feat: add modular vLLM post-install patch framework with hybrid KV cache fix

[m199369309]

Summary

• Add xinference/model/llm/vllm/patches/ module with auto-discovery registry for applying post-install patches to vLLM site-packages
• Implement hybrid KV cache page size patch (LCM-based padding) for Qwen3.5/3.6/3.6-MoE models, working around vllm-project/vllm#37121
• Use subprocess-based vLLM path resolution to correctly handle uv-managed virtual environments and .pth-injected parent site-packages
• Filter patches by model architectures (not model names), automatically covering fine-tuned and custom-registered models

Motivation

Qwen3.5/3.6 models use hybrid attention (linear_attention + full_attention) with non-divisible KV cache page sizes. vLLM 0.20.x raises NotImplementedError during unify_kv_cache_spec_page_size(). The official fix (vllm-project/vllm#40128) has not been merged.

Previous monkey-patching approach fails because vLLM's EngineCore uses spawn multiprocessing (due to CUDA initialization), which doesn't inherit in-memory patches. File-level patching solves this by modifying the source directly in site-packages.

Design

xinference/model/llm/vllm/patches/
├── __init__.py                          # Auto-discovery registry + apply_vllm_patches()
├── _base.py                             # VllmPatch dataclass
└── hybrid_kv_cache_page_size.py         # Qwen3.5/3.6 KV cache patch

• Auto-discovery: pkgutil.iter_modules scans for modules exposing PATCH attribute
• Architecture filtering: Each patch declares architectures set; uses _resolve_architectures() to cover fine-tuned models that inherit from built-in families
• Idempotent: Marker-based detection prevents double-patching
• Subprocess path resolution: Uses venv's own Python interpreter to locate vLLM via importlib.util.find_spec, handling .pth injections and uv-managed envs
• Fallback: Glob pattern for standard venv layouts

Changes

File	Change
xinference/model/llm/vllm/patches/__init__.py	New: registry + entry point
xinference/model/llm/vllm/patches/_base.py	New: VllmPatch dataclass
xinference/model/llm/vllm/patches/hybrid_kv_cache_page_size.py	New: KV cache patch with subprocess path resolution
xinference/core/worker.py	Add model_name/architectures params to _prepare_virtual_env, call apply_vllm_patches after install

Removal condition

When vLLM officially merges the hybrid KV cache fix (tracking: vllm-project/vllm#37121, #38041), delete the patch file. The framework (__init__.py, _base.py) can remain for future patches or be removed entirely (worker.py has ImportError guard).

Test plan

• Deploy Qwen3.6-35B-A3B-FP8 with vLLM 0.20.x on multi-GPU worker
• Verify worker logs show: Applied hybrid KV cache patch to .../kv_cache_utils.py
• Verify logs show: Hybrid KV cache: padding max page size 1073152 -> 1081344
• Verify no NotImplementedError during model initialization
• Verify non-hybrid models (e.g., Qwen2.5) are unaffected (patch skipped)
• Verify fine-tuned model with model_family="qwen3.6" gets patch applied via architecture inheritance

[qinxuye]

LGTM

[qinxuye]

Tests failed, seemed that the other model types are effected.