Spaces:
Running
on
A100
Running
on
A100
File size: 10,377 Bytes
174ae06 |
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 |
# Copyright (c) 2025 NVIDIA CORPORATION.
# Licensed under the MIT license.
# Adapted from https://github.com/NVlabs/VILA/tree/main under the Apache 2.0 license.
# LICENSE is in incl_licenses directory.
# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
#
# 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.
#
# SPDX-License-Identifier: Apache-2.0
"""
Layer Normalization
====================
In this tutorial, you will write a high-performance layer normalization
kernel that runs faster than the PyTorch implementation.
In doing so, you will learn about:
* Implementing backward pass in Triton.
* Implementing parallel reduction in Triton.
"""
import torch
import triton
import triton.language as tl
from ._division_transpose import fp8_division_transpose
from .common import FP8_MAX_VALUE, SCALE_MIN_THRES
"""RMSNorm Forward + Backward"""
"""Forward: Input uses 1 * 16 group quantization"""
"""Forward: Output uses per-tensor quantization"""
"""Backward: Input uses full-precision/BF16."""
"""Backward: Output uses full-precision/BF16."""
"""The input can be 2D or 3D, but the calculation is performed in 2D"""
@triton.heuristics(
{
"BLOCK_SN": lambda args: args["N"] // args["QB"],
"BLOCK_SN2": lambda args: args["N2"] // args["QB"],
}
)
@triton.jit
def _rms_norm_fwd_fused(
X, # pointer to the input
SX, # pointer to the scale of input
Y, # pointer to the output
SY, # pointer to the scale of output
W, # Weight
Rstd, # pointer to the 1/std
stride, # how much to increase the pointer when moving by 1 row
scale_stride, # how much to increase the pointer when moving by 1 row
N: tl.constexpr, # number of columns in X,
N2: tl.constexpr, # number of columns in X,
SN: tl.constexpr,
SN2: tl.constexpr,
QB: tl.constexpr,
eps, # epsilon to avoid division by zero
fp8_max,
SCALE_MIN_THRES: tl.constexpr,
BLOCK_SN: tl.constexpr,
BLOCK_SN2: tl.constexpr,
):
# Map the program id to the row of X and Y it should compute.
row = tl.program_id(0)
Y += row * stride
X += row * stride
SX += row * scale_stride
cols = tl.arange(0, N2)
scale_cols = tl.arange(0, SN2)
x = tl.load(X + cols, mask=cols < N, other=0.0).to(tl.float32)
scale_x = tl.load(SX + scale_cols, mask=scale_cols < SN, other=0.0).to(tl.float32)
# Dequantize and swish calculation
scale_x = tl.reshape(scale_x, (BLOCK_SN2, 1))
x = tl.reshape(x, (BLOCK_SN2, QB))
x = x * scale_x
x = tl.reshape(x, N2)
# Compute variance
_var = x * x
var = tl.sum(_var, axis=0) / N
rstd = 1 / tl.sqrt(var + eps)
# Write mean / rstd
tl.store(Rstd + row, rstd)
# Normalize and apply linear transformation
w = tl.load(W + cols, mask=cols < N, other=0.0)
x_hat = x * rstd
y = x_hat * w
# Scale calculation
abs_y = tl.abs(y)
max_val = tl.max(abs_y) + SCALE_MIN_THRES
scale_output = max_val / fp8_max
tl.store(SY + row, scale_output)
# Write output
tl.store(Y + cols, y, mask=cols < N)
@triton.heuristics(
{
"BLOCK_SN": lambda args: args["N"] // args["QB"],
"BLOCK_SN2": lambda args: args["N2"] // args["QB"],
}
)
@triton.jit
def _rms_norm_bwd_dx_fused(
DX, # pointer to the input gradient
DY, # pointer to the output gradient
DW,
X, # pointer to the input
SX, # pointer to the input
W, # weight
Rstd, # pointer to the 1/std
Lock,
stride, # how much to increase the pointer when moving by 1 row
scale_stride, # how much to increase the pointer when moving by 1 row
N: tl.constexpr, # number of columns in X,
N2: tl.constexpr, # number of columns in X,
SN: tl.constexpr,
SN2: tl.constexpr,
QB: tl.constexpr,
SCALE_MIN_THRES: tl.constexpr,
BLOCK_SN: tl.constexpr,
BLOCK_SN2: tl.constexpr,
GROUP_SIZE_M: tl.constexpr,
):
