diff --git a/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py b/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py index 1e4a268306..945a7df07c 100644 --- a/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py +++ b/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py @@ -347,9 +347,7 @@ class ROCMAiterMLASparseMetadata(AttentionMetadata): class ROCMAiterMLASparseMetadataBuilder( AttentionMetadataBuilder[ROCMAiterMLASparseMetadata] ): - _cudagraph_support: ClassVar[AttentionCGSupport] = ( - AttentionCGSupport.UNIFORM_SINGLE_TOKEN_DECODE - ) + _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.UNIFORM_BATCH def __init__( self, @@ -365,6 +363,9 @@ class ROCMAiterMLASparseMetadataBuilder( self.device = device max_num_batched_tokens = vllm_config.scheduler_config.max_num_batched_tokens + self.vllm_config = vllm_config + self._init_reorder_batch_threshold(1, supports_spec_as_decode=True) + self.num_heads = self.model_config.get_num_attention_heads(parallel_config) self.mla_dims = get_mla_dims(self.model_config) self.topk_tokens = vllm_config.model_config.hf_config.index_topk