vllm

de830d3c - [Core] Extensible KV cache: VMM driver probe/fallback, manual size support

[Core] Extensible KV cache: VMM driver probe/fallback, manual size support - Extract the driver ctypes bindings into vllm/utils/vmm_driver.py behind a small VmmDriver interface (CUDA implementation; struct layouts and call signatures are shared with HIP for a future ROCm backend). - Probe VMM support on the workers (driver loads, VA reservation works) and fall back to standard KV cache allocation with a warning instead of failing on platforms without VMM (e.g. WSL2, non-GPU workers). - Support kv_cache_memory_bytes: the requested size is committed as-is after warmup (single sizing pass), still avoiding warmup-time OOM. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0133xqsNmqLHG9Pyhr5wSp1D Signed-off-by: Nick Hill <nickhill@us.ibm.com>