Skip to content

Commit d2c83b3

Browse files
seungrokjclaude
andcommitted
fix(agentic-mtp): restore GPU_MAX_HW_QUEUES=5 for ROCm overlap
Re-add GPU_MAX_HW_QUEUES=5 under DPA to enable compute/memory op overlap on ROCm, matching the old run configuration. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
1 parent 57a0245 commit d2c83b3

1 file changed

Lines changed: 2 additions & 0 deletions

File tree

benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -93,6 +93,8 @@ if [ "$DP_ATTENTION" = "true" ]; then
9393
export SGLANG_DP_SHARED_EXPERT_LOCAL=1
9494
export SGLANG_DP_USE_GATHERV=1
9595
export SGLANG_DP_USE_REDUCE_SCATTER=1
96+
# 0719
97+
export GPU_MAX_HW_QUEUES=5
9698

9799
# SGLang divides the configured chunk across DP schedulers. Use a 16K
98100
# per-scheduler chunk so long agentic prefill tails drain within the

0 commit comments

Comments
 (0)