Centralize runner model path mappings in runners.yaml / 在 runners.yaml 中集中管理 runner 模型路径映射

configs/runners.yaml

@@ -330,3 +330,317 @@ hardware:
   cluster:mi355x-amds:
     available-cpu-dram-mib: 3_095_781
     gpus-per-node: 8
+models:
+  # Cluster-local checkpoints used by launchers instead of downloading the
+  # portable HuggingFace model IDs in the benchmark configs. When multiple
+  # paths are listed, the first existing directory wins and the first entry is
+  # the deterministic fallback. Framework-specific mappings take precedence
+  # over the default mapping.
+  cluster:b200-dgxc:
+    default:
+      dsr1:
+        fp4:
+          # Per-node /raid is not populated reliably across the cluster.
+          model-paths:
+            - /scratch/fsw/models/DeepSeek-R1-0528-NVFP4-v2
+          srt-slurm-model-prefix: dsr1
+        fp8:
+          model-paths:
+            - /lustre/fsw/models/dsr1-0528-fp8
+          srt-slurm-model-prefix: dsr1-fp8
+      dsv4:
+        fp4:
+          model-paths:
+            - /lustre/fsw/models/deepseek-v4-pro
+            - /lustre/fsw/models/dsv4-pro
+            - /lustre/fsw/models/DeepSeek-V4-Pro
+          srt-slurm-model-prefix: deepseek-v4-pro
+          # Preserve the operator override supported by the previous launcher.
+          allow-model-path-override: true
+      qwen3.5:
+        bf16:
+          model-paths:
+            - /lustre/fsw/models/Qwen3.5-397B-A17B
+          srt-slurm-model-prefix: qwen3.5
+        fp8:
+          model-paths:
+            - /lustre/fsw/models/Qwen3.5-397B-A17B-FP8
+          srt-slurm-model-prefix: qwen3.5-fp8
+        fp4:
+          model-paths:
+            - /lustre/fsw/models/Qwen3.5-397B-A17B-NVFP4
+          srt-slurm-model-prefix: qwen3.5-fp4
+      glm5:
+        fp8:
+          model-paths:
+            - /lustre/fsw/models/GLM-5-FP8
+          srt-slurm-model-prefix: glm5-fp8
+        fp4:
+          model-paths:
+            - /lustre/fsw/models/GLM-5-NVFP4
+          srt-slurm-model-prefix: glm5-fp4
+      kimik2.5:
+        int4:
+          model-paths:
+            - /lustre/fsw/models/Kimi-K2.5
+          srt-slurm-model-prefix: kimik2.5
+        fp4:
+          model-paths:
+            - /lustre/fsw/models/Kimi-K2.5-NVFP4
+          srt-slurm-model-prefix: kimik2.5-fp4
+      minimaxm2.5:
+        fp8:
+          model-paths:
+            - /lustre/fsw/models/MiniMax-M2.5
+          srt-slurm-model-prefix: minimax-m2.5-fp8
+        fp4:
+          model-paths:
+            - /lustre/fsw/models/MiniMax-M2.5-NVFP4
+          srt-slurm-model-prefix: minimax-m2.5-nvfp4
+      gptoss:
+        fp4:
+          model-paths:
+            - /lustre/fsw/models/gpt-oss-120b
+          srt-slurm-model-prefix: gptoss
+      minimaxm3:
+        fp8:
+          # Day-zero model staged in the sa-shared-writable gharunners tree.
+          model-paths:
+            - /lustre/fsw/gharunners/models/MiniMax-M3-MXFP8
+          srt-slurm-model-prefix: minimax-m3-mxfp8
+        fp4:
+          model-paths:
+            - /scratch/fsw/models/MiniMax-M3-NVFP4
+          srt-slurm-model-prefix: minimax-m3-nvfp4
+  cluster:b300-nv:
+    default:
+      dsr1:
+        fp4:
+          model-paths:
+            - /data/models/dsr1-fp4
+          served-model-name: deepseek-r1-fp4
+          srt-slurm-model-prefix: dsr1
+        fp8:
+          model-paths:
+            - /data/models/dsr1-fp8
+          served-model-name: deepseek-r1-fp8
