mistralrs-quant 0.8.1

Fast, flexible LLM inference.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152

use float8::F8E4M3;
use half::{bf16, f16};

pub(crate) const HAVE_BLOCKWISE_DEQUANT_KERNELS: bool = cfg!(has_blockwise_fp8_kernels);
pub(crate) const HAVE_BLOCKWISE_QUANT_KERNELS: bool = cfg!(has_blockwise_fp8_kernels);
pub(crate) const HAVE_BLOCKWISE_GEMM_KERNELS: bool = cfg!(has_blockwise_fp8_kernels);

extern "C" {
    pub(crate) fn launch_dequant_fp8_blockwise_kernel_f32(
        d_weight: *const F8E4M3,
        d_scale: *const f32,
        d_output: *mut f32,
        weight_height: i32,
        weight_width: i32,
        weight_row_stride: i32,
        scale_stride: i32,
        weight_block_size_y: i32,
        weight_block_size_x: i32,
        stream: candle_core::cuda::cudarc::driver::sys::CUstream,
    );

    pub(crate) fn launch_dequant_fp8_blockwise_kernel_f16(
        d_weight: *const F8E4M3,
        d_scale: *const f32,
        d_output: *mut f16,
        weight_height: i32,
        weight_width: i32,
        weight_row_stride: i32,
        scale_stride: i32,
        weight_block_size_y: i32,
        weight_block_size_x: i32,
        stream: candle_core::cuda::cudarc::driver::sys::CUstream,
    );

    pub(crate) fn launch_dequant_fp8_blockwise_kernel_bf16(
        d_weight: *const F8E4M3,
        d_scale: *const f32,
        d_output: *mut bf16,
        weight_height: i32,
        weight_width: i32,
        weight_row_stride: i32,
        scale_stride: i32,
        weight_block_size_y: i32,
        weight_block_size_x: i32,
        stream: candle_core::cuda::cudarc::driver::sys::CUstream,
    );

    pub(crate) fn launch_quant_fp8_blockwise_kernel_f32(
        d_input: *const f32,
        d_weight: *mut F8E4M3,
        d_scale: *mut f32,
        weight_height: i32,
        weight_width: i32,
        weight_row_stride: i32,
        scale_stride: i32,
        weight_block_size_y: i32,
        weight_block_size_x: i32,
        stream: candle_core::cuda::cudarc::driver::sys::CUstream,
    );

    pub(crate) fn launch_quant_fp8_blockwise_kernel_f16(
        d_input: *const f16,
        d_weight: *mut F8E4M3,
        d_scale: *mut f32,
        weight_height: i32,
        weight_width: i32,
        weight_row_stride: i32,
        scale_stride: i32,
        weight_block_size_y: i32,
        weight_block_size_x: i32,
        stream: candle_core::cuda::cudarc::driver::sys::CUstream,
    );

    pub(crate) fn launch_quant_fp8_blockwise_kernel_bf16(
        d_input: *const bf16,
        d_weight: *mut F8E4M3,
        d_scale: *mut f32,
        weight_height: i32,
        weight_width: i32,
        weight_row_stride: i32,
        scale_stride: i32,
        weight_block_size_y: i32,
        weight_block_size_x: i32,
        stream: candle_core::cuda::cudarc::driver::sys::CUstream,
    );

    // FP8 Matmul kernels (for forward method)
    pub(crate) fn launch_fp8_matmul_f16(
        input: *const f16,
        weight: *const F8E4M3,
        weight_scale: *const f32,
        output: *mut f16,
        m: i32,
        n: i32,
        k: i32,
        scale_row_stride: i32,
        block_size_y: i32,
        block_size_x: i32,
        stream: candle_core::cuda::cudarc::driver::sys::CUstream,
    );

    pub(crate) fn launch_fp8_matmul_bf16(
        input: *const bf16,
        weight: *const F8E4M3,
        weight_scale: *const f32,
        output: *mut bf16,
        m: i32,
        n: i32,
        k: i32,
        scale_row_stride: i32,
        block_size_y: i32,
        block_size_x: i32,
        stream: candle_core::cuda::cudarc::driver::sys::CUstream,
    );

    // FP8 Indexed MoE GEMM kernels (for gather_forward method)
    pub(crate) fn launch_fp8_indexed_moe_gemm_f16(
        input: *const f16,
        weights: *const F8E4M3,
        weight_scales: *const f32,
        indices: *const u32,
        output: *mut f16,
        num_tokens: i32,
        topk: i32,
        num_experts: i32,
        n: i32,
        k: i32,
        scale_row_stride: i32,
        block_size_y: i32,
        block_size_x: i32,
        input_has_topk_dim: bool,
        stream: candle_core::cuda::cudarc::driver::sys::CUstream,
    );

    pub(crate) fn launch_fp8_indexed_moe_gemm_bf16(
        input: *const bf16,
        weights: *const F8E4M3,
        weight_scales: *const f32,
        indices: *const u32,
        output: *mut bf16,
        num_tokens: i32,
        topk: i32,
        num_experts: i32,
        n: i32,
        k: i32,
        scale_row_stride: i32,
        block_size_y: i32,
        block_size_x: i32,
        input_has_topk_dim: bool,
        stream: candle_core::cuda::cudarc::driver::sys::CUstream,
    );
}