boostr 0.1.0

ML framework built on numr - attention, quantization, model architectures
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
251
252
253
254
255
256
257
258
259
260
261
262
263

//! CUDA INT4 GEMM dispatch helpers (AWQ, GPTQ, Marlin)

use crate::error::{Error, Result};
use cudarc::driver::PushKernelArg;
use cudarc::driver::safe::LaunchConfig;
use numr::runtime::Device;
use numr::runtime::cuda::{CudaClient, CudaRuntime};
use numr::tensor::Tensor;

use super::kernels::{self, INT4_GEMM_GPTQ_MODULE, INT4_GEMM_MODULE, MARLIN_GEMM_MODULE};

/// M threshold: use GEMV for M <= this, tiled GEMM for M > this
const GEMV_THRESHOLD: u32 = 4;

#[allow(clippy::too_many_arguments)]
pub fn launch_int4_gemm(
    client: &CudaClient,
    input: &Tensor<CudaRuntime>,
    qweight: &Tensor<CudaRuntime>,
    scales: &Tensor<CudaRuntime>,
    zeros: &Tensor<CudaRuntime>,
    output: &Tensor<CudaRuntime>,
    m: u32,
    k: u32,
    n: u32,
    group_size: u32,
) -> Result<()> {
    let device_index = input.device().id();
    let module = kernels::get_or_load_module(client.context(), device_index, INT4_GEMM_MODULE)?;

    if m <= GEMV_THRESHOLD {
        tracing::debug!(
            m,
            k,
            n,
            path = "awq_int4_gemv",
            "CUDA AWQ kernel: INT4 GEMV (optimized)"
        );
        // GEMV path: 4 warps/block (128 threads), each warp handles 8 cols (one packed u32)
        // 32 output columns per block. Shared memory caches input row.
        let func = kernels::get_kernel_function(&module, "int4_gemv_f32")?;
        let n_packed = n / 8;
        let cfg = LaunchConfig {
            grid_dim: (n_packed.div_ceil(4), m, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };

        let input_ptr = input.ptr();
        let qweight_ptr = qweight.ptr();
        let scales_ptr = scales.ptr();
        let zeros_ptr = zeros.ptr();
        let output_ptr = output.ptr();

        unsafe {
            let mut builder = client.stream().launch_builder(&func);
            builder.arg(&input_ptr);
            builder.arg(&qweight_ptr);
            builder.arg(&scales_ptr);
            builder.arg(&zeros_ptr);
            builder.arg(&output_ptr);
            builder.arg(&m);
            builder.arg(&k);
            builder.arg(&n);
            builder.arg(&group_size);
            builder.launch(cfg).map_err(|e| Error::QuantError {
                reason: format!("CUDA int4_gemv launch failed: {:?}", e),
            })?;
        }
    } else {
        tracing::debug!(
            m,
            k,
            n,
            path = "awq_int4_gemm",
            "CUDA AWQ kernel: INT4 tiled GEMM (optimized)"
        );
        // Tiled GEMM path: BM=32, BN=32, BK=32
        // Block: (32, 4) = 128 threads, each thread handles 8 rows
        let func = kernels::get_kernel_function(&module, "int4_gemm_f32")?;
        let cfg = LaunchConfig {
            grid_dim: (n.div_ceil(32), m.div_ceil(32), 1),
            block_dim: (32, 4, 1),
            shared_mem_bytes: 0,
        };

        let input_ptr = input.ptr();
        let qweight_ptr = qweight.ptr();
        let scales_ptr = scales.ptr();
        let zeros_ptr = zeros.ptr();
        let output_ptr = output.ptr();

        unsafe {
            let mut builder = client.stream().launch_builder(&func);
            builder.arg(&input_ptr);
            builder.arg(&qweight_ptr);
            builder.arg(&scales_ptr);
            builder.arg(&zeros_ptr);
            builder.arg(&output_ptr);
            builder.arg(&m);
            builder.arg(&k);
            builder.arg(&n);
            builder.arg(&group_size);
            builder.launch(cfg).map_err(|e| Error::QuantError {
                reason: format!("CUDA int4_gemm launch failed: {:?}", e),
            })?;
        }
    }
    Ok(())
}

