boostr 0.2.0

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
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302

//! Variable-length (ragged) attention CUDA launchers
//!
//! Packed sequences with cu_seqlens indexing. Supports F32 and F16,
//! head dims 64, 128, and 256, with GQA (num_kv_heads <= num_heads).
//! Both forward and backward are implemented for all supported head dims.

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

use super::flash::set_smem_attribute;
use crate::ops::cuda::kernels::{
    self, VARLEN_ATTENTION_BWD_FP16_MODULE, VARLEN_ATTENTION_BWD_MODULE, VARLEN_ATTENTION_MODULE,
};

/// Return `(BLOCK_M, BLOCK_N)` for the given head_dim and dtype.
///
/// Constraints: smem = (BLOCK_M + 2*BLOCK_N) * head_dim * dtype_size ≤ 48 KB,
/// and block_dim = BLOCK_M (one thread per Q row).
///
/// | head_dim | dtype | BLOCK_M | BLOCK_N | smem    |
/// |----------|-------|---------|---------|---------|
/// | 64/128   | any   | 128     | 64      | ≤ 48 KB |
/// | 256      | F32   | 16      | 16      | 48 KB   |
/// | 256      | F16   | 32      | 32      | 48 KB   |
#[inline]
fn block_config(head_dim: usize, dtype: DType) -> (usize, usize) {
    match head_dim {
        256 => match dtype {
            DType::F16 => (32, 32),
            _ => (16, 16), // F32 and any future type default to 16/16
        },
        _ => (128, 64), // 64 and 128 use the proven large-tile config
    }
}

impl VarLenAttentionOps<CudaRuntime> for CudaClient {
    fn varlen_attention_fwd(
        &self,
        q: &Tensor<CudaRuntime>,
        k: &Tensor<CudaRuntime>,
        v: &Tensor<CudaRuntime>,
        cu_seqlens_q: &Tensor<CudaRuntime>,
        cu_seqlens_k: &Tensor<CudaRuntime>,
        batch_size: usize,
        num_heads: usize,
        num_kv_heads: usize,
        max_seqlen_q: usize,
        max_seqlen_k: usize,
        head_dim: usize,
        causal: bool,
    ) -> Result<(Tensor<CudaRuntime>, Tensor<CudaRuntime>)> {
        if head_dim != 64 && head_dim != 128 && head_dim != 256 {
            return Err(Error::KernelError {
                reason: format!(
                    "varlen attention: unsupported head_dim {head_dim}, only 64/128/256"
                ),
            });
        }

        let dtype = q.dtype();
        let dtype_suffix = match dtype {
            DType::F32 => "fp32",
            DType::F16 => "fp16",
            _ => {
                return Err(Error::KernelError {
                    reason: format!("varlen attention: unsupported dtype {dtype:?}, only F32/F16"),
                });
            }
        };

        let kernel_name = format!("varlen_flash_attention_fwd_{head_dim}_{dtype_suffix}");
        let device = q.device();
        let device_index = device.id();

        let module =
            kernels::get_or_load_module(self.context(), device_index, VARLEN_ATTENTION_MODULE)?;
        let func = kernels::get_kernel_function(&module, &kernel_name)?;

        // Total tokens from Q shape: [total_tokens_q, num_heads, head_dim]
        let total_tokens_q = q.shape()[0];

        let output =
            Tensor::<CudaRuntime>::empty(&[total_tokens_q, num_heads, head_dim], dtype, device);
        let lse = Tensor::<CudaRuntime>::empty(&[total_tokens_q, num_heads], DType::F32, device);

        // Block sizes are chosen so that shared memory fits within 48 KB
        // (after set_smem_attribute opts in to larger dynamic smem).
        //   head_dim 64/128: BLOCK_M=128, BLOCK_N=64  (proven config)
        //   head_dim 256 fp32: BLOCK_M=16, BLOCK_N=16
        //     smem = (16 + 2*16) * 256 * 4 = 49 152 B = 48 KB
        //   head_dim 256 fp16: BLOCK_M=32, BLOCK_N=32
        //     smem = (32 + 2*32) * 256 * 2 = 49 152 B = 48 KB
        // block_dim == BLOCK_M (one thread per Q row) — invariant kept.
        let (block_m, block_n) = block_config(head_dim, dtype);

