Spaces:
Paused
Paused
File size: 18,836 Bytes
8e8a34f | 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 | """Lockstep batched branch decoding for the Transformers backend."""
from __future__ import annotations
from collections import deque
from collections.abc import Iterator
from dataclasses import dataclass, field
from typing import Any, Protocol
import torch
import torch.nn.functional as F
class _Engine(Protocol):
model: Any
processor: Any
tokenizer: Any
device: torch.device
eos_token_id: int
fork_token_map: dict[int, list[int]]
max_new_tokens: int
max_branch_tokens: int
max_concurrent_branches: int
max_total_branches: int
strict: bool
_branch_forbidden_ids: list[int]
def _position_ids(
self,
logical_positions: torch.LongTensor,
*,
multimodal: bool,
) -> torch.LongTensor: ...
def _next_token(self, logits: torch.Tensor, *, branch: bool) -> int: ...
CacheState = list[dict[str, Any]]
@dataclass(slots=True)
class _Stream:
past: CacheState
cache_len: int
next_input: int | None
branch_index: int | None = None
fork_position: int | None = None
injected_tokens: list[int] = field(default_factory=list)
generated_tokens: list[int] = field(default_factory=list)
state: str = "active"
def _cache_to_state(cache: Any) -> CacheState:
"""Extract Qwen DynamicCache tensors into independently batchable state."""
try:
from transformers.cache_utils import LinearAttentionLayer
except ImportError as exc: # pragma: no cover - pinned Transformers provides it
raise RuntimeError("Parallel decoding requires Transformers DynamicCache support.") from exc
if not hasattr(cache, "layers"):
raise TypeError(
"Parallel decoding requires a Transformers DynamicCache; "
f"received {type(cache).__name__}."
)
state: CacheState = []
for layer in cache.layers:
if isinstance(layer, LinearAttentionLayer):
if not layer.is_conv_states_initialized or not layer.is_recurrent_states_initialized:
raise RuntimeError("A linear-attention cache layer was not initialized by prefill.")
state.append(
{
"kind": "linear",
"conv": layer.conv_states.contiguous(),
"recur": layer.recurrent_states.contiguous(),
"has_previous_state": bool(layer.has_previous_state),
}
)
continue
keys = getattr(layer, "keys", None)
values = getattr(layer, "values", None)
if keys is None or values is None:
raise TypeError(
"Parallel decoding only supports initialized dynamic or linear cache layers; "
f"received {type(layer).__name__}."
)
state.append(
{
"kind": "full",
"keys": keys.contiguous(),
"values": values.contiguous(),
}
)
return state
def _clone_state(state: CacheState, *, truncate_to: int | None = None) -> CacheState:
cloned: CacheState = []
for layer in state:
if layer["kind"] == "full":
keys = layer["keys"]
values = layer["values"]
if truncate_to is not None:
keys = keys[:, :, :truncate_to, :]
values = values[:, :, :truncate_to, :]
cloned.append(
{
"kind": "full",
"keys": keys.contiguous().clone(),
"values": values.contiguous().clone(),
}
)
else:
cloned.append(
{
"kind": "linear",
"conv": layer["conv"].contiguous().clone(),
"recur": layer["recur"].contiguous().clone(),
"has_previous_state": layer["has_previous_state"],
}
)
return cloned
def _state_to_cache(state: CacheState):
from transformers.cache_utils import DynamicCache, DynamicLayer, LinearAttentionLayer
cache = DynamicCache()
layers = []
for entry in state:
if entry["kind"] == "full":
layer = DynamicLayer()
layer.update(entry["keys"], entry["values"])
else:
layer = LinearAttentionLayer()
conv = entry["conv"]
recur = entry["recur"]
layer.lazy_initialization(conv_states=conv, recurrent_states=recur)
layer.conv_states.copy_(conv)
layer.recurrent_states.copy_(recur)
layer.has_previous_state = entry["has_previous_state"]
layers.append(layer)
cache.layers = layers
return cache
def _batch_cache(streams: list[_Stream], max_len: int):
from transformers.cache_utils import DynamicCache, DynamicLayer, LinearAttentionLayer
cache = DynamicCache()
layers = []
for layer_index in range(len(streams[0].past)):
first = streams[0].past[layer_index]
if first["kind"] == "full":
keys = []
values = []
for stream in streams:
entry = stream.past[layer_index]
pad = max_len - stream.cache_len
keys.append(F.pad(entry["keys"], (0, 0, pad, 0)))
