minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[BugFix] Fix EAGLE speculative decoding missing grammar-based finish … (#21723)

Browse Source
This commit is contained in:
mingyue300
2026-04-14 12:43:50 -07:00
committed by GitHub
parent d2f479e544
commit b4616dcbf5
1 changed files with 9 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/speculative/eagle_info.py
+9 -9
View File
@@ -426,20 +426,20 @@ class EagleVerifyInput(SpecInput, EagleVerifyInputV2Mixin):
if req.require_reasoning and think_end_id is not None: if req.require_reasoning and think_end_id is not None:
req.update_reasoning_tokens(id, think_end_id) req.update_reasoning_tokens(id, think_end_id)
req.check_finished() req.check_finished()
if not req.finished() and req.grammar is not None:
try:
req.grammar.accept_token(id)
except ValueError as e:
logger.info(
f"{i=}, {req=}\n" f"{accept_index=}\n" f"{predict=}\n"
)
raise e
req.check_finished()
if req.finished(): if req.finished():
has_finished = True has_finished = True
# set all tokens after finished token to -1 and break # set all tokens after finished token to -1 and break
accept_index[i, j + 1 :] = -1 accept_index[i, j + 1 :] = -1
break break
else:
if req.grammar is not None:
try:
req.grammar.accept_token(id)
except ValueError as e:
logger.info(
f"{i=}, {req=}\n" f"{accept_index=}\n" f"{predict=}\n"
)
raise e
# Update KV cache tracking for the accepted tokens # Update KV cache tracking for the accepted tokens
req.kv_committed_len += num_accepted req.kv_committed_len += num_accepted
req.kv_allocated_len = req.kv_committed_len req.kv_allocated_len = req.kv_committed_len