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[0])); } // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3 @@ -1625,6 +1697,19 @@ void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const ggml_gemv_iq4_nl_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); } +void ggml_gemv_mxfp4_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { +#if defined(__AVX2__) + __m256i signextendlut = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*)kvalues_mxfp4)); + signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0); + + gemv_q4_b32_8x8_q8_0_lut_avx(n, s, bs, vx, vy, nr, nc, signextendlut); + + return; +#endif + + ggml_gemv_mxfp4_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); +} + void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK_K; const int nb = n / qk; @@ -3423,6 +3508,21 @@ void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc); } +void ggml_gemm_mxfp4_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { +#if defined(__AVX2__) || defined(__AVX512F__) + { + __m256i signextendlut = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*)kvalues_mxfp4)); + signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0); + + gemm_q4_b32_8x8_q8_0_lut_avx(n, s, bs, vx, vy, nr, nc, signextendlut); + + return; + } +#endif // defined(__AVX2__) || defined(__AVX512F__) + + ggml_gemm_mxfp4_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); +} + void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK_K; const int nb = n / qk; diff --git a/ggml/src/ggml-cpu/repack.cpp b/ggml/src/ggml-cpu/repack.cpp index 1b3d23cbe..5edba4212 100644 --- a/ggml/src/ggml-cpu/repack.cpp +++ b/ggml/src/ggml-cpu/repack.cpp @@ -1098,6 +1098,82 @@ void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs } } +void ggml_gemv_mxfp4_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 4; + + assert(nr == 1); + assert(n % qk == 0); + assert(nc % ncols_interleaved == 0); + + UNUSED(bs); + UNUSED(nr); + + float sumf[4]; + int sumi; + + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_mxfp4x4 * b_ptr = (const block_mxfp4x4 *) vx + (x * nb); + + for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; + const int v1 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; + sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); + } + sumf[j] += sumi * GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); + } + } + } + for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; + } +} + +void ggml_gemv_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 8; + const int blocklen = 8; + + assert(nr == 1); + assert(n % qk == 0); + assert(nc % ncols_interleaved == 0); + + UNUSED(bs); + UNUSED(nr); + + float sumf[8]; + int sumi; + + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_mxfp4x8 * b_ptr = (const block_mxfp4x8 *) vx + (x * nb); + + for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; + const int v1 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; + sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); + } + sumf[j] += sumi * GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); + } + } + } + for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; + } +} + void ggml_gemv_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, @@ -1726,6 +1802,94 @@ void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs } } +void ggml_gemm_mxfp4_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 4; + + assert(n % qk == 0); + assert(nr % 4 == 0); + assert(nc % ncols_interleaved == 0); + + float sumf[4][4]; + int sumi; + + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_mxfp4x4 * b_ptr = (const block_mxfp4x4 *) vx + (x * nb); + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; + } + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; + const int v1 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])); + } + sumf[m][j] += sumi * GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); + } + } + } + } + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) + s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; + } + } + } +} + +void ggml_gemm_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 8; + const int blocklen = 8; + + assert(n % qk == 0); + assert(nr % 4 == 0); + assert(nc % ncols_interleaved == 0); + + float sumf[4][8]; + int sumi; + + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_mxfp4x8 * b_ptr = (const block_mxfp4x8 *) vx + (x * nb); + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; + } + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; + const int v1 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])); + } + sumf[m][j] += sumi * GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); + } + } + } + } + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) + s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; + } + } + } +} + void ggml_gemm_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, @@ -2510,6 +2674,121 @@ static int repack_iq4_nl_to_iq4_nl_8_bl(struct ggml_tensor * t, int interleave_b GGML_UNUSED(data_size); } + +static block_mxfp4x4 make_block_mxfp4x4(block_mxfp4 * in, unsigned int