values.append(F.pad(entry["values"], (0, 0, pad, 0)))
layer = DynamicLayer()
layer.update(torch.cat(keys, dim=0), torch.cat(values, dim=0))
else:
conv = torch.cat([stream.past[layer_index]["conv"] for stream in streams], dim=0)
recur = torch.cat(
[stream.past[layer_index]["recur"] for stream in streams], dim=0
)
layer = LinearAttentionLayer()
layer.lazy_initialization(conv_states=conv, recurrent_states=recur)
layer.conv_states.copy_(conv)
layer.recurrent_states.copy_(recur)
layer.has_previous_state = any(
stream.past[layer_index]["has_previous_state"] for stream in streams
)
layers.append(layer)
cache.layers = layers
return cache
def _split_batch_cache(cache: Any, streams: list[_Stream]) -> None:
from transformers.cache_utils import LinearAttentionLayer
new_lengths = [stream.cache_len + 1 for stream in streams]
for batch_index, stream in enumerate(streams):
state: CacheState = []
for layer in cache.layers:
if isinstance(layer, LinearAttentionLayer):
state.append(
{
"kind": "linear",
"conv": layer.conv_states[batch_index : batch_index + 1].contiguous(),
"recur": layer.recurrent_states[
batch_index : batch_index + 1
].contiguous(),
"has_previous_state": bool(layer.has_previous_state),
}
)
continue
length = new_lengths[batch_index]
state.append(
{
"kind": "full",
"keys": layer.keys[
batch_index : batch_index + 1, :, -length:, :
].contiguous(),
"values": layer.values[
batch_index : batch_index + 1, :, -length:, :
].contiguous(),
}
)
stream.past = state
stream.cache_len = new_lengths[batch_index]
def _scheduler_event(
main: _Stream,
branches: list[_Stream],
*,
phase: str,
batch_size: int = 0,
) -> dict[str, Any]:
return {
"type": "scheduler",
"execution_mode": "parallel",
"phase": phase,
"main_active": main.state == "active",
"active_branches": sum(branch.state == "active" for branch in branches),
"queued_branches": sum(branch.state == "queued" for branch in branches),
"completed_branches": sum(branch.state == "done" for branch in branches),
"batch_size": batch_size,
}
def stream_parallel(engine: _Engine, messages: list[dict[str, Any]]) -> Iterator[dict[str, Any]]:
"""Decode main and active branches together in a per-step GPU batch."""
with torch.inference_mode():
yield from _stream_parallel(engine, messages)
def _stream_parallel(engine: _Engine, messages: list[dict[str, Any]]) -> Iterator[dict[str, Any]]:
yield {"type": "accepted", "execution_mode": "parallel"}
inputs = engine.processor.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_dict=True,
return_tensors="pt",
).to(engine.device)
prompt_ids = inputs["input_ids"][0].tolist()
prompt_length = len(prompt_ids)
multimodal = inputs.get("image_grid_thw") is not None or inputs.get("video_grid_thw") is not None
output = engine.model(**inputs, use_cache=True)
first_token = engine._next_token(output.logits, branch=False)
main = _Stream(
past=_cache_to_state(output.past_key_values),
cache_len=prompt_length,
next_input=first_token,
generated_tokens=[first_token],
)
del output
if first_token == engine.eos_token_id or engine.max_new_tokens == 1:
main.state = "done"
branches: list[_Stream] = []
pending: deque[int] = deque()
active: set[int] = set()
dropped_triggers = 0
peak_batch_size = 1
yield _scheduler_event(main, branches, phase="prefill_complete", batch_size=1)
yield {
"type": "main",
"token_ids": [first_token],
"delta_text": engine.tokenizer.decode([first_token], skip_special_tokens=False),
"total": 1,
}
if main.state == "done":
yield {"type": "main_done", "total": 1}
def register_fork(token_id: int) -> dict[str, Any] | None:
nonlocal dropped_triggers
injected = engine.fork_token_map.get(token_id)
if not injected:
return None
if len(branches) >= engine.max_total_branches:
dropped_triggers += 1
return None
fork_position = main.cache_len
expected_position = prompt_length + len(main.generated_tokens) - 1
if fork_position != expected_position:
raise RuntimeError(
"Main cache is not aligned with its fork token: "
f"cache={fork_position}, token={expected_position}."
)
branch_index = len(branches)
injected_tokens = list(injected)
if not injected_tokens:
raise ValueError("A fork target must contain at least one token.")
