File size: 20,351 Bytes
be59ae6 ae901d9 be59ae6 ae901d9 be59ae6 ae901d9 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 | #!/usr/bin/env python3
"""
Triton-fused Chunked Sparse Backward Pass.
Replaces the Python for-loop over active chunks with fused Triton kernels:
1. sparse_bwd_dW: grad_W[c*CS:(c+1)*CS, :] = grad_Y[:, c*CS:(c+1)*CS].T @ X for active c
2. sparse_bwd_dX: grad_X += grad_Y[:, c*CS:(c+1)*CS] @ W[c*CS:(c+1)*CS, :] for active c
3. sparse_fwd: Y[:, c*CS:(c+1)*CS] = X @ W[c*CS:(c+1)*CS, :].T for active c
Benchmark against the Python-loop baseline at various d_model sizes.
"""
import math
import os
import random
import time
import urllib.request
from dataclasses import dataclass
import torch
import torch.nn as nn
import torch.nn.functional as F
import triton
import triton.language as tl
try:
import tiktoken
except ImportError:
raise ImportError("pip install tiktoken")
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# TRITON KERNELS
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# ββ Kernel 1: Sparse dW ββββββββββββββββββββββββββββββββββββββββββ
# For each active chunk c:
# grad_W[c*CS:(c+1)*CS, :] = grad_Y[:, c*CS:(c+1)*CS].T @ X
#
# In terms of shapes:
# grad_Y: (M, d_out), X: (M, d_in), W: (d_out, d_in)
# For chunk c: rows c*CS..(c+1)*CS of W get grad from cols c*CS..(c+1)*CS of grad_Y
#
# Grid: (num_active * ceil(CS/BN), ceil(d_in/BK))
# pid0 encodes (active_chunk_linear_id, N-block within CS)
# pid1 encodes K-block within d_in
@triton.autotune(
configs=[
triton.Config({'BN': 32, 'BK': 64, 'BM': 32}, num_stages=3, num_warps=4),
triton.Config({'BN': 64, 'BK': 64, 'BM': 32}, num_stages=3, num_warps=4),
triton.Config({'BN': 64, 'BK': 128, 'BM': 32}, num_stages=3, num_warps=4),
triton.Config({'BN': 32, 'BK': 128, 'BM': 64}, num_stages=3, num_warps=4),
triton.Config({'BN': 64, 'BK': 64, 'BM': 64}, num_stages=4, num_warps=4),
],
key=['M', 'd_in', 'CS'],
)
@triton.jit
def _sparse_bwd_dW_kernel(
X_ptr, dY_ptr, dW_ptr, chunk_ids_ptr,
M, d_in, d_out, num_active,
stride_xm, stride_xk,
stride_dym, stride_dyn,
stride_dwn, stride_dwk,
CS: tl.constexpr,
BN: tl.constexpr,
BK: tl.constexpr,
BM: tl.constexpr,
):
"""Compute dW tiles for active chunks. Each program writes one [BN, BK] tile."""
pid0 = tl.program_id(0)
pid1 = tl.program_id(1)
N_BLOCKS_PER_CHUNK = tl.cdiv(CS, BN)
chunk_linear_id = pid0 // N_BLOCKS_PER_CHUNK
n_block_id = pid0 % N_BLOCKS_PER_CHUNK
k_block_id = pid1
if chunk_linear_id >= num_active:
return
chunk_idx = tl.load(chunk_ids_ptr + chunk_linear_id)
chunk_start = chunk_idx * CS
# Tile ranges
rn = n_block_id * BN + tl.arange(0, BN) # rows of dW (= cols of chunk in dY)
rk = k_block_id * BK + tl.arange(0, BK) # cols of dW (= cols of X)
n_abs = chunk_start + rn # absolute column indices in dY
n_mask = rn < CS
k_mask = rk < d_in
# Accumulate dY[:, chunk_cols].T @ X[:, k_cols] over M-tiles
acc = tl.zeros((BN, BK), dtype=tl.float32)
for m_start in range(0, M, BM):
rm = m_start + tl.arange(0, BM)
m_mask = rm < M
# Load X tile: (BM, BK)
x = tl.load(
X_ptr + rm[:, None] * stride_xm + rk[None, :] * stride_xk,
mask=m_mask[:, None] & k_mask[None, :],
other=0.0,
)
# Load dY tile: (BM, BN)
dy = tl.load(
dY_ptr + rm[:, None] * stride_dym + n_abs[None, :] * stride_dyn,
mask=m_mask[:, None] & n_mask[None, :],
other=0.0,
)
# dY.T @ X -> (BN, BK)
acc = tl.dot(tl.trans(dy), x, acc=acc)
# Write to dW: row = chunk_start + rn, col = rk
# dW layout: (d_out, d_in)
dw_ptrs = dW_ptr + n_abs[:, None] * stride_dwn + rk[None, :] * stride_dwk
tl.store(dw_ptrs, acc.to(dW_ptr.dtype.element_ty), mask=n_mask[:, None] & k_mask[None, :])
def sparse_bwd_dW(X, dY, active_chunks, chunk_size, d_out):
"""Fused Triton kernel for sparse dW computation."""
