xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-15 18:06:03 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fix handling of max_num_batched_tokens for pooling tasks (#23004)

Browse Source 
Signed-off-by: Max de Bayser <mbayser@br.ibm.com>
This commit is contained in:
Maximilien de Bayser 2025-08-16 14:36:30 -03:00 committed by GitHub
parent 829bbd7882
commit 52ce1420e9
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 5 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
vllm/config/__init__.py
View File

@ -3600,9 +3600,6 @@ class VllmConfig:
logger.info(reason) logger.info(reason)
self.scheduler_config.chunked_prefill_enabled = False self.scheduler_config.chunked_prefill_enabled = False
self.scheduler_config.long_prefill_token_threshold = 0 self.scheduler_config.long_prefill_token_threshold = 0
self.scheduler_config.max_num_batched_tokens = max(
self.scheduler_config.max_model_len,
DEFAULT_MAX_NUM_BATCHED_TOKENS)
if self.cache_config is not None: if self.cache_config is not None:
self.cache_config.enable_prefix_caching = False self.cache_config.enable_prefix_caching = False

10
vllm/engine/arg_utils.py
View File

@ -1602,9 +1602,6 @@ class EngineArgs:
self.enable_prefix_caching = incremental_prefill_supported self.enable_prefix_caching = incremental_prefill_supported
logger.info("(%s) prefix caching by default", action) logger.info("(%s) prefix caching by default", action)
if not self.enable_chunked_prefill:
self.max_num_batched_tokens = model_config.max_model_len
# V1 should use the new scheduler by default. # V1 should use the new scheduler by default.
# Swap it only if this arg is set to the original V0 default # Swap it only if this arg is set to the original V0 default
if self.scheduler_cls == EngineArgs.scheduler_cls: if self.scheduler_cls == EngineArgs.scheduler_cls:
@ -1692,8 +1689,11 @@ class EngineArgs:
self.max_num_batched_tokens = \ self.max_num_batched_tokens = \
default_max_num_batched_tokens[usage_context] default_max_num_batched_tokens[usage_context]
else: else:
self.max_num_batched_tokens = default_max_num_batched_tokens[ if not self.enable_chunked_prefill:
usage_context] self.max_num_batched_tokens = model_config.max_model_len
else:
self.max_num_batched_tokens = \
default_max_num_batched_tokens[usage_context]
logger.debug( logger.debug(
"Setting max_num_batched_tokens to %d for %s usage context.", "Setting max_num_batched_tokens to %d for %s usage context.",
self.max_num_batched_tokens, use_context_value) self.max_num_batched_tokens, use_context_value)