Commits · 9f9886d8056615478fdcdab89fe866fce3064e89 · OpenDAS / vllm_cscc

02 Dec, 2025 2 commits
- add VLLM_USE_V32_ENCODE to use encoding_dsv32.py · 9f9886d8
  zhuwenwen authored Dec 02, 2025
  
  9f9886d8
- [fix]解决dp模式mtp卡住问题 · 90afe2e4
  zhuwenwen authored Dec 02, 2025
  
  90afe2e4
01 Dec, 2025 2 commits
- if kv_cache_detype=="fp8_e4m3", use non fused cat + mla · ba0cd35c
  zhuwenwen authored Dec 01, 2025
  
  ba0cd35c
- feat: pp 增加输出每个micro batch调度的请求数功能，VLLM_PP_DEBUG打开，默认关闭 · 92058666
  jujl1 authored Dec 01, 2025
  
  92058666
28 Nov, 2025 2 commits
- update bs390 of qwen3-30b-tp2(bw) · 4c92e64a
  zhuwenwen authored Nov 28, 2025
```
[Qwen3-30B] MoE configs for BW1000 TP=2 (bs390)
```
  4c92e64a
- change the scale of convert_fp8 to 1 · b0e99e82
  zhuwenwen authored Nov 28, 2025
  
  b0e99e82
27 Nov, 2025 13 commits
- update torch.float16 · 511eeccd
  zhuwenwen authored Nov 27, 2025
  
  511eeccd
- 解决宽松mtp request_ids报错找不到问题 · 6446f76c
  王敏 authored Nov 27, 2025
  
  6446f76c
- update topk_softmax · 0c060491
  zhuwenwen authored Nov 27, 2025
  
  0c060491
- set VLLM_USE_TOPK_RENORM=1 · 4c167900
  zhuwenwen authored Nov 27, 2025
  
  4c167900
- fix: 修改attetion这里的import bug · 8cfbe041
  wujl5 authored Nov 27, 2025
  
  8cfbe041
- update reshape_and_cache_cuda import · b550cf96
  zhuwenwen authored Nov 27, 2025
  
  b550cf96
- fix: 修复融合moe.quant对其他模型的多余传参影响 · 4f11b099
  wujl5 authored Nov 27, 2025
  
  4f11b099
- 去掉宽松mtp中的隐式同步 · 03e822d1
  王敏 authored Nov 27, 2025
  
  03e822d1
- add VLLM_USE_TOPK_RENORM to use optimized topk_softmax + renormalize · 7ff04b72
  zhuwenwen authored Nov 27, 2025
  
  7ff04b72
- 增加VLLM_USE_OPT_RESHAPE_AND_CACHE 环境变量用来控制优化reshape and cache kernel · fe054987
  laibao authored Nov 27, 2025
  
  fe054987
- [fix]解决宽松mtp1报错 · deae0a22
  王敏 authored Nov 27, 2025
  
  deae0a22
- update VLLM_USE_FUSE_SILU_AND_MUL init value · afd18c9c
  zhuwenwen authored Nov 27, 2025
  
  afd18c9c
- 增加 VLLM_USE_FUSE_SILU_AND_MUL 环境变量，用来控制 fused MoE 里 silu+mul 的融合 kernel。 · a921f34c
  laibao authored Nov 27, 2025
  
  a921f34c
26 Nov, 2025 6 commits
- 更新 vllm/_custom_ops.py, vllm/model_executor/layers/quantization/utils/w8a8_utils.py · 2164aab4
  wanglong3 authored Nov 26, 2025
  
  2164aab4
- [feat]宽松mtp默认关掉 · b9bc84e2
  王敏 authored Nov 26, 2025
  
  b9bc84e2
- [feat]支持宽松mtp · 02b4ec43
  王敏 authored Nov 26, 2025
  
  02b4ec43
- [pref]: DS_v2_w8a8模型融掉moe.quant · 68972532
  wujl5 authored Nov 26, 2025
  
  68972532
- [fix]修复不开mtp精度异常问题 · a363d2c3
  zhuwenwen authored Nov 26, 2025
  
  a363d2c3
- set VLLM_USE_ZERO_MTP=0 · 7ff48a6c
  zhuwenwen authored Nov 26, 2025
  
  7ff48a6c
23 Nov, 2025 2 commits
- update VLLM_USE_CAT_MLA · 83c1f04a
  zhuwenwen authored Nov 23, 2025
  
  83c1f04a
- add VLLM_USE_CAT_MLA to use fused cat and mla · 4b00d1ba
  zhuwenwen authored Nov 23, 2025
  
  4b00d1ba
21 Nov, 2025 6 commits
- deepseek_v2_w8a8 增加 silu_mul_quant融合 · 216e414b
  wujl5 authored Nov 21, 2025
  
  216e414b
- add VLLM_USE_CUDA_GRAPH_SIZES(1) to use 1-24... (not only 1 2 4 8 16) · 2b47bce9
  zhuwenwen authored Nov 21, 2025
```
set VLLM_USE_LIGHTOP_FILL_MOE_ALIGN=1, VLLM_USE_OPT_ZEROS=1 and VLLM_USE_PP_SYNC=1
```
  2b47bce9
- feat: pp mtp加入零消耗调度，加入环境变量VLLM_USE_ZERO_MTP，默认打开 · c50f084a
  jujl1 authored Nov 21, 2025
  
  c50f084a
- add VLLM_USE_CUDA_GRAPH_SIZES to use 1-18... (not only 1 2 4 8 16) · d126ce21
  zhuwenwen authored Nov 21, 2025
  
  d126ce21
- update Is_fuse_fill · 77fccdf4
  zhuwenwen authored Nov 21, 2025
  
  77fccdf4
- update fill + moe align interface of lightop · e3d1dfdd
  zhuwenwen authored Nov 21, 2025
  
  e3d1dfdd
20 Nov, 2025 6 commits
- feat: pipeline_parallel新增pp域请求数均衡,VLLM_USE_PP_BALANCE控制，默认开启 · 48cf947c
  jujl1 authored Oct 31, 2025
  
  48cf947c
- 根据不同场景，更新默认调度和分离调度的选择 · 244b534f
  zhuwenwen authored Nov 20, 2025
  
  244b534f
- use schedule_split_pd (dpsk mtp + fullgraph) · dbbc0b2e
  zhuwenwen authored Nov 20, 2025
  
  dbbc0b2e
- USE_FUSED_SILU_MUL_QUANT默认改成开启 · 421850cb
  wujl5 authored Nov 20, 2025
  
  421850cb
- DS_v2_w4a8_CRQ增加slilu_mul_quant支持 · 075841f3
  wujl5 authored Nov 20, 2025
  
  075841f3
- MLA模块CRQ融合分支增加rms_cuda_opt · 8c646ebe
  wujl5 authored Nov 20, 2025
  
  8c646ebe
19 Nov, 2025 1 commit
- 修复pd分离开cp引起的校检问题 · 22a14b84
  zhuwenwen authored Nov 19, 2025
```
修改schedule_default和schedule_split_pd的选择
```
  22a14b84