Commits · e606d4b24d5daa3b690d5305360952fa946bb826 · OpenDAS / vllm_cscc

07 May, 2025 1 commit
- [fix]修复cudagraph和eager分段模式下，开启mla后报错问题 · e0c80c12
  王敏 authored May 07, 2025
  
  e0c80c12
06 May, 2025 1 commit
- [fix]修复并行解码多机推理报错 · 0858880a
  王敏 authored May 06, 2025
  
  0858880a
29 Apr, 2025 4 commits
- 需改mtp 16卡性能差问题 · 5e77e9b1
  xiabo authored Apr 29, 2025
  
  5e77e9b1
- debug and fix error · 333e3374
  lizhigong authored Apr 29, 2025
  
  333e3374
- fix，去掉对base类的耦合 · 3bbb6e9d
  lizhigong authored Apr 29, 2025
  
  3bbb6e9d
- add triton_key · 62920e37
  zhuwenwen authored Apr 29, 2025
  
  62920e37
27 Apr, 2025 2 commits
- resolve the issue of requiring HIP_VISIBLEDEVICES to be set in order to run multiple cards · abf008ef
  zhuwenwen authored Apr 27, 2025
  
  abf008ef
- update version · 469a3a88
  zhuwenwen authored Apr 27, 2025
  
  469a3a88
24 Apr, 2025 2 commits
- update utils.py · f51086de
  zhuwenwen authored Apr 24, 2025
  
  f51086de
- fix llama and qwen layout · db94f061
  zhuwenwen authored Apr 24, 2025
  
  db94f061
23 Apr, 2025 1 commit
- fix nccl 报超时的问题，原因是双线程调度导致提前关闭远端worker，warmup后远端worker执行异常，引发rank0超时。 · 14f46a65
  lizhg1 authored Apr 23, 2025
```
http://hpczentao.sugon.com/bug-view-93029.html
```
  14f46a65
22 Apr, 2025 1 commit
- [fix]修复开并行解码后，speculative-disable-by-batch-size设的比测试的batch小的话可能出现的数组越界问题 · 8776c63c
  王敏 authored Apr 22, 2025
  
  8776c63c
18 Apr, 2025 1 commit
- deepseek_v3/r1 int8 量化首字调优 · 8fc15e04
  gaoqiong authored Apr 18, 2025
  
  8fc15e04
17 Apr, 2025 2 commits
- [fix]修复开启并行解码后，在极端测试情况下，由于设置了speculative-disable-by-batch-size导致不跑并行解码导致previo... · 7a23da92
  王敏 authored Apr 17, 2025
```
[fix]修复开启并行解码后，在极端测试情况下，由于设置了speculative-disable-by-batch-size导致不跑并行解码导致previous_hidden_states不断增加，最终导致显存用尽服务无响应问题
```
  7a23da92
- add K100AI custom allreduce · ec078ef1
  xiabo authored Apr 17, 2025
  
  ec078ef1
15 Apr, 2025 4 commits
- [feat]添加VLLM_ENFORCE_EAGER_BS_THRESHOLD环境变量，支持cudagraph模式下，当bs大于阈值时，强制切换为eager模式，对大bs有效果 · 3ce3e3d9
  王敏 authored Apr 15, 2025
  
  3ce3e3d9
- back to mla v2 · 6f49c1ed
  zhuwenwen authored Apr 15, 2025
  
  6f49c1ed
- vdim pad 32 · cf28e5a4
  zhuwenwen authored Apr 15, 2025
  
  cf28e5a4
- set vdim=128 · 9c3190d0
  zhuwenwen authored Apr 15, 2025
  
  9c3190d0
14 Apr, 2025 3 commits
- support numa bind · ec2e17d8
  zhuwenwen authored Apr 14, 2025
  
  ec2e17d8
- update vdim pad · 1d57ec3d
  zhuwenwen authored Apr 14, 2025
  
  1d57ec3d
- update prefix cache interface and vdim padding strategy · 4564b2f5
  zhuwenwen authored Apr 14, 2025
  
  4564b2f5
11 Apr, 2025 1 commit
- fix glm register error · b99b5676
  zhuwenwen authored Apr 11, 2025
  
  b99b5676
10 Apr, 2025 3 commits
- add auto finish thread when use LLM object · c78f6594
  lizhigong authored Apr 10, 2025
  
  c78f6594
- support vdim 128 · b62fe3c1
  zhuwenwen authored Apr 10, 2025
  
  b62fe3c1
- fix bug for decode len > max model len · fdf9bf98
  lizhigong authored Apr 10, 2025
  
  fdf9bf98
09 Apr, 2025 2 commits
- update triton_decode_attention.py · 085c3454
  zhuwenwen authored Apr 09, 2025
  
  085c3454
- add info log reminder zero overhead start · 09553647
  lizhigong authored Apr 09, 2025
  
  09553647
08 Apr, 2025 2 commits
- update triton_decode_attention.py · 8b5a09f6
  zhuwenwen authored Apr 08, 2025
  
  8b5a09f6
- fix bug when scheduler_outputs changed cause assert error · d9789532
  lizhigong authored Apr 08, 2025
  
  d9789532
07 Apr, 2025 2 commits
- update layout · f6044f1a
  zhuwenwen authored Apr 07, 2025
  
  f6044f1a
- update layout · 525d9d7e
  zhuwenwen authored Apr 07, 2025
  
  525d9d7e
03 Apr, 2025 3 commits
- debug and fix error · 20e9c3eb
  lizhigong authored Apr 03, 2025
  
  20e9c3eb
- update triton_decode_attention.py · a30b3ce2
  zhuwenwen authored Apr 03, 2025
  
  a30b3ce2
- [update] MLA stage=2 并启动pipelinev2，修复pipeline subop 操作类型不一致 · 343a10fa
  zhuwenwen authored Apr 03, 2025
  
  343a10fa
02 Apr, 2025 1 commit
- Fix reasoning_content for chat_template include <think> tag as input · 2c35b6cd
  zhuwenwen authored Apr 02, 2025
  
  2c35b6cd
31 Mar, 2025 2 commits
- add VLLM_USE_FLUX · f7f16f29
  zhuwenwen authored Mar 31, 2025
  
  f7f16f29
- suppoet nn layout · 3ab7f0ef
  zhuwenwen authored Mar 31, 2025
  
  3ab7f0ef
28 Mar, 2025 2 commits
- fix pending error in zero overhead · e6bbec9e
  lizhigong authored Mar 28, 2025
  
  e6bbec9e
- [Kernel] Prototype integration of flux kernels · 5e766c85
  zhuwenwen authored Mar 28, 2025
  
  5e766c85