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

5
vllm_kunlun/models/mimo_v2_flash.py
View File

@@ -21,7 +21,7 @@ from vllm.distributed import (
tensor_model_parallel_all_gather,
)
from vllm.logger import init_logger
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 (
MergedColumnParallelLinear,
@@ -185,8 +185,7 @@ class MiMoV2MoE(nn.Module):
gate_input = hidden_states
router_logits = self.gate(gate_input)
final_hidden_states = self.experts(
hidden_states=hidden_states, router_logits=router_logits, linear_weights=self.gate.weight
)
hidden_states=hidden_states, router_logits=router_logits)
return final_hidden_states.squeeze(0) if is_input_1d else final_hidden_states