[WIP] Update vllm B300 Minimax M3

benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b300.sh

@@ -3,7 +3,13 @@
 # MiniMax-M3 NVFP4 B300 single-node vLLM recipe.
 # Same shape as minimaxm3_fp8_b300.sh but uses the nvidia/MiniMax-M3-NVFP4
 # checkpoint. MiniMax-M3 modelopt NVFP4 support (vllm-project/vllm PR #46380) is
-# baked into the perf container image, so no runtime patch is needed.
+# baked into the perf container image.
+#
+# At runtime the recipe swaps the image's FlashInfer for the pinned nightly
+# release and applies the CuTeDSL split-K gemm patch on top, mirroring vLLM
+# commits 82a00090a (nightly install) and 0a16bea6b (patch). It also patches
+# the installed vLLM with the VLLM_MINIMAX_M3_BF16_ROUTER_GEMM knob (bf16
+# router gate weights, cuBLAS bf16 x bf16 -> fp32 logits) and enables it.
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
@@ -19,6 +25,51 @@ check_env_vars \
     RANDOM_RANGE_RATIO \
     RESULT_FILENAME
 
+# --- FlashInfer nightly + CuTeDSL split-K gemm patch -------------------------
+# Pinned to nightly release nightly-v0.6.14-20260702. The wheels are not yet on
+# the flashinfer.ai/whl/nightly flat index, so install by direct release URL.
+FLASHINFER_VERSION=0.6.14.dev20260702
+FLASHINFER_NIGHTLY_TAG=nightly-v0.6.14-20260702
+FLASHINFER_RELEASE_URL="https://github.com/flashinfer-ai/flashinfer/releases/download/${FLASHINFER_NIGHTLY_TAG}"
+
+# The flashinfer-jit-cache wheel is CUDA-version-specific; derive cuXYZ from
+# the torch build inside the container (e.g. 13.0 -> cu130).
+CUDA_MAJOR_MINOR="cu$(python3 -c 'import torch; print(torch.version.cuda.replace(".", ""))')" \
+    || { echo "Failed to determine CUDA version from torch" >&2; exit 1; }
+
+python3 -m pip uninstall -y flashinfer-python flashinfer-cubin flashinfer-jit-cache
+
+python3 -m pip install --pre \
+    "${FLASHINFER_RELEASE_URL}/flashinfer_python-${FLASHINFER_VERSION}-py3-none-any.whl" \
+    "${FLASHINFER_RELEASE_URL}/flashinfer_cubin-${FLASHINFER_VERSION}-py3-none-any.whl" \
+    "${FLASHINFER_RELEASE_URL}/flashinfer_jit_cache-${FLASHINFER_VERSION}+${CUDA_MAJOR_MINOR}-cp39-abi3-manylinux_2_28_$(uname -m).whl" \
+    || { echo "FlashInfer nightly install failed" >&2; exit 1; }
+
+# Apply CuTeDSL split-K gemm patch (jiahanc/flashinfer branch cutedsl-splitK,
+# flashinfer/gemm changes only) on top of the pinned nightly. The reinstall
+# above guarantees pristine files, so the patch always applies cleanly.
+FLASHINFER_PATCH="$(dirname "$0")/patches/flashinfer-cutedsl-splitk-gemm.patch"
+if ! command -v patch >/dev/null 2>&1; then
+    apt-get update -y && apt-get install -y --no-install-recommends patch \
+        || { echo "Failed to install patch(1)" >&2; exit 1; }
+fi
+SITE_PACKAGES=$(dirname "$(python3 -c "import importlib.util; print(importlib.util.find_spec('flashinfer').submodule_search_locations[0])")") \
+    || { echo "Could not locate the installed flashinfer package" >&2; exit 1; }
+patch -p1 -d "${SITE_PACKAGES}" < "$FLASHINFER_PATCH" \
+    || { echo "FlashInfer CuTeDSL split-K gemm patch failed" >&2; exit 1; }
+
+# Patch the installed vLLM with the VLLM_MINIMAX_M3_BF16_ROUTER_GEMM knob
+# (bf16 router gate weights + cuBLAS bf16 x bf16 -> fp32 router logits).
+# vLLM is not reinstalled above, so guard against re-applying on reruns.
+VLLM_ROUTER_PATCH="$(dirname "$0")/patches/minimax-m3-bf16-router-gemm.patch"
+if patch -R -p1 -s -f --dry-run -d "${SITE_PACKAGES}" < "$VLLM_ROUTER_PATCH" >/dev/null 2>&1; then
+    echo "[minimax-m3-bf16-router-gemm] already applied"
+else
+    patch -p1 -d "${SITE_PACKAGES}" < "$VLLM_ROUTER_PATCH" \
+        || { echo "MiniMax M3 bf16 router gemm patch failed" >&2; exit 1; }
+fi
+# -----------------------------------------------------------------------------
+
 if [[ -n "${MODEL_PATH:-}" ]]; then
     if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
         hf download "$MODEL" --local-dir "$MODEL_PATH"
@@ -38,7 +89,9 @@ SERVER_LOG=/workspace/server.log
 
 export VLLM_ENGINE_READY_TIMEOUT_S=3600
 export VLLM_FLOAT32_MATMUL_PRECISION=high
+export VLLM_MINIMAX_M3_BF16_ROUTER_GEMM=1
 export VLLM_FLASHINFER_ALLREDUCE_BACKEND=trtllm
+export VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=1800
 
 if [ "${DP_ATTENTION}" = "true" ]; then
   PARALLEL_ARGS="--tensor-parallel-size=1 --data-parallel-size=$TP --enable-expert-parallel"
@@ -57,6 +110,7 @@ start_gpu_monitor
 set -x
 vllm serve "$MODEL_PATH" --served-model-name "$MODEL" --host 0.0.0.0 --port $PORT \
 $PARALLEL_ARGS \
+--attention_config.indexer_kv_dtype fp8 \
 --gpu-memory-utilization 0.95 \
 --max-model-len $MAX_MODEL_LEN \
 --kv-cache-dtype fp8 \