File size: 17,588 Bytes
31dc8dc | 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 | import os
from dataclasses import dataclass, field
from transformers import AutoConfig
from diffulex.distributed.parallel_state import get_world_size
from diffulex.logger import get_logger
logger = get_logger(__name__)
SUPPORTED_PAGE_BLOCK_SIZES = (4, 8, 16, 32)
EDIT_SAMPLING_MODEL_NAMES = {
"llada2",
"llada2_moe",
"llada2_mini",
"llada2dot1_mini",
"llada2_mini_dmax",
}
DMAX_MODEL_NAMES = EDIT_SAMPLING_MODEL_NAMES
@dataclass
class DecodingThresholds:
add_block_threshold: float # whether add a new block
semi_complete_threshold: float # whether unleash the decoding of the next block
accept_threshold: float # whether the token should be decoded (M2T)
edit_threshold: float = 0.0 # whether a filled token should be edited (T2T)
remask_threshold: float = 0.4 # whether a filled token should be remasked
token_stability_threshold: float = 0.0 # whether decoded tokens are stable enough to add a new block
@dataclass
class Config:
model: str
lora_path: str = ""
model_name: str = "dream"
decoding_strategy: str = "d2f" # "d2f", "multi_bd"
# Sampling Harness
hf_config: AutoConfig | None = None
tokenizer_vocab_size: int | None = None
eos: int = -1
mask_token_id: int = 151666
block_size: int = 32
buffer_size: int = 4
multi_block_prefix_full: bool = True
token_merge_mode: str = "dmax_topk" # "dmax_topk" or "iter_smooth_topk"
token_merge_top_k: int = 1
token_merge_renormalize: bool = True
token_merge_weight: float = 1.0
sampling_mode: str = "naive" # "naive" or "edit"
max_post_edit_steps: int = 16 # max refinement steps after all masks filled (JointThreshold)
penalty_lambda: float = 0.0 # repetition penalty coefficient (JointThreshold)
use_lora: bool = False
pre_merge_lora: bool = False
max_num_batched_tokens: int = 4096
max_num_reqs: int = 128
max_model_len: int = 2048
gpu_memory_utilization: float = 0.9
decoding_thresholds: DecodingThresholds | dict | None = None
# TODO: Should be deprecated in the future
add_block_threshold: float | None = None
semi_complete_threshold: float | None = None
accept_threshold: float | None = None
remask_threshold: float | None = None
token_stability_threshold: float | None = None
# Truncation
auto_max_nfe_warmup_steps: int = 8
auto_max_nfe_tpf_floor: float = 1.0
# Parallelism
data_parallel_size: int = 1
tensor_parallel_size: int = 2
expert_parallel_size: int = 1
master_addr: str = "localhost"
master_port: int = 2333
distributed_timeout_seconds: int = 600
shm_name: str = "diffulex_shm"
device_start: int = 0
device_ids: list[int] = field(default_factory=lambda: [])
# CUDA Graph
enforce_eager: bool = False
attn_impl: str = "triton" # "triton" or "naive"
enable_prefill_cudagraph: bool = True
enable_full_static_runner: bool = True
prefill_cudagraph_max_len: int = 0
enable_torch_compile: bool = True
enable_cudagraph_torch_compile: bool = False
torch_compile_mode: str = "reduce-overhead"
# MoE
moe_dispatcher_backend: str = "standard" # "standard", "naive", or "deepep"
moe_gemm_impl: str = "triton" # "triton", "vllm", or "naive"
deepep_mode: str = "auto" # "normal", "low_latency", or "auto"
deepep_num_max_dispatch_tokens_per_rank: int = 256
# KV Cache Page Table
page_size: int = 32
enable_prefix_caching: bool = True
num_pages: int = -1
k_cache_hdim_split_factor_x: int = 8
kv_cache_layout: str = "unified" # "unified" or "distinct"
def _validate_sampling_mode(self) -> None:
if self.sampling_mode == "edit" and self.model_name not in EDIT_SAMPLING_MODEL_NAMES:
allowed = ", ".join(sorted(EDIT_SAMPLING_MODEL_NAMES))
raise ValueError(
f"sampling_mode='edit' is only supported for model_name in {{{allowed}}}, "
f"got: {self.model_name}"
)
if self.model_name == "llada2dot1_mini" and self.sampling_mode != "edit":
raise ValueError("model_name='llada2dot1_mini' requires sampling_mode='edit'.")
