xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 00:25:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix max seq len (#489)

Browse Source
This commit is contained in:
Lily Liu 2023-07-17 23:20:20 -07:00 committed by GitHub
parent 20b0d88d16
commit b4b195b360
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 8 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/config.py
View File

@ -204,10 +204,10 @@ class SchedulerConfig:
"""
def __init__(self, max_num_batched_tokens: int, max_num_seqs: int,
max_seq_len: int) -> None:
max_model_len: int) -> None:
self.max_num_batched_tokens = max_num_batched_tokens
self.max_num_seqs = max_num_seqs
self.max_seq_len = max_seq_len
self.max_model_len = max_model_len
_STR_DTYPE_TO_TORCH_DTYPE = {

4
vllm/core/scheduler.py
View File

@ -190,7 +190,9 @@ class Scheduler:
break
num_prompt_tokens = seq_group.get_seqs()[0].get_len()
if num_prompt_tokens > self.scheduler_config.max_seq_len:
if num_prompt_tokens > min(
self.scheduler_config.max_model_len,
self.scheduler_config.max_num_batched_tokens):
logger.warning(
f"Input prompt ({num_prompt_tokens} tokens) is too long"
" and exceeds limit of "

5
vllm/engine/arg_utils.py
View File

@ -155,11 +155,10 @@ class EngineArgs:
parallel_config = ParallelConfig(self.pipeline_parallel_size,
self.tensor_parallel_size,
self.worker_use_ray)
model_max_len = getattr(model_config.hf_config,
max_model_len = getattr(model_config.hf_config,
'max_position_embeddings', float('inf'))
max_seq_len = min(self.max_num_batched_tokens, model_max_len)
scheduler_config = SchedulerConfig(self.max_num_batched_tokens,
self.max_num_seqs, max_seq_len)
self.max_num_seqs, max_model_len)
return model_config, cache_config, parallel_config, scheduler_config

3
vllm/engine/llm_engine.py
View File

@ -300,8 +300,7 @@ class LLMEngine:
continue
# Check if the sequence has reached max_seq_len.
if (seq.get_len() >
self.scheduler.scheduler_config.max_seq_len):
if seq.get_len() > self.scheduler_config.max_model_len:
self.scheduler.free_seq(
seq, SequenceStatus.FINISHED_LENGTH_CAPPED)
continue