        // Grid must cover all batches × heads × Q blocks per batch
        let num_q_blocks_per_batch = max_seqlen_q.div_ceil(block_m);
        let num_q_blocks = num_q_blocks_per_batch * batch_size;

        let dtype_size = dtype.size_in_bytes();
        // smem layout: Q tile + K tile + V tile, each with +1 padding on the column
        // stride to eliminate bank conflicts (HEAD_STRIDE = HEAD_DIM + 1).
        let head_stride = head_dim + 1;
        let smem_size =
            (block_m * head_stride + block_n * head_stride + block_n * head_stride) * dtype_size;
        set_smem_attribute(&func, smem_size)?;

        let cfg = LaunchConfig {
            grid_dim: ((num_q_blocks * num_heads) as u32, 1, 1),
            block_dim: (block_m as u32, 1, 1),
            shared_mem_bytes: smem_size as u32,
        };

        let q_ptr = q.ptr();
        let k_ptr = k.ptr();
        let v_ptr = v.ptr();
        let cu_q_ptr = cu_seqlens_q.ptr();
        let cu_k_ptr = cu_seqlens_k.ptr();
        let o_ptr = output.ptr();
        let l_ptr = lse.ptr();
        let scale = (head_dim as f32).sqrt().recip();
        let batch_i32 = batch_size as i32;
        let nh_i32 = num_heads as i32;
        let nkv_i32 = num_kv_heads as i32;
        let msq_i32 = max_seqlen_q as i32;
        let msk_i32 = max_seqlen_k as i32;
        let causal_i32 = if causal { 1i32 } else { 0i32 };

        unsafe {
            let mut builder = self.stream().launch_builder(&func);
            builder.arg(&q_ptr);
            builder.arg(&k_ptr);
            builder.arg(&v_ptr);
            builder.arg(&cu_q_ptr);
            builder.arg(&cu_k_ptr);
            builder.arg(&o_ptr);
            builder.arg(&l_ptr);
            builder.arg(&batch_i32);
            builder.arg(&nh_i32);
            builder.arg(&nkv_i32);
            builder.arg(&msq_i32);
            builder.arg(&msk_i32);
            builder.arg(&scale);
            builder.arg(&causal_i32);
            builder.launch(cfg).map_err(|e| Error::KernelError {
                reason: format!("varlen attention fwd launch failed: {e:?}"),
            })?;
        }

        // No sync needed: same-stream ordering guarantees correctness.

        Ok((output, lse))
    }

    fn varlen_attention_bwd(
        &self,
        dout: &Tensor<CudaRuntime>,
        q: &Tensor<CudaRuntime>,
        k: &Tensor<CudaRuntime>,
        v: &Tensor<CudaRuntime>,
        output: &Tensor<CudaRuntime>,
        lse: &Tensor<CudaRuntime>,
        cu_seqlens_q: &Tensor<CudaRuntime>,
        cu_seqlens_k: &Tensor<CudaRuntime>,
        batch_size: usize,
        num_heads: usize,
        num_kv_heads: usize,
        max_seqlen_q: usize,
        max_seqlen_k: usize,
        head_dim: usize,
        causal: bool,
    ) -> Result<(
        Tensor<CudaRuntime>,
        Tensor<CudaRuntime>,
        Tensor<CudaRuntime>,
    )> {
        if head_dim != 64 && head_dim != 128 && head_dim != 256 {
            return Err(Error::KernelError {
                reason: format!(
                    "varlen attention bwd: unsupported head_dim {head_dim}, only 64/128/256"
                ),
            });
        }

        let dtype = q.dtype();
        let dtype_suffix = match dtype {
            DType::F32 => "fp32",
            DType::F16 => "fp16",
            _ => {
                return Err(Error::KernelError {
                    reason: format!("varlen attention bwd: unsupported dtype {dtype:?}"),
                });
            }
        };

        let kernel_name = format!("varlen_flash_attention_bwd_{head_dim}_{dtype_suffix}");
        let device = q.device();
        let device_index = device.id();