if self.decoding_strategy == "dmax":
allowed = ", ".join(sorted(DMAX_MODEL_NAMES))
if self.model_name not in DMAX_MODEL_NAMES:
raise ValueError(
f"decoding_strategy='dmax' is only supported for model_name in {{{allowed}}}, "
f"got: {self.model_name}"
)
if self.sampling_mode != "edit":
raise ValueError("decoding_strategy='dmax' requires sampling_mode='edit'.")
def __post_init__(self):
if not os.path.isdir(self.model):
raise ValueError(f"model must be an existing directory, got: {self.model}")
if self.decoding_strategy == "d2f":
if not self.multi_block_prefix_full:
logger.warning("Forcing multi_block_prefix_full=True for decoding_strategy=d2f.")
if self.enable_prefix_caching:
logger.warning("Disabling prefix caching for decoding_strategy=d2f.")
self.multi_block_prefix_full = True
self.enable_prefix_caching = False
elif self.decoding_strategy in ("multi_bd", "dmax"):
if self.multi_block_prefix_full:
logger.warning(f"Forcing multi_block_prefix_full=False for decoding_strategy={self.decoding_strategy}.")
self.multi_block_prefix_full = False
if self.enable_prefix_caching:
logger.info(f"Enabling prefix caching for decoding_strategy={self.decoding_strategy}.")
if self.page_size not in SUPPORTED_PAGE_BLOCK_SIZES:
raise ValueError(
f"page_size must be one of {SUPPORTED_PAGE_BLOCK_SIZES}, got: {self.page_size}"
)
if self.block_size not in SUPPORTED_PAGE_BLOCK_SIZES:
raise ValueError(
f"block_size must be one of {SUPPORTED_PAGE_BLOCK_SIZES}, got: {self.block_size}"
)
if self.block_size > self.page_size:
raise ValueError(
"block_size must be <= page_size, "
f"got: block_size={self.block_size}, page_size={self.page_size}"
)
if not 1 <= self.tensor_parallel_size <= 8:
raise ValueError(
"tensor_parallel_size must be in [1, 8], "
f"got: {self.tensor_parallel_size}"
)
if not 1 <= self.data_parallel_size <= 1024:
raise ValueError(
"data_parallel_size must be in [1, 1024], "
f"got: {self.data_parallel_size}"
)
if not 1 <= self.expert_parallel_size <= 32768:
raise ValueError(
"expert_parallel_size must be in [1, 32768], "
f"got: {self.expert_parallel_size}"
)
if self.token_merge_top_k <= 0:
raise ValueError(f"token_merge_top_k must be positive, got: {self.token_merge_top_k}")
if self.token_merge_mode not in ("dmax_topk", "iter_smooth_topk"):
raise ValueError(
"token_merge_mode must be one of {'dmax_topk', 'iter_smooth_topk'}, "
f"got: {self.token_merge_mode}"
)
if self.token_merge_weight < 0:
raise ValueError(f"token_merge_weight must be non-negative, got: {self.token_merge_weight}")
if self.sampling_mode not in ("naive", "edit"):
raise ValueError(
"sampling_mode must be one of {'naive', 'edit'}, "
f"got: {self.sampling_mode}"
)
self._validate_sampling_mode()
if self.kv_cache_layout not in {"unified", "distinct"}:
raise ValueError(
"kv_cache_layout must be one of {'unified', 'distinct'}, "
f"got: {self.kv_cache_layout}"
)
if self.attn_impl not in {"triton", "naive"}:
raise ValueError(
"attn_impl must be one of {'triton', 'naive'}, "
f"got: {self.attn_impl}"
)
if self.moe_dispatcher_backend not in {"standard", "naive", "deepep"}:
raise ValueError(
"moe_dispatcher_backend must be one of {'standard', 'naive', 'deepep'}, "
f"got: {self.moe_dispatcher_backend}"
)
if self.moe_gemm_impl not in {"triton", "vllm", "vllm_modular", "naive"}:
raise ValueError(
"moe_gemm_impl must be one of {'triton', 'vllm', 'vllm_modular', 'naive'}, "
f"got: {self.moe_gemm_impl}"
