biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Convert formatting to use ruff instead of yapf + isort (#26247)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-10-05 15:06:22 +01:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115244 additions and 94146 deletions
Download Patch File Download Diff File Expand all files Collapse all files

336
vllm/config/speculative.py
View File

@@ -24,23 +24,41 @@ else:
PretrainedConfig = Any
ModelConfig = Any
me_quant = LazyLoader("model_executor", globals(),
"vllm.model_executor.layers.quantization")
me_quant = LazyLoader(
"model_executor", globals(), "vllm.model_executor.layers.quantization"
)
logger = init_logger(__name__)
SpeculativeMethod = Literal["ngram", "eagle", "eagle3", "medusa",
"mlp_speculator", "draft_model", "deepseek_mtp",
"ernie_mtp", "qwen3_next_mtp", "mimo_mtp",
"longcat_flash_mtp", "mtp"]
MTP_MODEL_TYPES = ("deepseek_mtp", "mimo_mtp", "glm4_moe_mtp", "ernie_mtp",
"qwen3_next_mtp", "longcat_flash_mtp")
SpeculativeMethod = Literal[
"ngram",
"eagle",
"eagle3",
"medusa",
"mlp_speculator",
"draft_model",
"deepseek_mtp",
"ernie_mtp",
"qwen3_next_mtp",
"mimo_mtp",
"longcat_flash_mtp",
"mtp",
]
MTP_MODEL_TYPES = (
"deepseek_mtp",
"mimo_mtp",
"glm4_moe_mtp",
"ernie_mtp",
"qwen3_next_mtp",
"longcat_flash_mtp",
)
@config
@dataclass
class SpeculativeConfig:
"""Configuration for speculative decoding."""
enforce_eager: Optional[bool] = None
"""Override the default enforce_eager from model_config"""
# General speculative decoding control
@@ -107,8 +125,7 @@ class SpeculativeConfig:
# required configuration params passed from engine
target_model_config: SkipValidation[ModelConfig] = None # type: ignore
"""The configuration of the target model."""
target_parallel_config: SkipValidation[
ParallelConfig] = None # type: ignore
target_parallel_config: SkipValidation[ParallelConfig] = None # type: ignore
"""The parallel configuration for the target model."""
enable_chunked_prefill: SkipValidation[bool] = None # type: ignore
"""Whether vLLM is configured to use chunked prefill or not. Used for
@@ -120,8 +137,7 @@ class SpeculativeConfig:
# params generated in the post-init stage
draft_model_config: SkipValidation[ModelConfig] = None # type: ignore
"""The configuration of the draft model initialized internal."""
draft_parallel_config: SkipValidation[
ParallelConfig] = None # type: ignore
draft_parallel_config: SkipValidation[ParallelConfig] = None # type: ignore
"""The parallel configuration for the draft model initialized internal."""
def compute_hash(self) -> str:
@@ -140,8 +156,7 @@ class SpeculativeConfig:
# Eagle3 affects the computation graph because it returns intermediate
# hidden states in addition to the final hidden state.
factors.append(self.method == "eagle3")
hash_str = hashlib.md5(str(factors).encode(),
usedforsecurity=False).hexdigest()
hash_str = hashlib.md5(str(factors).encode(), usedforsecurity=False).hexdigest()
return hash_str
@staticmethod
@@ -150,58 +165,57 @@ class SpeculativeConfig:
hf_config.model_type = "deepseek_mtp"
if hf_config.model_type == "deepseek_mtp":
n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
hf_config.update({
"n_predict": n_predict,
"architectures": ["DeepSeekMTPModel"]
})
hf_config.update(
{"n_predict": n_predict, "architectures": ["DeepSeekMTPModel"]}
)
if hf_config.architectures[0] == "MiMoForCausalLM":
hf_config.model_type = "mimo_mtp"
n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
hf_config.update({
"num_hidden_layers": 0,
"n_predict": n_predict,
"architectures": ["MiMoMTPModel"]
})
hf_config.update(
{
"num_hidden_layers": 0,
"n_predict": n_predict,
"architectures": ["MiMoMTPModel"],
}
)
if hf_config.architectures[0] == "Glm4MoeForCausalLM":
hf_config.model_type = "glm4_moe_mtp"
n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
hf_config.update({
"num_hidden_layers": 0,
"n_predict": n_predict,
"architectures": ["Glm4MoeMTPModel"]
})
hf_config.update(
{
"num_hidden_layers": 0,
"n_predict": n_predict,
"architectures": ["Glm4MoeMTPModel"],
}
)
if hf_config.model_type == "ernie4_5_moe":
hf_config.model_type = "ernie_mtp"
if hf_config.model_type == "ernie_mtp":
n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
hf_config.update({
"n_predict": n_predict,
"architectures": ["ErnieMTPModel"]
})
hf_config.update(
{"n_predict": n_predict, "architectures": ["ErnieMTPModel"]}
)
if hf_config.model_type == "qwen3_next":
hf_config.model_type = "qwen3_next_mtp"
if hf_config.model_type == "qwen3_next_mtp":
n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
hf_config.update({
"n_predict": n_predict,
"architectures": ["Qwen3NextMTP"]
})
hf_config.update(
{"n_predict": n_predict, "architectures": ["Qwen3NextMTP"]}
)
if hf_config.model_type == "longcat_flash":
hf_config.model_type = "longcat_flash_mtp"
n_predict = getattr(hf_config, "num_nextn_predict_layers", 1)
hf_config.update({
"n_predict": n_predict,
"architectures": ["LongCatFlashMTPModel"]
})
hf_config.update(
{"n_predict": n_predict, "architectures": ["LongCatFlashMTPModel"]}
)
return hf_config
def __post_init__(self):
# Note: "method" is a new parameter that helps to extend the
# configuration of non-model-based proposers, and the "model" parameter
# will be used to set the draft model, eagle head, or additional weight
@@ -211,17 +225,17 @@ class SpeculativeConfig:
# default.
if self.method in MTP_MODEL_TYPES:
logger.warning("method `%s` is deprecated and replaced with mtp.",
self.method)
logger.warning(
"method `%s` is deprecated and replaced with mtp.", self.method
)
self.method = "mtp"
if self.model is None and self.num_speculative_tokens is not None:
if self.method == "mtp":
assert (
self.target_model_config
is not None), "target_model_config must be present for mtp"
if self.target_model_config.hf_text_config.model_type \
== "deepseek_v32":
assert self.target_model_config is not None, (
"target_model_config must be present for mtp"
)
if self.target_model_config.hf_text_config.model_type == "deepseek_v32":
# FIXME(luccafong): cudgraph with v32 MTP is not supported,
# remove this when the issue is fixed.
self.enforce_eager = True
@@ -235,21 +249,21 @@ class SpeculativeConfig:
self.model = "ngram"
else:
raise ValueError(
"num_speculative_tokens was provided but without "
"speculative model.")
"num_speculative_tokens was provided but without speculative model."
)
# Automatically configure the method for ngram when "model" is used
# instead of "method"
if self.method is None and (self.model is not None
and self.model in ("ngram", "[ngram]")):
if self.method is None and (
self.model is not None and self.model in ("ngram", "[ngram]")
):
self.method = "ngram"
if self.method in ("ngram", "[ngram]"):
# Unified to "ngram" internally
self.method = "ngram"
# Set default values if not provided
if (self.prompt_lookup_min is None
and self.prompt_lookup_max is None):
if self.prompt_lookup_min is None and self.prompt_lookup_max is None:
# TODO(woosuk): Tune these values. They are arbitrarily chosen.
self.prompt_lookup_min = 5
self.prompt_lookup_max = 5
@@ -263,14 +277,17 @@ class SpeculativeConfig:
# Validate values
if self.prompt_lookup_min < 1:
raise ValueError(
f"prompt_lookup_min={self.prompt_lookup_min} must be > 0")
f"prompt_lookup_min={self.prompt_lookup_min} must be > 0"
)
if self.prompt_lookup_max < 1:
raise ValueError(
f"prompt_lookup_max={self.prompt_lookup_max} must be > 0")
f"prompt_lookup_max={self.prompt_lookup_max} must be > 0"
)
if self.prompt_lookup_min > self.prompt_lookup_max:
raise ValueError(
f"prompt_lookup_min={self.prompt_lookup_min} must "
f"be <= prompt_lookup_max={self.prompt_lookup_max}")
f"be <= prompt_lookup_max={self.prompt_lookup_max}"
)
# TODO: current we still need extract vocab_size from target model
# config, in future, we may try refactor it out, and set
@@ -285,25 +302,21 @@ class SpeculativeConfig:
