Coverage for src/flag_gems/runtime/backend/_ascend/fla/chunk_delta_h.py: 0%
112 statements
« prev ^ index » next coverage.py v7.6.9, created at 2026-03-21 14:31 +0800
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# ruff: noqa: E501
10# mypy: ignore-errors
11import torch
12import triton
13import triton.language as tl
15from .utils import prepare_chunk_indices, prepare_chunk_offsets, safe_exp
17_CONDITIONS = ("seq7168",)
20@triton.heuristics(
21 {
22 "USE_G": lambda args: args["g"] is not None,
23 "USE_INITIAL_STATE": lambda args: args["h0"] is not None,
24 "STORE_FINAL_STATE": lambda args: args["ht"] is not None,
25 "SAVE_NEW_VALUE": lambda args: args["v_new"] is not None,
26 "IS_VARLEN": lambda args: args["cu_seqlens"] is not None,
27 }
28)
29@triton.jit(do_not_specialize=["T"])
30def chunk_gated_delta_rule_fwd_kernel_h_blockdim64(
31 k,
32 v,
33 w,
34 v_new,
35 g,
36 h,
37 h0,
38 ht,
39 cu_seqlens,
40 chunk_offsets,
41 T,
42 H: tl.constexpr,
43 Hg: tl.constexpr,
44 K: tl.constexpr,
45 V: tl.constexpr,
46 BT: tl.constexpr,
47 USE_G: tl.constexpr,
48 USE_INITIAL_STATE: tl.constexpr,
49 STORE_FINAL_STATE: tl.constexpr,
50 SAVE_NEW_VALUE: tl.constexpr,
51 IS_VARLEN: tl.constexpr,
52):
53 i_nh = tl.program_id(1)
54 i_n, i_h = i_nh // H, i_nh % H
55 T_max = 1 * T
56 if IS_VARLEN:
57 bos, eos = (
58 tl.load(cu_seqlens + i_n).to(tl.int32),
59 tl.load(cu_seqlens + i_n + 1).to(tl.int32),
60 )
61 T = eos - bos
62 NT = tl.cdiv(T, BT)
63 boh = tl.load(chunk_offsets + i_n).to(tl.int32)
64 else:
65 bos, eos = i_n * T, i_n * T + T
66 NT = tl.cdiv(T, BT)
67 boh = i_n * NT
69 stride_v = H * V
70 stride_k = Hg * K
71 stride_w = H * K
73 b_h1_bv1 = tl.zeros([128, 64], dtype=tl.float32)
74 b_h1_bv2 = tl.zeros([128, 64], dtype=tl.float32)
76 v_start1 = 0
77 v_start2 = 64
79 offs_k = tl.arange(0, 128)[:, None]
80 offs_v1 = v_start1 + tl.arange(0, 64)[None, :]
81 offs_v2 = v_start2 + tl.arange(0, 64)[None, :]
82 mask_kv1 = (offs_k < K) & (offs_v1 < V)
83 mask_kv2 = (offs_k < K) & (offs_v2 < V)
85 # load initial state
86 if USE_INITIAL_STATE:
87 h0_ptr = h0 + i_nh * K * V
88 ptr_h0_bv1 = h0_ptr + offs_k * V + offs_v1 * 1
89 b_h1_bv1 += tl.load(ptr_h0_bv1, mask=mask_kv1, other=0.0).to(tl.float32)
91 ptr_h0_bv2 = h0_ptr + offs_k * V + offs_v2 * 1
92 b_h1_bv2 += tl.load(ptr_h0_bv2, mask=mask_kv2, other=0.0).to(tl.float32)
94 # main recurrence
95 for i_t in range(NT):
96 h_base = h + (boh + i_t) * H * K * V + i_h * K * V
98 p_h1_bv1 = tl.make_block_ptr(
99 h_base, (K, V), (V, 1), (0, v_start1), (128, 64), (1, 0)
100 )
101 tl.store(
102 p_h1_bv1, b_h1_bv1.to(p_h1_bv1.dtype.element_ty), boundary_check=(0, 1)
103 )
105 p_h1_bv2 = tl.make_block_ptr(
106 h_base, (K, V), (V, 1), (0, v_start2), (128, 64), (1, 0)
107 )
108 tl.store(
109 p_h1_bv2, b_h1_bv2.to(p_h1_bv2.dtype.element_ty), boundary_check=(0, 1)
110 )
112 offs_t_wv = (i_t * BT + tl.arange(0, BT))[:, None]
113 offs_k_wv = tl.arange(0, 128)[None, :]
114 mask_w = (offs_t_wv < T) & (offs_k_wv < K)
116 w_base = w + bos * H * K + i_h * K
117 ptr_w = w_base + offs_t_wv * stride_w + offs_k_wv * 1
118 b_w = tl.load(ptr_w, mask=mask_w, other=0.0)
120 k_base = k + bos * Hg * K + (i_h // (H // Hg)) * K
121 p_k = tl.make_block_ptr(
122 k_base, (K, T), (1, stride_k), (0, i_t * BT), (128, BT), (0, 1)
123 )
124 b_k = tl.load(p_k, boundary_check=(0, 1))
126 v_new_base = v_new + bos * H * V + i_h * V
128 last_idx = min((i_t + 1) * BT, T) - 1
129 b_g_last = tl.load(g + bos + i_h * T_max + last_idx)
131 offs_t = i_t * BT + tl.arange(0, BT)
132 mask_t = offs_t < T
133 g_ptr = g + bos + i_h * T_max
134 b_g = tl.load(g_ptr + offs_t, mask=mask_t, other=0.0)
136 b_g = safe_exp(b_g_last - b_g)
137 b_g_last = tl.exp(b_g_last)
