File size: 19,229 Bytes
d8bc908 | 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 | """
FlashVQ: Custom Vector Quantization with dual Triton GPU + PyTorch CPU path.
Replaces vector_quantize_pytorch entirely (D-100). FlashVQCodebook is a standalone
nn.Module implementing all VQ operations:
- Cosine similarity codebook lookup
- EMA codebook update
- Dead code reset
- Rotation trick (gradient through quantization)
- Commitment loss
Dispatch pattern (following tscale.py):
if x.is_cuda and _HAS_TRITON β _TritonFlashVQFn.apply()
else β self._cpu_forward()
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
_HAS_TRITON = False
try:
import triton
import triton.language as tl
_HAS_TRITON = True
except ImportError:
pass
class _RotationTrickFn(torch.autograd.Function):
"""
Rotation trick gradient through vector quantization.
Instead of straight-through estimator (STE), rotate the encoder output
gradient toward the quantized vector direction. This helps the encoder
learn to produce outputs that align with codebook entries.
"""
@staticmethod
def forward(ctx, x, quantized):
ctx.save_for_backward(x.detach(), quantized.detach())
return quantized
@staticmethod
def backward(ctx, grad_output):
x, quantized = ctx.saved_tensors
# Normalize in fp32 for numerical stability
x_norm = F.normalize(x.float(), dim=-1)
q_norm = F.normalize(quantized.float(), dim=-1)
# Gradient deflection: subtract projection onto (x_norm - q_norm)
# This rotates the gradient toward the quantized direction
diff = x_norm - q_norm
proj = (grad_output.float() * x_norm).sum(dim=-1, keepdim=True)
grad_x = grad_output.float() - proj * diff
return grad_x.to(grad_output.dtype), None
class FlashVQCodebook(nn.Module):
"""
Vector quantization codebook with dual GPU (Triton) / CPU (PyTorch) paths.
Interface matches vector_quantize_pytorch.VectorQuantize:
forward(x) β (quantized, indices, commitment_loss)
All VQ operations are self-contained:
- Cosine similarity codebook lookup
- Straight-through estimator (STE) with optional rotation trick
- EMA codebook update (decay=0.99)
- Dead code reset (threshold_ema_dead_code=2)
- Commitment loss
"""
def __init__(
self,
codebook_size: int = 8192,
codebook_dim: int = 32,
decay: float = 0.99,
commitment_weight: float = 1.0,
threshold_ema_dead_code: int = 2,
kmeans_init: bool = True,
kmeans_iters: int = 10,
rotation_trick: bool = True,
):
super().__init__()
self.codebook_size = codebook_size
self.codebook_dim = codebook_dim
self.decay = decay
self.commitment_weight = commitment_weight
self.threshold_ema_dead_code = threshold_ema_dead_code
self.kmeans_init = kmeans_init
self.kmeans_iters = kmeans_iters
self.rotation_trick = rotation_trick
# Codebook buffers
self.register_buffer('embed', torch.randn(codebook_size, codebook_dim) * 0.02)
self.register_buffer('cluster_size', torch.zeros(codebook_size))
self.register_buffer('embed_avg', torch.zeros(codebook_size, codebook_dim))
# Tile sizes for Triton kernel (set on first GPU forward)
self._triton_block_bt = 16
self._triton_tile_k = 1024
def _compute_tile_sizes(self):
"""
Dynamic tile sizing per D-102.
