Activate FlashHead under vllm serve

[WilhelmTr]

Summary

• Add patch_async_llm targeting AsyncLLM.__init__, so the FlashHead metadata-load runs under both the Python LLM(...) API and vllm serve. The existing patch_llm only covers LLMEngine.from_engine_args, which vllm serve never reaches in vLLM 0.19 (the OpenAI entrypoint goes through AsyncLLM.from_vllm_config then AsyncLLM.__init__).
• Drop the negative-result cache in logits_processor._get_flash_head so a metadata file that appears after server startup is still picked up on the next decode step.
• Bump to 0.1.10 to trigger the PyPI release.

What went wrong today (repro)

With flash-head==0.1.9 installed against vllm==0.19.1, running

vllm serve embedl/Cosmos-Reason2-2B-W4A16-Edge2-FlashHead \
    --max-model-len 8192 --gpu-memory-utilization 0.75 --max-num-seqs 2

starts up and serves correctly, but /tmp/flashhead_metadata.pt is never written. get_flash_head() returns None, and the patched LogitsProcessor._get_logits falls straight through to the original dense path on every decode step, so FlashHead is effectively disabled under vllm serve, silently.

Traced through: vllm.entrypoints.openai.api_server.build_async_engine_client_from_engine_args calls AsyncLLM.from_vllm_config which calls AsyncLLM.__init__. LLMEngine.from_engine_args (the legacy class the current patch targets) is never called. The Python LLM(...) API still works because LLM.__init__ does call LLMEngine.from_engine_args.

Verification

Before: no [FlashHead] Loaded lazily... log from either process after startup, no /tmp/flashhead_metadata.pt, dense-head fallback.

After, with this PR:

flash_head.patches.async_llm INFO [FlashHead] Patched AsyncLLM.__init__
flash_head.patches INFO [FlashHead] All patches applied
flash_head INFO [FlashHead] Plugin registered
flash_head.loading INFO [FlashHead] Metadata prepared for lazy loading from flash_head_assets
flash_head.patches.async_llm INFO [FlashHead] Metadata saved for model: embedl/Cosmos-Reason2-2B-W4A16-Edge2-FlashHead
...
flash_head.loading INFO [FlashHead] Loaded lazily on GPU using 'lm_head.weight'

Exact curl from the README returns a coherent detailed video description.

Note (not fixed here)

vLLM's DEFAULT_LOGGING_CONFIG only attaches a handler to the vllm logger, so every [FlashHead] ... INFO line is dropped unless the user sets VLLM_LOGGING_CONFIG_PATH to a config that includes a flash_head logger. Worth either adding a handler inside register(), or mentioning in the README that the activation banner won't appear under vllm serve by default.

[JonnaMat]

This is needed since get_flash_head() may be None (e.g., when running non-FlashHead models).

[JonnaMat]

I think we should add a guard for idempotence (similar to what we do in logits_processor.py) [if _flash_head is None:...]

While AsyncLLm is run only once per engine construction (not per decode / request) there may be other parts of vllm that call it. We could add a _FLASH_HEAD_NOT_LOADED.