[Bugfix] Disable sequence parallelism for Dots3 NOTE (#52172)
Signed-off-by: KurodaKanbei <mistergalahad@gmail.com>
范裕达 committed
Aug 13, 2026 at 22:13 UTC
170592a931d92572f65ee808b3b03e6eed2f54a3
1 file changed
+3
-5
vllm/models/dots3_note/nvidia/model.py
+3
-5
@@ -497,6 +497,7 @@ class Dots3NoteDecoderLayer(DeepseekV32DecoderLayer):
497
layer_idx = int(prefix.split(sep=".")[-1])
498
self.layer_idx = layer_idx
499
self.use_mha = False
500
+ self.use_sequence_parallel = False
501
attention_cls = (
502
Dots3NoteSlidingAttention
503
if config.layer_types[layer_idx] == "sliding_attention"
@@ -536,11 +537,7 @@ class Dots3NoteDecoderLayer(DeepseekV32DecoderLayer):
537
prefix=f"{prefix}.mlp",
538
reduce_results=False,
539
)
539
- self.use_sequence_parallel_moe = (
540
- parallel_config.use_sequence_parallel_moe
541
- and parallel_config.pipeline_parallel_size == 1
542
- and isinstance(self.mlp, DeepseekV2MoE)
543
- )
540
+ self.use_sequence_parallel_moe = False
541
self.tp_size = parallel_config.tensor_parallel_size
542
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
543
self.post_attention_layernorm = RMSNorm(
@@ -559,6 +556,7 @@ class Dots3NoteModel(DeepseekV32Model):
556
self.config = config
557
self.device = current_platform.device_type
558
self.vocab_size = config.vocab_size
559
+ self.use_sequence_parallel = False
560
self.is_v32 = True
561
self._weight_block_size = getattr(quant_config, "weight_block_size", None)
562
topk_indices_buffer = torch.empty(