Skip to content

Commit 57a0245

Browse files
seungrokjclaude
andcommitted
fix(agentic-mtp): reduce chunked prefill back to 8K per scheduler
Revert per-scheduler chunked prefill from 16K to 8K to match the old run configuration (8192*TP = 65536 total vs 16384*TP = 131072). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
1 parent 5989120 commit 57a0245

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
@@ -97,7 +97,8 @@ if [ "$DP_ATTENTION" = "true" ]; then
9797
# SGLang divides the configured chunk across DP schedulers. Use a 16K
9898
# per-scheduler chunk so long agentic prefill tails drain within the
9999
# standard 600-second warmup grace period.
100-
CHUNKED_PREFILL_SIZE=$((16384 * TP))
100+
# 0719
101+
CHUNKED_PREFILL_SIZE=$((8192 * TP))
101102
PARALLEL_ARGS+=(
102103
--dp "$TP"
103104
--enable-dp-attention

0 commit comments

Comments
 (0)