From d66409b34679f62dfe8892ecbac8cfc2b0517623 Mon Sep 17 00:00:00 2001
From: jiacao-amd <jiahui.cao@amd.com>
Date: Wed, 29 Apr 2026 11:28:27 -0500
Subject: [PATCH] Adjust MiniMax block size for TP8 EP8

---
 benchmarks/single_node/minimaxm2.5_fp8_mi355x.sh | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/minimaxm2.5_fp8_mi355x.sh b/benchmarks/single_node/minimaxm2.5_fp8_mi355x.sh
index 53cffceee..f5979dedb 100755
--- a/benchmarks/single_node/minimaxm2.5_fp8_mi355x.sh
+++ b/benchmarks/single_node/minimaxm2.5_fp8_mi355x.sh
@@ -27,6 +27,13 @@ fi
 export VLLM_ROCM_USE_AITER=1
 export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4
 export VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=1
+VLLM_BLOCK_SIZE=16
+
+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."
+fi
 
 SERVER_LOG=/workspace/server.log
 PORT=${PORT:-8888}
@@ -52,7 +59,7 @@ $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" \
 --trust-remote-code > $SERVER_LOG 2>&1 &