Spaces:
Sleeping
Sleeping
File size: 11,693 Bytes
e45d058 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 |
# [2022-10-23] Downloaded from https://github.com/openai/triton/blob/master/python/tutorials/06-fused-attention.py
# for benchmarking.
# We fixed a few dtype cast to make it work for bf16
"""
Fused Attention
===============
This is a Triton implementation of the Flash Attention algorithm
(see: Dao et al., https://arxiv.org/pdf/2205.14135v2.pdf; Rabe and Staats https://arxiv.org/pdf/2112.05682v2.pdf)
"""
import pytest
import torch
import triton
import triton.language as tl
@triton.jit
def _fwd_kernel(
Q,
K,
V,
sm_scale,
TMP,
L,
M, # NOTE: TMP is a scratchpad buffer to workaround a compiler bug
Out,
stride_qz,
stride_qh,
stride_qm,
stride_qk,
stride_kz,
stride_kh,
stride_kn,
stride_kk,
stride_vz,
stride_vh,
stride_vk,
stride_vn,
stride_oz,
stride_oh,
stride_om,
stride_on,
Z,
H,
N_CTX,
BLOCK_M: tl.constexpr,
BLOCK_DMODEL: tl.constexpr,
BLOCK_N: tl.constexpr,
):
start_m = tl.program_id(0)
off_hz = tl.program_id(1)
# initialize offsets
offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
offs_n = tl.arange(0, BLOCK_N)
offs_d = tl.arange(0, BLOCK_DMODEL)
off_q = off_hz * stride_qh + offs_m[:, None] * stride_qm + offs_d[None, :] * stride_qk
off_k = off_hz * stride_qh + offs_n[:, None] * stride_kn + offs_d[None, :] * stride_kk
off_v = off_hz * stride_qh + offs_n[:, None] * stride_qm + offs_d[None, :] * stride_qk
# Initialize pointers to Q, K, V
q_ptrs = Q + off_q
k_ptrs = K + off_k
v_ptrs = V + off_v
# initialize pointer to m and l
t_ptrs = TMP + off_hz * N_CTX + offs_m
m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
l_i = tl.zeros([BLOCK_M], dtype=tl.float32)
acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
# load q: it will stay in SRAM throughout
q = tl.load(q_ptrs)
# loop over k, v and update accumulator
for start_n in range(0, (start_m + 1) * BLOCK_M, BLOCK_N):
start_n = tl.multiple_of(start_n, BLOCK_N)
# -- compute qk ----
k = tl.load(k_ptrs + start_n * stride_kn)
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
qk += tl.dot(q, k, trans_b=True)
qk *= sm_scale
qk += tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), 0, float("-inf"))
# -- compute m_ij, p, l_ij
m_ij = tl.max(qk, 1)
p = tl.exp(qk - m_ij[:, None])
l_ij = tl.sum(p, 1)
# -- update m_i and l_i
m_i_new = tl.maximum(m_i, m_ij)
alpha = tl.exp(m_i - m_i_new)
beta = tl.exp(m_ij - m_i_new)
l_i_new = alpha * l_i + beta * l_ij
# -- update output accumulator --
# scale p
p_scale = beta / l_i_new
p = p * p_scale[:, None]
# scale acc
acc_scale = l_i / l_i_new * alpha
tl.store(t_ptrs, acc_scale)
acc_scale = tl.load(t_ptrs) # BUG: have to store and immediately load
acc = acc * acc_scale[:, None]
# update acc
v = tl.load(v_ptrs + start_n * stride_vk)
p = p.to(v.dtype)
acc += tl.dot(p, v)
# update m_i and l_i
l_i = l_i_new
m_i = m_i_new
# rematerialize offsets to save registers
start_m = tl.program_id(0)
offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