)
if self.deepep_mode not in {"normal", "low_latency", "auto"}:
raise ValueError(
"deepep_mode must be one of {'normal', 'low_latency', 'auto'}, "
f"got: {self.deepep_mode}"
)
if self.deepep_num_max_dispatch_tokens_per_rank <= 0:
raise ValueError(
"deepep_num_max_dispatch_tokens_per_rank must be positive, "
f"got: {self.deepep_num_max_dispatch_tokens_per_rank}"
)
if self.expert_parallel_size != 1 or self.moe_dispatcher_backend != "standard":
raise ValueError(
"MoE A2A backends are currently unsupported. Require expert_parallel_size == 1 "
"and moe_dispatcher_backend == 'standard', "
f"got expert_parallel_size={self.expert_parallel_size}, "
f"moe_dispatcher_backend={self.moe_dispatcher_backend}."
)
if not (isinstance(self.master_port, int) and 0 < self.master_port < 65536):
raise ValueError(
"master_port must be an int in (0, 65536), "
f"got: {self.master_port}"
)
if not (
isinstance(self.distributed_timeout_seconds, int)
and self.distributed_timeout_seconds > 0
):
raise ValueError(
"distributed_timeout_seconds must be a positive int, "
f"got: {self.distributed_timeout_seconds}"
)
if not (isinstance(self.device_start, int) and self.device_start >= 0):
raise ValueError(
"device_start must be a non-negative int, "
f"got: {self.device_start}"
)
# LoRA validation
if self.use_lora:
if not self.lora_path:
raise ValueError("lora_path must be provided when use_lora is True")
if not os.path.exists(self.lora_path):
logger.warning(f"LoRA path {self.lora_path} does not exist")
self.hf_config = AutoConfig.from_pretrained(self.model, trust_remote_code=True)
for name in (
"attn_impl",
"moe_dispatcher_backend",
"moe_gemm_impl",
"deepep_mode",
"deepep_num_max_dispatch_tokens_per_rank",
"expert_parallel_size",
"tensor_parallel_size",
"data_parallel_size",
"mask_token_id",
):
setattr(self.hf_config, name, getattr(self, name))
cfg_max_model_len = (
self.hf_config.max_position_embeddings
if hasattr(self.hf_config, "max_position_embeddings")
else self.hf_config.max_sequence_length
)
self.max_model_len = min(self.max_model_len, cfg_max_model_len)
if self.max_num_batched_tokens < self.max_model_len:
raise ValueError(
"max_num_batched_tokens must be >= max_model_len after HF config clamp, "
f"got max_num_batched_tokens={self.max_num_batched_tokens}, "
f"max_model_len={self.max_model_len}"
)
if self.prefill_cudagraph_max_len < 0:
raise ValueError(
"prefill_cudagraph_max_len must be non-negative, "
f"got: {self.prefill_cudagraph_max_len}"
)
if self.auto_max_nfe_warmup_steps <= 0:
raise ValueError(
"auto_max_nfe_warmup_steps must be positive, "
f"got: {self.auto_max_nfe_warmup_steps}"
)
if self.auto_max_nfe_tpf_floor <= 0:
raise ValueError(
"auto_max_nfe_tpf_floor must be positive, "
f"got: {self.auto_max_nfe_tpf_floor}"
)
if not self.device_ids:
import torch
# When CUDA_VISIBLE_DEVICES is set, PyTorch maps physical devices to logical device 0, 1, ...
# So we should use logical device indices (0, 1, ...) instead of physical device IDs
self.device_ids = list(range(torch.cuda.device_count()))
logger.info(f"Using CUDA devices: {self.device_ids}")
required_world_size = get_world_size(
self.tensor_parallel_size,
self.expert_parallel_size,
self.data_parallel_size,
)
if len(self.device_ids) < required_world_size:
raise ValueError(
"Requested parallel world size exceeds available CUDA devices, "
f"required={required_world_size}, device_ids={self.device_ids}."