# Map the program id to the elements of X, DX, and DY it should compute.
row = tl.program_id(0)
cols = tl.arange(0, N2)
scale_cols = tl.arange(0, SN2)
mask = cols < N
X += row * stride
DY += row * stride
DX += row * stride
SX += row * scale_stride
# Offset locks and weights/biases gradient pointer for parallel reduction
lock_id = row % GROUP_SIZE_M
Lock += lock_id
Count = Lock + GROUP_SIZE_M
DW = DW + lock_id * N + cols
# Load data to SRAM
x = tl.load(X + cols, mask=mask, other=0.0).to(tl.float32)
scale_x = tl.load(SX + scale_cols, mask=scale_cols < SN, other=0.0).to(tl.float32)
scale_x = tl.reshape(scale_x, (BLOCK_SN2, 1))
x = tl.reshape(x, (BLOCK_SN2, QB))
x = x * scale_x
x = tl.reshape(x, N2)
dy = tl.load(DY + cols, mask=mask, other=0.0).to(tl.float32)
# Load weight
w = tl.load(W + cols, mask=cols < N, other=0.0).to(tl.float32)
rstd = tl.load(Rstd + row).to(tl.float32)
# Compute dx
xhat = x * rstd
wdy = w * dy
xhat = tl.where(mask, xhat, 0.0)
wdy = tl.where(mask, wdy, 0.0)
c1 = tl.sum(xhat * wdy, axis=0) / N
dx = (wdy - (xhat * c1)) * rstd # layer norm have c2 term, rmsnorm do not
dx = dx.to(DX.type.element_ty)
# Write dx
tl.store(DX + cols, dx, mask=mask)
# Accumulate partial sums for dw/db
partial_dw = (dy * xhat).to(w.dtype)
while tl.atomic_cas(Lock, 0, 1) == 1:
pass
count = tl.load(Count)
# First store doesn't accumulate
if count == 0:
tl.atomic_xchg(Count, 1)
else:
partial_dw += tl.load(DW, mask=mask)
tl.store(DW, partial_dw, mask=mask)
# Release the lock
tl.atomic_xchg(Lock, 0)
@triton.jit
def _rms_norm_bwd_dwdb(
DW, # pointer to the partial sum of weights gradient
FINAL_DW, # pointer to the weights gradient
M, # GROUP_SIZE_M
N, # number of columns
BLOCK_SIZE_M: tl.constexpr,
BLOCK_SIZE_N: tl.constexpr,
):
# Map the program id to the elements of DW and DB it should compute.
pid = tl.program_id(0)
cols = pid * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
dw = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
# Iterate through the rows of DW and DB to sum the partial sums.
for i in range(0, M, BLOCK_SIZE_M):
rows = i + tl.arange(0, BLOCK_SIZE_M)
mask = (rows[:, None] < M) & (cols[None, :] < N)
offs = rows[:, None] * N + cols[None, :]
dw += tl.load(DW + offs, mask=mask, other=0.0)
