File size: 11,132 Bytes
d91766b | 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 | """DllmBlock and DllmBlockBuffer - extracted to avoid circular import with mixin."""
from __future__ import annotations
import weakref
import torch
from dataclasses import dataclass, field
from diffulex.config import DecodingThresholds
from diffulex.engine.status import DllmBlockStatus, DllmBlockType
from typing import TYPE_CHECKING
if TYPE_CHECKING:
from diffulex.engine.request import DllmReq
from diffulex.engine.dllm_block import DllmBlockBuffer
weakref_fn = lambda x: weakref.ref(x) if x is not None else None
@dataclass
class DllmBlock:
block_id: int
start: int
end: int
block_size: int
mask_token_id: int
thresholds: DecodingThresholds
status: DllmBlockStatus = None
prev_block: "DllmBlock" = None
block_type: DllmBlockType = DllmBlockType.IN_CONTEXT
editable_start: int = 0
commit_ready: bool = False
same_as_previous: bool = False
same_token_ratio: float = 0.0
all_confident: bool = False
post_edit_steps: int = 0
total_steps: int = 0
def __repr__(self):
prev_block_id = self.prev_block.block_id if self.prev_block is not None else None
return f"DllmBlock(block_id={self.block_id}, start={self.start}, end={self.end}, block_size={self.block_size}, mask_token_id={self.mask_token_id}, thresholds={self.thresholds}, status={self.status}, prev_block_id={prev_block_id}, is_last_in_context={self.is_last_in_context})"
def post_init_dllm_block(self, req: "DllmReq", dllm_block_buffer: "DllmBlockBuffer"):
assert self.end - self.start == self.block_size
if not 0 <= int(self.editable_start) <= self.block_size:
raise ValueError(
f"editable_start must be in [0, {self.block_size}], got: {self.editable_start}"
)
if req is not None:
self.bind_req(req)
if dllm_block_buffer is not None:
self.bind_buffer(dllm_block_buffer)
if self.status is None:
self.status = DllmBlockStatus.TO_CACHE if self.is_complete else DllmBlockStatus.ACTIVE
if self.is_complete:
self.commit_ready = True
self.make_in_context()
def bind_req(self, req: "DllmReq" | None):
self._req = weakref_fn(req)
def bind_buffer(self, dllm_block_buffer: "DllmBlockBuffer" | None):
self._dllm_block_buffer = weakref_fn(dllm_block_buffer)
def __getstate__(self):
state = self.__dict__.copy()
state.pop("_req", None)
state.pop("_dllm_block_buffer", None)
return state
def __setstate__(self, state):
self.__dict__.update(state)
@property
def req(self) -> "DllmReq":
ref = getattr(self, "_req", None)
return ref() if ref else None
@property
def dllm_block_buffer(self) -> "DllmBlockBuffer":
ref = getattr(self, "_dllm_block_buffer", None)
return ref() if ref else None
@property
def rel_page_id(self) -> int:
return self.start // self.req.page_size
@property
def token_ids(self) -> list[int]:
return self.req[self.start : self.end]
@property
def editable_relative_ids(self) -> list[int]:
return list(range(int(self.editable_start), self.block_size))
@property
def mask_token_relative_ids(self) -> list[int]:
editable_start = int(self.editable_start)
return [
i
for i, token_id in enumerate(self.token_ids)
if i >= editable_start and token_id == self.mask_token_id
]
@property
def mask_token_global_ids(self) -> list[int]:
editable_start = int(self.editable_start)
return [
i + self.start
for i, token_id in enumerate(self.token_ids)
if i >= editable_start and token_id == self.mask_token_id
]
@property
def in_buffer_block_id(self) -> int:
return self.dllm_block_buffer.block_ids.index(self.block_id)
@property
def num_mask_tokens(self):
return sum(token_id == self.mask_token_id for token_id in self.token_ids)
@property
def progress(self):
return (self.block_size - self.num_mask_tokens) / self.block_size
@property
def is_complete(self):
return self.progress == 1.0
@property
def is_semi_complete(self):
return self.progress >= self.thresholds.semi_complete_threshold
@property
def should_force_decode_topk(self):
return self.prev_block is not None and self.prev_block.is_semi_complete
@property
def should_add_block(self):
return (
self.progress >= self.thresholds.add_block_threshold
and self.same_token_ratio >= self.thresholds.token_stability_threshold
and not self.is_last_in_context
)
@property
def is_dummy(self):
return self.status == DllmBlockStatus.DUMMY
@property
def is_active(self):
return self.status == DllmBlockStatus.ACTIVE
@property
def is_to_cache(self):
return self.status == DllmBlockStatus.TO_CACHE
@property
def is_in_cache(self):
return self.status == DllmBlockStatus.IN_CACHE
@property
def is_in_context(self):
return self.block_type == DllmBlockType.IN_CONTEXT
@property
def is_out_of_context(self):
return self.block_type == DllmBlockType.OUT_OF_CONTEXT
@property
def is_last_in_context(self):
