AlvisGong ef8157a5f2 fixed fused alltoall execute all reduce (#5109) ...

### What this PR does / why we need it?
fixed fused alltoall execute all reduce, when moe_comm_type is
MoECommType.FUSED_ALLTOALL

if moe_comm_type in {MoECommType.ALLTOALL, MoECommType.MC2,
MoECommType.FUSED_ALLTOALL} \
                    and not shared_expert_dp_enabled():
shared_out = tensor_model_parallel_all_reduce(shared_out)


- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: AlvisGong <gwly0401@163.com>
Co-authored-by: Jade Zheng <zheng.shoujian@outlook.com>

2025-12-18 15:07:40 +08:00

..

fused_moe

fixed fused alltoall execute all reduce (#5109)

2025-12-18 15:07:40 +08:00

triton

qwen3_next add triton ops : fused_qkvzba_split_reshape (#4788)

2025-12-18 11:31:04 +08:00

__init__.py

[Fusion] [Graph] Add qknorm rope fusion operator (#4711)

2025-12-17 08:53:44 +08:00

activation.py

[refact] unified soc_version code (#4359)

2025-11-26 14:28:55 +08:00

expert_load_balancer.py

eplb redundant expert bugfix (#4291)

2025-11-21 14:24:35 +08:00

layernorm.py

Adopt inductor fusion and define quantization fusion pass (#4168)

2025-12-04 10:29:48 +08:00

linear_op.py

[feat] apply flashcomm1 on bailing (#4868)

2025-12-11 17:02:21 +08:00

linear.py

[Bugfix] Remove ModelSlim-"M4 Quantization". (#4589)

2025-12-01 23:45:02 +08:00

mla.py

upgrade vLLM to main (#4608)

2025-12-02 22:10:52 +08:00

register_custom_ops.py

[Refactor] Remove the process patches of Qwen2.5-VL and Qwen2.5-Omni (#5035)

2025-12-16 11:43:52 +08:00

rotary_embedding.py

[main] rename device type (#5099)

2025-12-17 14:08:19 +08:00

shared_weight_layer.py

[Feat] Flashcomm2 use o_shared linear (#4188)

2025-12-11 12:43:04 +08:00

vocab_parallel_embedding.py

[Feat] Add custom Embedding tensor model parallel (#2616)

2025-12-12 14:41:20 +08:00

weight_prefetch.py

Update torch-npu version to 2.7.1 (#3896)

2025-10-31 17:16:31 +08:00