File size: 7,849 Bytes
a3d804a |
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 |
# Copyright (c) 2023, Tri Dao.
# Copyright 2024 CATIE. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# Modifications to the orignal file
# - support for torch.compile
import triton
import triton.language as tl
import torch
import math
import torch.nn.functional as F
from torch.amp import custom_fwd, custom_bwd
@triton.jit
def _rmsnorm_fwd_kernel(
X, # pointer to the input
Y, # pointer to the output
W, # pointer to the weights
Rstd, # pointer to the 1/std
stride_x_row, # how much to increase the pointer when moving by 1 row
stride_y_row,
N, # number of columns in X
eps, # epsilon to avoid division by zero
BLOCK_N: tl.constexpr,
IS_EVEN_N: tl.constexpr
):
row = tl.program_id(0)
X += row * stride_x_row
Y += row * stride_y_row
# Compute mean and variance
cols = tl.arange(0, BLOCK_N)
x = tl.load(X + cols, mask=cols < N, other=0.0).to(tl.float32)
xbar = tl.where(cols < N, x, 0.0)
var = tl.sum(xbar * xbar, axis=0) / N
rstd = 1 / tl.sqrt(var + eps)
tl.store(Rstd + row, rstd)
# Normalize and apply linear transformation
mask = cols < N
if IS_EVEN_N:
w = tl.load(W + cols).to(tl.float32)
else:
w = tl.load(W + cols, mask=mask).to(tl.float32)
x_hat = x * rstd
y = x_hat * w
# Write output
if IS_EVEN_N:
tl.store(Y + cols, y)
else:
tl.store(Y + cols, y, mask=mask)
@triton.jit
def _rmsnorm_bwd_kernel(
X, # pointer to the input
W, # pointer to the weights
DY, # pointer to the output gradient
DX, # pointer to the input gradient
DW, # pointer to the partial sum of weights gradient
Rstd, # pointer to the 1/std
stride_x_row, # how much to increase the pointer when moving by 1 row
stride_dy_row,
stride_dx_row,
M, # number of rows in X
N, # number of columns in X
eps, # epsilon to avoid division by zero
rows_per_program,
BLOCK_N: tl.constexpr,
IS_EVEN_N: tl.constexpr
):
# Map the program id to the elements of X, DX, and DY it should compute.
row_block_id = tl.program_id(0)
row_start = row_block_id * rows_per_program
cols = tl.arange(0, BLOCK_N)
mask = cols < N
X += row_start * stride_x_row
DY += row_start * stride_dy_row
DX += row_start * stride_dx_row
w = tl.load(W + cols, mask=mask).to(tl.float32)
dw = tl.zeros((BLOCK_N,), dtype=tl.float32)
row_end = min((row_block_id + 1) * rows_per_program, M)
for row in range(row_start, row_end):
# Load data to SRAM
if IS_EVEN_N:
x = tl.load(X + cols).to(tl.float32)
dy = tl.load(DY + cols).to(tl.float32)
else:
x = tl.load(X + cols, mask=mask, other=0).to(tl.float32)
dy = tl.load(DY + cols, mask=mask, other=0).to(tl.float32)
rstd = tl.load(Rstd + row)
# Compute dx
xhat = x * rstd
if not IS_EVEN_N:
xhat = tl.where(mask, xhat, 0.0)
wdy = w * dy
dw += dy * xhat
c1 = tl.sum(xhat * wdy, axis=0) / N
dx = (wdy - xhat * c1) * rstd
tl.store(DX + cols, dx, mask=mask)
X += stride_x_row
DY += stride_dy_row
DX += stride_dx_row
tl.store(DW + row_block_id * N + cols, dw, mask=mask)
