Skip to content

Tune MiniMax MI355X vLLM scheduling thresholds #1276

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
chunfangamd merged 10 commits into from
May 6, 2026
+39 −2
Merged

Tune MiniMax MI355X vLLM scheduling thresholds #1276

Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
a9a3cef
Tune MiniMax MI355X vLLM scheduling thresholds
jiacao-amd May 4, 2026
4b89f40
Clarify MiniMax 8k1k scheduling branches
jiacao-amd May 4, 2026
91595f7
Refactor MiniMax MI355X scheduling policy
jiacao-amd May 5, 2026
807bf64
Merge branch 'main' into minimax-mi355x-scheduler-thresholds
chunfangamd May 5, 2026
e2524a2
Disable AITER MoE for MiniMax 8k1k TP8
jiacao-amd May 6, 2026
804eba4
Disable async for MiniMax 8k1k TP8
jiacao-amd May 6, 2026
fbf4738
Make MiniMax 1k1k TP8 no-async explicit
jiacao-amd May 6, 2026
8d8b1e0
Remove MiniMax perf changelog entry
jiacao-amd May 6, 2026
8bbdc81
Remove MiniMax default policy log
jiacao-amd May 6, 2026
b4df99e
Merge branch 'main' into minimax-mi355x-scheduler-thresholds
chunfangamd May 6, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
41 changes: 39 additions & 2 deletions benchmarks/single_node/minimaxm2.5_fp8_mi355x.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,7 +26,43 @@ fi

export VLLM_ROCM_USE_AITER=1
export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
export VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=1
export VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=0
VLLM_BLOCK_SIZE=32
ASYNC_SCHEDULING_ARGS=""

if [[ "$ISL" == "1024" && "$OSL" == "1024" ]]; then
if [[ "$TP" == "8" && "$EP_SIZE" == "8" ]]; then
ASYNC_SCHEDULING_ARGS="--no-async-scheduling"
echo "1k1k TP8/EP8: using block size 32, shuffle disabled, async scheduling disabled."
elif (( CONC <= 128 )); then
export VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=1
VLLM_BLOCK_SIZE=16
ASYNC_SCHEDULING_ARGS="--no-async-scheduling"
echo "1k1k c${CONC}: using block size 16, shuffle enabled, async scheduling disabled."
else
export VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=1
VLLM_BLOCK_SIZE=16
echo "1k1k c${CONC}: using block size 16, shuffle enabled, async scheduling enabled."
fi
elif [[ "$ISL" == "8192" && "$OSL" == "1024" ]]; then
if [[ "$TP" == "8" && "$EP_SIZE" == "8" ]]; then
export VLLM_ROCM_USE_AITER_MOE=0
ASYNC_SCHEDULING_ARGS="--no-async-scheduling"
echo "8k1k TP8/EP8: using block size 32, shuffle disabled, AITER MoE disabled, async scheduling disabled."
elif (( CONC < 64 )); then
ASYNC_SCHEDULING_ARGS="--no-async-scheduling"
echo "8k1k c${CONC}: using block size 32, shuffle disabled, async scheduling disabled."
elif (( CONC == 64 )); then
ASYNC_SCHEDULING_ARGS="--no-async-scheduling"
export VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=1
VLLM_BLOCK_SIZE=16
echo "8k1k c64: using block size 16, shuffle enabled, async scheduling disabled."
else
export VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=1
VLLM_BLOCK_SIZE=16
echo "8k1k c${CONC}: using block size 16, shuffle enabled, async scheduling enabled."
fi
fi

SERVER_LOG=/workspace/server.log
PORT=${PORT:-8888}
Expand All @@ -52,9 +88,10 @@ $EP \
--gpu-memory-utilization 0.95 \
--max-model-len $MAX_MODEL_LEN \
--kv-cache-dtype fp8 \
--block-size=32 \
--block-size=$VLLM_BLOCK_SIZE \
--no-enable-prefix-caching \
--attention-backend "ROCM_AITER_FA" \
$ASYNC_SCHEDULING_ARGS \
--trust-remote-code > $SERVER_LOG 2>&1 &

SERVER_PID=$!
Expand Down
Loading