Add xpu case

.buildkite/generate_pipeline.py

@@ -273,9 +273,10 @@ def make_step(path, data, profiles):
         setup_commands = [setup_command("'lm-eval[api]' pyyaml bfcl-eval soundfile")]
     else:
         setup_commands = FULL_SETUP_COMMANDS
+    agents = {"queue": queue, **(profile.get("agent_tags") or {})}
     step = {
         "label": f"{emoji} {name}",
-        "agents": {"queue": queue},
+        "agents": agents,
         "timeout_in_minutes": timeout,
         "commands": setup_commands + [RUN_TEMPLATE.format(path=path)],
         "artifact_paths": ["results/**/*"],

lib/gpu_profiles.yaml

@@ -33,3 +33,13 @@ MI355X:
   k8s_plugin: amd
   # See MI300X — HF cache defaults to an emptyDir-backed container path; override
   # the volume source per-cluster with MI355X_HF_CACHE_VOLUME.
+
+XPU:
+  queue: intel-gpu
+  agent_tags:
+    label: performance
+    gpu: 1+
+    mem: 32+
+  image_repo: vllm/vllm-openai-xpu
+  server_runtime: xpu_docker
+  hf_home: /data/huggingface

lib/server.sh

@@ -38,6 +38,24 @@ start_server() {
   local docker_args=(--gpus all --ipc=host --ulimit nofile=65536:65536
                      -e VLLM_ENGINE_READY_TIMEOUT_S=3600
                      -p "${port}:${port}")
+  local container_command=("$image" "$model" --port "$port" $serve_args)
+  if [[ "$runtime" == "xpu_docker" ]]; then
+    docker_args=(
+      --device /dev/dri:/dev/dri
+      --net=host
+      --ipc=host
+      --privileged
+      --ulimit nofile=65536:65536
+      -e VLLM_ENGINE_READY_TIMEOUT_S=3600
+    )
+    [[ -d /dev/dri/by-path ]] && docker_args+=(-v /dev/dri/by-path:/dev/dri/by-path)
+    [[ -n "${ZE_AFFINITY_MASK:-}" ]] && docker_args+=(-e "ZE_AFFINITY_MASK=${ZE_AFFINITY_MASK}")
+    [[ -n "${HF_TOKEN:-}" ]] && docker_args+=(-e "HF_TOKEN=${HF_TOKEN}")
+    local bash_serve_cmd
+    printf -v bash_serve_cmd 'source /root/.bashrc >/dev/null 2>&1; exec vllm serve %q --port %q %s' \
+      "$model" "$port" "$serve_args"
+    container_command=(--entrypoint /bin/bash "$image" -ic "$bash_serve_cmd")
+  fi
   local hf_home=""
   while IFS= read -r kv; do
     [[ -z "$kv" ]] && continue
@@ -52,8 +70,7 @@ start_server() {
   # vllm/vllm-openai's entrypoint takes the model as the first positional
   # arg; do not prepend `vllm` or `serve`.
   docker run -d --rm --name "$container" "${docker_args[@]}" \
-    "$image" \
-    "$model" --port "$port" $serve_args
+    "${container_command[@]}"
 
   # Install pytest to avoid cupy.testing import failure during torch.compile
   docker exec "$container" pip install -q pytest 2>/dev/null || true

workloads/gpt_oss_120b_B70.yaml

@@ -0,0 +1,48 @@
+# GPT-OSS 120B on XPU (8xXPU, TP=4)
+# Tracking perf regression: https://github.com/vllm-project/vllm/issues/40838
+name: gpt_oss_120b-xpu
+gpu: XPU
+num_gpus: 4
+nightly: true
+
+vllm:
+  model: openai/gpt-oss-120b
+  serve_args: >-
+    --tensor-parallel-size 4
+    --max-model-len 32768
+    --max-num-seqs 512
+    --enforce-eager
+    --trust-remote-code
+    --no-enable-prefix-caching
+    --tool-call-parser openai
+    --enable-auto-tool-choice
+
+lm_eval:
+  model_args:
+    tokenized_requests: false
+    tokenizer_backend: null
+    timeout: 6000
+  tasks:
+    - name: gsm8k
+      num_fewshot: 5
+      model_args:
+        num_concurrent: 64
+        max_length: 32768
+        max_gen_toks: 8192
+
+bfcl:
+  test_categories:
+    - simple_python
+    - multiple
+    - multi_turn
+  num_threads: 8
+
+vllm_bench:
+  configs:
+    - name: 4k-in-1k-out-conc-32
+      backend: openai
+      dataset: random
+      input_len: 4096
+      output_len: 1024
+      num_prompts: 128
+      max_concurrency: 32