muh-engine
|
c0395ade14
|
[ENGINE] muh_cc_dispatch.py: CCCL cc_dispatch.cuh Python port
Unified kernel policy dispatch — single entry point for ALL kernel configs.
Architecture directly mirrors CCCL cc_dispatch.cuh:
dispatch_compute_cap(policy_selector, cc, functor)
→ policy_getter<PolicySelector, CC>{}()
→ concrete policy struct
Our equivalent:
dispatch_kernel_config('attention', hw=BI_V100)
→ pre-computed AttentionConfig (frozen dataclass)
Includes:
- HardwareCapability (mirrors hardware.cuh bi_v100())
- AttentionConfig (mirrors ReducePolicy for V1/V2 dispatch)
- MoEConfig (mirrors TopkPolicy for fused_moe BLOCK_SIZE_M)
- TransformConfig (mirrors transform bytes_in_flight)
- CacheConfig (mirrors batch_memcpy threads)
- grid_even_share() (Python port of GridEvenShare::DispatchInit)
- check_smem() (SMEM constraint checker used by all policies)
- Pre-computed configs for Qwen3.6 at import time
(lowest_cc_resolver pattern: compute once, lookup always)
CCCL files read: cc_dispatch.cuh, dispatch_reduce.cuh,
dispatch_transform.cuh, dispatch_topk.cuh, dispatch_common.cuh,
grid_even_share.cuh, agent_reduce.cuh
|
2026-08-05 09:25:48 +00:00 |
|