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

[Bugfix] Fix v1/spec_decode/test_ngram.py (#16895)

Browse Source 
Signed-off-by: qizixi <qizixi@meta.com>
This commit is contained in:
qizixi
2025-04-20 20:54:29 -07:00
committed by GitHub
parent fe742aef5a
commit bb3605db85
2 changed files with 30 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
vllm/config.py
View File

@@ -120,7 +120,7 @@ def get_attr_docs(cls: type[Any]) -> dict[str, str]:
def pairwise(iterable):
"""
Manually implement https://docs.python.org/3/library/itertools.html#itertools.pairwise
Can be removed when Python 3.9 support is dropped.
"""
iterator = iter(iterable)
@@ -266,7 +266,7 @@ class ModelConfig:
config_format: The config format which shall be loaded.
Defaults to 'auto' which defaults to 'hf'.
hf_token: The token to use as HTTP bearer authorization for remote files
. If `True`, will use the token generated when running
. If `True`, will use the token generated when running
`huggingface-cli login` (stored in `~/.huggingface`).
hf_overrides: If a dictionary, contains arguments to be forwarded to the
HuggingFace config. If a callable, it is called to update the
@@ -1624,7 +1624,7 @@ class ParallelConfig:
"""The full name of the worker class to use. If "auto", the worker class
will be determined based on the platform."""
sd_worker_cls: str = "auto"
"""The full name of the worker class to use for speculative decofing.
"""The full name of the worker class to use for speculative decofing.
If "auto", the worker class will be determined based on the platform."""
worker_extension_cls: str = ""
"""The full name of the worker extension class to use. The worker extension
@@ -1815,13 +1815,13 @@ class SchedulerConfig:
max_num_batched_tokens: int = None # type: ignore
"""Maximum number of tokens to be processed in a single iteration.
This config has no static default. If left unspecified by the user, it will
be set in `EngineArgs.create_engine_config` based on the usage context."""
max_num_seqs: int = None # type: ignore
"""Maximum number of sequences to be processed in a single iteration.
This config has no static default. If left unspecified by the user, it will
be set in `EngineArgs.create_engine_config` based on the usage context."""
@@ -1867,7 +1867,7 @@ class SchedulerConfig:
# TODO (ywang96): Make this configurable.
max_num_encoder_input_tokens: int = field(init=False)
"""Multimodal encoder compute budget, only used in V1.
NOTE: This is not currently configurable. It will be overridden by
max_num_batched_tokens in case max multimodal embedding size is larger."""
@@ -2306,7 +2306,8 @@ class SpeculativeConfig:
if self.model is None and self.num_speculative_tokens is not None:
# TODO(Shangming): Refactor mtp configuration logic when supporting
# mtp acceleration for more models besides deepseek_v3
if self.target_model_config.hf_text_config.model_type \
if self.target_model_config and \
self.target_model_config.hf_text_config.model_type \
== "deepseek_v3":
# use the draft model from the same model:
self.model = self.target_model_config.model