Queries GPU device properties to determine SRAM budget, then computes
BLOCK_BT and TILE_K such that:
BLOCK_BT * codebook_dim * 2 + TILE_K * codebook_dim * 2 < SRAM * 0.9
For sm_89 (RTX 4060, 99KB SRAM per SM):
codebook_size=8192, codebook_dim=32 β BLOCK_BT=16, TILE_K=1024 (65KB)
codebook_size=4096, codebook_dim=32 β BLOCK_BT=16, TILE_K=512 (33KB)
"""
if not torch.cuda.is_available():
return
try:
props = torch.cuda.get_device_properties(0)
sram_budget = 99 * 1024 # SM 8.9: 99KB per SM
# Conservative estimate: each element is 2 bytes (bf16) in SRAM
elem_bytes = 2
# Find largest TILE_K that fits with BLOCK_BT=16
bt = 16
for tk in [2048, 1024, 512, 256, 128]:
sram_usage = bt * self.codebook_dim * elem_bytes + tk * self.codebook_dim * elem_bytes
if sram_usage < sram_budget * 0.9:
self._triton_block_bt = bt
self._triton_tile_k = tk
return
# Fallback for very constrained SRAM or large codebook_dim
self._triton_block_bt = 8
self._triton_tile_k = 256
except Exception:
# Default values
self._triton_block_bt = 16
self._triton_tile_k = 1024
def forward(self, x: torch.Tensor):
"""
Args:
x: Input tensor of shape [*, codebook_dim]
Returns:
quantized: Tensor of same shape as x
indices: Tensor of shape [*] with codebook indices
commitment_loss: Scalar tensor
"""
orig_shape = x.shape
x_flat = x.reshape(-1, self.codebook_dim)
if x.is_cuda and _HAS_TRITON:
quantized, indices, commitment_loss = self._triton_forward(x_flat)
else:
quantized, indices, commitment_loss = self._cpu_forward(x_flat)
quantized = quantized.reshape(orig_shape)
indices = indices.reshape(orig_shape[:-1])
return quantized, indices, commitment_loss
def _triton_forward(self, x_flat: torch.Tensor):
"""Triton GPU path β dispatched when CUDA + Triton available."""
# Use _TritonFlashVQFn for forward + backward via autograd
quantized, indices, commitment_loss = _TritonFlashVQFn.apply(
x_flat, self.embed, self.cluster_size, self.embed_avg,
self.codebook_size, self.codebook_dim,
self.commitment_weight, self.rotation_trick,
)
# EMA update and dead code reset (under torch.no_grad)
with torch.no_grad():
self._ema_update(x_flat, indices)
self._dead_code_reset(x_flat)
return quantized, indices, commitment_loss
def _cpu_forward(self, x_flat: torch.Tensor):
"""
Pure PyTorch CPU path β implements all VQ operations.
Steps:
1. Cosine similarity lookup β nearest codebook entry indices
2. Quantize via straight-through estimator (or rotation trick)
3. Compute commitment loss
4. EMA update codebook (under torch.no_grad)
5. Dead code reset (under torch.no_grad)
"""
# ββ Step 1: Cosine similarity lookup ββ
x_norm = F.normalize(x_flat.float(), dim=-1)
embed_norm = F.normalize(self.embed.float(), dim=-1)
sim = x_norm @ embed_norm.T # [N, codebook_size]
indices = sim.argmax(dim=-1) # [N]
# ββ Step 2: Quantize with STE or rotation trick ββ
with torch.no_grad():
quantized = self.embed[indices] # [N, D]
if self.rotation_trick:
quantized = _RotationTrickFn.apply(x_flat, quantized)
else:
# Straight-through estimator
quantized = x_flat + (quantized - x_flat).detach()
# ββ Step 3: Commitment loss ββ
commitment_loss = self.commitment_weight * F.mse_loss(
x_flat.float(), quantized.detach().float()
)
# ββ Step 4: EMA update ββ
with torch.no_grad():
self._ema_update(x_flat, indices)
# ββ Step 5: Dead code reset ββ
self._dead_code_reset(x_flat)
return quantized, indices, commitment_loss
def _ema_update(self, x_flat: torch.Tensor, indices: torch.Tensor):
"""
Exponential moving average codebook update.
