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

[RL] [Spec v2] Use stop-aware seqlen for returned topk metadata (#26126)

Browse Source
This commit is contained in:
Yueming Yuan
2026-05-22 18:13:16 -07:00
committed by GitHub
parent c112f7623a
commit 208397affc
1 changed files with 8 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/managers/scheduler_components/batch_result_processor.py
+8 -5
View File
@@ -117,25 +117,27 @@ class SchedulerBatchResultProcessor:
if capturer is None: if capturer is None:
return return
start_len = req.routed_experts_start_len start_len = req.routed_experts_start_len
seqlen = len(req.origin_input_ids) + len(req.output_ids_through_stop)
req.routed_experts = capturer.get_topk( req.routed_experts = capturer.get_topk(
req_pool_idx=req.req_pool_idx, req_pool_idx=req.req_pool_idx,
seqlen=req.seqlen, seqlen=seqlen,
req_to_token_pool=self.req_to_token_pool, req_to_token_pool=self.req_to_token_pool,
start_len=start_len, start_len=start_len,
) )
expected_rows = max(0, req.seqlen - 1 - start_len) expected_rows = max(0, seqlen - 1 - start_len)
if ( if (
req.routed_experts is not None req.routed_experts is not None
and req.routed_experts.shape[0] != expected_rows and req.routed_experts.shape[0] != expected_rows
): ):
logger.warning( logger.warning(
"routed_experts row-count mismatch for req %s: got %d, " "routed_experts row-count mismatch for req %s: got %d, expected %d "
"expected %d (seqlen=%d, cached_tokens=%d, start_len=%s). " "(seqlen=%d, raw_seqlen=%d, cached_tokens=%d, start_len=%s). "
"This indicates a silent bug.", "This indicates a silent bug.",
req.rid, req.rid,
req.routed_experts.shape[0], req.routed_experts.shape[0],
expected_rows, expected_rows,
seqlen,
req.seqlen, req.seqlen,
req.cached_tokens, req.cached_tokens,
req.routed_experts_start_len, req.routed_experts_start_len,
@@ -145,9 +147,10 @@ class SchedulerBatchResultProcessor:
capturer = get_global_indexer_capturer() capturer = get_global_indexer_capturer()
if capturer is None: if capturer is None:
return return
seqlen = len(req.origin_input_ids) + len(req.output_ids_through_stop)
req.indexer_topk = capturer.get_topk( req.indexer_topk = capturer.get_topk(
req_pool_idx=req.req_pool_idx, req_pool_idx=req.req_pool_idx,
seqlen=req.seqlen, seqlen=seqlen,
req_to_token_pool=self.req_to_token_pool, req_to_token_pool=self.req_to_token_pool,
) )