M, d_in = X.shape
num_active = active_chunks.shape[0]
CS = chunk_size
dW = torch.zeros(d_out, d_in, device=X.device, dtype=X.dtype)
if num_active == 0:
return dW
chunk_ids = active_chunks.to(torch.int32).contiguous()
grid = lambda META: (
num_active * triton.cdiv(CS, META['BN']),
triton.cdiv(d_in, META['BK']),
)
_sparse_bwd_dW_kernel[grid](
X, dY, dW, chunk_ids,
M, d_in, d_out, num_active,
X.stride(0), X.stride(1),
dY.stride(0), dY.stride(1),
dW.stride(0), dW.stride(1),
CS=CS,
)
return dW
# ββ Kernel 2: Sparse dX ββββββββββββββββββββββββββββββββββββββββββ
# For each active chunk c:
# grad_X += grad_Y[:, c*CS:(c+1)*CS] @ W[c*CS:(c+1)*CS, :]
#
# Grid: (ceil(M/BM), ceil(d_in/BK))
# Each program accumulates contributions from ALL active chunks.
@triton.autotune(
configs=[
triton.Config({'BM': 32, 'BK': 64, 'BN': 32}, num_stages=3, num_warps=4),
triton.Config({'BM': 64, 'BK': 64, 'BN': 32}, num_stages=3, num_warps=4),
triton.Config({'BM': 64, 'BK': 128, 'BN': 64}, num_stages=3, num_warps=4),
triton.Config({'BM': 32, 'BK': 128, 'BN': 32}, num_stages=4, num_warps=4),
],
key=['M', 'd_in', 'CS'],
)
@triton.jit
def _sparse_bwd_dX_kernel(
dY_ptr, W_ptr, dX_ptr, chunk_ids_ptr,
M, d_in, d_out, num_active,
stride_dym, stride_dyn,
stride_wn, stride_wk,
stride_dxm, stride_dxk,
CS: tl.constexpr,
BM: tl.constexpr,
BK: tl.constexpr,
BN: tl.constexpr,
):
"""Compute dX tiles by summing over active chunks."""
pid_m = tl.program_id(0)
pid_k = tl.program_id(1)
rm = pid_m * BM + tl.arange(0, BM)
rk = pid_k * BK + tl.arange(0, BK)
m_mask = rm < M
k_mask = rk < d_in
acc = tl.zeros((BM, BK), dtype=tl.float32)
# Sum over all active chunks
for i in range(num_active):
chunk_idx = tl.load(chunk_ids_ptr + i)
chunk_start = chunk_idx * CS
# Tile over BN within the chunk
for n_start in range(0, CS, BN):
rn = n_start + tl.arange(0, BN)
n_abs = chunk_start + rn
n_mask = rn < CS
# Load dY tile: (BM, BN)
dy = tl.load(
dY_ptr + rm[:, None] * stride_dym + n_abs[None, :] * stride_dyn,
mask=m_mask[:, None] & n_mask[None, :],
other=0.0,
)
# Load W tile: (BN, BK) β W[chunk_start+rn, rk]
w = tl.load(
W_ptr + n_abs[:, None] * stride_wn + rk[None, :] * stride_wk,
mask=n_mask[:, None] & k_mask[None, :],
other=0.0,
)
# dY @ W -> (BM, BK)
acc = tl.dot(dy, w, acc=acc)
# Write dX
dx_ptrs = dX_ptr + rm[:, None] * stride_dxm + rk[None, :] * stride_dxk
tl.store(dx_ptrs, acc.to(dX_ptr.dtype.element_ty), mask=m_mask[:, None] & k_mask[None, :])
def sparse_bwd_dX(dY, W, active_chunks, chunk_size, M, d_in):
"""Fused Triton kernel for sparse dX computation."""