blck_size_interleave) { + block_mxfp4x4 out; + + for (int i = 0; i < 4; i++) { + out.e[i] = in[i].e; + } + + const int end = QK_MXFP4 * 2 / blck_size_interleave; + + if (blck_size_interleave == 4) { + for (int i = 0; i < end; ++i) { + int src_id = i % 4; + int src_offset = (i / 4) * blck_size_interleave; + int dst_offset = i * blck_size_interleave; + + memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint32_t)); + } + } else { + GGML_ASSERT(false); + } + + return out; +} + +static int repack_mxfp4_to_mxfp4_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { + GGML_ASSERT(t->type == GGML_TYPE_MXFP4); + GGML_ASSERT(interleave_block == 4); + + const block_mxfp4 * src = (const block_mxfp4 *)data; + block_mxfp4x4 * dst = ( block_mxfp4x4 *)t->data; + + block_mxfp4 dst_tmp[4]; + + int nrow = ggml_nrows(t); + int nrows_interleaved = 4; + int nblocks = t->ne[0] / QK_MXFP4; + + GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_mxfp4)); + + if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { + return -1; + } + + for (int b = 0; b < nrow; b += nrows_interleaved) { + for (int64_t x = 0; x < nblocks; x++) { + for (int i = 0; i < nrows_interleaved; i++) { + dst_tmp[i] = src[x + i * nblocks]; + } + *dst++ = make_block_mxfp4x4(dst_tmp, interleave_block); + } + src += nrows_interleaved * nblocks; + } + return 0; + + GGML_UNUSED(data_size); +} + +static block_mxfp4x8 make_block_mxfp4x8(block_mxfp4 * in, unsigned int blck_size_interleave) { + block_mxfp4x8 out; + + for (int i = 0; i < 8; i++) { + out.e[i] = in[i].e; + } + + const int end = QK_MXFP4 * 4 / blck_size_interleave; + + if (blck_size_interleave == 8) { + for (int i = 0; i < end; ++i) { + int src_id = i % 8; + int src_offset = (i / 8) * blck_size_interleave; + int dst_offset = i * blck_size_interleave; + + memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t)); + } + } else { + GGML_ASSERT(false); + } + + return out; +} + +static int repack_mxfp4_to_mxfp4_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { + GGML_ASSERT(t->type == GGML_TYPE_MXFP4); + GGML_ASSERT(interleave_block == 8); + + const block_mxfp4 * src = (const block_mxfp4 *)data; + block_mxfp4x8 * dst = ( block_mxfp4x8 *)t->data; + + block_mxfp4 dst_tmp[8]; + + int nrow = ggml_nrows(t); + int nrows_interleaved = 8; + int nblocks = t->ne[0] / QK_MXFP4; + + GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_mxfp4)); + + if (t->ne[1] % nrows_interleaved != 0) { + return -1; + } + + for (int b = 0; b < nrow; b += nrows_interleaved) { + for (int64_t x = 0; x < nblocks; x++) { + for (int i = 0; i < nrows_interleaved; i++) { + dst_tmp[i] = src[x + i * nblocks]; + } + *dst++ = make_block_mxfp4x8(dst_tmp, interleave_block); + } + src += nrows_interleaved * nblocks; + } + return 0; + + GGML_UNUSED(data_size); +} + namespace ggml::cpu::repack { // repack template @@ -2569,6 +2848,14 @@ template <> int repack(struct ggml_tensor * t, const void * return repack_iq4_nl_to_iq4_nl_8_bl(t, 8, data, data_size); } +template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { + return repack_mxfp4_to_mxfp4_4_bl(t, 4, data, data_size); +} + +template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { + return repack_mxfp4_to_mxfp4_8_bl(t, 8, data, data_size); +} + template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { return repack_q8_0_to_q8_0_4_bl(t, 4, data, data_size); } @@ -2636,6 +2923,14 @@ template <> void gemv(int n, float * s, size ggml_gemv_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc); } +template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemv_mxfp4_4x4_q8_0(n, s, bs, vx, vy, nr, nc); +} + +template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemv_mxfp4_8x8_q8_0(n, s, bs, vx, vy, nr, nc); +} + template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { ggml_gemv_q8_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc); } @@ -2703,6 +2998,14 @@ template <> void gemm(int n, float * s, size ggml_gemm_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc); } +template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemm_mxfp4_4x4_q8_0(n, s, bs, vx, vy, nr, nc); +} + +template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemm_mxfp4_8x8_q8_0(n, s, bs, vx, vy, nr, nc); +} + template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { ggml_gemm_q8_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc); } @@ -3111,6 +3414,10 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons static const ggml::cpu::repack::tensor_traits iq4_nl_4x4_q8_0; static const ggml::cpu::repack::tensor_traits iq4_nl_8x8_q8_0; + // instance for MXFP4 + static const ggml::cpu::repack::tensor_traits mxfp4_4x4_q8_0; + static const ggml::cpu::repack::tensor_traits mxfp4_8x8_q8_0; + // instance for Q8_0 static const ggml::cpu::repack::tensor_traits q8_0_4x4_q8_0; static const ggml::cpu::repack::tensor_traits q8_0_4x8_q8_0; @@ -3187,6 +3494,17 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons return &iq4_nl_4x4_q8_0; } } + } else if (cur->type == GGML_TYPE_MXFP4) { + if (ggml_cpu_has_avx2()) { + if (cur->ne[1] % 8 == 0) { + return &mxfp4_8x8_q8_0; + } + } + if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { + if (cur->ne[1] % 4 == 0) { + return &mxfp4_4x4_q8_0; + } + } } else if (cur->type == GGML_TYPE_Q8_0) { if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { if (cur->ne[1] % 4 == 0) { diff --git a/ggml/src/ggml-cpu/repack.h b/ggml/src/ggml-cpu/repack.h index ddf03d764..b9f821630 100644 --- a/ggml/src/ggml-cpu/repack.h +++ b/ggml/src/ggml-cpu/repack.h @@ -97,6 +97,19 @@ struct block_iq4_nlx8 { static_assert(sizeof(block_iq4_nlx8) == 8 * sizeof(ggml_half) + QK4_NL * 4, "wrong iq4_nlx8 block size/padding"); +struct block_mxfp4x4 { + uint8_t e[4]; + uint8_t qs[QK_MXFP4 * 2]; +}; +static_assert(sizeof(block_mxfp4x4) == 4 + QK_MXFP4 * 2, "wrong mxfp4x4 block size/padding"); + +struct block_mxfp4x8 { + uint8_t e[8]; + uint8_t qs[QK_MXFP4 * 4]; +}; +static_assert(sizeof(block_mxfp4x8) == 8 + QK_MXFP4 * 4, "wrong mxfp4x8 block size/padding"); + + #if defined(__cplusplus) extern "C" { #endif @@ -117,6 +130,8 @@ void ggml_gemv_q6_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo void ggml_gemv_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_mxfp4_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_mxfp4_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); @@ -129,6 +144,8 @@ void ggml_gemm_q6_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo void ggml_gemm_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_mxfp4_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_mxfp4_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q8_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); @@ -151,6 +168,8 @@ void ggml_gemv_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, void ggml_gemv_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_mxfp4_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); @@ -163,6 +182,8 @@ void ggml_gemm_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, void ggml_gemm_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_mxfp4_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); From 8d3b962f47cdb05df8ab8801da38cb3c55d50337 Mon Sep 17 00:00:00 2001 From: Daniel Bevenius Date: Fri, 27 Feb 2026 14:42:24 +0100 Subject: [PATCH 2/4] ci : use ubuntu-latest for gguf-publish workflow (#19951) This commit changes the runner for the gguf-publish workflow from ubuntu-slim back to ubuntu-latest, which was updated in Commit 142cbe2ac68978e5dec3a2e19c1b64ef1c5740b1 ("ci : use new 1vCPU runner for lightweight jobs (#19107)"). The motivation for this is that the action used in the workflow depends on the docker daemon, which does not seem not available in the ubuntu-slim runner. This is currently causing an error in the workflow and preventing the gguf-publish workflow from running successfully. Today was the the first time since the original change (I think) that publish task has been run which may be why the issue was not noticed before. Refs: https://github.com/ggml-org/llama.cpp/actions/runs/22481900566 --- .github/workflows/gguf-publish.yml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/gguf-publish.yml b/.github/workflows/gguf-publish.yml index 0e9576645..5bdab0f15 100644 --- a/.github/workflows/gguf-publish.yml +++ b/.github/workflows/gguf-publish.yml @@ -21,7 +21,7 @@ on: jobs: deploy: - runs-on: ubuntu-slim + runs-on: ubuntu-latest steps: - uses: actions/checkout@v6 From 5596a35791806ec0222b7e0d5ecb80b48b24585a Mon Sep 17 00:00:00 2001 From: Sami Kama Date: Fri, 27 Feb 2026 08:44:42 -0800 Subject: [PATCH 3/4] server: Mirroring /v1/responses to /responses to match /v1/chat/completions pattern (#19873) --- tools/server/server.cpp | 1 + 1 file changed, 1 insertion(+) diff --git a/tools/server/server.cpp b/tools/server/server.cpp index 542b98453..f353dcdde 100644 --- a/tools/server/server.cpp +++ b/tools/server/server.cpp @@ -178,6 +178,7 @@ int main(int argc, char ** argv) { ctx_http.post("/v1/chat/completions", ex_wrapper(routes.post_chat_completions)); ctx_http.post("/api/chat", ex_wrapper(routes.post_chat_completions)); // ollama specific endpoint ctx_http.post("/v1/responses", ex_wrapper(routes.post_responses_oai)); + ctx_http.post("/responses", ex_wrapper(routes.post_responses_oai)); ctx_http.post("/v1/messages", ex_wrapper(routes.post_anthropic_messages)); // anthropic messages API ctx_http.post("/v1/messages/count_tokens", ex_wrapper(routes.post_anthropic_count_tokens)); // anthropic token counting ctx_http.post("/infill", ex_wrapper(routes.post_infill)); From 3e6ab244ad32a2fa9c36372dc741079fa4c0c8e8 Mon Sep 17 00:00:00 2001 From: Roj234 <82699138+roj234@users.noreply.github.com> Date: Sat, 28 Feb 2026 01:28:36 +0800 Subject: [PATCH 4/4] server: Add pragma once to server-context.h (#19944) --- tools/server/server-context.h | 2 ++ 1 file changed, 2 insertions(+) diff --git a/tools/server/server-context.h b/tools/server/server-context.h index 631d573fc..75f3d2de5 100644 --- a/tools/server/server-context.h +++ b/tools/server/server-context.h @@ -1,3 +1,5 @@ +#pragma once + #include "server-http.h" #include "server-task.h" #include "server-queue.h"