# TODO: Move this import to the top once `ModelConfig`
# lives in `vllm.config.model`.
from vllm.config import ModelConfig
self.draft_model_config = ModelConfig(
model=self.model,
runner="draft",
tokenizer=self.target_model_config.tokenizer,
tokenizer_mode=self.target_model_config.tokenizer_mode,
trust_remote_code=self.target_model_config.
trust_remote_code,
allowed_local_media_path=self.target_model_config.
allowed_local_media_path,
allowed_media_domains=self.target_model_config.
allowed_media_domains,
trust_remote_code=self.target_model_config.trust_remote_code,
allowed_local_media_path=self.target_model_config.allowed_local_media_path,
allowed_media_domains=self.target_model_config.allowed_media_domains,
dtype=self.target_model_config.dtype,
seed=self.target_model_config.seed,
revision=self.revision,
code_revision=self.code_revision,
tokenizer_revision=self.target_model_config.
tokenizer_revision,
spec_target_max_model_len=self.target_model_config.
max_model_len,
tokenizer_revision=self.target_model_config.tokenizer_revision,
spec_target_max_model_len=self.target_model_config.max_model_len,
quantization=self.quantization,
enforce_eager=self.target_model_config.enforce_eager,
max_logprobs=self.target_model_config.max_logprobs,
@@ -311,7 +324,7 @@ class SpeculativeConfig:
)
# Automatically detect the method
if self.method in ('eagle', 'eagle3'):
if self.method in ("eagle", "eagle3"):
pass
# examples:
# yuhuili/EAGLE-LLaMA3-Instruct-8B
@@ -323,94 +336,101 @@ class SpeculativeConfig:
self.method = "eagle3"
elif self.draft_model_config.hf_config.model_type == "medusa":
self.method = "medusa"
elif (self.draft_model_config.hf_config.model_type ==
"mlp_speculator"):
elif self.draft_model_config.hf_config.model_type == "mlp_speculator":
self.method = "mlp_speculator"
elif (self.draft_model_config.hf_config.model_type
in MTP_MODEL_TYPES):
elif self.draft_model_config.hf_config.model_type in MTP_MODEL_TYPES:
self.method = "mtp"
if self.num_speculative_tokens > 1:
logger.warning(
"Enabling num_speculative_tokens > 1 will run" \
"multiple times of forward on same MTP layer" \
",which may result in lower acceptance rate" \
)
elif (self.draft_model_config.hf_config.model_type
in ("longcat_flash_mtp")):
"Enabling num_speculative_tokens > 1 will run"
"multiple times of forward on same MTP layer"
",which may result in lower acceptance rate"
)
elif self.draft_model_config.hf_config.model_type in (
"longcat_flash_mtp"
):
self.method = "longcat_flash_mtp"
if self.num_speculative_tokens > 1:
logger.warning(
"LongCat MTP models only have " \
"one layer. Might need some code changes " \
"to support multiple layers."
)
"LongCat MTP models only have "
"one layer. Might need some code changes "
"to support multiple layers."
)
else:
self.method = "draft_model"
raise NotImplementedError(
"Speculative decoding with draft model is not "
"supported yet. Please consider using other "
"speculative decoding methods such as ngram, medusa, "
"eagle, or mtp.")
"eagle, or mtp."
)
# Replace hf_config for EAGLE draft_model
if self.method in ("eagle", "eagle3"):
if self.enable_chunked_prefill and not envs.VLLM_USE_V1:
raise ValueError(
"Chunked prefill and EAGLE are not compatible "
"when using V0.")
"when using V0."
