EngineX/xc-llm-kunlun
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity
142 Commits 2 Branches 0 Tags
b82b6026d65c69a2a65e4d55ffe2065c86e561dd
Commit Graph

36 Commits

Author SHA1 Message Date
Xinyu Dong
bf9369f733 Migrate XTorch operations to Kunlun operations (accelerating iteration) (#177) 
Signed-off-by: dongxinyu03 <dongxinyu03@baidu.com>
 2026-02-12 18:13:00 +08:00
Li Wei
744719587e [Feature] Support glmx (#194) 
Signed-off-by: Li Wei <liwei.109@outlook.com>
Co-authored-by: tangshiwen <tangshiwen@baidu.com>
Co-authored-by: Xinyu Dong <dongxinyu03@baidu.com>
 2026-02-12 15:40:42 +08:00
WANG HAO
6f30bc439d clean pr for ds.2 mtp support (#164) 
* Add MTP support in eagle.py

Signed-off-by: wanghao129 <wanghao129@baidu.com>

* new pr for mtp

Signed-off-by: wanghao129 <wanghao129@baidu.com>

* Revert formatting changes in deepseek_v2.py

Signed-off-by: wanghao129 <wanghao129@baidu.com>

---------

Signed-off-by: wanghao129 <wanghao129@baidu.com>
Co-authored-by: wanghao129 <wanghao129@baidu.com>
 2026-02-02 15:23:33 +08:00
Li Wei
71bd70ad6c [Feature] support compressed-tensors w4a16 quantization (#154) 
- native int4 kimi model inference is supported

Signed-off-by: Li Wei <liwei.109@outlook.com>
 2026-01-27 19:56:22 +08:00
Shiwen Tang
0711c1abfa [Feature] Support AWQ MoE W4A16 Quantization (#142) 
Signed-off-by: tangshiwen <tangshiwen@baidu.com>
Co-authored-by: Li Wei <liwei.109@outlook.com>
 2026-01-26 18:56:05 +08:00
baoqian426
1eaa1336ac [Bugfix]remove mla patch, server args no need --compilation-config for ds v3.1 (#145) 
Signed-off-by: baoqian426 <1354987947@qq.com>
 2026-01-23 15:59:43 +08:00
fromck
0ce5f1a3f7 Add kernels to optimize RoPE and the decoding stage (#143) 
Co-authored-by: chengxiaokang <chengxiaokang@baidu.com>
 2026-01-23 10:29:52 +08:00
fromck
74d4f804e8 add 2 kernels and optimize the calculation of topk_indices (#134) 
Co-authored-by: chengxiaokang <chengxiaokang@baidu.com>
 2026-01-22 10:29:28 +08:00
Li Wei
2a2d773ad0 [fix]bias bug in kunlun_scale_mm (#126) 2026-01-20 13:24:52 +08:00
Li Wei
8f56cbf3ed [refactor]update Kunlun classes with monkey patch (#122) 
Signed-off-by: Li Wei <liwei.109@outlook.com>
 2026-01-19 20:24:19 +08:00
baoqian426
2512259944 longcontext chunk make attention crash, fix it (#117) 
Co-authored-by: root <root@rdtest-node1150.bcc-zwlt.baidu.com>
 2026-01-17 18:38:23 +08:00
fromck
71a5a04e0a [Misc]Specify that DS32 only supports --kv-cache-dtype bfloat16 (#119) 
* [Kernel] add kernels to torch.ops

* [Misc]Specify that DS only supports --kv-cache-dtype bfloat16

---------

Co-authored-by: chengxiaokang <chengxiaokang@baidu.com>
 2026-01-17 16:52:02 +08:00
Shiwen Tang
8988ad08b2 [Feature] Support Mixed-Precision Quantization for MoE (#112) 2026-01-14 18:42:18 +08:00
baoqian426
eb40e8a07a [Bugfix] fix can not import compressed_tensors (#87) 
Co-authored-by: root <root@rdtest-node1150.bcc-zwlt.baidu.com>
 2026-01-07 11:32:10 +08:00
Li Wei
1c1b84d78c [fix]update compressed-tensors scheme 
Deepseek v3.2 is supported now

