numr 0.5.2

High-performance numerical computing with multi-backend GPU acceleration (CPU/CUDA/WebGPU)
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
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250

//! FP8 matmul CUDA kernel launchers
//!
//! Launches FP8 GEMM kernels with per-tensor scaling and F32 accumulation.
//! Output can be F32, F16, or BF16.

use cudarc::driver::PushKernelArg;
use cudarc::driver::safe::{CudaContext, CudaStream};
use std::sync::Arc;

use super::loader::{get_kernel_function, get_or_load_module, launch_config};
use crate::dtype::DType;
use crate::error::{Error, Result};

const FP8_MATMUL_MODULE: &str = "fp8_matmul";

// Tile config matching the .cu defines
const TILE_M: u32 = 64;
const TILE_N: u32 = 64;
const THREAD_M: u32 = 4;
const THREAD_N: u32 = 4;

fn fp8_matmul_launch_cfg(m: usize, n: usize, batch: usize) -> super::loader::LaunchConfig {
    let grid_x = ((n as u32) + TILE_N - 1) / TILE_N;
    let grid_y = ((m as u32) + TILE_M - 1) / TILE_M;
    let threads_x = TILE_N / THREAD_N;
    let threads_y = TILE_M / THREAD_M;
    launch_config(
        (grid_x, grid_y, (batch as u32).max(1)),
        (threads_x, threads_y, 1),
        0,
    )
}

fn out_dtype_suffix(out_dtype: DType) -> Result<&'static str> {
    match out_dtype {
        DType::F32 => Ok("f32"),
        DType::F16 => Ok("f16"),
        DType::BF16 => Ok("bf16"),
        _ => Err(Error::UnsupportedDType {
            dtype: out_dtype,
            op: "fp8_matmul output",
        }),
    }
}

/// Launch FP8 E4M3 x E4M3 matmul kernel.
///
/// # Safety
///
/// All pointers must be valid device memory with correct sizes.
pub unsafe fn launch_fp8_matmul_e4m3(
    context: &Arc<CudaContext>,
    stream: &CudaStream,
    device_index: usize,
    out_dtype: DType,
    a_ptr: u64,
    b_ptr: u64,
    c_ptr: u64,
    scale_a: f32,
    scale_b: f32,
    m: usize,
    n: usize,
    k: usize,
) -> Result<()> {
    let module = get_or_load_module(context, device_index, FP8_MATMUL_MODULE)?;
    let suffix = out_dtype_suffix(out_dtype)?;
    let func_name = format!("fp8_matmul_e4m3_{}", suffix);
    let func = get_kernel_function(&module, &func_name)?;

    let cfg = fp8_matmul_launch_cfg(m, n, 1);
    let m_u32 = m as u32;
    let n_u32 = n as u32;
    let k_u32 = k as u32;

    unsafe {
        let mut builder = stream.launch_builder(&func);
        builder.arg(&a_ptr);
        builder.arg(&b_ptr);
        builder.arg(&c_ptr);
        builder.arg(&scale_a);
        builder.arg(&scale_b);
        builder.arg(&m_u32);
        builder.arg(&n_u32);
        builder.arg(&k_u32);
        builder.launch(cfg).map_err(|e| {
            Error::Internal(format!(
                "CUDA fp8_matmul_e4m3 kernel launch failed: {:?}",
                e
            ))
        })?;
    }

    Ok(())
}

/// Launch FP8 E5M2 x E4M3 matmul kernel (backward pass).
///
/// # Safety
///
/// All pointers must be valid device memory with correct sizes.
pub unsafe fn launch_fp8_matmul_e5m2(
    context: &Arc<CudaContext>,
    stream: &CudaStream,
    device_index: usize,
    out_dtype: DType,
    a_ptr: u64,
    b_ptr: u64,
    c_ptr: u64,
    scale_a: f32,
    scale_b: f32,
    m: usize,
    n: usize,
    k: usize,
) -> Result<()> {
    let module = get_or_load_module(context, device_index, FP8_MATMUL_MODULE)?;
    let suffix = out_dtype_suffix(out_dtype)?;
    let func_name = format!("fp8_matmul_e5m2_{}", suffix);
    let func = get_kernel_function(&module, &func_name)?;

