File size: 5,939 Bytes
0094a2a |
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 |
# -*- coding: utf-8 -*-
import torch
import triton
import triton.language as tl
@triton.autotune(
configs=[
triton.Config({}, num_warps=1),
triton.Config({}, num_warps=2),
triton.Config({}, num_warps=4),
triton.Config({}, num_warps=8),
triton.Config({}, num_warps=16),
triton.Config({}, num_warps=32),
],
key=["N"],
)
# @triton.heuristics({"HAS_BIAS": lambda args: args["B"] is not None})
# @triton.heuristics({"HAS_RESIDUAL": lambda args: args["RESIDUAL"] is not None})
@triton.jit
def _l2_norm_fwd_1pass_kernel(
X, # pointer to the input
Y, # pointer to the output
stride_x_row, # how much to increase the pointer when moving by 1 row
N, # number of columns in X
eps, # epsilon to avoid division by zero
BLOCK_N: tl.constexpr,
):
# Map the program id to the row of X and Y it should compute.
row = tl.program_id(0)
X += row * stride_x_row
Y += row * stride_x_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)
rstd = 1 / tl.sqrt(var + eps)
# tl.store(Rstd + row, rstd)
# Normalize and apply linear transformation
mask = cols < N
y = x * rstd
# Write output
tl.store(Y + cols, y, mask=mask)
@triton.autotune(
configs=[
triton.Config({}, num_warps=1),
triton.Config({}, num_warps=2),
triton.Config({}, num_warps=4),
triton.Config({}, num_warps=8),
triton.Config({}, num_warps=16),
triton.Config({}, num_warps=32),
],
key=["N"],
)
# @triton.heuristics({"HAS_BIAS": lambda args: args["B"] is not None})
# @triton.heuristics({"HAS_DRESIDUAL": lambda args: args["DRESIDUAL"] is not None})
# @triton.heuristics({"STORE_DRESIDUAL": lambda args: args["DRESIDUAL_IN"] is not None})
# @triton.heuristics({"RECOMPUTE_OUTPUT": lambda args: args["Y"] is not None})
@triton.jit
def _l2_norm_bwd_kernel(
X, # pointer to the input
# Y, # pointer to the output to be recomputed
DY, # pointer to the output gradient
DX, # pointer to the input gradient
stride_x_row, # how much to increase the pointer when moving by 1 row
N, # number of columns in X
eps, # epsilon to avoid division by zero
BLOCK_N: tl.constexpr,
):
# Map the program id to the elements of X, DX, and DY it should compute.
# Map the program id to the row of X and Y it should compute.
row = tl.program_id(0)
X += row * stride_x_row
DX += row * stride_x_row
DY += row * stride_x_row
# Y += row * stride_y_row
cols = tl.arange(0, BLOCK_N)
x = tl.load(X + cols, mask=cols < N, other=0.0).to(tl.float32)
x = tl.where(cols < N, x, 0.0)
var = tl.sum(x * x)
rstd = 1 / tl.sqrt(var + eps)
# tl.store(Rstd + row, rstd)
# Normalize and apply linear transformation
mask = cols < N
# y = x * rstd
dy = tl.load(DY + cols, mask=cols < N, other=0.0).to(tl.float32)
dy = tl.where(cols < N, dy, 0.0)
# dx = dy * rstd - tl.sum(dy * x) * (1 / (var+eps)) * rstd * x
dx = dy * rstd - tl.sum(dy * x) * (1 / (var+eps)) * rstd * x
tl.store(DX + cols, dx, mask=mask)
def _l2_norm_fwd(
x, eps=1e-6
):
x_shape_og = x.shape
x = x.reshape(-1, x.shape[-1])
if x.stride(-1) != 1:
x = x.contiguous()
M, N = x.shape
assert x.stride(-1) == 1
# allocate output
y = torch.empty_like(x)
assert y.stride(-1) == 1
N = x.shape[-1]
M = x.shape[0]
# rstd = torch.empty((M,), dtype=torch.float32, device="cuda")
# 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))
if N > BLOCK_N:
raise RuntimeError(
"This layer norm doesn't support feature dim >= 64KB.")
# heuristics for number of warps
with torch.cuda.device(x.device.index):
_l2_norm_fwd_1pass_kernel[(M,)](
x,
y,
x.stride(0),
N,
eps,
# is_rms_norm,
BLOCK_N,
# residual is not None,
# residual_out is not None,
# bias is not None,
)
return y.reshape(x_shape_og)
def _l2_norm_bwd(
x, dy, eps=1e-5,
):
x_shape_og = x.shape
x = x.reshape(-1, dy.shape[-1])
dy = dy.reshape(-1, dy.shape[-1])
if dy.stride(-1) != 1:
dy = dy.contiguous()
assert dy.shape == x.shape
# allocate output
dx = torch.empty_like(x)
N = x.shape[-1]
M = x.shape[0]
assert x.stride(-1) == 1
assert dy.stride(-1) == 1
# rstd = torch.empty((M,), dtype=torch.float32, device="cuda")
# 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))
if N > BLOCK_N:
raise RuntimeError(
"This layer norm doesn't support feature dim >= 64KB.")
# heuristics for number of warps
with torch.cuda.device(x.device.index):
_l2_norm_bwd_kernel[(M,)](
x,
dy,
dx,
x.stride(0),
N,
eps,
BLOCK_N,
)
return dx.reshape(x_shape_og)
class L2NormFunction(torch.autograd.Function):
@staticmethod
def forward(
ctx,
x,
eps=1e-6,
):
# reshape input data into 2D tensor
y = _l2_norm_fwd(x, eps)
ctx.eps = eps
ctx.x_dtype = x.dtype
ctx.save_for_backward(x)
return y
@staticmethod
def backward(ctx, dy, *args):
x, = ctx.saved_tensors
dx = _l2_norm_bwd(
x,
dy,
ctx.eps,
)
return (
dx,
None
)
l2_norm = L2NormFunction.apply
|