Add run_vllm.sh: reference dual-Spark launch script with tuned env vars + flag rationale

run_vllm.sh

@@ -0,0 +1,67 @@
+#!/usr/bin/env bash
+# run_vllm.sh — reference vLLM launch for MiniMax-M2.7-NVFP4-GB10-AC on dual-Spark TP=2.
+#
+# Intended as a drop-in starter. Adjust MODEL_PATH, HOST, PORT, and the TP/cluster args
+# for your own Ray topology. See README.md §"Running on 2× DGX Spark (Tensor Parallel)"
+# for rationale behind every flag.
+#
+# Assumes:
+#   - Ray head + worker are already running (one per Spark)
+#   - Model is on a shared path or bind-mounted into both containers at MODEL_PATH
+#   - vllm ≥ 0.19 with FlashInfer MoE FP4 support
+
+set -euo pipefail
+
+MODEL_PATH="${MODEL_PATH:-/models/MiniMax-M2.7-NVFP4-GB10-AC}"
+SERVED_NAME="${SERVED_NAME:-minimax-m2.7-ac}"
+HOST="${HOST:-0.0.0.0}"
+PORT="${PORT:-30000}"
+GPU_MEM_UTIL="${GPU_MEM_UTIL:-0.88}"
+MAX_MODEL_LEN="${MAX_MODEL_LEN:-196608}"
+MAX_NUM_SEQS="${MAX_NUM_SEQS:-12}"
+MAX_NUM_BATCHED_TOKENS="${MAX_NUM_BATCHED_TOKENS:-32768}"
+TP_SIZE="${TP_SIZE:-2}"
+
+# --- Tuned environment variables ----------------------------------------------
+# Forum + vendor-recipe validated for MiniMax-M2.7 NVFP4 on Blackwell/GB10.
+# The single biggest win is VLLM_USE_FLASHINFER_MOE_FP4=1 — many older images
+# default this to 0, which falls back to the slower CUTLASS MoE path.
+
+export VLLM_USE_FLASHINFER_MOE_FP4=1
+export VLLM_NVFP4_GEMM_BACKEND=flashinfer-cutlass
+export VLLM_FLASHINFER_MOE_BACKEND=throughput
+export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
+export VLLM_FLOAT32_MATMUL_PRECISION=high
+export VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=1
+export SAFETENSORS_FAST_GPU=1
+export OMP_NUM_THREADS=8
+
+# --- Compilation config -------------------------------------------------------
+# cudagraph_mode=none is INTENTIONAL for dual-Spark Ray TP.
+# PIECEWISE captures cleanly in current vLLM builds (historical deadlock is fixed)
+# but measurably regresses decode throughput 12–20% on multi-node TP because each
+# piece boundary forces a cross-node sync whose cost exceeds the launch-overhead
+# savings. Revisit if you change the distributed_executor_backend away from Ray
+# or run on a single Spark.
+
+COMPILATION_CONFIG='{"cudagraph_mode":"none","inductor_compile_config":{"combo_kernels":false,"benchmark_combo_kernel":false,"max_autotune":false,"max_autotune_gemm":false}}'
+
+# --- vLLM serve ---------------------------------------------------------------
+exec vllm serve "$MODEL_PATH" \
+  --host "$HOST" --port "$PORT" \
+  --served-model-name "$SERVED_NAME" \
+  --tensor-parallel-size "$TP_SIZE" \
+  --distributed-executor-backend ray \
+  --gpu-memory-utilization "$GPU_MEM_UTIL" \
+  --max-model-len "$MAX_MODEL_LEN" \
+  --max-num-seqs "$MAX_NUM_SEQS" \
+  --max-num-batched-tokens "$MAX_NUM_BATCHED_TOKENS" \
+  --kv-cache-dtype fp8_e4m3 \
+  --attention-backend flashinfer \
+  --attention-config.use_trtllm_attention=0 \
+  --enable-prefix-caching \
+  --enable-chunked-prefill \
+  --trust-remote-code \
+  --enable-auto-tool-choice --tool-call-parser minimax_m2 \
+  --reasoning-parser minimax_m2_append_think \
+  --compilation-config "$COMPILATION_CONFIG"