num_active = active_chunks.shape[0]
CS = chunk_size
dX = torch.zeros(M, d_in, device=dY.device, dtype=dY.dtype)
if num_active == 0:
return dX
chunk_ids = active_chunks.to(torch.int32).contiguous()
grid = lambda META: (
triton.cdiv(M, META['BM']),
triton.cdiv(d_in, META['BK']),
)
_sparse_bwd_dX_kernel[grid](
dY, W, dX, chunk_ids,
M, d_in, dY.shape[1], num_active,
dY.stride(0), dY.stride(1),
W.stride(0), W.stride(1),
dX.stride(0), dX.stride(1),
CS=CS,
)
return dX
# ββ Kernel 3: Sparse dBias ββββββββββββββββββββββββββββββββββββββββ
# Simple: bias_grad[c*CS:(c+1)*CS] = dY[:, c*CS:(c+1)*CS].sum(dim=0)
@triton.jit
def _sparse_bwd_dbias_kernel(
dY_ptr, dB_ptr, chunk_ids_ptr,
M, d_out, num_active,
stride_dym, stride_dyn,
CS: tl.constexpr,
BM: tl.constexpr,
):
pid = tl.program_id(0) # one per (active_chunk, col_within_chunk)
chunk_linear = pid // CS
col_in_chunk = pid % CS
if chunk_linear >= num_active:
return
chunk_idx = tl.load(chunk_ids_ptr + chunk_linear)
col_abs = chunk_idx * CS + col_in_chunk
acc = 0.0
for m_start in range(0, M, BM):
rm = m_start + tl.arange(0, BM)
m_mask = rm < M
vals = tl.load(dY_ptr + rm * stride_dym + col_abs * stride_dyn, mask=m_mask, other=0.0)
acc += tl.sum(vals)
tl.store(dB_ptr + col_abs, acc.to(dB_ptr.dtype.element_ty))
def sparse_bwd_dbias(dY, active_chunks, chunk_size, d_out):
M = dY.shape[0]
num_active = active_chunks.shape[0]
dB = torch.zeros(d_out, device=dY.device, dtype=dY.dtype)
if num_active == 0:
return dB
chunk_ids = active_chunks.to(torch.int32).contiguous()
BM = 128
grid = (num_active * chunk_size,)
_sparse_bwd_dbias_kernel[grid](
dY, dB, chunk_ids,
M, d_out, num_active,
dY.stride(0), dY.stride(1),
CS=chunk_size, BM=BM,
)
return dB
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# AUTOGRAD FUNCTION: Triton-fused
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class TritonChunkedSparseLinear(torch.autograd.Function):
@staticmethod
def forward(ctx, x, weight, bias, active_chunks, chunk_size, sparse_dx):
ctx.save_for_backward(x, weight, active_chunks)
ctx.has_bias = bias is not None
ctx.sparse_dx = sparse_dx
ctx.chunk_size = chunk_size
return F.linear(x, weight, bias)
@staticmethod
def backward(ctx, grad_y):
x, weight, active_chunks = ctx.saved_tensors
cs = ctx.chunk_size
d_out, d_in = weight.shape
x_flat = x.reshape(-1, d_in)
gy_flat = grad_y.reshape(-1, d_out)
M = x_flat.shape[0]
# grad_W via Triton
grad_w = sparse_bwd_dW(x_flat, gy_flat, active_chunks, cs, d_out)
# grad_bias via Triton
grad_b = sparse_bwd_dbias(gy_flat, active_chunks, cs, d_out) if ctx.has_bias else None
# grad_X
if ctx.sparse_dx:
grad_x_flat = sparse_bwd_dX(gy_flat, weight, active_chunks, cs, M, d_in)
else:
grad_x_flat = gy_flat @ weight # dense dX
return grad_x_flat.reshape(x.shape), grad_w, grad_b, None, None, None
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# AUTOGRAD FUNCTION: Python-loop baseline (for comparison)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class PythonLoopSparseLinear(torch.autograd.Function):
@staticmethod
def forward(ctx, x, weight, bias, active_chunks, chunk_size, sparse_dx):
ctx.save_for_backward(x, weight, active_chunks)