139 offs_t_v = (i_t * BT + tl.arange(0, BT))[:, None]
140 mask_v1 = (offs_t_v < T) & (offs_v1 < V)
142 v_base = v + bos * H * V + i_h * V
143 ptr_v1 = v_base + offs_t_v * stride_v + offs_v1 * 1
144 b_v1 = tl.load(ptr_v1, mask=mask_v1, other=0.0)
145 b_v_new1 = b_v1.to(tl.float32)
146 b_v_new1 -= tl.dot(b_w, b_h1_bv1.to(b_w.dtype))
148 if SAVE_NEW_VALUE:
149 p_v_new1 = tl.make_block_ptr(
150 v_new_base,
151 (T, V),
152 (stride_v, 1),
153 (i_t * BT, v_start1),
154 (BT, 64),
155 (1, 0),
156 )
157 tl.store(
158 p_v_new1, b_v_new1.to(p_v_new1.dtype.element_ty), boundary_check=(0, 1)
159 )
161 if USE_G:
162 b_v_new1 = b_v_new1 * b_g[:, None]
163 b_h1_bv1 = b_h1_bv1 * b_g_last
165 b_v_new1 = b_v_new1.to(k.dtype.element_ty)
166 b_h1_bv1 += tl.dot(b_k, b_v_new1)
168 mask_v2 = (offs_t_v < T) & (offs_v2 < V)
169 ptr_v2 = v_base + offs_t_v * stride_v + offs_v2 * 1
170 b_v2 = tl.load(ptr_v2, mask=mask_v2, other=0.0)
171 b_v_new2 = b_v2.to(tl.float32)
172 b_v_new2 -= tl.dot(b_w, b_h1_bv2.to(b_w.dtype))
174 if SAVE_NEW_VALUE:
175 p_v_new2 = tl.make_block_ptr(
176 v_new_base,
177 (T, V),
178 (stride_v, 1),
179 (i_t * BT, v_start2),
180 (BT, 64),
181 (1, 0),
182 )
183 tl.store(
184 p_v_new2, b_v_new2.to(p_v_new2.dtype.element_ty), boundary_check=(0, 1)
185 )
187 if USE_G:
188 b_v_new2 = b_v_new2 * b_g[:, None]
189 b_h1_bv2 = b_h1_bv2 * b_g_last
191 b_v_new2 = b_v_new2.to(k.dtype.element_ty)
192 b_h1_bv2 += tl.dot(b_k, b_v_new2)
194 # epilogue
195 if STORE_FINAL_STATE:
196 ht_ptr = ht + i_nh * K * V
198 p_ht1_bv1 = tl.make_block_ptr(
199 ht_ptr, (K, V), (V, 1), (0, v_start1), (128, 64), (1, 0)
200 )
201 tl.store(
202 p_ht1_bv1, b_h1_bv1.to(p_ht1_bv1.dtype.element_ty), boundary_check=(0, 1)
203 )
205 p_ht1_bv2 = tl.make_block_ptr(
206 ht_ptr, (K, V), (V, 1), (0, v_start2), (128, 64), (1, 0)
207 )
208 tl.store(
209 p_ht1_bv2, b_h1_bv2.to(p_ht1_bv2.dtype.element_ty), boundary_check=(0, 1)
210 )
213def chunk_gated_delta_rule_fwd_h(
214 k: torch.Tensor,
215 w: torch.Tensor,
216 u: torch.Tensor,
217 g: torch.Tensor | None = None,
218 initial_state: torch.Tensor | None = None,
219 output_final_state: bool = False,
220 chunk_size: int = 64, # SY: remove this argument and force chunk size 64?
221 save_new_value: bool = True,
222 cu_seqlens: torch.LongTensor | None = None,
223) -> tuple[torch.Tensor, torch.Tensor]:
224 # This kernel is slightly different from fla to support Q/K with different head numbers.
225 # In fla, Q/K always have the same head number, so Hg is always equal to H.
226 B, T, Hg, K, V = *k.shape, u.shape[-1]
227 H = u.shape[-2]
228 BT = chunk_size
230 chunk_indices = (
231 prepare_chunk_indices(cu_seqlens, chunk_size)
232 if cu_seqlens is not None
233 else None
234 )
235 # N: the actual number of sequences in the batch with either equal or variable lengths
236 if cu_seqlens is None:
237 N, NT, chunk_offsets = B, triton.cdiv(T, BT), None
238 else:
239 N, NT, chunk_offsets = (
240 len(cu_seqlens) - 1,
241 len(chunk_indices),
242 prepare_chunk_offsets(cu_seqlens, BT),
243 )
244 assert K <= 256, "current kernel does not support head dimension larger than 256."
246 h = k.new_empty(B, NT, H, K, V)
247 final_state = (
248 k.new_empty(N, H, K, V, dtype=torch.float32) if output_final_state else None
249 )
251 v_new = torch.empty_like(u) if save_new_value else None
252 g = g.transpose(1, 2).contiguous()
254 def grid(meta):
255 return (1, N * H)
257 chunk_gated_delta_rule_fwd_kernel_h_blockdim64[grid](
258 k=k,
259 v=u,
260 w=w,
261 v_new=v_new,
262 g=g,
263 h=h,
264 h0=initial_state,
265 ht=final_state,
266 cu_seqlens=cu_seqlens,
267 chunk_offsets=chunk_offsets,
268 T=T,
269 H=H,
270 Hg=Hg,
271 K=K,
272 V=V,
273 BT=BT,
274 num_warps=4,
275 num_stages=2,
276 )
277 return h, v_new, final_state