+          srt-slurm-model-prefix: dsr1-fp8
+    dynamo-vllm:
+      dsv4:
+        fp4:
+          model-paths:
+            - /data/models/dsv4-pro
+            - /data/models/deepseek-v4-pro
+            - /data/models/DeepSeek-V4-Pro
+          srt-slurm-model-prefix: deepseek-v4-pro
+          allow-model-path-override: true
+      minimaxm2.5:
+        fp4:
+          model-paths:
+            - /data/models/MiniMax-M2.5-NVFP4
+          srt-slurm-model-prefix: minimax-m2.5-nvfp4
+        fp8:
+          model-paths:
+            - /data/models/MiniMax-M2.5
+          srt-slurm-model-prefix: minimax-m2.5-fp8
+      minimaxm3:
+        fp4:
+          model-paths:
+            - /scratch/models/MiniMax-M3-NVFP4
+          srt-slurm-model-prefix: nvidia/MiniMax-M3-NVFP4
+        fp8:
+          model-paths:
+            - /data/models/MiniMax-M3-MXFP8
+          srt-slurm-model-prefix: MiniMaxAI/MiniMax-M3-MXFP8
+  cluster:gb200-nv:
+    llmd-vllm:
+      dsv4:
+        fp4:
+          model-paths:
+            - /mnt/numa1/models/DeepSeek-V4-Pro
+          model-name: deepseek-ai/DeepSeek-V4-Pro
+    dynamo-sglang:
+      dsr1:
+        fp8:
+          model-paths:
+            - /mnt/lustre01/models/deepseek-r1-0528
+          srt-slurm-model-prefix: dsr1-fp8
+        fp4:
+          model-paths:
+            - /mnt/lustre01/models/deepseek-r1-0528-fp4-v2/
+          srt-slurm-model-prefix: dsr1-fp4
+      dsv4:
+        fp4:
+          model-paths:
+            - /mnt/lustre01/models/deepseek-v4-pro
+          srt-slurm-model-prefix: deepseek-v4-pro
+      glm5.1:
+        fp4:
+          model-paths:
+            - /mnt/lustre01/models/GLM-5.1-NVFP4
+          srt-slurm-model-prefix: glm-5-fp4
+        fp8:
+          model-paths:
+            - /mnt/lustre01/models/GLM-5.1-FP8
+          srt-slurm-model-prefix: glm-5.1-fp8
+      qwen3.5:
+        fp8:
+          model-paths:
+            - /mnt/lustre01/models/Qwen3.5-397B-A17B-FP8
+          srt-slurm-model-prefix: qwen3.5-fp8
+    dynamo-trt:
+      gptoss:
+        default:
+          model-paths:
+            - /mnt/lustre01/models/gpt-oss-120b
+          served-model-name: gpt-oss-120b
+      dsr1:
+        fp4:
+          model-paths:
+            - /mnt/numa1/models/DeepSeek-R1-0528-NVFP4-v2
+          served-model-name: deepseek-r1-fp4
+          srt-slurm-model-prefix: dsr1
+        fp8:
+          model-paths:
+            - /mnt/numa1/models/DeepSeek-R1-0528
+          served-model-name: deepseek-r1-fp8
+          srt-slurm-model-prefix: dsr1-fp8
+      kimik2.5:
+        fp4:
+          model-paths:
+            - /mnt/lustre01/models/kimi-k2.5-nvfp4
+          served-model-name: kimi-k2.5-nvfp4
+          srt-slurm-model-prefix: nvidia/Kimi-K2.5-NVFP4
+      glm5:
+        fp4:
+          model-paths:
+            - /mnt/lustre01/slurm-shared/glm-model/GLM-5-NVFP4
+          served-model-name: glm-5-nvfp4
+          srt-slurm-model-prefix: nvidia/GLM-5-NVFP4
+    dynamo-vllm:
+      kimik2.5:
+        fp4:
+          model-paths:
+            - /mnt/lustre01/models/kimi-k2.5-nvfp4
+          srt-slurm-model-prefix: kimi-k2.5-nvfp4
+      dsv4:
+        fp4:
+          # Compute-visible NVFP4 checkpoint; the lowercase sibling is FP8.
+          model-paths:
+            - /mnt/lustre01/models/DeepSeek-V4-Pro-NVFP4/
+          srt-slurm-model-prefix: deepseek-v4-pro
+      minimaxm2.5:
+        fp4:
+          model-paths:
+            - /mnt/lustre01/models/MiniMax-M2.5-NVFP4
+          srt-slurm-model-prefix: minimax-m2.5-nvfp4
+        fp8:
+          model-paths:
+            - /mnt/lustre01/models/MiniMax-M2.5
+          srt-slurm-model-prefix: minimax-m2.5-fp8
+      minimaxm3:
+        fp8:
+          model-paths:
+            - /mnt/lustre01/models/MiniMax-M3-MXFP8
+          srt-slurm-model-prefix: minimax-m3-mxfp8
+  cluster:gb300-nv:
+    default:
+      dsr1:
+        fp4:
+          model-paths:
+            - /scratch/models/DeepSeek-R1-0528-NVFP4-v2
+          served-model-name: deepseek-r1-fp4
+          srt-slurm-model-prefix: dsr1
+        fp8:
+          model-paths:
+            - /scratch/models/DeepSeek-R1-0528
+          served-model-name: deepseek-r1-fp8
+          srt-slurm-model-prefix: dsr1-fp8
+      dsv4:
+        fp4:
+          # Compute-node-local path; srtctl preflight must be skipped.
+          model-paths:
+            - /scratch/models/DeepSeek-V4-Pro
+          srt-slurm-model-prefix: deepseek-v4-pro
+      glm5.1:
+        fp4:
+          model-paths:
+            - /scratch/models/GLM-5.1-NVFP4
+          srt-slurm-model-prefix: glm-5-fp4
+      glm5:
+        fp4:
+          model-paths:
+            - /scratch/models/GLM-5-NVFP4
+          srt-slurm-model-prefix: glm-5-fp4
+        fp8:
+          model-paths:
+            - /scratch/models/GLM-5-FP8
+          srt-slurm-model-prefix: glm-5-fp8
+      minimaxm2.5:
+        fp4:
+          model-paths:
+            - /data/models/MiniMax-M2.5-NVFP4
+          srt-slurm-model-prefix: minimax-m2.5-nvfp4
+        fp8:
+          model-paths:
+            - /data/models/MiniMax-M2.5
+          srt-slurm-model-prefix: minimax-m2.5-fp8
+      minimaxm3:
+        fp8:
+          model-paths:
+            - /data/models/MiniMax-M3-MXFP8
+          srt-slurm-model-prefix: minimax-m3-mxfp8
+      kimik2.5:
+        fp4:
+          model-paths:
+            - /scratch/models/Kimi-K2.5-NVFP4
+          srt-slurm-model-prefix: nvidia/Kimi-K2.5-NVFP4
+      qwen3.5:
+        fp4:
+          model-paths:
+            - /scratch/models/Qwen3.5-397B-A17B-NVFP4
+          srt-slurm-model-prefix: qwen3.5-fp4
+        fp8:
+          model-paths:
+            - /scratch/models/Qwen3.5-397B-A17B-FP8
+          srt-slurm-model-prefix: qwen3.5-fp8
+    dynamo-trt:
+      dsv4:
+        fp4:
+          model-paths:
+            - /scratch/models/DeepSeek-V4-Pro
+          srt-slurm-model-prefix: deepseek-ai/DeepSeek-V4-Pro
+      glm5:
+        fp4:
+          model-paths:
+            - /scratch/models/GLM-5-NVFP4
+          served-model-name: glm-5-nvfp4
+          srt-slurm-model-prefix: nvidia/GLM-5-NVFP4
+  cluster:h100-dgxc:
+    dynamo-sglang:
+      dsr1:
+        fp8:
+          model-paths:
+            - /mnt/nfs/lustre/models/dsr1-fp8
+          srt-slurm-model-prefix: dsr1-fp8
+    dynamo-trt:
+      dsr1:
+        fp8:
+          model-paths:
+            - /mnt/nfs/lustre/models/dsr1-fp8
+          served-model-name: DeepSeek-R1-0528
+          srt-slurm-model-prefix: DeepSeek-R1-0528
+  cluster:h200-dgxc:
+    dynamo-sglang:
+      dsr1:
+        fp8:
+          model-paths:
+            - /models/DeepSeek-R1-0528
+          srt-slurm-model-prefix: dsr1-fp8
+    dynamo-trt:
+      dsr1:
+        fp8:
+          model-paths:
+            - /models/DeepSeek-R1-0528
+          served-model-name: DeepSeek-R1-0528
+          srt-slurm-model-prefix: DeepSeek-R1-0528