# write back l and m
l_ptrs = L + off_hz * N_CTX + offs_m
m_ptrs = M + off_hz * N_CTX + offs_m
tl.store(l_ptrs, l_i)
tl.store(m_ptrs, m_i)
# initialize pointers to output
offs_n = tl.arange(0, BLOCK_DMODEL)
off_o = off_hz * stride_oh + offs_m[:, None] * stride_om + offs_n[None, :] * stride_on
out_ptrs = Out + off_o
tl.store(out_ptrs, acc)
@triton.jit
def _bwd_preprocess(
Out,
DO,
L,
NewDO,
Delta,
BLOCK_M: tl.constexpr,
D_HEAD: tl.constexpr,
):
off_m = tl.program_id(0) * BLOCK_M + tl.arange(0, BLOCK_M)
off_n = tl.arange(0, D_HEAD)
# load
o = tl.load(Out + off_m[:, None] * D_HEAD + off_n[None, :]).to(tl.float32)
do = tl.load(DO + off_m[:, None] * D_HEAD + off_n[None, :]).to(tl.float32)
denom = tl.load(L + off_m).to(tl.float32)
# compute
do = do / denom[:, None]
delta = tl.sum(o * do, axis=1)
# write-back
tl.store(NewDO + off_m[:, None] * D_HEAD + off_n[None, :], do)
tl.store(Delta + off_m, delta)
@triton.jit
def _bwd_kernel(
Q,
K,
V,
sm_scale,
Out,
DO,
DQ,
DK,
DV,
L,
M,
D,
stride_qz,
stride_qh,
stride_qm,
stride_qk,
stride_kz,
stride_kh,
stride_kn,
stride_kk,
stride_vz,
stride_vh,
stride_vk,
stride_vn,
Z,
H,
N_CTX,
num_block,
BLOCK_M: tl.constexpr,
BLOCK_DMODEL: tl.constexpr,
BLOCK_N: tl.constexpr,
):
off_hz = tl.program_id(0)
off_z = off_hz // H
off_h = off_hz % H
# offset pointers for batch/head
Q += off_z * stride_qz + off_h * stride_qh
K += off_z * stride_qz + off_h * stride_qh
V += off_z * stride_qz + off_h * stride_qh
DO += off_z * stride_qz + off_h * stride_qh
DQ += off_z * stride_qz + off_h * stride_qh
DK += off_z * stride_qz + off_h * stride_qh
DV += off_z * stride_qz + off_h * stride_qh
for start_n in range(0, num_block):
lo = start_n * BLOCK_M
# initialize row/col offsets
offs_qm = lo + tl.arange(0, BLOCK_M)
offs_n = start_n * BLOCK_M + tl.arange(0, BLOCK_M)
offs_m = tl.arange(0, BLOCK_N)
offs_k = tl.arange(0, BLOCK_DMODEL)
# initialize pointers to value-like data
q_ptrs = Q + (offs_qm[:, None] * stride_qm + offs_k[None, :] * stride_qk)
k_ptrs = K + (offs_n[:, None] * stride_kn + offs_k[None, :] * stride_kk)
v_ptrs = V + (offs_n[:, None] * stride_qm + offs_k[None, :] * stride_qk)
do_ptrs = DO + (offs_qm[:, None] * stride_qm + offs_k[None, :] * stride_qk)
dq_ptrs = DQ + (offs_qm[:, None] * stride_qm + offs_k[None, :] * stride_qk)
# pointer to row-wise quantities in value-like data
D_ptrs = D + off_hz * N_CTX
m_ptrs = M + off_hz * N_CTX
# initialize dv amd dk
dv = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
dk = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
# k and v stay in SRAM throughout
k = tl.load(k_ptrs)
v = tl.load(v_ptrs)
# loop over rows
for start_m in range(lo, num_block * BLOCK_M, BLOCK_M):
offs_m_curr = start_m + offs_m
# load q, k, v, do on-chip
q = tl.load(q_ptrs)
# recompute p = softmax(qk, dim=-1).T
# NOTE: `do` is pre-divided by `l`; no normalization here
qk = tl.dot(q, k, trans_b=True)
qk = tl.where(offs_m_curr[:, None] >= (offs_n[None, :]), qk, float("-inf"))
m = tl.load(m_ptrs + offs_m_curr)
