[GopeshKh]

What changed and why

vLLM v0.31.0 turned XPU graph capture on by default and removed VLLM_XPU_ENABLE_XPU_GRAPH. The vLLM skills, emit_launch.sh, and the model-config-recommend launch line still hardcoded --enforce-eager, a flag
that now disables exactly what the version enables. The current vllm/vllm-openai-xpu:latest is v0.32.0, so anyone following the skills today serves at roughly half the decode throughput they should get.

This PR makes graphs-on the default everywhere. Eager stays available as an explicit opt-in for pre-v0.31.0 images. The PR also corrects the bench and profile guidance, so users stop benchmarking and profiling a configuration
nobody should deploy.

• vllm-xpu-run: drop --enforce-eager from the quickstart, multi-GPU, and remote-deploy launch lines. Rewrite the flag-table row and the CUDA translation row to say when the flag is still warranted (pre-v0.31.0 images, or isolating a suspected graph-capture bug). Add a one-liner to check the image's vLLM version.

• emit_launch.sh / recommend.py: graphs on by default. New opt-in --enforce-eager flag for older images.

• vllm-xpu-bench / vllm-xpu-profile: remove the flag from the example launches. Note that an eager baseline understates real throughput, and that wide inter-step gaps on v0.31.0+ are a real finding rather than eager-mode
  noise.

• torch-xpu-bench: reframe the compile A/B as deciding whether to add eager back for a problem model, not whether to drop it.

• scripts/generate_agents.py: read and write with explicit UTF-8, so the generator runs on Windows Python. The cp1252 default raised UnicodeDecodeError on the non-ASCII glyphs in SKILL.md bodies. Output on Linux is unchanged.

Behavior check of the emitters:

$ emit_launch.sh --model Qwen/Qwen2.5-1.5B-Instruct --quant bf16
        --dtype bfloat16 \
        --block-size=64 \
        --max-model-len 4096 \
        --gpu-memory-utilization 0.85

$ emit_launch.sh --model Qwen/Qwen2.5-1.5B-Instruct --quant bf16 --enforce-eager
        --dtype bfloat16 \
        --enforce-eager \
        --block-size=64 \
        ...

recommend.py behaves the same with and without --enforce-eager.

Type of change

• New skill
• Fix to an existing skill
• Performance or accuracy improvement (include benchmark data below)
• Docs or tooling

Validation

• bash scripts/check-skills.sh passes locally
• python3 scripts/generate_agents.py was run after any SKILL.md
  frontmatter change (or the pre-commit hook handled it)
• For skill behavior changes, the relevant acceptance layer in
  HOW_TO_TEST.md was run on real hardware, or the reason it was not
  is explained above
• Every commit is signed off (git commit -s), per the Developer
  Certificate of Origin in CONTRIBUTING.md

Also passing: tests/static.sh, tests/secure-config.sh (22/22),
tests/model-fit.sh.

Benchmark data

Config	TPOT	TTFT
--enforce-eager (previous default)	64 ms	unchanged

4x Intel Arc Pro B70, concurrency 8, vLLM-XPU v0.31.0+. Same model, same image, same hardware; the only difference is the flag.

[YuningQiu]

LGTM! Thanks!