Skip to content

Adjust MiniMax MI355X block size for TP8 EP8 #1228

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

Closed
jiacao-amd wants to merge 3 commits into from
+42 −2
Closed

Adjust MiniMax MI355X block size for TP8 EP8 #1228

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
d66409b
Adjust MiniMax block size for TP8 EP8
jiacao-amd Apr 29, 2026
a34fb25
Use async scheduling only for MiniMax c128
jiacao-amd May 3, 2026
37962d4
Tune MiniMax MI355X shuffle and async thresholds
jiacao-amd May 4, 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
44 changes: 42 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 +26,46 @@ 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 [[ "$TP" == "8" && "$EP_SIZE" == "8" ]]; then
export VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=0
VLLM_BLOCK_SIZE=32
echo "Disabling shuffle KV cache layout and using block size 32 for TP8/EP8."
elif [[ "$ISL" == "1024" && "$OSL" == "1024" ]]; then
export VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=1
VLLM_BLOCK_SIZE=16

if (( CONC <= 128 )); then
ASYNC_SCHEDULING_ARGS="--no-async-scheduling"
echo "Using shuffle KV cache layout with block size 16 and disabling async scheduling for 1k1k c${CONC}."
else
echo "Using shuffle KV cache layout with block size 16 and async scheduling for 1k1k c${CONC}."
fi
elif [[ "$ISL" == "8192" && "$OSL" == "1024" ]]; then
if (( CONC <= 64 )); then
ASYNC_SCHEDULING_ARGS="--no-async-scheduling"
fi

if (( CONC >= 32 )); then
export VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=1
VLLM_BLOCK_SIZE=16

if [[ -n "$ASYNC_SCHEDULING_ARGS" ]]; then
echo "Using shuffle KV cache layout with block size 16 and disabling async scheduling for 8k1k c${CONC}."
else
echo "Using shuffle KV cache layout with block size 16 and async scheduling for 8k1k c${CONC}."
fi
elif [[ -n "$ASYNC_SCHEDULING_ARGS" ]]; then
echo "Using baseline block size 32, shuffle disabled, and disabling async scheduling for 8k1k c${CONC}."
else
echo "Using baseline block size 32, shuffle disabled, and async scheduling for 8k1k c${CONC}."
fi
else
echo "Using baseline block size 32, shuffle disabled, and async scheduling for ISL=${ISL}, OSL=${OSL}, c${CONC}."
fi

SERVER_LOG=/workspace/server.log
PORT=${PORT:-8888}
Expand All @@ -52,9 +91,10 @@ $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