char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Oliver SimonsCUDA: Improve NVFP4 W4A4 activation quantization (#25730)1a064ab09

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
3.4 KiB70 linesraw
1#pragma once
2
3#include "common.cuh"
4#include "mmq.cuh"
5
6#include <cstdint>
7
8#define CUDA_QUANTIZE_BLOCK_SIZE     256
9#define CUDA_QUANTIZE_BLOCK_SIZE_MMQ 128
10
11static_assert(MATRIX_ROW_PADDING %    CUDA_QUANTIZE_BLOCK_SIZE      == 0, "Risk of out-of-bounds access.");
12static_assert(MATRIX_ROW_PADDING % (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ) == 0, "Risk of out-of-bounds access.");
13
14typedef void (*quantize_cuda_t)(
15        const float * x, const int32_t * ids, void * vy,
16        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
17        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);
18
19void quantize_row_q8_1_cuda(
20        const float * x, const int32_t * ids, void * vy,
21        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
22        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);
23
24void quantize_mmq_q8_1_cuda(
25        const float * x, const int32_t * ids, void * vy,
26        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
27        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);
28
29void quantize_mmq_fp4_cuda(const float *   x,
30                             const int32_t * ids,
31                             void *          vy,
32                             float *         scale,
33                             ggml_type       type_src0,
34                             bool            use_aligned_float8,
35                             int64_t         ne00,
36                             int64_t         s01,
37                             int64_t         s02,
38                             int64_t         s03,
39                             int64_t         ne0,
40                             int64_t         ne1,
41                             int64_t         ne2,
42                             int64_t         ne3,
43                             cudaStream_t    stream);
44
45// quantize each token once and scatter the block to its compact rows (via the inverse map)
46void quantize_scatter_mmq_fp4_cuda(const float *   x,
47                                   const int32_t * ids_src1_inv,
48                                   void *          vy,
49                                   float *         scale,
50                                   ggml_type       type_src0,
51                                   bool            use_aligned_float8,
52                                   int64_t         ne00,
53                                   int64_t         stride_token,
54                                   int64_t         ne0,
55                                   int64_t         n_tokens,
56                                   int64_t         nrows_dst,
57                                   int             n_expert_used,
58                                   cudaStream_t    stream);
59
60void quantize_scatter_mmq_q8_1_cuda(const float *   x,
61                                    const int32_t * ids_src1_inv,
62                                    void *          vy,
63                                    ggml_type       type_src0,
64                                    int64_t         ne00,
65                                    int64_t         stride_token,
66                                    int64_t         ne0,
67                                    int64_t         n_tokens,
68                                    int64_t         nrows_dst,
69                                    int             n_expert_used,
70                                    cudaStream_t    stream);