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

[Minor] Rename no_eos_trim to no_stop_trim (#1661)

Browse Source
This commit is contained in:
Ying Sheng
2024-10-13 20:30:03 -07:00
committed by GitHub
parent da1ffed689
commit 2725f8da61
6 changed files with 16 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
python/sglang/srt/managers/detokenizer_manager.py
View File

@@ -75,8 +75,8 @@ class DetokenizerManager:
self.decode_status = LimitedCapacityDict()
def trim_eos(self, output: Union[str, List[int]], finished_reason, no_eos_trim):
if no_eos_trim:
def trim_eos(self, output: Union[str, List[int]], finished_reason, no_stop_trim):
if no_stop_trim:
return output
# Trim stop str. TODO(lmzheng): handle the case where multiple stop strs are hit
@@ -141,7 +141,7 @@ class DetokenizerManager:
self.trim_eos(
s.decode_ids[s.surr_offset :],
recv_obj.finished_reason[i],
recv_obj.no_eos_trim[i],
recv_obj.no_stop_trim[i],
)
)
surr_ids.append(s.decode_ids[s.surr_offset : s.read_offset])
@@ -177,7 +177,7 @@ class DetokenizerManager:
self.trim_eos(
s.decoded_text + new_text,
recv_obj.finished_reason[i],
recv_obj.no_eos_trim[i],
recv_obj.no_stop_trim[i],
)
)

2
python/sglang/srt/managers/io_struct.py
View File

@@ -295,7 +295,7 @@ class BatchTokenIDOut:
spaces_between_special_tokens: List[bool]
meta_info: List[Dict]
finished_reason: List[BaseFinishReason]
no_eos_trim: List[bool]
no_stop_trim: List[bool]
@dataclass

6
python/sglang/srt/managers/scheduler.py
View File

@@ -885,7 +885,7 @@ class Scheduler:
output_read_offsets = []
output_skip_special_tokens = []
output_spaces_between_special_tokens = []
output_no_eos_trim = []
output_no_stop_trim = []
else: # embedding or reward model
output_embeddings = []
unfinished_indices = []
@@ -917,7 +917,7 @@ class Scheduler:
output_spaces_between_special_tokens.append(
req.sampling_params.spaces_between_special_tokens
)
output_no_eos_trim.append(req.sampling_params.no_eos_trim)
output_no_stop_trim.append(req.sampling_params.no_stop_trim)
meta_info = {
"prompt_tokens": len(req.origin_input_ids),
@@ -968,7 +968,7 @@ class Scheduler:
output_spaces_between_special_tokens,
output_meta_info,
output_finished_reason,
output_no_eos_trim,
output_no_stop_trim,
)
)
else: # embedding or reward model