EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Remove inf value for chunked prefill size (#812)

Browse Source
This commit is contained in:
Liangsheng Yin
2024-07-29 18:34:25 -07:00
committed by GitHub
parent c8e9fed87a
commit 3520f75fb1
2 changed files with 6 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
python/sglang/srt/managers/controller/tp_worker.py
View File

@@ -442,8 +442,11 @@ class ModelTpServer:
else: else:
# Add this request to the running batch # Add this request to the running batch
if ( if (
new_batch_input_tokens + req.extend_input_len self.chunked_prefill_size is None
<= self.chunked_prefill_size or (
new_batch_input_tokens + req.extend_input_len
<= self.chunked_prefill_size
)
or ( or (
req.return_logprob and req.normalized_prompt_logprob is None req.return_logprob and req.normalized_prompt_logprob is None
) )

4
python/sglang/srt/server_args.py
View File

@@ -87,8 +87,6 @@ class ServerArgs:
node_rank: Optional[int] = None node_rank: Optional[int] = None
def __post_init__(self): def __post_init__(self):
if self.chunked_prefill_size is None:
self.chunked_prefill_size = 1 << 30
if self.tokenizer_path is None: if self.tokenizer_path is None:
self.tokenizer_path = self.model_path self.tokenizer_path = self.model_path
if self.mem_fraction_static is None: if self.mem_fraction_static is None:
@@ -414,7 +412,7 @@ class ServerArgs:
), "multi-node data parallel is not supported" ), "multi-node data parallel is not supported"
assert not ( assert not (
self.chunked_prefill_size < (1 << 30) and self.disable_radix_cache self.chunked_prefill_size is not None and self.disable_radix_cache
), "chunked prefill is not supported with radix cache disabled currently" ), "chunked prefill is not supported with radix cache disabled currently"