7 Commits

Author	SHA1	Message	Date
biondizzle	9d067add90	Fix device reference in full_attention_reference	2026-05-19 08:01:31 +00:00
biondizzle	3e3e998578	Fix attention: manual causal mask for batched single-query	2026-05-19 08:01:08 +00:00
biondizzle	1e675ccc9a	Fix causal mask shape for SDPA: (1,1,T,T) broadcast	2026-05-19 08:00:39 +00:00
biondizzle	57615029a4	Fix KV expand for SDPA: (T,HD) → (T*NH, T, HD)	2026-05-19 08:00:08 +00:00
biondizzle	dd3a12bbda	Fix full_attention_reference: broadcast KV to all heads+positions	2026-05-19 07:59:28 +00:00
biondizzle	910015c47e	Fix kv shape: expand to (T, NH, HD) before reshape	2026-05-19 07:58:42 +00:00
biondizzle	3de75c4e37	Add CSA/HCA attention kernel (PyTorch SDPA, Blackwell-safe) ... Replaces vLLM's broken FlashMLA sparse attention which doesn't work on SM100 (Blackwell). Uses torch.nn.functional.scaled_dot_product_attention which works on all GPUs. Architecture: - CSA (C128A): Batched sparse gather + SDPA on top-k positions - HCA (C4A): Same with compressed KV + per-layer indexer - SWA: Sliding window attention - Full reference: standard SDPA for testing without compression Also adds test_csa_attention_b200.py to verify the full attention path.	2026-05-19 07:58:10 +00:00