Args:
x_flat: [N, D] input vectors
indices: [N] codebook indices for each input vector
"""
one_hot = F.one_hot(indices, num_classes=self.codebook_size).float() # [N, codebook_size]
n_assign = one_hot.sum(dim=0) # [codebook_size]
# EMA on cluster_size (how many inputs assigned to each code)
self.cluster_size.mul_(self.decay).add_(n_assign * (1 - self.decay))
# EMA on embed_avg: weighted sum of assigned inputs
# embed_avg[c] = decay * embed_avg[c] + (1 - decay) * sum(x assigned to c)
x_float = x_flat.float()
for c in range(self.codebook_size):
mask = indices == c
count = mask.sum().item()
if count > 0:
assigned_sum = x_float[mask].sum(dim=0)
self.embed_avg[c].mul_(self.decay).add_(assigned_sum * (1 - self.decay))
# Normalize: embed = embed_avg / cluster_size (with epsilon)
cluster_size_safe = self.cluster_size.clamp(min=1e-5)
self.embed.copy_(self.embed_avg / cluster_size_safe.unsqueeze(1))
def _dead_code_reset(self, x_flat: torch.Tensor):
"""
Replace dead codebook entries (cluster_size < threshold) with
random vectors from the current input batch.
"""
dead_mask = self.cluster_size < self.threshold_ema_dead_code
n_dead = dead_mask.sum().item()
if n_dead == 0:
return
dead_indices = torch.where(dead_mask)[0]
# Replace with random input vectors
rand_idx = torch.randint(0, x_flat.shape[0], (n_dead,), device=x_flat.device)
self.embed[dead_indices] = x_flat[rand_idx].detach()
self.cluster_size[dead_indices] = 0.0
self.embed_avg[dead_indices] = 0.0
@torch.no_grad()
def kmeans_init_codebook(self, x: torch.Tensor):
"""Initialize codebook via k-means on first batch."""
x_flat = x.reshape(-1, self.codebook_dim).float()
centroids = x_flat[torch.randperm(x_flat.shape[0])[:self.codebook_size]].clone()
for _ in range(self.kmeans_iters):
dist = torch.cdist(x_flat, centroids)
assign = dist.argmin(dim=-1)
for i in range(self.codebook_size):
mask = assign == i
if mask.sum() > 0:
centroids[i] = x_flat[mask].mean(dim=0)
self.embed.copy_(centroids)
@torch.no_grad()
def get_codebook_utilization(self) -> float:
"""Fraction of codebook entries with any usage."""
return (self.cluster_size > 0).float().mean().item()
@torch.no_grad()
def get_dead_code_count(self) -> int:
"""Number of codebook entries below EMA dead threshold."""
return (self.cluster_size < self.threshold_ema_dead_code).sum().item()
# βββ Triton GPU Kernels βββ
# Only defined when Triton is available
if _HAS_TRITON:
@triton.jit
def _triton_flash_vq_lookup_kernel(
x_ptr, codebook_ptr, indices_ptr,
stride_xb, stride_xd,
stride_cb, stride_cd,
N_CTX: tl.constexpr,
CODEBOOK_SIZE: tl.constexpr,
CODEBOOK_DIM: tl.constexpr,
BLOCK_BT: tl.constexpr,
TILE_K: tl.constexpr,
):
"""
Tiled cosine similarity + argmax lookup for VQ codebook.
Architecture:
pid = batch tile index
Load input tile [BLOCK_BT, CODEBOOK_DIM]
Normalize in fp32
Tile over codebook in TILE_K chunks:
Load codebook tile [TILE_K, CODEBOOK_DIM]
Normalize in fp32
Compute dot product via tl.dot β [BLOCK_BT, TILE_K]
Update running argmax
Store best indices
SRAM: all arithmetic in fp32 with small tiles to fit 99KB budget.
