17720b5386
cc_dispatch.cuh is CCCL's runtime-hardware → compile-time-policy bridge: 1. Detect device compute_capability at runtime 2. policy_selector(cc) returns full kernel config 3. lowest_cc_resolver merges identical policies across CCs 4. dispatch_compute_cap bridges runtime → compile-time specialization Translated as _HardwarePolicy class in qwen3_5.py: 1. detect() probes BI-V100 capabilities once (SMEM, cuSOLVER, MoE ops) 2. Returns deltanet_chunk_size, solve_triangular_available, moe_native_* 3. All kernel code reads from _hw_policy instead of hardcoded constants 4. MoE forward skips native attempt if hasattr() shows ops missing Concrete changes: - DeltaNet chunk_size: hw_policy-selected (64 if solve_tri, 32 if not) - _forward_sub_lower: no per-call try/except, uses pre-detected flag - _DNN_CHUNK: reads from hw_policy - MoE native: hasattr() pre-check avoids exception on every layer init CCCL source: cub/cub/detail/cc_dispatch.cuh (full file translation) Maps to: qwen3_6_scripts/qwen3_5.py