@tduncan / Nvidia-Optimization-Stack-VLLM
Code Commits Issues Pulls Cloud Agent CI/CD
main
main
py 190 lines 5.27 KB
Raw
1 # SPDX-License-Identifier: Apache-2.0
2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3 # SPDX-FileCopyrightText: Songlin Yang, Yu Zhang
4 #
5 # This file contains code copied from the flash-linear-attention project.
6 # The original source code was licensed under the MIT license and included
7 # the following copyright notice:
8 # Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
9
10 # ruff: noqa: E501
11
12
13 import torch
14
15 from vllm.triton_utils import tl, triton
16
17 from .index import prepare_chunk_indices
18 from .op import exp
19 from .utils import FLA_CHUNK_SIZE, check_shared_mem, is_nvidia_hopper
20
21 BKV_LIST = [64, 128] if check_shared_mem() else [32, 64]
22 NUM_WARPS = [2, 4] if is_nvidia_hopper else [2, 4, 8]
23
24
25 @triton.heuristics(
26 {
27 "USE_G": lambda args: args["g"] is not None,
28 "IS_VARLEN": lambda args: args["cu_seqlens"] is not None,
29 }
30 )
31 @triton.autotune(
32 configs=[
33 triton.Config({"BK": BK, "BV": BV}, num_warps=num_warps, num_stages=num_stages)
34 for BK in BKV_LIST
35 for BV in BKV_LIST
36 for num_warps in NUM_WARPS
37 for num_stages in [2, 3, 4]
38 ],
39 key=["H", "K", "V", "BT"],
40 )
41 @triton.jit(do_not_specialize=["T"])
42 def chunk_fwd_kernel_o(
43 q,
44 k,
45 v,
46 h,
47 g,
48 o,
49 cu_seqlens,
50 chunk_indices,
51 scale,
52 T,
53 H: tl.constexpr,
54 Hg: tl.constexpr,
55 K: tl.constexpr,
56 V: tl.constexpr,
57 BT: tl.constexpr,
58 BK: tl.constexpr,
59 BV: tl.constexpr,
60 USE_G: tl.constexpr,
61 IS_VARLEN: tl.constexpr,
62 ):
63 i_v, i_t, i_bh = tl.program_id(0), tl.program_id(1), tl.program_id(2)
64 i_b, i_h = i_bh // H, i_bh % H
65
66 if IS_VARLEN:
67 i_tg = i_t
68 i_n, i_t = (
69 tl.load(chunk_indices + i_t * 2).to(tl.int32),
70 tl.load(chunk_indices + i_t * 2 + 1).to(tl.int32),
71 )
72 bos, eos = (
73 tl.load(cu_seqlens + i_n).to(tl.int32),
74 tl.load(cu_seqlens + i_n + 1).to(tl.int32),
75 )
76 T = eos - bos
77 NT = tl.cdiv(T, BT)
78 else:
79 NT = tl.cdiv(T, BT)
80 i_tg = i_b * NT + i_t
81 bos, eos = i_b * T, i_b * T + T
82
83 # offset calculation
84 q += (bos * Hg + i_h // (H // Hg)) * K
85 k += (bos * Hg + i_h // (H // Hg)) * K
86 v += (bos * H + i_h) * V
87 o += (bos * H + i_h) * V
88 h += (i_tg * H + i_h).to(tl.int64) * V * K
89
90 b_o = tl.zeros([BT, BV], dtype=tl.float32)
91 b_A = tl.zeros([BT, BT], dtype=tl.float32)
92
93 for i_k in range(tl.cdiv(K, BK)):
94 p_q = tl.make_block_ptr(
95 q, (T, K), (Hg * K, 1), (i_t * BT, i_k * BK), (BT, BK), (1, 0)
96 )
97 p_k = tl.make_block_ptr(
98 k, (K, T), (1, Hg * K), (i_k * BK, i_t * BT), (BK, BT), (0, 1)
99 )
100 p_h = tl.make_block_ptr(
101 h, (V, K), (K, 1), (i_v * BV, i_k * BK), (BV, BK), (1, 0)
102 )
103 # [BT, BK]
104 b_q = tl.load(p_q, boundary_check=(0, 1))
105 # [BK, BT]
106 b_k = tl.load(p_k, boundary_check=(0, 1))
107 # [BV, BK]
108 b_h = tl.load(p_h, boundary_check=(0, 1))
109
110 # [BT, BK] @ [BK, BV] -> [BT, BV]
111 b_o += tl.dot(b_q, tl.trans(b_h))
112 # [BT, BK] @ [BK, BT] -> [BT, BT]
113 b_A += tl.dot(b_q, b_k)
114
115 if USE_G:
116 g += bos * H + i_h
117 p_g = tl.make_block_ptr(g, (T,), (H,), (i_t * BT,), (BT,), (0,))
118 b_g = tl.load(p_g, boundary_check=(0,))
119 b_o = b_o * exp(b_g)[:, None]
120 b_A = b_A * exp(b_g[:, None] - b_g[None, :])
121
122 o_t = i_t * BT + tl.arange(0, BT)
123 m_t = o_t < T
124 m_A = (o_t[:, None] >= o_t[None, :]) & (m_t[:, None] & m_t)
125 b_A = tl.where(m_A, b_A, 0)
126
127 p_v = tl.make_block_ptr(
128 v, (T, V), (H * V, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0)
129 )
130 p_o = tl.make_block_ptr(
131 o, (T, V), (H * V, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0)
132 )
133 b_v = tl.load(p_v, boundary_check=(0, 1))
134
135 # to fix mma -> mma layout conversion
136 # already solved by triton v3.2 or higher
137 b_o = b_o * scale + tl.dot(b_A.to(b_v.dtype), b_v) * scale
138 tl.store(p_o, b_o.to(p_o.dtype.element_ty), boundary_check=(0, 1))
139
140
141 def chunk_fwd_o(
142 q: torch.Tensor,
143 k: torch.Tensor,
144 v: torch.Tensor,
145 h: torch.Tensor,
146 g: torch.Tensor | None = None, # cumsum of log decay
147 scale: float | None = None,
148 cu_seqlens: torch.Tensor | None = None,
149 chunk_indices: torch.Tensor | None = None,
150 chunk_size: int = FLA_CHUNK_SIZE,
151 core_attn_out: torch.Tensor | None = None,
152 ) -> torch.Tensor:
153 B, T, Hg, K, V = *q.shape, v.shape[-1]
154 H = v.shape[-2]
155 BT = chunk_size
156 if chunk_indices is None and cu_seqlens is not None:
157 chunk_indices = prepare_chunk_indices(cu_seqlens, BT)
158 NT = triton.cdiv(T, BT) if cu_seqlens is None else len(chunk_indices)
159 if scale is None:
160 scale = k.shape[-1] ** -0.5
161
162 if core_attn_out is not None:
163 assert core_attn_out.numel() >= v.numel(), (
164 f"core_attn_out too small: {core_attn_out.numel()} < {v.numel()}"
165 )
166 o = core_attn_out[: v.numel()].view(*v.shape)
167 else:
168 o = torch.empty_like(v)
169
170 def grid(meta):
171 return (triton.cdiv(V, meta["BV"]), NT, B * H)
172
173 chunk_fwd_kernel_o[grid](
174 q,
175 k,
176 v,
177 h,
178 g,
179 o,
180 cu_seqlens,
181 chunk_indices,
182 scale,
183 T=T,
184 H=H,
185 Hg=Hg,
186 K=K,
187 V=V,
188 BT=BT,
189 )
190 return o