Signed-off-by: Li Wei <liwei.109@outlook.com>
 2026-01-06 22:30:27 +08:00
Li Wei
9533f68e99 [fix]matmul not support cuda graph 2026-01-06 17:32:45 +08:00
Li Wei
515a4eeda9 [dev] support compressed-tensors w8a8 quantization (#75) 
* [dev] support compressed-tensors w8a8 quantization

Co-authored-by: Li Wei <liwei.109@outlook.com>

* [refact]update KunlunScaleMMKernel impl

* [rebase]resolve conflicts and remove redundant code

---------

Co-authored-by: tangshiwen <tangshiwen@baidu.com>
 2026-01-06 13:51:53 +08:00
baoqian426
ee0f50e68f [Feature] support deepseek v3/r1/v3.2 (#78) 
* [Feature] support deepseek v3/r1/v3.2

* fix gpt_oss

* update readme

* update readme

---------

Co-authored-by: hanhaowen <hanhaowen@baidu.com>
 2026-01-05 22:55:35 +08:00
Xinyu Dong
07bc24a555 [Bugs] Fix moe when without bias (#76) 2026-01-05 10:51:23 +08:00
callmelaoyi
b86953acf9 [Kernel] Qwen3-next 优化 recompute_w_u_fwd & chunk_fwd_o (#74) 
Co-authored-by: yuanjizhong <yuanjizhong@baidu.com>
 2026-01-05 10:24:51 +08:00
Xinyu Dong
fe666fb24f [Feature] Support gpt-oss and update model list (#71) 
* [Docs] Update Support Models

* [Feature] Support gpt-oss

* [Docs] fix model support list

* Fix Moe

* Fix

* Fix moe_ep

* remove gpt oss graph support , not yet

---------

Co-authored-by: hanhaowen <hanhaowen@baidu.com>
 2026-01-04 21:19:49 +08:00
hanhaowen
b015bb76fd remove qwen2.py llama.py fix llama output 2025-12-31 11:39:37 +08:00
Xinyu Dong
b3c30a3cb9 [Feature] Support XiaoMi MIMO Flash V2 (#62) 
* [Feature] Support MIMO Flash V2
 2025-12-31 10:16:33 +08:00
Li Wei
9cee025f41 Merge pull request #59 from liwei109/aicapx-quant 
[fix]remove weight_loader_v2 to suport cuda graph
 2025-12-29 19:56:24 +08:00
baoqian426
45c6b8e927 Merge pull request #52 from liwei109/awq_gptq 
[dev] support AWQ/GPTQ quantization for dense models
 2025-12-24 17:05:26 +08:00
Li Wei
6546323c71 [dev] support AWQ/GPTQ quantization for dense models 2025-12-24 13:46:06 +08:00
Li Wei
383eb5459a [refactor] remove redundant code in linear 2025-12-24 12:02:09 +08:00
Xinyu Dong
75d0bdae2f Merge pull request #40 from ldh2020/v0.11.0dev 
[Kernel] Optimize the performance of Qwen3-Next
 2025-12-22 21:50:27 +08:00
hanhaowen
a4b9e92ca1 [Kernel] Replace native torch solve_tril by solve_tril_fwd kernel op 2025-12-22 17:37:19 +08:00
ldh2020
004e164bdb [Kernel] Optimize the recurrent op 2025-12-21 11:18:00 +08:00
ldh2020
fce97df908 [Kernel] Use l2norm kernel op instead of triton op. 2025-12-16 16:24:47 +08:00
ldh2020
cff4727fbb [Kernel] Optimize the performance of causal_conv1d. 2025-12-12 17:22:35 +08:00
ldh2020
9bb2ee06a4 [Bugfix] fix the bug of torch_solve_tril 2025-12-12 17:01:50 +08:00
chenyili
7c22d621fb 提交vllm0.11.0开发分支 2025-12-10 17:51:24 +08:00
zhaoyingzhuo
b614823125 [chore] Remove obsolete comments 2025-12-10 15:52:23 +08:00
dongxinyu03
c728e52505 Initial commit for vLLM-Kunlun Plugin 2025-12-10 12:05:39 +08:00