fix: use max_context_len for stable num_splits in paged decode

- PagedKV::host_kv_len now returns max_context_len instead of max_seq_len - Eliminates grid-z instability for CUDA graph capture/replay - Restore skip_no_kernel re-export accidentally removed by ruff --fix
2026-08-07 14:42:53 +08:00
parent 55ee258e95
commit 0e7fe57d96
2 changed files with 2 additions and 1 deletions
@@ -111,7 +111,7 @@ struct PagedKV {
        return p.max_q_len;
    }
    HOST_DEV_FORCEINLINE int host_kv_len(const AttentionParams<bf16>& p) {
-        return p.max_seq_len;
+        return p.max_context_len;
    }

    // prefill: Q rows start at qo_indptr[batch] (ragged batch base)