From 2c17d33f4291a55b447317640c81eb61077b1b00 Mon Sep 17 00:00:00 2001 From: music-dino <111048524+music-dino@users.noreply.github.com> Date: Thu, 9 Jul 2026 04:09:42 +0200 Subject: [PATCH] [Bugfix][ROCm] Change AttentionCGSuppoort in TritonMLA to UNIFORM_SINGLE_TOKEN_DECODE (#47144) Signed-off-by: Dino Music Co-authored-by: Andreas Karatzas --- vllm/v1/attention/backends/mla/triton_mla.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/vllm/v1/attention/backends/mla/triton_mla.py b/vllm/v1/attention/backends/mla/triton_mla.py index 3b91c22516..acc9c9cb50 100644 --- a/vllm/v1/attention/backends/mla/triton_mla.py +++ b/vllm/v1/attention/backends/mla/triton_mla.py @@ -48,7 +48,9 @@ def _compute_num_kv_splits(max_seq_len: int, sm_count: int) -> int: class TritonMLAMetadataBuilder(MLACommonMetadataBuilder[MLACommonMetadata]): - _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.UNIFORM_BATCH + _cudagraph_support: ClassVar[AttentionCGSupport] = ( + AttentionCGSupport.UNIFORM_SINGLE_TOKEN_DECODE + ) def __init__(self, kv_cache_spec, layer_names, vllm_config, device): super().__init__(kv_cache_spec, layer_names, vllm_config, device)