3426d8185a
从 CCCL cc_dispatch.cuh (150行) 读入完整的 compute capability 分派架构: dispatch_compute_cap → dispatch_to_cc_list → policy_getter<PolicySelector, CC> C++20: policy_constant 做相同 policy 的 CC 去重 C++17: lowest_cc_resolver 找最低 CC with same policy 从 qwen3_5.py 确认 Qwen3.6-35B-A3B 实际参数: head_dim = 256 (NOT 128) num_heads = 24, num_kv_heads = 4 GQA ratio = 6 关键修正: 1. head_dim 128→256 旧: qwen36_config(head_dim=128) → BLOCK_N=64 → SMEM=64×128×2×2=32KB ✓ 实际: head_dim=256 → BLOCK_N=64 → SMEM=64×256×2×2=64KB > 48KB → CRASH 修正: BLOCK_N=32 → SMEM=32×256×2×2=32KB ≤ 48KB ✓ 2. 删除 _read_reduce_config (依赖 gen_patch, 容器内不可用) 3. 删除 reduce_threads/reduce_items (ixformer 有自己的 reduce, 我们控制不了) 4. 强制 V1 (v1_v2_threshold = max_seq_len + 1) 5. Pre-computed configs at import time (mirrors CCCL compile-time instantiation)