    let cfg = fp8_matmul_launch_cfg(m, n, 1);
    let m_u32 = m as u32;
    let n_u32 = n as u32;
    let k_u32 = k as u32;

    unsafe {
        let mut builder = stream.launch_builder(&func);
        builder.arg(&a_ptr);
        builder.arg(&b_ptr);
        builder.arg(&c_ptr);
        builder.arg(&scale_a);
        builder.arg(&scale_b);
        builder.arg(&m_u32);
        builder.arg(&n_u32);
        builder.arg(&k_u32);
        builder.launch(cfg).map_err(|e| {
            Error::Internal(format!(
                "CUDA fp8_matmul_e5m2 kernel launch failed: {:?}",
                e
            ))
        })?;
    }

    Ok(())
}

/// Launch batched FP8 E4M3 x E4M3 matmul kernel.
///
/// # Safety
///
/// All pointers must be valid device memory with correct sizes.
pub unsafe fn launch_fp8_matmul_e4m3_batched(
    context: &Arc<CudaContext>,
    stream: &CudaStream,
    device_index: usize,
    out_dtype: DType,
    a_ptr: u64,
    b_ptr: u64,
    c_ptr: u64,
    scale_a: f32,
    scale_b: f32,
    batch: usize,
    m: usize,
    n: usize,
    k: usize,
) -> Result<()> {
    let module = get_or_load_module(context, device_index, FP8_MATMUL_MODULE)?;
    let suffix = out_dtype_suffix(out_dtype)?;
    let func_name = format!("fp8_matmul_e4m3_batched_{}", suffix);
    let func = get_kernel_function(&module, &func_name)?;

    let cfg = fp8_matmul_launch_cfg(m, n, batch);
    let batch_u32 = batch as u32;
    let m_u32 = m as u32;
    let n_u32 = n as u32;
    let k_u32 = k as u32;

    unsafe {
        let mut builder = stream.launch_builder(&func);
        builder.arg(&a_ptr);
        builder.arg(&b_ptr);
        builder.arg(&c_ptr);
        builder.arg(&scale_a);
        builder.arg(&scale_b);
        builder.arg(&batch_u32);
        builder.arg(&m_u32);
        builder.arg(&n_u32);
        builder.arg(&k_u32);
        builder.launch(cfg).map_err(|e| {
            Error::Internal(format!(
                "CUDA fp8_matmul_e4m3_batched kernel launch failed: {:?}",
                e
            ))
        })?;
    }

    Ok(())
}

/// Launch batched FP8 E5M2 x E4M3 matmul kernel (backward pass).
///
/// # Safety
///
/// All pointers must be valid device memory with correct sizes.
pub unsafe fn launch_fp8_matmul_e5m2_batched(
    context: &Arc<CudaContext>,
    stream: &CudaStream,
    device_index: usize,
    out_dtype: DType,
    a_ptr: u64,
    b_ptr: u64,
    c_ptr: u64,
    scale_a: f32,
    scale_b: f32,
    batch: usize,
    m: usize,
    n: usize,
    k: usize,
) -> Result<()> {
    let module = get_or_load_module(context, device_index, FP8_MATMUL_MODULE)?;
    let suffix = out_dtype_suffix(out_dtype)?;
    let func_name = format!("fp8_matmul_e5m2_batched_{}", suffix);
    let func = get_kernel_function(&module, &func_name)?;

    let cfg = fp8_matmul_launch_cfg(m, n, batch);
    let batch_u32 = batch as u32;
    let m_u32 = m as u32;
    let n_u32 = n as u32;
    let k_u32 = k as u32;

    unsafe {
        let mut builder = stream.launch_builder(&func);
        builder.arg(&a_ptr);
        builder.arg(&b_ptr);
        builder.arg(&c_ptr);
        builder.arg(&scale_a);
        builder.arg(&scale_b);
        builder.arg(&batch_u32);
        builder.arg(&m_u32);
        builder.arg(&n_u32);
        builder.arg(&k_u32);
        builder.launch(cfg).map_err(|e| {
            Error::Internal(format!(
                "CUDA fp8_matmul_e5m2_batched kernel launch failed: {:?}",
                e
            ))
        })?;
    }

    Ok(())
}