diff --git a/tests/v1/attention/test_mla_backends.py b/tests/v1/attention/test_mla_backends.py index b90600ce40..4695810c36 100644 --- a/tests/v1/attention/test_mla_backends.py +++ b/tests/v1/attention/test_mla_backends.py @@ -827,13 +827,25 @@ def test_mock_mla_dcp_fp8_decode_gathers_quantized_query( ) -def test_tokenspeed_mla_dcp_single_token_decode_contract(monkeypatch): +def test_tokenspeed_mla_noncausal_capability(): + builder = tokenspeed_mla_module.TokenspeedMLAMetadataBuilder + assert builder.supports_non_causal_multi_token_decode + assert tokenspeed_mla_module.TokenspeedMLABackend.supports_non_causal() + + +@pytest.mark.parametrize( + ("causal", "tokens_per_decode", "dcp_world_size", "dcp_rank"), + [ + pytest.param(True, 1, 2, 1, id="causal-dcp"), + pytest.param(False, 3, 1, 0, id="noncausal-multi-token"), + ], +) +def test_tokenspeed_mla_decode_contract( + monkeypatch, causal, tokens_per_decode, dcp_world_size, dcp_rank +): decode_call = None num_decodes = 2 - tokens_per_decode = 1 num_decode_tokens = num_decodes * tokens_per_decode - dcp_world_size = 2 - dcp_rank = 1 num_heads = 128 kv_lora_rank = 512 qk_rope_head_dim = 64 @@ -879,6 +891,7 @@ def test_tokenspeed_mla_dcp_single_token_decode_contract(monkeypatch): num_decodes=num_decodes, num_decode_tokens=num_decode_tokens, max_seq_len=max_seq_len, + causal=causal, decode=SimpleNamespace( block_table=torch.empty((num_decodes, 1), dtype=torch.int32), seq_lens=torch.tensor([16, max_seq_len], dtype=torch.int32), @@ -913,9 +926,13 @@ def test_tokenspeed_mla_dcp_single_token_decode_contract(monkeypatch): ) torch.testing.assert_close(decode_call["seq_lens"], metadata.decode.seq_lens) torch.testing.assert_close(decode_call["block_tables"], metadata.decode.block_table) - torch.testing.assert_close( - decode_call["causal_seqs"], metadata.decode.dcp_tot_seq_lens - ) + if dcp_world_size > 1: + torch.testing.assert_close( + decode_call["causal_seqs"], metadata.decode.dcp_tot_seq_lens + ) + else: + assert decode_call["causal_seqs"] is None + assert decode_call["causal_mask"] is causal assert decode_call["return_lse"] is True assert decode_call["cp_world"] == dcp_world_size assert decode_call["cp_rank"] == dcp_rank diff --git a/vllm/v1/attention/backends/mla/tokenspeed_mla.py b/vllm/v1/attention/backends/mla/tokenspeed_mla.py index 2daf85ccb3..df4e5c45d1 100644 --- a/vllm/v1/attention/backends/mla/tokenspeed_mla.py +++ b/vllm/v1/attention/backends/mla/tokenspeed_mla.py @@ -58,6 +58,9 @@ def _get_workspace( class TokenspeedMLAMetadataBuilder(MLACommonMetadataBuilder[MLACommonMetadata]): _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.UNIFORM_BATCH query_len_support: ClassVar[QueryLenSupport] = QueryLenSupport.UNIFORM + # The kernel accepts an explicit causal mask, so a non-causal DSpark + # block can remain fused instead of being flattened to single tokens. + supports_non_causal_multi_token_decode: ClassVar[bool] = True def __init__( self, @@ -103,6 +106,10 @@ class TokenspeedMLABackend(MLACommonBackend): def supports_compute_capability(cls, capability: DeviceCapability) -> bool: return capability.major == 10 + @classmethod + def supports_non_causal(cls) -> bool: + return True + @classmethod def supports_combination( cls, @@ -302,6 +309,7 @@ class TokenspeedMLAImpl(MLACommonImpl[MLACommonMetadata]): output_scale=self.output_scale, enable_pdl=False, return_lse=return_lse, + causal_mask=attn_metadata.causal, causal_seqs=causal_seqs if self.dcp_world_size > 1 else None, cp_world=self.dcp_world_size, cp_rank=self.dcp_rank,