perf: flatten paged prefill tile dispatch

- remove the host-provided max_q_len argument - dispatch only the ragged prefill tile upper bound - validate the rebuilt CUDA backend end to end
2026-08-09 23:12:53 +08:00
parent cd31f1f62f
commit c5fba9c238
10 changed files with 69 additions and 33 deletions
@@ -590,7 +590,6 @@ class CudaBackend(AttentionBackend):
            kv_indptr,
            qo_indptr,
            attn_mask,
-            q_len,
            is_causal=is_causal,
        )
        return out.reshape(b, q_len, q.size(2), q.size(3)).flatten(2)