return self.block_type == DllmBlockType.LAST_IN_CONTEXT
def write_token(self, token_id: int, rel_idx: int):
if int(rel_idx) < int(self.editable_start):
raise ValueError(
f"Cannot write non-editable token in block {self.block_id}: "
f"rel_idx={rel_idx}, editable_start={self.editable_start}"
)
self.req.token_ids[self.start + rel_idx] = token_id
self.commit_ready = False
def write_tokens_parallel(self, token_ids: torch.Tensor, abs_ids: torch.Tensor):
token_ids_list = token_ids.tolist() if isinstance(token_ids, torch.Tensor) else list(token_ids)
abs_ids_list = abs_ids.tolist() if isinstance(abs_ids, torch.Tensor) else list(abs_ids)
for abs_idx, token_id in zip(abs_ids_list, token_ids_list):
if int(abs_idx) - self.start < int(self.editable_start):
raise ValueError(
f"Cannot write non-editable token in block {self.block_id}: "
f"abs_idx={abs_idx}, editable_start={self.editable_start}"
)
self.req.token_ids[int(abs_idx)] = int(token_id)
if abs_ids_list:
self.commit_ready = False
def to_cache(self):
if self.is_active:
self.status = DllmBlockStatus.TO_CACHE
def in_cache(self):
if self.is_to_cache:
self.status = DllmBlockStatus.IN_CACHE
def make_in_context(self):
self.block_type = DllmBlockType.IN_CONTEXT
def make_out_of_context(self):
self.block_type = DllmBlockType.OUT_OF_CONTEXT
def make_last_in_context(self):
if self.is_in_context:
self.block_type = DllmBlockType.LAST_IN_CONTEXT
@dataclass
class DllmBlockBuffer:
buffer_size: int
dllm_blocks: list[DllmBlock] = field(default_factory=list)
def __repr__(self):
return f"DllmBlockBuffer(buffer_size={self.buffer_size}, dllm_blocks={self.dllm_blocks})"
def post_init_dllm_block_buffer(self, req: "DllmReq"):
assert len(self.dllm_blocks) == self.buffer_size
if req is not None:
self.bind_req(req)
if len(self.dllm_blocks) > 0:
for block in self.dllm_blocks:
block.post_init_dllm_block(None, self)
def bind_req(self, req: "DllmReq" | None):
self._req = weakref_fn(req)
def __getstate__(self):
state = self.__dict__.copy()
state.pop("_req", None)
return state
def __setstate__(self, state):
self.__dict__.update(state)
@property
def req(self) -> "DllmReq":
ref = getattr(self, "_req", None)
return ref() if ref else None
@property
def buffer_sequence(self) -> list[int]:
return self.req[self.dllm_blocks[0].start : self.dllm_blocks[-1].end]
@property
def buffer_position_ids(self) -> list[int]:
return list(range(self.dllm_blocks[0].start, self.dllm_blocks[-1].end))
@property
def block_ids(self) -> list[int]:
return [block.block_id for block in self.dllm_blocks]
@property
def cursor_slot_idx(self) -> int:
return len(self.valid_blocks)
@property
def valid_blocks(self) -> list[DllmBlock]:
return [block for block in self.dllm_blocks if not block.is_dummy]
@property
def dummy_blocks(self) -> list[DllmBlock]:
return [block for block in self.dllm_blocks if block.is_dummy]
@property
def active_blocks(self) -> list[DllmBlock]:
return [block for block in self.dllm_blocks if block.is_active]
@property
def to_cache_blocks(self) -> list[DllmBlock]:
return [block for block in self.dllm_blocks if block.is_to_cache]
@property
def in_cache_blocks(self) -> list[DllmBlock]:
return [block for block in self.dllm_blocks if block.is_in_cache]
@property
def first_running_block(self) -> DllmBlock:
return self.dllm_blocks[0]
@property
def last_running_block(self) -> DllmBlock:
return self.dllm_blocks[-1]
@property
def first_valid_block(self) -> DllmBlock:
return self.dllm_blocks[0]
@property
def last_valid_block(self) -> DllmBlock:
return self.dllm_blocks[self.cursor_slot_idx - 1]
@property
def first_to_cache_block(self) -> DllmBlock:
return self.to_cache_blocks[0]
@property
def last_to_cache_block(self) -> DllmBlock:
return self.to_cache_blocks[-1]
@property
def slot_block(self) -> DllmBlock:
return self.dllm_blocks[self.cursor_slot_idx]
@property
def num_valid_blocks(self) -> int:
return self.cursor_slot_idx
@property
def num_running_blocks(self) -> int:
return self.buffer_size
@property
def should_add_block(self) -> bool:
return self.last_valid_block.should_add_block
@property
def is_overflow(self) -> bool:
return self.cursor_slot_idx >= self.buffer_size
@property
def prev_step_popped(self) -> bool:
if len(self.dllm_blocks) < 2:
return False
return self.dllm_blocks[-1].block_id == self.dllm_blocks[-2].block_id
def push_back(self, block: DllmBlock):
self.dllm_blocks[-1] = block
def pop_front(self):
for i in range(0, self.buffer_size - 1):
self.dllm_blocks[i] = self.dllm_blocks[i + 1]
def activate_cursor_slot_block(self):
self.slot_block.status = DllmBlockStatus.ACTIVE
def maybe_fix_context_management(self):
if self.first_valid_block.is_dummy and self.first_valid_block.is_last_in_context:
self.first_valid_block.prev_block.make_last_in_context()
|