xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 10:18:39 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Limit the default value of max_model_len when it is not specified by users (#27556)

Browse Source 
Signed-off-by: shen-shanshan <467638484@qq.com>
This commit is contained in:
Shanshan Shen 2025-10-27 18:16:20 +08:00 committed by GitHub
parent 7c2bdb83dc
commit a3e8611da5
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 28 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
vllm/config/model.py
View File

@ -2112,20 +2112,13 @@ def _get_and_verify_max_len(
if encoder_config and "max_seq_length" in encoder_config:
derived_max_model_len = encoder_config["max_seq_length"]
# If the user specified a max length, make sure it is smaller than the
# derived length from the HF model config.
# If the user didn't specify `max_model_len`, then use that derived from
# the model config as a default value.
if max_model_len is None:
max_model_len = int(derived_max_model_len)
if current_platform.is_tpu():
logger.warning(
"--max-model-len is not specified, "
"it's currently using model's default length %s, "
"which might be too large."
"Please input with --max-model-len based on your "
"request input length and output length, to avoid "
"unnecessary degradation.",
max_model_len,
)
max_model_len = current_platform.check_max_model_len(max_model_len)
# If the user specified a max length, make sure it is smaller than the
# derived length from the HF model config.
elif max_model_len > derived_max_model_len:
# Some models might have a separate key for specifying model_max_length
# that will be bigger than derived_max_model_len. We compare user input

7
vllm/platforms/interface.py
View File

@ -608,6 +608,13 @@ class Platform:
"""
return None
@classmethod
def check_max_model_len(cls, max_model_len: int) -> int:
"""
Check max_model_len for the current platform.
"""
return max_model_len
class UnspecifiedPlatform(Platform):
_enum = PlatformEnum.UNSPECIFIED

16
vllm/platforms/tpu.py
View File

@ -251,6 +251,22 @@ class TpuPlatform(Platform):
def use_sync_weight_loader(cls) -> bool:
return True
@classmethod
def check_max_model_len(cls, max_model_len: int) -> int:
"""
Check max_model_len for the current platform.
"""
logger.warning(
"--max-model-len is not specified, "
"it's currently using model's default length %d, "
"which might be too large."
"Please input with --max-model-len based on your "
"request input length and output length, to avoid "
"unnecessary degradation.",
max_model_len,
)
return max_model_len
try:
from tpu_inference.platforms import TpuPlatform as TpuInferencePlatform