# Wrapper for triton kernel for torch.compile - should be unecessary for PyTorch 2.3 ?
torch.library.define("flasht5::rmsnorm_triton_fwd", "(Tensor X, Tensor W, float eps) -> (Tensor, Tensor)")
@torch.library.impl("flasht5::rmsnorm_triton_fwd", "default")
def rmsnorm_triton_fwd(X, weight, eps):
M, N = X.shape
assert X.stride(-1) == 1
assert weight.shape == (N,)
assert weight.stride(-1) == 1
# allocate output
Y = torch.empty_like(X)
assert Y.stride(-1) == 1
rstd = torch.empty((M,), dtype=torch.float32, device=X.device)
# Less than 64KB per feature: enqueue fused kernel
MAX_FUSED_SIZE = 65536 // X.element_size()
BLOCK_N = min(MAX_FUSED_SIZE, triton.next_power_of_2(N))
assert N <= BLOCK_N
# heuristics for number of warps
with torch.cuda.device(X.device.index):
_rmsnorm_fwd_kernel[(M,)](
X,
Y,
weight,
rstd,
X.stride(0),
Y.stride(0),
N,
eps,
BLOCK_N,
(N % BLOCK_N == 0)
)
return Y, rstd
@torch.library.register_fake("flasht5::rmsnorm_triton_fwd", rmsnorm_triton_fwd)
def rmsnorm_triton_fwd_abstract(X, weight, eps):
M, N = X.shape
Y = torch.empty_like(X)
rstd = torch.empty((M,), dtype=torch.float32, device=X.device)
return Y, rstd
torch.library.define("flasht5::rmsnorm_triton_bwd", "(Tensor dY, Tensor X, Tensor W, Tensor rstd, float eps) -> (Tensor, Tensor)")
@torch.library.impl("flasht5::rmsnorm_triton_bwd", "default")
def rmsnorm_triton_bwd(
dy,
x,
weight,
rstd,
eps
):
M, N = x.shape
assert x.stride(-1) == 1
assert dy.stride(-1) == 1
assert dy.shape == (M, N)
assert weight.shape == (N,)
assert weight.stride(-1) == 1
# allocate output
dx = torch.empty_like(x)
# Less than 64KB per feature: enqueue fused kernel
MAX_FUSED_SIZE = 65536 // x.element_size()
BLOCK_N = min(MAX_FUSED_SIZE, triton.next_power_of_2(N))
assert N <= BLOCK_N
sm_count = torch.cuda.get_device_properties(x.device).multi_processor_count
_dw = torch.empty((sm_count, N), dtype=torch.float32, device=weight.device)
rows_per_program = math.ceil(M / sm_count)
grid = (sm_count,)
with torch.cuda.device(x.device.index):
_rmsnorm_bwd_kernel[grid](
x,
weight,
dy,
dx,
_dw,
rstd,
x.stride(0),
dy.stride(0),
dx.stride(0),
M,
N,
eps,
rows_per_program,
BLOCK_N,
(N % BLOCK_N == 0)
)
dw = _dw.sum(0).to(weight.dtype)
return dx, dw
@torch.library.register_fake("flasht5::rmsnorm_triton_bwd", rmsnorm_triton_bwd)
def rmsnorm_triton_bwd_abstract(dy, x, weight, rstd, eps):
M, N = x.shape
dx = torch.empty_like(x)
dw = torch.empty((1, N), dtype=torch.float32, device=weight.device)
return dx, dw
class Fast_RMS_Layernorm(torch.autograd.Function):
@staticmethod
@custom_fwd(device_type="cuda")
def forward(ctx, X, W, eps=1e-6):
X_orig_shape = X.shape
X = X.reshape(-1, X.shape[-1])
y, rstd, = torch.ops.flasht5.rmsnorm_triton_fwd(X, W, eps)
y = y.reshape(X_orig_shape)
# We don't store y, will be recomputed in the backward pass to save memory
ctx.save_for_backward(X, W, rstd)
ctx.x_shape_og = X_orig_shape
ctx.eps = eps
return y
@staticmethod
def backward(ctx, dY):
X, weight, rstd = ctx.saved_tensors
dY = dY.reshape(-1, dY.shape[-1])
assert dY.shape == X.shape
dx, dw = torch.ops.flasht5.rmsnorm_triton_bwd(
dY,
X,
weight,
rstd,
ctx.eps
)
return dx.reshape(ctx.x_shape_og), dw, None
def fast_rms_layernorm(X, W, eps):
out = Fast_RMS_Layernorm.apply(X, W, eps)
return out
|