ca3697f4b0
Eliminates per-decode-step torch.full/torch.zeros GPU allocations that cause OOM after thousands of generation steps (case_truncation max_tokens=8192). Three allocation sites replaced with staging buffer .fill_()/.zero_() reuse: - scores_padded: torch.full([H, padded_len], -inf) → _staging_scores slice - v_padded_kv: torch.zeros([kv_h, padded_len, d]) → _staging_v_kv slice - v_padded: torch.zeros([H, padded_len, d]) → _staging_v slice Pattern from CCCL cub/detail/warpspeed/resource/smem_resource.cuh: SmemResource pre-allocates stageCount buffers, nextStage() cycles through them. PyTorch translation: allocate once at function entry, slice per step.