Skip to content

Commit 5989120

Browse files
seungrokjclaude
andcommitted
fix(agentic-mtp): restore MAX_RUNNING_REQUESTS to 2*CONC
Revert MAX_RUNNING_REQUESTS from CONC back to 2*CONC to match the old (good) run configuration that achieved 5,985 completed records. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
1 parent 5baf0d5 commit 5989120

1 file changed

Lines changed: 2 additions & 1 deletion

File tree

benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -112,7 +112,8 @@ fi
112112
# SGLang treats max-running-requests as a global DPA limit and partitions it
113113
# internally. CUDA graph capture is per scheduler, so only its batch size is
114114
# divided across DP ranks.
115-
MAX_RUNNING_REQUESTS=$CONC
115+
# 0719
116+
MAX_RUNNING_REQUESTS=$((2 * CONC))
116117
# 0719
117118
CUDA_GRAPH_MAX_BS=$CONC
118119
[ "$CUDA_GRAPH_MAX_BS" -gt 128 ] && CUDA_GRAPH_MAX_BS=128

0 commit comments

Comments
 (0)