Commits · 18b4e6f3597313c7704173de5be5c6bc3cd5c5ef · OpenDAS / vllm_cscc

07 Dec, 2025 3 commits
- set USE_FUSED_RMS_QUANT=1, USE_FUSED_SILU_MUL_QUANT=1 and ... · 18b4e6f3
  zhuwenwen authored Dec 07, 2025
```
set USE_FUSED_RMS_QUANT=1, USE_FUSED_SILU_MUL_QUANT=1 and  VLLM_SCHED_ENABLE_MINIMAL_INJECTION=1 for dpsk-v3
```
  18b4e6f3
- sync flashmla get_mla_metadata · 62420b27
  zhuwenwen authored Dec 07, 2025
  
  62420b27
- add qwen3-480b tp32 tn moe json · 93f78899
  zhuwenwen authored Dec 07, 2025
  
  93f78899
05 Dec, 2025 1 commit
- update fused_rms_norm_rope_contiguous params · 0ac26aec
  zhuwenwen authored Dec 05, 2025
```
set VLLM_REJECT_SAMPLE_OPT=1 for dpsk-v3
```
  0ac26aec
04 Dec, 2025 5 commits
- skip cupy · e9532d9e
  zhuwenwen authored Dec 04, 2025
  
  e9532d9e
- Merge branch 'v0.9.2-dev-wm-1204' into 'v0.9.2-dev' · f7f0a9e0
  zhuwenwen authored Dec 04, 2025
```
[fix]解决mtp保存draft prob在例如pd分离场景下的OOM问题

See merge request dcutoolkit/deeplearing/vllm!284
```
  f7f0a9e0
- [fix]解决mtp保存draft prob在例如pd分离场景下的OOM问题 · 30b93f8e
  王敏 authored Dec 04, 2025
  
  30b93f8e
- add VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT when use USE_FUSED_RMS_QUANT and... · de87d606
  zhuwenwen authored Dec 04, 2025
```
add VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT when use USE_FUSED_RMS_QUANT and USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT
```
  de87d606
- fix SUPPORT_TC · 8118d493
  zhuwenwen authored Dec 04, 2025
  
  8118d493
03 Dec, 2025 1 commit
- set VLLM_USE_OPT_RESHAPE_AND_CACHE=1 for qwen3-30b · 8364249c
  zhuwenwen authored Dec 03, 2025
  
  8364249c
02 Dec, 2025 9 commits
- Merge branch 'v0.9.2-dev-wm-1202' into 'v0.9.2-dev' · d533c559
  zhuwenwen authored Dec 02, 2025
```
回退错误修改，避免dp模式漏掉的dispatch和combine通信

See merge request dcutoolkit/deeplearing/vllm!282
```
  d533c559
- Merge remote-tracking branch 'origin/v0.9.2-dev' into v0.9.2-dev · 14688ccd
  王敏 authored Dec 02, 2025
  
  14688ccd
- [fix]回退错误修改，避免dp模式漏掉的dispatch和combine通信 · 55310f4f
  王敏 authored Dec 02, 2025
  
  55310f4f
- update weight and cache · fd559b9f
  zhuwenwen authored Dec 02, 2025
  
  fd559b9f
- add VLLM_USE_V32_ENCODE to use encoding_dsv32.py · 2ae2d013
  zhuwenwen authored Dec 02, 2025
  
  2ae2d013
- Merge branch 'v0.9.2-dev-wm-1202' into 'v0.9.2-dev' · 272a4a7b
  zhuwenwen authored Dec 02, 2025
```
[fix]解决dp模式mtp卡住问题

See merge request dcutoolkit/deeplearing/vllm!280
```
  272a4a7b
- [fix]解决dp模式mtp卡住问题 · 55ee9d72
  王敏 authored Dec 02, 2025
  
  55ee9d72
- update opt2 · 73cbc9fe
  zhuwenwen authored Dec 02, 2025
  
  73cbc9fe
- add gfx938 · 4b4512fa
  zhuwenwen authored Dec 02, 2025
  
  4b4512fa
01 Dec, 2025 5 commits
- update VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT · d4d251fe
  zhuwenwen authored Dec 01, 2025
  
  d4d251fe
- Merge branch 'v0.9.2-dev' of http://10.16.6.30/dcutoolkit/deeplearing/vllm into v0.9.2-dev · 903a588f
  zhuwenwen authored Dec 01, 2025
  
  903a588f
- add VLLM_USE_OPT_RESHAPE_AND_CACHE (test) · 64e307c7
  zhuwenwen authored Dec 01, 2025
  
  64e307c7
- 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 14 commits
- update torch.float16 · 511eeccd
  zhuwenwen authored Nov 27, 2025
  
  511eeccd
- Merge branch 'v0.9.2-dev-wm-1127' into 'v0.9.2-dev' · 42a95309
  zhuwenwen authored Nov 27, 2025
```
解决宽松mtp request_ids报错找不到问题

See merge request dcutoolkit/deeplearing/vllm!276
```
  42a95309
- 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
- Merge branch 'v0.9.2-dev' of http://10.16.6.30/dcutoolkit/deeplearing/vllm into v0.9.2-dev · 03cdc9be
  zhuwenwen authored Nov 27, 2025
  
  03cdc9be
- set VLLM_USE_TOPK_RENORM=1 · 4c167900
  zhuwenwen authored Nov 27, 2025
  
  4c167900
- Merge branch 'v0.9.2-dev-fix_moe_quant' into 'v0.9.2-dev' · a6bed85b
  zhuwenwen authored Nov 27, 2025
```
fix: moe quant bug and attention bug

See merge request dcutoolkit/deeplearing/vllm!275
```
  a6bed85b
- fix: 修改attetion这里的import bug · 8cfbe041
  wujl5 authored Nov 27, 2025
  
  8cfbe041
- Merge branch 'v0.9.2-dev' of http://10.16.6.30/dcutoolkit/deeplearing/vllm into v0.9.2-dev · 2fc5b0bb
  zhuwenwen authored Nov 27, 2025
  
  2fc5b0bb
- 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
- Merge branch 'v0.9.2-dev-1127' into 'v0.9.2-dev' · 48742057
  zhuwenwen authored Nov 27, 2025
```
去掉宽松mtp中的隐式同步

See merge request dcutoolkit/deeplearing/vllm!274
```
  48742057
- 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