)
# Build decoding_thresholds: dict or flat keys -> DecodingThresholds
d = self.__dict__
if isinstance(self.decoding_thresholds, dict):
for key in (
"add_block_threshold",
"semi_complete_threshold",
"accept_threshold",
"edit_threshold",
"remask_threshold",
"token_stability_threshold",
):
if d.get(key) is not None:
self.decoding_thresholds[key] = d[key]
if "edit_threshold" not in self.decoding_thresholds:
self.decoding_thresholds["edit_threshold"] = 0.0
if "remask_threshold" not in self.decoding_thresholds:
self.decoding_thresholds["remask_threshold"] = 0.4
if "token_stability_threshold" not in self.decoding_thresholds:
self.decoding_thresholds["token_stability_threshold"] = 0.0
self.decoding_thresholds = DecodingThresholds(**self.decoding_thresholds)
elif self.decoding_thresholds is None:
add_block_threshold = d.get("add_block_threshold")
semi_complete_threshold = d.get("semi_complete_threshold")
accept_threshold = d.get("accept_threshold")
edit_threshold = d.get("edit_threshold")
remask_threshold = d.get("remask_threshold")
token_stability_threshold = d.get("token_stability_threshold")
self.decoding_thresholds = DecodingThresholds(
add_block_threshold=0.1 if add_block_threshold is None else add_block_threshold,
semi_complete_threshold=0.9 if semi_complete_threshold is None else semi_complete_threshold,
accept_threshold=0.9 if accept_threshold is None else accept_threshold,
edit_threshold=0.0 if edit_threshold is None else edit_threshold,
remask_threshold=0.4 if remask_threshold is None else remask_threshold,
token_stability_threshold=0.0
if token_stability_threshold is None
else token_stability_threshold,
)
if not 0.0 <= self.decoding_thresholds.accept_threshold <= 1.0:
raise ValueError(
"decoding_thresholds.accept_threshold must be in [0, 1], "
f"got: {self.decoding_thresholds.accept_threshold}"
)
if not 0.0 <= self.decoding_thresholds.edit_threshold <= 1.0:
raise ValueError(
"decoding_thresholds.edit_threshold must be in [0, 1], "
f"got: {self.decoding_thresholds.edit_threshold}"
)
if not 0.0 <= self.decoding_thresholds.remask_threshold <= 1.0:
raise ValueError(
"decoding_thresholds.remask_threshold must be in [0, 1], "
f"got: {self.decoding_thresholds.remask_threshold}"
)
if not 0.0 <= self.decoding_thresholds.token_stability_threshold <= 1.0:
raise ValueError(
"decoding_thresholds.token_stability_threshold must be in [0, 1], "
f"got: {self.decoding_thresholds.token_stability_threshold}"
)
if not isinstance(self.max_post_edit_steps, int) or self.max_post_edit_steps < 0:
raise ValueError(
"max_post_edit_steps must be a non-negative int, "
f"got: {self.max_post_edit_steps}"
)
if not isinstance(self.penalty_lambda, (int, float)) or self.penalty_lambda < 0:
raise ValueError(
"penalty_lambda must be a non-negative float, "
f"got: {self.penalty_lambda}"
)
self.accept_threshold = self.decoding_thresholds.accept_threshold
self.remask_threshold = self.decoding_thresholds.remask_threshold
@property
def kv_cache_page_size(self) -> int:
"""Alias for page_size, used by engine/model_runner."""
return self.page_size
@property
def add_new_block_threshold(self) -> float:
return self.decoding_thresholds.add_block_threshold
@property
def complete_threshold(self) -> float:
return self.decoding_thresholds.semi_complete_threshold
|