# Write the final sum to the output.
sum_dw = tl.sum(dw, axis=0)
tl.store(FINAL_DW + cols, sum_dw, mask=cols < N)
def fp8_rmsnorm_forward(x, s_x, w, QB, eps, transpose_output_2d=False):
# Change batched 3D input to 2D
batched = False
if len(x.shape) == 3:
assert len(s_x.shape) == 3
batched = True
BS = x.shape[0]
x = x.reshape(-1, x.shape[-1])
s_x = s_x.reshape(-1, s_x.shape[-1])
# allocate output
M, N = x.shape
_, SN = s_x.shape
y = torch.empty_like(x, dtype=torch.bfloat16)
s_y = torch.empty((M,), dtype=torch.bfloat16, device=x.device)
rstd = torch.empty((M,), dtype=torch.float32, device=x.device)
# heuristics for number of warps
num_warps = 8
fp8MaxValue = FP8_MAX_VALUE[x.dtype]
N2 = triton.next_power_of_2(N)
SN2 = N2 // QB
# import os
# if int(os.environ.get("LOCAL_RANK")) == 7:
# print(x.device, x.shape, x.dtype, s_x.shape, s_x.dtype, y.shape, y.dtype, s_y.shape, s_y.dtype, w.shape, w.dtype, rstd.shape, rstd.dtype, x.stride(0), s_x.stride(0), N, N2, SN, SN2, QB, "\n")
# enqueue kernel
_rms_norm_fwd_fused[(M,)]( #
x,
s_x,
y,
s_y,
w,
rstd, #
x.stride(0),
s_x.stride(0),
N,
N2,
SN,
SN2,
QB,
eps,
fp8MaxValue,
SCALE_MIN_THRES=SCALE_MIN_THRES, #
num_warps=num_warps,
num_ctas=1,
)
# reduction
s_y_max = s_y.max()
qy, s_y_max, qy_t = fp8_division_transpose(y, QB, x.dtype, s_y_max)
# Recover 2D to 3D
if batched:
y = y.reshape(BS, -1, y.shape[-1])
qy = qy.reshape(BS, -1, y.shape[-1])
if not transpose_output_2d:
qy_t = qy_t.reshape(BS, -1, y.shape[-1])
return qy, s_y_max, qy_t, (w.clone(), rstd, num_warps)
def fp8_rmsnorm_backward(x, s_x, g, w, v, QB, num_warps):
# Change batched 3D input to 2D
batched = False
if len(x.shape) == 3:
assert len(s_x.shape) == 3
batched = True
BS = x.shape[0]
x = x.reshape(-1, x.shape[-1])
s_x = s_x.reshape(-1, s_x.shape[-1])
g = g.reshape(-1, g.shape[-1])
# enqueue kernel using forward pass heuristics
# also compute partial sums for DW and DB
M, N = g.shape
_, SN = s_x.shape
GROUP_SIZE_M = 128
# heuristics for amount of parallel reduction stream for DW/DB
locks = torch.zeros(2 * GROUP_SIZE_M, dtype=torch.int32, device=w.device)
_dw = torch.zeros((GROUP_SIZE_M, N), dtype=w.dtype, device=w.device)
dw = torch.empty((N,), dtype=w.dtype, device=w.device)
dx = torch.empty_like(g, dtype=torch.bfloat16)
N2 = triton.next_power_of_2(N)
SN2 = triton.next_power_of_2(SN)
_rms_norm_bwd_dx_fused[(M,)]( #
dx,
g,
_dw,
x,
s_x,
w,
v,
locks, #
x.stride(0),
s_x.stride(0),
N,
N2,
SN,
SN2,
QB,
SCALE_MIN_THRES=SCALE_MIN_THRES,
num_warps=num_warps,
GROUP_SIZE_M=GROUP_SIZE_M,
)
if batched:
dx = dx.reshape(BS, -1, dx.shape[-1])
grid = lambda meta: [triton.cdiv(N, meta["BLOCK_SIZE_N"])]
# accumulate partial sums in separate kernel
_rms_norm_bwd_dwdb[grid](_dw, dw, min(GROUP_SIZE_M, M), N, BLOCK_SIZE_M=32, BLOCK_SIZE_N=128, num_ctas=1) # #
return dx, dw
|