From 4be3dcf0fc7a9086d978eaea3c5a4d78fb98e44a Mon Sep 17 00:00:00 2001 From: Lucas Wilkinson Date: Mon, 14 Sep 2026 10:20:46 -0400 Subject: [PATCH] [PCP][DCP] Declare FlashMLASparse MTP support at CP interleave > 1 (#56722) Co-authored-by: Claude Opus 5 (1M context) --- vllm/v1/attention/backends/mla/flashmla_sparse.py | 1 + 1 file changed, 1 insertion(+) diff --git a/vllm/v1/attention/backends/mla/flashmla_sparse.py b/vllm/v1/attention/backends/mla/flashmla_sparse.py index 433cded567..d8c4368401 100644 --- a/vllm/v1/attention/backends/mla/flashmla_sparse.py +++ b/vllm/v1/attention/backends/mla/flashmla_sparse.py @@ -634,6 +634,7 @@ class FlashMLASparseMetadataBuilder( class FlashMLASparseImpl(SparseMLACommonImpl[FlashMLASparseMetadata]): can_return_lse_for_decode: bool = True supports_dcp: bool = True + supports_mtp_with_cp_non_trivial_interleave_size: bool = True @staticmethod def _compute_fp8_decode_padded_heads(num_heads: int) -> int: