[Bugfix][IPEX] Add VLLM_CPU_MOE_PREPACK to allow disabling MoE prepack when CPU does not support it

[gau-nernst]

On my Ryzen 5600

vllm serve deepseek-ai/DeepSeek-V2-Lite-Chat --trust-remote-code --max-model-len 4000

curl http://0.0.0.0:8000/v1/completions -H "Content-Type: application/json" -d '{"model": "deepseek-ai/DeepSeek-V2-Lite-Chat", "prompt": "Hello"}'

ERROR 03-12 21:39:46 [engine.py:141]   File "/home/thien/code/vllm_cpu/.venv/lib/python3.11/site-packages/intel_extension_for_pytorch/frontend.py", line 557, in optimize
ERROR 03-12 21:39:46 [engine.py:141]     assert core.onednn_has_bf16_support(), (
ERROR 03-12 21:39:46 [engine.py:141]            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 03-12 21:39:46 [engine.py:141] AssertionError: BF16 weight prepack needs the cpu support avx_ne_convert or avx512bw, avx512vl and avx512dq, but the desired instruction sets are not available. Please set dtype to torch.float or set weights_prepack to False.

Hence, I use the value of core.onednn_has_bf16_support() to decide whether to enable weight prepack.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[gau-nernst]

@bigPYJ1151 Do you mind taking a look at this PR? Thank you

[bigPYJ1151]

Perhaps make this config as a environment variable is better, such as VLLM_CPU_MOE_PREPACK, by default is False.

You can refer to VLLM_CPU_OMP_THREADS_BIND for adding an env in vLLM.

[gau-nernst]

I have made the change. I set True by default to maintain current behavior, unless you want to change the default to False instead? (Then we probably need to document this flag somewhere e.g. CPU perf)

[bigPYJ1151]

No problem.
Please also update the section Related runtime environment variables in docs/source/getting_started/installation/cpu.md, thanks :)

[bigPYJ1151]

LGTM, thanks for your fix!
@Isotr0py Please help to take a look, thanks :)