runners/launch_b200-dgxc.sh

@@ -6,78 +6,9 @@ SLURM_ACCOUNT="benchmark"
 
 set -x
 
-# MODEL_PATH: Override with pre-downloaded paths on the shared Lustre tree.
-# Bench scripts and srt-slurm yaml configs specify HuggingFace model IDs for
-# portability, but we resolve to /lustre/fsw/models/* here to avoid repeated
-# downloading on every dgxc node. Runs for both single-node and multinode
-# launches.
-# NOTE: per-node /raid/models/* would be faster but is only populated on a
-# subset of dgxc nodes today, so we use Lustre for reliability.
-if [[ $MODEL_PREFIX == "dsr1" && $PRECISION == "fp4" ]]; then
-    export MODEL_PATH="/scratch/fsw/models/DeepSeek-R1-0528-NVFP4-v2"
-    export SRT_SLURM_MODEL_PREFIX="dsr1"
-elif [[ $MODEL_PREFIX == "dsr1" && $PRECISION == "fp8" ]]; then
-    export MODEL_PATH="/lustre/fsw/models/dsr1-0528-fp8"
-    export SRT_SLURM_MODEL_PREFIX="dsr1-fp8"
-elif [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" ]]; then
-    SELECTED_MODEL_PATH=""
-    if [[ -n "${MODEL_PATH:-}" && -d "${MODEL_PATH}" ]]; then
-        SELECTED_MODEL_PATH="$MODEL_PATH"
-    else
-        for candidate in /lustre/fsw/models/deepseek-v4-pro /lustre/fsw/models/dsv4-pro /lustre/fsw/models/DeepSeek-V4-Pro; do
-            if [[ -d "$candidate" ]]; then
-                SELECTED_MODEL_PATH="$candidate"
-                break
-            fi
-        done
-    fi
-    export MODEL_PATH="${SELECTED_MODEL_PATH:-/lustre/fsw/models/deepseek-v4-pro}"
-    export SRT_SLURM_MODEL_PREFIX="deepseek-v4-pro"
-elif [[ $MODEL_PREFIX == "qwen3.5" && $PRECISION == "bf16" ]]; then
-    export MODEL_PATH="/lustre/fsw/models/Qwen3.5-397B-A17B"
-    export SRT_SLURM_MODEL_PREFIX="qwen3.5"
-elif [[ $MODEL_PREFIX == "qwen3.5" && $PRECISION == "fp8" ]]; then
-    export MODEL_PATH="/lustre/fsw/models/Qwen3.5-397B-A17B-FP8"
-    export SRT_SLURM_MODEL_PREFIX="qwen3.5-fp8"
-elif [[ $MODEL_PREFIX == "qwen3.5" && $PRECISION == "fp4" ]]; then
-    export MODEL_PATH="/lustre/fsw/models/Qwen3.5-397B-A17B-NVFP4"
-    export SRT_SLURM_MODEL_PREFIX="qwen3.5-fp4"
-elif [[ $MODEL_PREFIX == "glm5" && $PRECISION == "fp8" ]]; then
-    export MODEL_PATH="/lustre/fsw/models/GLM-5-FP8"
-    export SRT_SLURM_MODEL_PREFIX="glm5-fp8"
-elif [[ $MODEL_PREFIX == "glm5" && $PRECISION == "fp4" ]]; then
-    export MODEL_PATH="/lustre/fsw/models/GLM-5-NVFP4"
-    export SRT_SLURM_MODEL_PREFIX="glm5-fp4"
-elif [[ $MODEL_PREFIX == "kimik2.5" && $PRECISION == "int4" ]]; then
-    export MODEL_PATH="/lustre/fsw/models/Kimi-K2.5"
-    export SRT_SLURM_MODEL_PREFIX="kimik2.5"
-elif [[ $MODEL_PREFIX == "kimik2.5" && $PRECISION == "fp4" ]]; then
-    export MODEL_PATH="/lustre/fsw/models/Kimi-K2.5-NVFP4"
-    export SRT_SLURM_MODEL_PREFIX="kimik2.5-fp4"
-elif [[ $MODEL_PREFIX == "minimaxm2.5" && $PRECISION == "fp8" ]]; then
-    export MODEL_PATH="/lustre/fsw/models/MiniMax-M2.5"
-    export SRT_SLURM_MODEL_PREFIX="minimax-m2.5-fp8"
-elif [[ $MODEL_PREFIX == "minimaxm2.5" && $PRECISION == "fp4" ]]; then
-    export MODEL_PATH="/lustre/fsw/models/MiniMax-M2.5-NVFP4"
-    export SRT_SLURM_MODEL_PREFIX="minimax-m2.5-nvfp4"
-elif [[ $MODEL_PREFIX == "gptoss" && $PRECISION == "fp4" ]]; then
-    export MODEL_PATH="/lustre/fsw/models/gpt-oss-120b"
-    export SRT_SLURM_MODEL_PREFIX="gptoss"
-elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" ]]; then
-    # Day-zero: MiniMax-M3-MXFP8 is not in the SRE-staged /lustre/fsw/models
-    # tree (root-owned); it lives in the sa-shared-writable gharunners tree.
-    export MODEL_PATH="/lustre/fsw/gharunners/models/MiniMax-M3-MXFP8"
-    export SRT_SLURM_MODEL_PREFIX="minimax-m3-mxfp8"
-elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" ]]; then
-    # NVFP4 checkpoint, pre-staged on the b200-dgxc scratch tree.
-    export MODEL_PATH="/scratch/fsw/models/MiniMax-M3-NVFP4"
-    export SRT_SLURM_MODEL_PREFIX="minimax-m3-nvfp4"
-else
-    echo "Unsupported model prefix/precision: $MODEL_PREFIX/$PRECISION"
-    echo "Available models under /lustre/fsw/models:"
-    ls -la /lustre/fsw/models
-    exit 1
-fi
+# shellcheck source=runners/runner_model_utils.sh
+source "$(dirname "${BASH_SOURCE[0]}")/runner_model_utils.sh"
+resolve_runner_model_config "cluster:b200-dgxc" || exit 1
 
 export AIPERF_MMAP_CACHE_HOST_PATH="/lustre/fsw/gharunners/aiperf-cache"