)
from vllm.transformers_utils.configs import (
SpeculatorsConfig)
from vllm.transformers_utils.configs.eagle import (
EAGLEConfig)
from vllm.transformers_utils.configs import SpeculatorsConfig
from vllm.transformers_utils.configs.eagle import EAGLEConfig
if isinstance(self.draft_model_config.hf_config,
(EAGLEConfig, SpeculatorsConfig)):
if isinstance(
self.draft_model_config.hf_config,
(EAGLEConfig, SpeculatorsConfig),
):
pass
else:
eagle_config = EAGLEConfig(
self.draft_model_config.hf_config,
method=self.method,
model_type="eagle")
model_type="eagle",
)
self.draft_model_config.hf_config = eagle_config
if (self.num_speculative_tokens is not None
and hasattr(self.draft_model_config.hf_config,
"num_lookahead_tokens")):
self.draft_model_config.hf_config.num_lookahead_tokens = \
self.num_speculative_tokens
if self.num_speculative_tokens is not None and hasattr(
self.draft_model_config.hf_config, "num_lookahead_tokens"
):
self.draft_model_config.hf_config.num_lookahead_tokens = (
self.num_speculative_tokens
)
n_predict = getattr(self.draft_model_config.hf_config,
"n_predict", None)
n_predict = getattr(
self.draft_model_config.hf_config, "n_predict", None
)
if n_predict is not None:
if self.num_speculative_tokens is None:
# Default to max value defined in draft model config.
self.num_speculative_tokens = n_predict
elif self.num_speculative_tokens > n_predict and \
self.num_speculative_tokens % n_predict != 0:
elif (
self.num_speculative_tokens > n_predict
and self.num_speculative_tokens % n_predict != 0
):
# Ensure divisibility for MTP module reuse.
raise ValueError(
f"num_speculative_tokens:{self.num_speculative_tokens}"
f" must be divisible by {n_predict=}")
f" must be divisible by {n_predict=}"
)
if self.speculative_token_tree is None:
# Generate chain of tokens.
self.speculative_token_tree = str([
(i + 1) * (0, )
for i in range(self.num_speculative_tokens)
])
self.speculative_token_tree = str(
[(i + 1) * (0,) for i in range(self.num_speculative_tokens)]
)
else:
# Sort the token tree breadth-first.
tree_choices = ast.literal_eval(
self.speculative_token_tree)
tree_choices = ast.literal_eval(self.speculative_token_tree)
self.speculative_token_tree = str(
sorted(tree_choices, key=lambda t: (len(t), t)))
sorted(tree_choices, key=lambda t: (len(t), t))
)
self.draft_tensor_parallel_size = \
self.draft_tensor_parallel_size = (
SpeculativeConfig._verify_and_get_draft_tp(
self.target_parallel_config,
self.draft_tensor_parallel_size,
self.draft_model_config.hf_config
self.draft_model_config.hf_config,
)
)
self.draft_model_config.max_model_len = (
@@ -418,12 +438,14 @@ class SpeculativeConfig:
self.max_model_len,
self.draft_model_config.max_model_len,
self.target_model_config.max_model_len,
))
)
)
self.draft_parallel_config = (
SpeculativeConfig.create_draft_parallel_config(
self.target_parallel_config,
self.draft_tensor_parallel_size))
self.target_parallel_config, self.draft_tensor_parallel_size
)
)
@staticmethod
def _maybe_override_draft_max_model_len(
@@ -444,14 +466,17 @@ class SpeculativeConfig:
"""
if speculative_max_model_len is not None:
if speculative_max_model_len > draft_max_model_len:
raise ValueError(f"{speculative_max_model_len=} cannot be "
f"larger than {draft_max_model_len=}")
raise ValueError(
f"{speculative_max_model_len=} cannot be "
f"larger than {draft_max_model_len=}"
)
if speculative_max_model_len > target_max_model_len:
raise ValueError(f"{speculative_max_model_len=} cannot be "
f"larger than {target_max_model_len=}")
raise ValueError(
f"{speculative_max_model_len=} cannot be "
f"larger than {target_max_model_len=}"
)
return speculative_max_model_len
@@ -462,9 +487,10 @@ class SpeculativeConfig:
@staticmethod
def _verify_and_get_draft_tp(
target_parallel_config: ParallelConfig,
speculative_draft_tensor_parallel_size: Optional[int],
draft_hf_config: PretrainedConfig) -> int:
target_parallel_config: ParallelConfig,
speculative_draft_tensor_parallel_size: Optional[int],
draft_hf_config: PretrainedConfig,
) -> int:
"""
Verifies and adjusts the tensor parallel size for a draft model
specified using speculative_draft_tensor_parallel_size.