"""
pid = tl.program_id(0)
offs_bt = pid * BLOCK_BT + tl.arange(0, BLOCK_BT)
offs_d = tl.arange(0, CODEBOOK_DIM)
# ββ Load input tile ββ
x_ptrs = x_ptr + offs_bt[:, None] * stride_xb + offs_d[None, :] * stride_xd
x = tl.load(x_ptrs, mask=offs_bt[:, None] < N_CTX, other=0.0)
# ββ Normalize input in fp32 (no keepdims in Triton tl.sum) ββ
x_f32 = x.to(tl.float32)
x_sq = tl.sum(x_f32 * x_f32, axis=1) # [BLOCK_BT]
x_norm_f32 = x_f32 / tl.sqrt(x_sq[:, None] + 1e-8)
# ββ Running argmax over tiled codebook ββ
best_sim = tl.full([BLOCK_BT], -float('inf'), dtype=tl.float32)
best_idx = tl.zeros([BLOCK_BT], dtype=tl.int32)
for k_start in range(0, CODEBOOK_SIZE, TILE_K):
offs_k = k_start + tl.arange(0, TILE_K)
k_mask = offs_k < CODEBOOK_SIZE
# Load codebook tile into fp32 directly for normalization
cb_ptrs = (codebook_ptr
+ offs_k[:, None] * stride_cb
+ offs_d[None, :] * stride_cd)
cb = tl.load(cb_ptrs, mask=k_mask[:, None], other=0.0)
# Normalize codebook tile in fp32
cb_f32 = cb.to(tl.float32)
cb_sq = tl.sum(cb_f32 * cb_f32, axis=1) # [TILE_K]
cb_norm_f32 = cb_f32 / tl.sqrt(cb_sq[:, None] + 1e-8)
# Cosine similarity via tl.dot (tf32 on sm_89)
sim = tl.dot(x_norm_f32, tl.trans(cb_norm_f32)) # [BLOCK_BT, TILE_K]
# Running argmax within this tile
tile_max = tl.max(sim, axis=1)
tile_argmax = tl.argmax(sim, axis=1)
tile_idx = k_start + tile_argmax
# Merge with best across tiles using element-wise mask
update_mask = tile_max > best_sim
best_sim = tl.where(update_mask, tile_max, best_sim)
best_idx = tl.where(update_mask, tile_idx, best_idx)
# ββ Store results ββ
tl.store(indices_ptr + offs_bt, best_idx, mask=offs_bt < N_CTX)
@triton.jit
def _triton_flash_vq_quantize_kernel(
codebook_ptr, indices_ptr, quantized_ptr,
stride_cb, stride_cd,
stride_qb, stride_qd,
N_CTX: tl.constexpr,
CODEBOOK_DIM: tl.constexpr,
BLOCK_BT: tl.constexpr,
):
"""
Gather quantized vectors from codebook at given indices.
Kernel form of: quantized[i] = codebook[indices[i]]
"""
pid = tl.program_id(0)
offs_bt = pid * BLOCK_BT + tl.arange(0, BLOCK_BT)
offs_d = tl.arange(0, CODEBOOK_DIM)
# Load indices for this batch tile
idx = tl.load(indices_ptr + offs_bt, mask=offs_bt < N_CTX, other=0)
# Gather: for each i in BLOCK_BT, load codebook[idx[i], :]
# Pointer arithmetic with broadcasting
gather_ptrs = (codebook_ptr
+ idx[:, None] * stride_cb
+ offs_d[None, :] * stride_cd)
quantized = tl.load(gather_ptrs,
mask=offs_bt[:, None] < N_CTX,
other=0.0)
# Store quantized output
out_ptrs = (quantized_ptr
+ offs_bt[:, None] * stride_qb
+ offs_d[None, :] * stride_qd)
tl.store(out_ptrs, quantized, mask=offs_bt[:, None] < N_CTX)
def _triton_lookup(x, embed, block_bt=None, tile_k=None):
"""
Launch Triton VQ lookup kernel with SRAM-safe tile sizes.