p = tl.exp(qk * sm_scale - m[:, None])
# compute dv
do = tl.load(do_ptrs)
dv += tl.dot(p.to(do.dtype), do, trans_a=True)
# compute dp = dot(v, do)
Di = tl.load(D_ptrs + offs_m_curr)
dp = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32) - Di[:, None]
dp += tl.dot(do, v, trans_b=True)
# compute ds = p * (dp - delta[:, None])
ds = p * dp * sm_scale
# compute dk = dot(ds.T, q)
dk += tl.dot(ds.to(q.dtype), q, trans_a=True)
# # compute dq
dq = tl.load(dq_ptrs, eviction_policy="evict_last")
dq += tl.dot(ds.to(k.dtype), k)
tl.store(dq_ptrs, dq, eviction_policy="evict_last")
# # increment pointers
dq_ptrs += BLOCK_M * stride_qm
q_ptrs += BLOCK_M * stride_qm
do_ptrs += BLOCK_M * stride_qm
# write-back
dv_ptrs = DV + (offs_n[:, None] * stride_qm + offs_k[None, :] * stride_qk)
dk_ptrs = DK + (offs_n[:, None] * stride_kn + offs_k[None, :] * stride_kk)
tl.store(dv_ptrs, dv)
tl.store(dk_ptrs, dk)
class _attention(torch.autograd.Function):
@staticmethod
def forward(ctx, q, k, v, sm_scale):
BLOCK = 128
# shape constraints
Lq, Lk, Lv = q.shape[-1], k.shape[-1], v.shape[-1]
assert Lq == Lk and Lk == Lv
assert Lk in {16, 32, 64, 128}
o = torch.empty_like(q)
grid = (triton.cdiv(q.shape[2], BLOCK), q.shape[0] * q.shape[1])
tmp = torch.empty(
(q.shape[0] * q.shape[1], q.shape[2]), device=q.device, dtype=torch.float32
)
L = torch.empty((q.shape[0] * q.shape[1], q.shape[2]), device=q.device, dtype=torch.float32)
m = torch.empty((q.shape[0] * q.shape[1], q.shape[2]), device=q.device, dtype=torch.float32)
num_warps = 4 if Lk <= 64 else 8
_fwd_kernel[grid](
q,
k,
v,
sm_scale,
tmp,
L,
m,
o,
q.stride(0),
q.stride(1),
q.stride(2),
q.stride(3),
k.stride(0),
k.stride(1),
k.stride(2),
k.stride(3),
v.stride(0),
v.stride(1),
v.stride(2),
v.stride(3),
o.stride(0),
o.stride(1),
o.stride(2),
o.stride(3),
q.shape[0],
q.shape[1],
q.shape[2],
BLOCK_M=BLOCK,
BLOCK_N=BLOCK,
BLOCK_DMODEL=Lk,
num_warps=num_warps,
num_stages=1,
)
ctx.save_for_backward(q, k, v, o, L, m)
ctx.BLOCK = BLOCK
ctx.grid = grid
ctx.sm_scale = sm_scale
ctx.BLOCK_DMODEL = Lk
return o
@staticmethod
def backward(ctx, do):
q, k, v, o, l, m = ctx.saved_tensors
do = do.contiguous()
dq = torch.zeros_like(q, dtype=torch.float32)
dk = torch.empty_like(k)
dv = torch.empty_like(v)
do_scaled = torch.empty_like(do)
delta = torch.empty_like(l)
_bwd_preprocess[(ctx.grid[0] * ctx.grid[1],)](
o,
do,
l,
do_scaled,
delta,
BLOCK_M=ctx.BLOCK,
D_HEAD=ctx.BLOCK_DMODEL,
)
# NOTE: kernel currently buggy for other values of `num_warps`
num_warps = 8
_bwd_kernel[(ctx.grid[1],)](
q,
k,
v,
ctx.sm_scale,
o,
do_scaled,
dq,
dk,
dv,
l,
m,
delta,
q.stride(0),
q.stride(1),
q.stride(2),
q.stride(3),
k.stride(0),
k.stride(1),
k.stride(2),
k.stride(3),
v.stride(0),
v.stride(1),
v.stride(2),
v.stride(3),
q.shape[0],
q.shape[1],
q.shape[2],
ctx.grid[0],
BLOCK_M=ctx.BLOCK,
BLOCK_N=ctx.BLOCK,
BLOCK_DMODEL=ctx.BLOCK_DMODEL,
num_warps=num_warps,
num_stages=1,
)
return dq.to(q.dtype), dk, dv, None
attention = _attention.apply
|