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

//! CUDA Paged Attention — vLLM-style non-contiguous KV cache
//!
//! Fused kernel — PRIMITIVE op. Block table indirection for KV blocks.
//! Supports: F32, F16, BF16, FP8E4M3, FP8E5M2. Head dimensions: 64, 128.
//!
//! Implementation is split across:
//! - `paged_attention_fwd.rs`: forward (standard + FP8)
//! - `paged_attention_bwd.rs`: backward
//! - `paged_decode.rs`: S_q=1 decode fast path

use crate::error::{Error, Result};
use crate::ops::traits::PagedAttentionOps;
use numr::dtype::DType;
use numr::runtime::cuda::{CudaClient, CudaRuntime};
use numr::tensor::Tensor;

use super::paged_attention_bwd::paged_attention_bwd_impl;
use super::paged_attention_fwd::{paged_attention_fwd_fp8_impl, paged_attention_fwd_impl};

/// Get block sizes for paged attention forward.
/// Uses smaller blocks that fit in 48KB shared memory.
pub(super) fn fwd_block_config(head_dim: usize, dtype: DType) -> Result<(usize, usize)> {
    match (dtype, head_dim) {
        // FP32: 4 bytes per element
        (DType::F32, 64) => Ok((64, 32)),  // (64+32+32)*64*4 = 32KB
        (DType::F32, 128) => Ok((32, 32)), // (32+32+32)*128*4 = 48KB
        // FP16/BF16: 2 bytes per element
        (DType::F16 | DType::BF16, 64) => Ok((64, 32)), // (64+32+32)*64*2 = 16KB
        (DType::F16 | DType::BF16, 128) => Ok((32, 32)), // (32+32+32)*128*2 = 24KB
        // FP8: uses FP32 smem for compute
        (DType::FP8E4M3 | DType::FP8E5M2, 64) => Ok((64, 32)), // (64+64)*64*4 = 32KB
        (DType::FP8E4M3 | DType::FP8E5M2, 128) => Ok((32, 32)), // (32+64)*128*4 = 49KB
        _ => Err(Error::InvalidArgument {
            arg: "head_dim",
            reason: format!(
                "unsupported head_dim={} for paged attention. Supported: 64, 128",
                head_dim
            ),
        }),
    }
}

/// Get block sizes for paged attention backward.
/// Backward needs more smem: Q + K + V + dO + O = (3*BLOCK_M + 2*BLOCK_N) * HD * dtype_size
pub(super) fn bwd_block_config(head_dim: usize, dtype: DType) -> Result<(usize, usize)> {
    match (dtype, head_dim) {
        (DType::F32, 64) => Ok((32, 32)),  // (96+64)*64*4 = 40KB
        (DType::F32, 128) => Ok((16, 16)), // (48+32)*128*4 = 40KB
        (DType::F16 | DType::BF16, 64) => Ok((64, 32)), // (192+64)*64*2 = 32KB
        (DType::F16 | DType::BF16, 128) => Ok((32, 32)), // (96+64)*128*2 = 40KB
        _ => Err(Error::InvalidArgument {
            arg: "head_dim",
            reason: format!(
                "unsupported head_dim={} for paged attention backward",
                head_dim
            ),
        }),
    }
}

impl PagedAttentionOps<CudaRuntime> for CudaClient {
    fn paged_attention_fwd(
        &self,
        q: &Tensor<CudaRuntime>,
        k_blocks: &Tensor<CudaRuntime>,
        v_blocks: &Tensor<CudaRuntime>,
        block_table: &Tensor<CudaRuntime>,
        num_heads: usize,
        num_kv_heads: usize,
        seq_len_q: usize,
        seq_len_k: usize,
        head_dim: usize,
        block_size: usize,
        causal: bool,
    ) -> Result<(Tensor<CudaRuntime>, Tensor<CudaRuntime>)> {
        paged_attention_fwd_impl(
            self,
            q,
            k_blocks,
            v_blocks,
            block_table,
            num_heads,
            num_kv_heads,
            seq_len_q,
            seq_len_k,
            head_dim,
            block_size,
            causal,
        )
    }

    fn paged_attention_fwd_fp8(
        &self,
        q: &Tensor<CudaRuntime>,
        k_blocks: &Tensor<CudaRuntime>,
        v_blocks: &Tensor<CudaRuntime>,
        block_table: &Tensor<CudaRuntime>,
        num_heads: usize,
        num_kv_heads: usize,
        seq_len_q: usize,
        seq_len_k: usize,
        head_dim: usize,
        block_size: usize,
        causal: bool,
        q_scale: f32,
        k_scale: f32,
        v_scale: f32,
        o_scale: f32,
    ) -> Result<(Tensor<CudaRuntime>, Tensor<CudaRuntime>)> {
        paged_attention_fwd_fp8_impl(
            self,
            q,
            k_blocks,
            v_blocks,
            block_table,
            num_heads,
            num_kv_heads,
            seq_len_q,
            seq_len_k,
            head_dim,
            block_size,
            causal,
            q_scale,
            k_scale,
            v_scale,
            o_scale,
        )
    }

    fn paged_attention_bwd(
        &self,
        dout: &Tensor<CudaRuntime>,
        q: &Tensor<CudaRuntime>,
        k_blocks: &Tensor<CudaRuntime>,
        v_blocks: &Tensor<CudaRuntime>,
        output: &Tensor<CudaRuntime>,
        lse: &Tensor<CudaRuntime>,
        block_table: &Tensor<CudaRuntime>,
        num_heads: usize,
        num_kv_heads: usize,
        seq_len_q: usize,
        seq_len_k: usize,
        head_dim: usize,
        block_size: usize,
        causal: bool,
    ) -> Result<(
        Tensor<CudaRuntime>,
        Tensor<CudaRuntime>,
        Tensor<CudaRuntime>,
    )> {
        paged_attention_bwd_impl(
            self,
            dout,
            q,
            k_blocks,
            v_blocks,
            output,
            lse,
            block_table,
            num_heads,
            num_kv_heads,
            seq_len_q,
            seq_len_k,
            head_dim,
            block_size,
            causal,
        )
    }
}