Args:
x: [N, D] input tensor (cuda, contiguous)
embed: [codebook_size, D] codebook (cuda, contiguous)
block_bt: BLOCK_BT tile size (auto-computed if None)
tile_k: TILE_K tile size (auto-computed if None)
Returns:
indices: [N] int64 tensor of argmax indices
"""
N, D = x.shape
codebook_size = embed.shape[0]
assert embed.shape[1] == D, f"Codebook dim {embed.shape[1]} != input dim {D}"
# SRAM-safe tile sizes: kernel uses tf32 (fp32 math), and Triton
# pipelines data through shared memory. Conservative sizing ensures
# fits within ~99KB (sm_89) even with default num_stages=3.
#
# fp32 codebook tile: TILE_K * D * 4 β 128*32*4 = 16KB
# fp32 input tile: BLOCK_BT * D * 4 β 8*32*4 = 1KB
# Accumulator: BLOCK_BT*TILE_K*4 β 8*128*4 = 4KB
# Per stage: ~21KB. With 3 pipeline stages: ~63KB (fits in 99KB).
#
# Larger tiles oversubscribe SRAM (tested: TILE_K=1024 β 321KB needed).
if block_bt is None or tile_k is None:
BLOCK_BT = 8
TILE_K = 128
else:
BLOCK_BT, TILE_K = block_bt, tile_k
grid = (triton.cdiv(N, BLOCK_BT),)
indices = torch.empty(N, dtype=torch.int32, device=x.device)
_triton_flash_vq_lookup_kernel[grid](
x, embed, indices,
x.stride(0), x.stride(1),
embed.stride(0), embed.stride(1),
N, codebook_size, D,
BLOCK_BT=BLOCK_BT, TILE_K=TILE_K,
)
return indices.long()
class _TritonFlashVQFn(torch.autograd.Function):
"""
Custom autograd Function wrapping Triton VQ kernels.
Forward: Triton tiled cosine similarity + argmax lookup
Backward: Rotation trick gradient or straight-through estimator
"""
@staticmethod
def forward(ctx, x_flat, embed, cluster_size, embed_avg,
codebook_size, codebook_dim,
commitment_weight, rotation_trick):
# Triton tiled lookup for indices
with torch.no_grad():
indices = _triton_lookup(x_flat.contiguous(), embed.contiguous())
quantized = embed[indices]
commitment_loss = commitment_weight * F.mse_loss(x_flat.float(), quantized.detach().float())
# Clone saved tensors to avoid version conflicts with in-place EMA updates
ctx.save_for_backward(
x_flat.detach().clone(),
quantized.detach().clone(),
embed.detach().clone(),
)
ctx.codebook_dim = codebook_dim
ctx.rotation_trick = rotation_trick
return quantized, indices, commitment_loss
@staticmethod
def backward(ctx, grad_quantized, grad_indices, grad_commitment):
x_flat, quantized, embed = ctx.saved_tensors
if ctx.rotation_trick:
# Rotation trick gradient
x_norm = F.normalize(x_flat.float(), dim=-1)
q_norm = F.normalize(quantized.float(), dim=-1)
diff = x_norm - q_norm
proj = (grad_quantized.float() * x_norm).sum(dim=-1, keepdim=True)
grad_x = grad_quantized.float() - proj * diff
else:
# Straight-through estimator
grad_x = grad_quantized.float()
return grad_x.to(grad_quantized.dtype), None, None, None, None, None, None, None
# When Triton is not available, define a fallback lookup
if not _HAS_TRITON:
def _triton_lookup(x, embed):
"""Fallback: torch-based cosine similarity lookup (CPU or CUDA without Triton)."""
with torch.no_grad():
x_norm = F.normalize(x.float(), dim=-1)
embed_norm = F.normalize(embed.float(), dim=-1)
sim = x_norm @ embed_norm.T
indices = sim.argmax(dim=-1)
return indices
|