        // FP16 backward kernels live in their own compiled module (split out to
        // keep each .cu within the file-size budget); FP32 stays in the base module.
        let bwd_module = match dtype {
            DType::F16 => VARLEN_ATTENTION_BWD_FP16_MODULE,
            _ => VARLEN_ATTENTION_BWD_MODULE,
        };
        let module = kernels::get_or_load_module(self.context(), device_index, bwd_module)?;
        let func = kernels::get_kernel_function(&module, &kernel_name)?;

        let total_tokens_q = q.shape()[0];
        let total_tokens_k = k.shape()[0];

        // dq: same head layout as Q (num_heads)
        // dk/dv: kv head layout (num_kv_heads) — GQA: fewer heads than Q
        let dq =
            Tensor::<CudaRuntime>::zeros(&[total_tokens_q, num_heads, head_dim], dtype, device);
        let dk =
            Tensor::<CudaRuntime>::zeros(&[total_tokens_k, num_kv_heads, head_dim], dtype, device);
        let dv =
            Tensor::<CudaRuntime>::zeros(&[total_tokens_k, num_kv_heads, head_dim], dtype, device);

        // Use same block sizes as fwd (matches the kernel template instantiations)
        let (block_m, block_n) = block_config(head_dim, dtype);
        let num_q_blocks_per_batch = max_seqlen_q.div_ceil(block_m);
        let num_q_blocks = num_q_blocks_per_batch * batch_size;

        // Shared memory layout (with +1 HEAD_STRIDE padding, same as the bwd kernel):
        //   Q tile   : BLOCK_M * (HEAD_DIM+1) elements
        //   K tile   : BLOCK_N * (HEAD_DIM+1) elements
        //   V tile   : BLOCK_N * (HEAD_DIM+1) elements
        //   dO tile  : BLOCK_M * (HEAD_DIM+1) elements
        // Total: (2*BLOCK_M + 2*BLOCK_N) * HEAD_STRIDE * dtype_size bytes
        let dtype_size = dtype.size_in_bytes();
        let head_stride = head_dim + 1;
        let smem_size = (2 * block_m + 2 * block_n) * head_stride * dtype_size;
        set_smem_attribute(&func, smem_size)?;

        let cfg = LaunchConfig {
            grid_dim: ((num_q_blocks * num_heads) as u32, 1, 1),
            block_dim: (block_m as u32, 1, 1),
            shared_mem_bytes: smem_size as u32,
        };

        let q_ptr = q.ptr();
        let k_ptr = k.ptr();
        let v_ptr = v.ptr();
        let o_ptr = output.ptr();
        let l_ptr = lse.ptr();
        let do_ptr = dout.ptr();
        let cu_q_ptr = cu_seqlens_q.ptr();
        let cu_k_ptr = cu_seqlens_k.ptr();
        let dq_ptr = dq.ptr();
        let dk_ptr = dk.ptr();
        let dv_ptr = dv.ptr();
        let scale = (head_dim as f32).sqrt().recip();
        let batch_i32 = batch_size as i32;
        let nh_i32 = num_heads as i32;
        let nkv_i32 = num_kv_heads as i32;
        let msq_i32 = max_seqlen_q as i32;
        let msk_i32 = max_seqlen_k as i32;
        let causal_i32 = if causal { 1i32 } else { 0i32 };

        unsafe {
            let mut builder = self.stream().launch_builder(&func);
            builder.arg(&q_ptr);
            builder.arg(&k_ptr);
            builder.arg(&v_ptr);
            builder.arg(&o_ptr);
            builder.arg(&l_ptr);
            builder.arg(&do_ptr);
            builder.arg(&cu_q_ptr);
            builder.arg(&cu_k_ptr);
            builder.arg(&dq_ptr);
            builder.arg(&dk_ptr);
            builder.arg(&dv_ptr);
            builder.arg(&batch_i32);
            builder.arg(&nh_i32);
            builder.arg(&nkv_i32);
            builder.arg(&msq_i32);
            builder.arg(&msk_i32);
            builder.arg(&scale);
            builder.arg(&causal_i32);
            builder.launch(cfg).map_err(|e| Error::KernelError {
                reason: format!("varlen attention bwd launch failed: {e:?}"),
            })?;
        }

        // Sync stream: BWD uses atomicAdd so must complete before results are read
        self.stream()
            .synchronize()
            .map_err(|e| Error::KernelError {
                reason: format!("varlen attention bwd sync failed: {e:?}"),
            })?;

        Ok((dq, dk, dv))
    }
}