branches.append(
_Stream(
past=_clone_state(main.past, truncate_to=fork_position),
cache_len=fork_position,
next_input=None,
branch_index=branch_index,
fork_position=fork_position,
injected_tokens=injected_tokens,
state="queued",
)
)
pending.append(branch_index)
return {
"type": "fork",
"branch_index": branch_index,
"fork_position": fork_position,
"trigger_token_id": token_id,
"injected_token_ids": injected_tokens,
"injected_text": engine.tokenizer.decode(
injected_tokens, skip_special_tokens=False
),
"branch_state": "queued",
}
fork_event = register_fork(first_token)
if fork_event is not None:
yield fork_event
def activate_pending() -> Iterator[dict[str, Any]]:
while pending and len(active) < engine.max_concurrent_branches:
branch_index = pending.popleft()
branch = branches[branch_index]
branch.state = "active"
active.add(branch_index)
injected = torch.tensor(
[branch.injected_tokens], dtype=torch.long, device=engine.device
)
logical = torch.arange(
branch.fork_position,
branch.fork_position + len(branch.injected_tokens),
dtype=torch.long,
device=engine.device,
).unsqueeze(0)
output = engine.model(
input_ids=injected,
position_ids=engine._position_ids(logical, multimodal=multimodal),
past_key_values=_state_to_cache(branch.past),
use_cache=True,
)
branch.past = _cache_to_state(output.past_key_values)
branch.cache_len += len(branch.injected_tokens)
first_branch_token = engine._next_token(output.logits, branch=True)
branch.generated_tokens.append(first_branch_token)
branch.next_input = first_branch_token
del output
yield {
"type": "branch",
"branch_index": branch_index,
"fork_position": branch.fork_position,
"token_ids": [first_branch_token],
"delta_text": engine.tokenizer.decode(
[first_branch_token], skip_special_tokens=False
),
"total": len(branch.injected_tokens) + 1,
}
if (
first_branch_token == engine.eos_token_id
or len(branch.generated_tokens) >= engine.max_branch_tokens
):
branch.state = "done"
active.discard(branch_index)
yield {
"type": "branch_done",
"branch_index": branch_index,
"total": len(branch.injected_tokens) + len(branch.generated_tokens),
}
yield _scheduler_event(main, branches, phase="branch_started")
yield from activate_pending()
while main.state == "active" or active or pending:
yield from activate_pending()
streams: list[_Stream] = []
if main.state == "active":
streams.append(main)
streams.extend(branches[index] for index in sorted(active))
if not streams:
continue
batch_size = len(streams)
peak_batch_size = max(peak_batch_size, batch_size)
yield _scheduler_event(main, branches, phase="decoding", batch_size=batch_size)
max_len = max(stream.cache_len for stream in streams)
input_ids = torch.tensor(
[[stream.next_input] for stream in streams],
dtype=torch.long,
device=engine.device,
)
logical_positions = torch.tensor(
[[stream.cache_len] for stream in streams],
dtype=torch.long,
device=engine.device,
)
attention_mask = torch.zeros(
len(streams), max_len + 1, dtype=torch.long, device=engine.device
)
for row, stream in enumerate(streams):
attention_mask[row, max_len - stream.cache_len :] = 1
output = engine.model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=engine._position_ids(logical_positions, multimodal=multimodal),
past_key_values=_batch_cache(streams, max_len),
use_cache=True,
)
logits = output.logits[:, -1, :]
if engine.strict and engine._branch_forbidden_ids:
logits = logits.clone()
forbidden = torch.tensor(
engine._branch_forbidden_ids, dtype=torch.long, device=logits.device
)
for row, stream in enumerate(streams):
if stream.branch_index is not None:
logits[row, forbidden] = torch.finfo(logits.dtype).min
token_ids = logits.argmax(dim=-1).cpu().tolist()
_split_batch_cache(output.past_key_values, streams)
del output, logits
new_forks: list[dict[str, Any]] = []
for stream, token_id_raw in zip(streams, token_ids, strict=True):
token_id = int(token_id_raw)
stream.generated_tokens.append(token_id)
stream.next_input = token_id
if stream.branch_index is None:
yield {
"type": "main",
"token_ids": [token_id],
"delta_text": engine.tokenizer.decode(
[token_id], skip_special_tokens=False
),
"total": len(main.generated_tokens),
}
fork_event = register_fork(token_id)
if fork_event is not None:
new_forks.append(fork_event)
if (
token_id == engine.eos_token_id
or len(main.generated_tokens) >= engine.max_new_tokens
):
main.state = "done"
yield {"type": "main_done", "total": len(main.generated_tokens)}
continue
branch_index = stream.branch_index
yield {
"type": "branch",
"branch_index": branch_index,
"fork_position": stream.fork_position,
"token_ids": [token_id],
"delta_text": engine.tokenizer.decode([token_id], skip_special_tokens=False),
"total": len(stream.injected_tokens) + len(stream.generated_tokens),
}
if (
token_id == engine.eos_token_id
or len(stream.generated_tokens) >= engine.max_branch_tokens
):
stream.state = "done"
active.discard(branch_index)
yield {
"type": "branch_done",
"branch_index": branch_index,
"total": len(stream.injected_tokens) + len(stream.generated_tokens),
}
yield from new_forks
result_branches = []
for branch in branches:
token_ids = branch.injected_tokens + branch.generated_tokens
result_branches.append(
{
"branch_index": branch.branch_index,
"fork_position": branch.fork_position,
"injected_token_ids": branch.injected_tokens,
"token_ids": token_ids,
"text": engine.tokenizer.decode(token_ids, skip_special_tokens=False),
}
)
yield {
"type": "done",
"execution_mode": "parallel",
"main": engine.tokenizer.decode(main.generated_tokens, skip_special_tokens=False),
"main_token_ids": main.generated_tokens,
"prompt_tokens": prompt_length,
"branches": result_branches,
"branch_limit_reached": dropped_triggers > 0,
"dropped_branch_triggers": dropped_triggers,
"peak_batch_size": peak_batch_size,
}
|