#[allow(clippy::too_many_arguments)]
pub fn launch_int4_gemm_gptq(
    client: &CudaClient,
    input: &Tensor<CudaRuntime>,
    qweight: &Tensor<CudaRuntime>,
    qzeros: &Tensor<CudaRuntime>,
    scales: &Tensor<CudaRuntime>,
    g_idx: &Tensor<CudaRuntime>,
    output: &Tensor<CudaRuntime>,
    m: u32,
    k: u32,
    n: u32,
) -> Result<()> {
    let device_index = input.device().id();
    let module =
        kernels::get_or_load_module(client.context(), device_index, INT4_GEMM_GPTQ_MODULE)?;

    let input_ptr = input.ptr();
    let qweight_ptr = qweight.ptr();
    let qzeros_ptr = qzeros.ptr();
    let scales_ptr = scales.ptr();
    let g_idx_ptr = g_idx.ptr();
    let output_ptr = output.ptr();

    if m <= GEMV_THRESHOLD {
        tracing::debug!(
            m,
            k,
            n,
            path = "gptq_int4_gemv",
            "CUDA GPTQ kernel: INT4 GEMV (optimized)"
        );
        // GEMV: 128 threads (one per output col), tiled over K in chunks of 128
        // Grid: (ceil(K/128), ceil(N/128), M). Uses atomicAdd → output must be zeroed.
        let func = kernels::get_kernel_function(&module, "int4_gemv_gptq_f32")?;
        let k_blocks = k.div_ceil(128);
        let cfg = LaunchConfig {
            grid_dim: (k_blocks, n.div_ceil(128), m),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };

        // Zero output before atomicAdd accumulation
        let output_bytes = (m * n) as usize * std::mem::size_of::<f32>();
        unsafe {
            cudarc::driver::result::memset_d8_async(
                output.ptr(),
                0,
                output_bytes,
                client.stream().cu_stream(),
            )
            .map_err(|e| Error::QuantError {
                reason: format!("CUDA memset failed: {:?}", e),
            })?;
        }

        unsafe {
            let mut builder = client.stream().launch_builder(&func);
            builder.arg(&input_ptr);
            builder.arg(&qweight_ptr);
            builder.arg(&qzeros_ptr);
            builder.arg(&scales_ptr);
            builder.arg(&g_idx_ptr);
            builder.arg(&output_ptr);
            builder.arg(&m);
            builder.arg(&k);
            builder.arg(&n);
            builder.launch(cfg).map_err(|e| Error::QuantError {
                reason: format!("CUDA int4_gemv_gptq launch failed: {:?}", e),
            })?;
        }
    } else {
        tracing::debug!(
            m,
            k,
            n,
            path = "gptq_int4_gemm",
            "CUDA GPTQ kernel: INT4 tiled GEMM (optimized)"
        );
        // Tiled GEMM: BM=32, BN=32, BK=32
        let func = kernels::get_kernel_function(&module, "int4_gemm_gptq_f32")?;
        let cfg = LaunchConfig {
            grid_dim: (n.div_ceil(32), m.div_ceil(32), 1),
            block_dim: (32, 4, 1),
            shared_mem_bytes: 0,
        };

        unsafe {
            let mut builder = client.stream().launch_builder(&func);
            builder.arg(&input_ptr);
            builder.arg(&qweight_ptr);
            builder.arg(&qzeros_ptr);
            builder.arg(&scales_ptr);
            builder.arg(&g_idx_ptr);
            builder.arg(&output_ptr);
            builder.arg(&m);
            builder.arg(&k);
            builder.arg(&n);
            builder.launch(cfg).map_err(|e| Error::QuantError {
                reason: format!("CUDA int4_gemm_gptq launch failed: {:?}", e),
            })?;
        }
    }
    Ok(())
}

#[allow(clippy::too_many_arguments)]
pub fn launch_marlin_gemm(
    client: &CudaClient,
    input: &Tensor<CudaRuntime>,
    weight: &Tensor<CudaRuntime>,
    scales: &Tensor<CudaRuntime>,
    zeros: &Tensor<CudaRuntime>,
    output: &Tensor<CudaRuntime>,
    m: u32,
    k: u32,
    n: u32,
    group_size: u32,
) -> Result<()> {
    let device_index = input.device().id();
    let module = kernels::get_or_load_module(client.context(), device_index, MARLIN_GEMM_MODULE)?;
    let func = kernels::get_kernel_function(&module, "marlin_gemm_f32")?;

    let cfg = LaunchConfig {
        grid_dim: (n.div_ceil(16), m.div_ceil(16), 1),
        block_dim: (16, 16, 1),
        shared_mem_bytes: 0,
    };

    let input_ptr = input.ptr();
    let weight_ptr = weight.ptr();
    let scales_ptr = scales.ptr();
    let zeros_ptr = zeros.ptr();
    let output_ptr = output.ptr();

    unsafe {
        let mut builder = client.stream().launch_builder(&func);
        builder.arg(&input_ptr);
        builder.arg(&weight_ptr);
        builder.arg(&scales_ptr);
        builder.arg(&zeros_ptr);
        builder.arg(&output_ptr);
        builder.arg(&m);
        builder.arg(&k);
        builder.arg(&n);
        builder.arg(&group_size);
        builder.launch(cfg).map_err(|e| Error::QuantError {
            reason: format!("CUDA marlin_gemm launch failed: {:?}", e),
        })?;
    }
    Ok(())
}