mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2026-06-27 23:50:20 -05:00
5eae9cb1d9
* WIP: add NVFP4 quantization support * tests * improve NVFP4 dot product implementation performance and fix bad super call * typo * Use nvfp4 kvalues * vulkan : fix NVFP4 shader compilation by including kvalues_mxfp4 lookup table * vulcal and perf fixes * wip * Fix metal * fix vulcan * Rename threshold & fix wrong scale * Fix MOE * Shelf backend implementations (CUDA, Metal, Vulkan, arch-specific SIMD) Remove NVFP4 support from GPU backends and architecture-specific optimized dot products. These should be added in separate PRs so backend specialists can review them independently. Reverted files: - ggml-cuda: common.cuh, convert.cu, mmq.cu/cuh, mmvq.cu, vecdotq.cuh, quantize.cu/cuh, mma.cuh, ggml-cuda.cu, fattn-tile.cuh - ggml-metal: ggml-metal.metal, ggml-metal-device.cpp, ggml-metal-impl.h, ggml-metal-ops.cpp - ggml-vulkan: ggml-vulkan.cpp, all vulkan-shaders/* - ggml-cpu arch: arm/quants.c, x86/quants.c, powerpc/quants.c, s390/quants.c Core NVFP4 support (type definition, CPU fallback dot product, quantization, dequantization, conversion) is retained. * Fix arch-fallback.h: add NVFP4 generic fallback for all platforms After shelving backend-specific SIMD implementations, the generic CPU dot product needs to be aliased on ARM, x86, PowerPC, and s390 platforms that previously relied on arch-specific versions. * quantize: add NVFP4 as a quantization type option * Fix ggml_fp32_to_ue4m3: handle subnormal values Previously, values with ue4m3_exp <= 0 were clamped to 0, causing all small scales to underflow. This made NVFP4 quantization via llama-quantize produce garbage (PPL = 5.8M) since typical transformer weights have amax/6.0 in the range 0.001-0.01, which falls in the UE4M3 subnormal range. Now subnormals are properly encoded as man * 2^-9 (exp=0, man=1..7), matching the decode path in ggml_ue4m3_to_fp32. Result: NVFP4 requantization now produces PPL = 15.25 (vs F16 = 14.33), comparable to Q4_1 (PPL = 15.81) at slightly lower BPW (4.70 vs 5.15). * Restore ARM NEON NVFP4 dot product implementation Restores the optimized ggml_vec_dot_nvfp4_q8_0 for ARM NEON using vqtbl1q_s8 lookup and ggml_vdotq_s32 dot products. tg128 performance: 4.37 t/s (generic) -> 13.66 t/s (NEON) = 3.1x speedup * Optimize ARM NEON NVFP4 dot product: LUT + vpaddq + vfmaq - Add ue4m3_scale_lut[128] to ggml-common.h replacing branch-heavy ggml_ue4m3_to_fp32() in the hot loop - Use vpaddq_s32 for pairwise int32 reduction instead of vaddvq_s32 - Accumulate with vfmaq_f32 into float32x4_t vector accumulators tg128: 8.1 -> 31.0 t/s (3.8x speedup, 77% of Q4_1 speed) * ARM NEON NVFP4: rearrange q8 to match nibble layout Alternative approach: rearrange q8 data to match the NVFP4 lo/hi nibble layout instead of rearranging the looked-up NVFP4 values. Eliminates vcombine_s8(vget_low, vget_low) shuffles. Performance is equivalent (~18.5 t/s) - the bottleneck is the 2x block overhead from QK=16 vs QK=32, not the shuffle instructions. * CPU only backend 64 super-block layout * cleanup * Remove unused LUT * int * exclude NVFP4 from unsupported ops in metal build * remove quantization for now * store scales as native UE4M3, preserve original model bits when possible * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * correct comment * format * reduce duplication and cleanup * Address comments * move detection to prepare_tensors * Use math instead of const * Move * fix comment * Shelf quantize tests * Rebase and move check * cleanup * lint * Update gguf-py/gguf/scripts/gguf_convert_endian.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Use fallback quant config * Simplify Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * organize * Refactor * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * add quantize_nvfp4 (required for test_quants.py) * add quantize_nvfp4 (required for test_quants.py) * add quantize_nvfp4 (required for test_quants.py) * fix return type --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
110 lines
8.9 KiB
C
110 lines
8.9 KiB
C
#pragma once
|
|
|
|
#define GGML_COMMON_DECL_C
|
|
#include "ggml-common.h"
|
|
|
|
#include "ggml.h"
|
|
|
|
// GGML internal header
|
|
|
|
#ifdef __cplusplus
|
|
extern "C" {
|
|
#endif
|
|
|
|
// NOTE: these functions are defined as GGML_API because they used by the CPU backend
|
|
|
|
// Quantization
|
|
GGML_API void quantize_row_q4_0_ref(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void quantize_row_q4_1_ref(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void quantize_row_q5_0_ref(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void quantize_row_q5_1_ref(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void quantize_row_q8_0_ref(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void quantize_row_q8_1_ref(const float * GGML_RESTRICT x, block_q8_1 * 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);
|
|
GGML_API void quantize_row_nvfp4_ref(const float * GGML_RESTRICT x, block_nvfp4 * GGML_RESTRICT y, int64_t k);
|
|
|
|
GGML_API void quantize_row_q2_K_ref(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void quantize_row_q3_K_ref(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void quantize_row_q4_K_ref(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void quantize_row_q5_K_ref(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void quantize_row_q6_K_ref(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void quantize_row_q8_K_ref(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k);
|
|
|
|
GGML_API void quantize_row_tq1_0_ref(const float * GGML_RESTRICT x, block_tq1_0 * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void quantize_row_tq2_0_ref(const float * GGML_RESTRICT x, block_tq2_0 * GGML_RESTRICT y, int64_t k);
|
|
|
|
GGML_API void quantize_row_iq3_xxs_ref(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void quantize_row_iq4_nl_ref (const float * GGML_RESTRICT x, block_iq4_nl * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void quantize_row_iq4_xs_ref (const float * GGML_RESTRICT x, block_iq4_xs * GGML_RESTRICT y, int64_t k);
|
|
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);
|
|
|
|
// 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);
|
|
GGML_API void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
//GGML_API void dequantize_row_q8_1(const block_q8_1 * 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);
|
|
GGML_API void dequantize_row_nvfp4(const block_nvfp4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
|
|
GGML_API void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
|
|
GGML_API void dequantize_row_tq1_0(const block_tq1_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void dequantize_row_tq2_0(const block_tq2_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
|
|
GGML_API void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void dequantize_row_iq2_xs (const block_iq2_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void dequantize_row_iq2_s (const block_iq2_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void dequantize_row_iq1_s (const block_iq1_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void dequantize_row_iq1_m (const block_iq1_m * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
GGML_API void dequantize_row_iq4_nl (const block_iq4_nl * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
|
|
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);
|
|
|
|
// 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);
|
|
GGML_API size_t quantize_iq2_s (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_iq3_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_iq1_s (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_iq1_m (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_iq4_nl (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_iq4_xs (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_iq3_s (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_tq1_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_tq2_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_q2_K(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_q3_K(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_q4_K(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_q5_K(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_q6_K(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_q4_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_q4_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_q5_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_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 size_t quantize_nvfp4(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);
|
|
GGML_API void iq3xs_free_impl(int grid_size);
|
|
|
|
#ifdef __cplusplus
|
|
}
|
|
#endif
|