Revert "[Rocm][Kimi-k3] Fix pipeline_parallel support for the kimik3 DCP mode (#53664)" (#56429)
Signed-off-by: Shanshan Shen <87969357+shen-shanshan@users.noreply.github.com>
Shanshan Shen committed
Sep 12, 2026 at 00:17 UTC
127143e27f2ba15e3b0ee6f4d73eaeaa60097bc2
2 files changed
-12
tests/models/test_registry.py
-2
@@ -129,8 +129,6 @@ def test_registry_model_property(model_arch, is_mm, init_cuda, score_type):
129
# in V1.
130
# ("MLPSpeculatorPreTrainedModel", False, False),
131
("DeepseekV2ForCausalLM", True, False),
132
- ("KimiLinearForCausalLM", True, False),
133
- ("KimiK3ForConditionalGeneration", True, False),
132
("Qwen2VLForConditionalGeneration", True, True),
133
],
134
)
vllm/model_executor/layers/attention/mla_attention.py
-10
@@ -2543,16 +2543,6 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
2543
if num_decodes > 0:
2544
dcp_tot_seq_lens_device = None
2545
if self.dcp_world_size > 1:
2546
- assert seq_lens is not None, (
2547
- "MLA DCP decode requires seq_lens on CommonAttentionMetadata"
2548
- )
2549
- if dcp_local_seq_lens is None:
2550
- dcp_local_seq_lens = get_dcp_local_seq_lens(
2551
- seq_lens,
2552
- dcp_size=self.dcp_world_size,
2553
- dcp_rank=get_dcp_group().rank_in_group,
2554
- cp_kv_cache_interleave_size=self.cp_kv_cache_interleave_size,
2555
- )
2546
dcp_tot_seq_lens_device = seq_lens[:num_decodes]
2547
seq_lens = dcp_local_seq_lens
2548