ctx.has_bias = bias is not None
ctx.sparse_dx = sparse_dx
ctx.chunk_size = chunk_size
return F.linear(x, weight, bias)
@staticmethod
def backward(ctx, grad_y):
x, weight, active_chunks = ctx.saved_tensors
cs = ctx.chunk_size
x_flat = x.reshape(-1, x.shape[-1])
gy_flat = grad_y.reshape(-1, grad_y.shape[-1])
grad_w = torch.zeros_like(weight)
grad_b = torch.zeros(weight.shape[0], device=weight.device, dtype=weight.dtype) if ctx.has_bias else None
if ctx.sparse_dx:
grad_x_flat = torch.zeros_like(x_flat)
else:
grad_x_flat = gy_flat @ weight
for c in active_chunks.tolist():
s, e = c * cs, (c + 1) * cs
gy_slice = gy_flat[:, s:e]
grad_w[s:e, :] = gy_slice.t() @ x_flat
if ctx.has_bias:
grad_b[s:e] = gy_slice.sum(0)
if ctx.sparse_dx:
grad_x_flat += gy_slice @ weight[s:e, :]
return grad_x_flat.reshape(x.shape), grad_w, grad_b, None, None, None
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# CORRECTNESS TEST
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def test_correctness():
print("Testing correctness...")
torch.manual_seed(42)
device = "cuda"
for d_in, d_out, cs in [(512, 2048, 64), (1024, 4096, 64), (256, 1024, 32)]:
M = 2048 # B*T
n_chunks = d_out // cs
n_active = max(1, int(0.1 * n_chunks))
active = torch.randperm(n_chunks, device=device)[:n_active].sort().values
x = torch.randn(M, d_in, device=device, requires_grad=False)
w = torch.randn(d_out, d_in, device=device, requires_grad=False)
b = torch.randn(d_out, device=device, requires_grad=False)
gy = torch.randn(M, d_out, device=device, requires_grad=False)
# Reference: Python loop
ref_dw = torch.zeros_like(w)
ref_db = torch.zeros_like(b)
ref_dx = gy @ w # dense dX
for c in active.tolist():
s, e = c * cs, (c + 1) * cs
ref_dw[s:e] = gy[:, s:e].t() @ x
ref_db[s:e] = gy[:, s:e].sum(0)
# Triton
tri_dw = sparse_bwd_dW(x, gy, active, cs, d_out)
tri_db = sparse_bwd_dbias(gy, active, cs, d_out)
tri_dx_sparse = sparse_bwd_dX(gy, w, active, cs, M, d_in)
# Compare
dw_err = (tri_dw - ref_dw).abs().max().item()
db_err = (tri_db - ref_db).abs().max().item()
# For sparse dX, reference
ref_dx_sparse = torch.zeros_like(x)
for c in active.tolist():
s, e = c * cs, (c + 1) * cs
ref_dx_sparse += gy[:, s:e] @ w[s:e]
dx_err = (tri_dx_sparse - ref_dx_sparse).abs().max().item()
status = "β" if dw_err < 1e-2 and db_err < 1e-2 and dx_err < 1e-2 else "β"
print(f" {status} d_in={d_in}, d_out={d_out}, cs={cs}: dW_err={dw_err:.6f}, dB_err={db_err:.6f}, dX_err={dx_err:.6f}")
print()
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# BENCHMARK
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def benchmark():
print("="*80)
print("BENCHMARK: Triton Fused vs Python Loop vs Dense")
print("="*80)
device = "cuda"
B, T = 8, 256
M = B * T
cs = 64
af = 0.10
warmup_iters = 10
bench_iters = 50
print(f"\nM={M} (B={B}, T={T}), chunk_size={cs}, active_frac={af}")
print(f"{'d_model':>7} | {'d_out':>7} | {'active':>6} | {'Dense':>10} | {'PyLoop':>10} | {'Triton':>10} | {'Tri/Dense':>10} | {'Tri/PyLoop':>10}")
print("-" * 95)
for d_in in [256, 512, 768, 1024, 1536, 2048]:
d_out = 4 * d_in
n_chunks = d_out // cs
n_active = max(1, int(af * n_chunks))
active = torch.randperm(n_chunks, device=device)[:n_active].sort().values
x = torch.randn(M, d_in, device=device)
w = torch.randn(d_out, d_in, device=device)
b = torch.randn(d_out, device=device)
gy = torch.randn(M, d_out, device=device)
# Dense backward (dW + dX + dB)
def dense_bwd():
dw = gy.t() @ x
dx = gy @ w
db = gy.sum(0)
return dw, dx, db
# Python loop backward
def pyloop_bwd():
dw = torch.zeros_like(w)
db = torch.zeros_like(b)
dx = gy @ w # dense dX
for c in active.tolist():
s, e = c * cs, (c + 1) * cs
dw[s:e] = gy[:, s:e].t() @ x
db[s:e] = gy[:, s:e].sum(0)
return dw, dx, db
# Triton fused backward
def triton_bwd():
dw = sparse_bwd_dW(x, gy, active, cs, d_out)
dx = gy @ w # dense dX (same as pyloop)
db = sparse_bwd_dbias(gy, active, cs, d_out)
return dw, dx, db
# Warmup
for _ in range(warmup_iters):
dense_bwd(); pyloop_bwd(); triton_bwd()
torch.cuda.synchronize()
# Bench dense
torch.cuda.synchronize(); t0 = time.perf_counter()
for _ in range(bench_iters): dense_bwd()
torch.cuda.synchronize(); dense_time = (time.perf_counter() - t0) / bench_iters
# Bench pyloop
torch.cuda.synchronize(); t0 = time.perf_counter()
for _ in range(bench_iters): pyloop_bwd()
torch.cuda.synchronize(); pyloop_time = (time.perf_counter() - t0) / bench_iters
# Bench triton
torch.cuda.synchronize(); t0 = time.perf_counter()
for _ in range(bench_iters): triton_bwd()
torch.cuda.synchronize(); triton_time = (time.perf_counter() - t0) / bench_iters
tri_vs_dense = dense_time / triton_time
tri_vs_pyloop = pyloop_time / triton_time
print(f"{d_in:>7} | {d_out:>7} | {n_active:>6} | {dense_time*1000:>9.2f}ms | {pyloop_time*1000:>9.2f}ms | {triton_time*1000:>9.2f}ms | {tri_vs_dense:>9.2f}x | {tri_vs_pyloop:>9.2f}x")
# Also benchmark with sparse_dX (Triton dX kernel)
print(f"\n{'='*80}")
print("With Triton sparse_dX (both dW and dX are sparse):")
print(f"{'d_model':>7} | {'Dense':>10} | {'Triton_all':>10} | {'Speedup':>10}")
print("-" * 50)
for d_in in [512, 1024, 2048]:
d_out = 4 * d_in
n_chunks = d_out // cs
n_active = max(1, int(af * n_chunks))
active = torch.randperm(n_chunks, device=device)[:n_active].sort().values
x = torch.randn(M, d_in, device=device)
w = torch.randn(d_out, d_in, device=device)
gy = torch.randn(M, d_out, device=device)
def dense_full():
dw = gy.t() @ x; dx = gy @ w; return dw, dx
def triton_full():
dw = sparse_bwd_dW(x, gy, active, cs, d_out)
dx = sparse_bwd_dX(gy, w, active, cs, M, d_in)
return dw, dx
for _ in range(warmup_iters): dense_full(); triton_full()
torch.cuda.synchronize()
torch.cuda.synchronize(); t0 = time.perf_counter()
for _ in range(bench_iters): dense_full()
torch.cuda.synchronize(); dt = (time.perf_counter() - t0) / bench_iters
torch.cuda.synchronize(); t0 = time.perf_counter()
for _ in range(bench_iters): triton_full()
torch.cuda.synchronize(); tt = (time.perf_counter() - t0) / bench_iters
print(f"{d_in:>7} | {dt*1000:>9.2f}ms | {tt*1000:>9.2f}ms | {dt/tt:>9.2f}x")
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# MAIN
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
if __name__ == "__main__":
test_correctness()
benchmark()
|