Add b300-cw (CoreWeave B300) runner launch script and pool

.github/configs/runners.yaml

@@ -145,6 +145,8 @@ b300:
 - 'b300-nv_8'
 b300-p1:
 - 'b300-p1'
+b300-cw:
+- 'b300-cw_0'
 gb300:
 - 'gb300-nv_0'
 - 'gb300-nv_1'

runners/launch_b300-cw.sh

@@ -0,0 +1,72 @@
+#!/usr/bin/env bash
+
+# Single-node launcher for the CoreWeave B300 cluster (runner b300-cw).
+# Follows the launch_b200-cw.sh CoreWeave template: allocate one node, import
+# the container to that node's local /tmp under flock, then srun the benchmark
+# in the same allocation. Importing to node-local /tmp (rather than shared NFS)
+# avoids the enroot aufs-whiteout failures that root-squash NFS triggers
+# (see launch_b300-nv.sh).
+
+export HF_HUB_CACHE_MOUNT="/tmp/gharunner/hf-hub-cache"
+export PORT=8888
+
+MODEL_CODE="${EXP_NAME%%_*}"
+FRAMEWORK_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
+SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
+# Prefer a framework-tagged script (e.g. dsv4_fp4_b300_vllm.sh) so models
+# with multiple inference engines can coexist; fall back to the historical
+# name without an engine suffix (`_trt` for trt, bare for everyone else).
+BENCH_BASE="benchmarks/single_node/${SCENARIO_SUBDIR}${MODEL_CODE}_${PRECISION}_b300"
+BENCH_SCRIPT="${BENCH_BASE}_${FRAMEWORK}${SPEC_SUFFIX}.sh"
+if [[ ! -f "$BENCH_SCRIPT" ]]; then
+    BENCH_SCRIPT="${BENCH_BASE}${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh"
+fi
+
+PARTITION="b300"
+SQUASH_FILE="/tmp/gharunner/squash/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
+LOCK_FILE="${SQUASH_FILE}.lock"
+
+CONTAINER_MOUNT_DIR=/workspace
+
+set -x
+
+JOB_ID=$(salloc --partition=$PARTITION --gres=gpu:b300:$TP --time=180 --no-shell --job-name="$RUNNER_NAME" 2>&1 | tee /dev/stderr | grep -oP 'Granted job allocation \K[0-9]+')
+
+if [ -z "$JOB_ID" ]; then
+    echo "ERROR: salloc failed to allocate a job"
+    exit 1
+fi
+
+# Use Docker image directly for openai/gpt-oss-120b with trt, otherwise use squash file
+if [[ "$MODEL" == "openai/gpt-oss-120b" && "$FRAMEWORK" == "trt" ]]; then
+    CONTAINER_IMAGE=$IMAGE
+else
+    # Use flock to serialize concurrent imports to the same squash file.
+    # mkdir on the worker first: /tmp/gharunner is node-local and may not
+    # exist on a freshly allocated node.
+    srun --jobid=$JOB_ID --job-name="$RUNNER_NAME" bash -c "
+        mkdir -p /tmp/gharunner/squash \"$HF_HUB_CACHE_MOUNT\"
+        exec 9>\"$LOCK_FILE\"
+        flock -w 600 9 || { echo 'Failed to acquire lock for $SQUASH_FILE'; exit 1; }
+        if unsquashfs -l \"$SQUASH_FILE\" > /dev/null 2>&1; then
+            echo 'Squash file already exists and is valid, skipping import'
+        else
+            rm -f \"$SQUASH_FILE\"
+            enroot import -o \"$SQUASH_FILE\" docker://$IMAGE
+        fi
+    "
+    # Squash file lives on the allocated worker node's /tmp, which is not
+    # visible from the host, so realpath on the host would return empty.
+    # Pass the path as-is; srun resolves it inside the job.
+    CONTAINER_IMAGE=$SQUASH_FILE
+fi
+
+srun --jobid=$JOB_ID \
+--container-image=$CONTAINER_IMAGE \
+--container-mounts=$GITHUB_WORKSPACE:$CONTAINER_MOUNT_DIR,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE \
+--container-mount-home \
+--container-workdir=$CONTAINER_MOUNT_DIR \
+--no-container-entrypoint --export=ALL \
+bash "$BENCH_SCRIPT"
+
+scancel $JOB_ID