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

9
vllm_kunlun/ops/rotary_embedding.py
View File

@@ -19,7 +19,9 @@ import torch
import xspeedgate_ops
import os
from vllm.model_executor.layers.rotary_embedding import (
RotaryEmbedding, YaRNScalingRotaryEmbedding, DynamicNTKScalingRotaryEmbedding, MRotaryEmbedding)
RotaryEmbedding, YaRNScalingRotaryEmbedding,
DynamicNTKScalingRotaryEmbedding, MRotaryEmbedding,
DeepseekScalingRotaryEmbedding)
from typing import Optional, Tuple
def vllm_kunlun_compute_cos_sin_cache(self) -> torch.Tensor:
@@ -143,12 +145,15 @@ def vllm_kunlun_mrope_forward_cuda(
return query, key
DeepseekScalingRotaryEmbedding_forward = DeepseekScalingRotaryEmbedding.forward
DeepseekScalingRotaryEmbedding_forward_cuda = DeepseekScalingRotaryEmbedding.forward_cuda
RotaryEmbedding.forward_cuda = vllm_kunlun_forward_cuda
RotaryEmbedding.forward = vllm_kunlun_forward_cuda
DeepseekScalingRotaryEmbedding.forward = DeepseekScalingRotaryEmbedding_forward
DeepseekScalingRotaryEmbedding.forward_cuda = DeepseekScalingRotaryEmbedding_forward_cuda
MRotaryEmbedding.forward_cuda = vllm_kunlun_mrope_forward_cuda
MRotaryEmbedding.forward = vllm_kunlun_mrope_forward_cuda
def Split_Norm_Rope(
qkv: torch.Tensor,
cos_sin_cache: torch.Tensor,