EngineX/xc-llm-kunlun
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[Feature] support deepseek v3/r1/v3.2 (#78)

Browse Source 
* [Feature] support deepseek v3/r1/v3.2

* fix gpt_oss

* update readme

* update readme

---------

Co-authored-by: hanhaowen <hanhaowen@baidu.com>
This commit is contained in:
baoqian426
2026-01-05 22:55:35 +08:00
committed by GitHub
parent 07bc24a555
commit ee0f50e68f
27 changed files with 5760 additions and 621 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm_kunlun/models/gpt_oss.py
View File

@@ -16,7 +16,7 @@ from vllm.distributed import (get_ep_group, get_pp_group,
get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
tensor_model_parallel_all_gather)
from vllm_kunlun.ops.fused_moe.layer import FusedMoE
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (QKVParallelLinear,
RowParallelLinear)
@@ -176,7 +176,7 @@ class MLPBlock(torch.nn.Module):
x = sequence_parallel_chunk(x)
g = self.router(x)
x = self.experts(hidden_states=x, router_logits=g, linear_weights=self.router.weight)
x = self.experts(hidden_states=x, router_logits=g)
if self.is_sequence_parallel:
x = tensor_model_parallel_all_gather(x.contiguous(), 0)