feat: enable CB paged decode fast path for MPS hub kernels + sensible block_size default
Two changes to make continuous batching work well on Apple Silicon (MPS):
1. initialization.py - ensure_decode_fast_path_is_available:
Previously only FA3/CUDA and FA2/XPU could use the paged decode fast path
(block tables + flash_attn_with_kvcache). Hub kernels like
kernels-community/metal-flash-sdpa were excluded entirely, forcing
max_blocks_per_request = 0 and losing the decode fast path even when
the kernel ships flash_attn_with_kvcache.
Now checks is_kernel(config._attn_implementation) so any hub kernel
that provides flash_attn_with_kvcache is automatically picked up,
regardless of device. This unblocks the metal-flash-sdpa paged-decode
kernel (HF discussion #4 on kernels-community/metal-flash-sdpa) once
its bindings land.
2. serve.py - MPS block_size default:
The default block_size=256 causes a ~17x slowdown on MPS because the
paged SDPA fallback (used when the kernel lacks block-table support)
builds a full attention mask proportional to block_size, and MPS SDPA
does not have the fused kernel that makes large blocks efficient on CUDA.
Default to block_size=32 on MPS when not explicitly set. This is a
temporary measure until flash_attn_with_kvcache ships in the
metal-flash-sdpa kernel, at which point the block-table decode fast path
will bypass the mask entirely.
Tested on M3 Max with Qwen2.5-0.5B-Instruct: 5 sequential CB requests
complete in ~4.3s each (vs 38-57s with block_size=256, or indefinite
hang with the stale kernel revision on previous main).