@@ -478,15 +504,20 @@ class SpeculativeConfig:
logger.warning(
"%s cannot currently be run with tp>1; "
"setting speculative_draft_tensor_parallel_size=1",
draft_hf_config.model_type)
draft_hf_config.model_type,
)
else:
speculative_draft_tensor_parallel_size = \
speculative_draft_tensor_parallel_size = (
target_parallel_config.tensor_parallel_size
)
elif speculative_draft_tensor_parallel_size not in (
1, target_parallel_config.tensor_parallel_size):
1,
target_parallel_config.tensor_parallel_size,
):
raise ValueError(
f"{speculative_draft_tensor_parallel_size=} cannot be "
f"other value than 1 or target model tensor_parallel_size")
f"other value than 1 or target model tensor_parallel_size"
)
return speculative_draft_tensor_parallel_size
@staticmethod
@@ -499,52 +530,57 @@ class SpeculativeConfig:
This is mostly a copy of the target parallel config, except the tp_size.
"""
draft_parallel_config = ParallelConfig(
pipeline_parallel_size=target_parallel_config.
pipeline_parallel_size,
pipeline_parallel_size=target_parallel_config.pipeline_parallel_size,
tensor_parallel_size=speculative_draft_tensor_parallel_size,
distributed_executor_backend=target_parallel_config.
distributed_executor_backend,
max_parallel_loading_workers=target_parallel_config.
max_parallel_loading_workers,
disable_custom_all_reduce=target_parallel_config.
disable_custom_all_reduce,
ray_workers_use_nsight=target_parallel_config.
ray_workers_use_nsight,
distributed_executor_backend=target_parallel_config.distributed_executor_backend,
max_parallel_loading_workers=target_parallel_config.max_parallel_loading_workers,
disable_custom_all_reduce=target_parallel_config.disable_custom_all_reduce,
ray_workers_use_nsight=target_parallel_config.ray_workers_use_nsight,
placement_group=target_parallel_config.placement_group,
)
return draft_parallel_config
@model_validator(mode='after')
@model_validator(mode="after")
def _verify_args(self) -> Self:
if self.num_speculative_tokens is None:
raise ValueError(
"num_speculative_tokens must be provided with "
"speculative model unless the draft model config contains an "
"n_predict parameter.")
"n_predict parameter."
)
if self.num_speculative_tokens <= 0:
raise ValueError("Expected num_speculative_tokens to be greater "
f"than zero ({self.num_speculative_tokens}).")
raise ValueError(
"Expected num_speculative_tokens to be greater "
f"than zero ({self.num_speculative_tokens})."
)
if self.draft_model_config:
self.draft_model_config.verify_with_parallel_config(
self.draft_parallel_config)
self.draft_parallel_config
)
if (self.disable_by_batch_size is not None
and self.disable_by_batch_size < 2):
raise ValueError("Expect the batch size threshold of disabling "
"speculative decoding is > 1, but got "
f"{self.disable_by_batch_size=}")
if self.disable_by_batch_size is not None and self.disable_by_batch_size < 2:
raise ValueError(
"Expect the batch size threshold of disabling "
"speculative decoding is > 1, but got "
f"{self.disable_by_batch_size=}"
)
eagle3_target_supported = ["llama", "qwen", "minicpm", "gpt_oss"]
if self.method == "eagle3" and self.target_model_config and not any(
supported_model in
self.target_model_config.hf_text_config.model_type
for supported_model in eagle3_target_supported):
if (
self.method == "eagle3"
and self.target_model_config
and not any(
supported_model in self.target_model_config.hf_text_config.model_type
for supported_model in eagle3_target_supported
)
):
raise ValueError(
f"Eagle3 is only supported for {eagle3_target_supported} models. " # noqa: E501
f"Got {self.target_model_config.hf_text_config.model_type=}")
f"Got {self.target_model_config.hf_text_config.model_type=}"
)
return self