Commits · 2799735a8fbc4bbf6c08eb87de5695f03d2cf9a9 · OpenDAS / vllm_cscc

13 Feb, 2026 1 commit
- [feat]支持glm4_moe_mtp使用torch compile，实现mtp cudagraph模式 · dc027d74
  王敏 authored Feb 13, 2026
  
  dc027d74
11 Feb, 2026 3 commits

remove unused impl and envs · b5ea93d4
zhuwenwen authored Feb 11, 2026

b5ea93d4

feat(moe): 补齐 v0.15 中 Marlin W16A16 MoE 端到端接入 · 1e2fe58f

laibao authored Feb 11, 2026

参考并移植 011/vllm 的关键提交逻辑
新增 VLLM_USE_MOE_W16A16_TRITON 开关，并接入基于 lightop 的运行时能力探测与启用结果缓存。
在权重加载后对 w13 与 w2 执行 W16A16 Marlin 预打包。
W16A16 Marlin 启用时保留 monolithic 执行路径，并在 fused_experts_impl 中增加 packed 权重 fast-path。
保持 Marlin 或 lightop 不可用时的回退行为不变。

1e2fe58f

接入channel、block triton 及channel-wise marlin · 7771c0ac
SAC_fanth authored Feb 11, 2026

7771c0ac

10 Feb, 2026 4 commits
- fix qwen3-next nn layout · 8cdc3a30
  zhuwenwen authored Feb 10, 2026
  
  8cdc3a30
- feat：适配Blaslt Channelwise gemm · 0636f239
  lixh6 authored Feb 10, 2026
  
  0636f239
- feat: implement FP8 blockwise GEMM with hipblaslt · 8510c10c
  lixh authored Feb 09, 2026
  
  8510c10c
- fix qwen3-next run error · 440222e9
  zhuwenwen authored Feb 10, 2026
  
  440222e9
09 Feb, 2026 1 commit
- pd separation uses default scheduling and set VLLM_USE_PD_SPLIT=1 · a4bcf959
  zhuwenwen authored Feb 09, 2026
  
  a4bcf959
08 Feb, 2026 2 commits
- 去掉调试代码 · d08e3d52
  王敏 authored Feb 08, 2026
  
  d08e3d52
- [feat]适配w8a8 deepep,接入lightop版deepgemm · 289f98c6
  王敏 authored Feb 08, 2026
  
  289f98c6
06 Feb, 2026 9 commits
- perf(qwen3): 融合 q/k RMSNorm + RoPE · e807ec39
  zhuwenwen authored Feb 06, 2026
```
set fp8_e4m3 only supported on nmz and support q&kvcache fp8
set VLLM_PCIE_USE_CUSTOM_ALLREDUCE=1
```
  e807ec39
- 删除重复代码 · aec90b84
  王敏 authored Feb 06, 2026
  
  aec90b84
- fix nn_moe run error · 4a4fb3de
  zhuwenwen authored Feb 06, 2026
  
  4a4fb3de
- [fix]修复nn_moe启动报错 · 74306deb
  王敏 authored Feb 06, 2026
  
  74306deb
- [feat]W8A8适配deepseek以及mtp · 530e785f
  zhuwenwen authored Feb 06, 2026
  
  530e785f
- fix indentation of self.quant_method · 3c35c466
  zhuwenwen authored Feb 06, 2026
  
  3c35c466
- [feat]W8A8适配deepseek以及mtp · ee58c1bf
  王敏 authored Feb 06, 2026
  
  ee58c1bf
- sync v0.15.1 (models) · df704163
  zhuwenwen authored Feb 06, 2026
  
  df704163
- 修复awq/w4a16的triton支持，以及fuse_moe的接口对齐，以及awq_moe_marlin推理的相关bug，并解决awq/w4a16的精度问题 · d7db129a
  zhuwenwen authored Feb 06, 2026
  
  d7db129a
05 Feb, 2026 2 commits
- sync v0.15.1 (fused_moe) · bc387d5a
  zhuwenwen authored Feb 05, 2026
  
  bc387d5a
- sync v0.15.1(ex fused_moe&models) · 899a2db4
  zhuwenwen authored Feb 05, 2026
  
  899a2db4
04 Feb, 2026 11 commits
- fix load error · 99fc9fc3
  zhuwenwen authored Feb 04, 2026
  
  99fc9fc3
- [perf] use optimized topk_softmax + renormalize (lightop) · e9e95d0f
  zhuwenwen authored Feb 04, 2026
  
  e9e95d0f
- [perf] update op.moe_fused_gate · 06e16a27
  zhuwenwen authored Feb 04, 2026
  
  06e16a27
- update VLLM_USE_OPT_RESHAPE_AND_CACHE to support bf16 and qwen3-dense · 263f45a4
  zhuwenwen authored Feb 04, 2026
  
  263f45a4
- [perf] add VLLM_USE_FLASH_ATTN_FP8 to use fa fp8 attention · ac28ab22
  zhuwenwen authored Feb 04, 2026
  
  ac28ab22
- [perf] add VLLM_USE_FUSED_FILL_RMS_CAT to use lightop for dpsk mtp fill + rms*2 + cat · 5fe03549
  zhuwenwen authored Feb 04, 2026
  
  5fe03549
- skip SPLIT_K · b8c7ba0a
  zhuwenwen authored Feb 04, 2026
  
  b8c7ba0a
- remove remove VLLM_USE_OPT_MOE_SUM · 2703e2e9
  zhuwenwen authored Feb 04, 2026
  
  2703e2e9
- skip AiterInt8ScaledMMLinearKernel · 4599e05f
  zhuwenwen authored Feb 04, 2026
  
  4599e05f
- cherry pick · 7d98f09b
  Michael Goin authored Feb 03, 2026
```
Signed-off-by: Robert Shaw <rshaw@neuralmagic.com>
```
  7d98f09b
- [Bugfix] Disable RoutingMethodType.[Renormalize,RenormalizeNaive] TRTLLM... · daa2784b
  Michael Goin authored Feb 03, 2026
```
[Bugfix] Disable RoutingMethodType.[Renormalize,RenormalizeNaive] TRTLLM per-tensor FP8 MoE (#33620)
Signed-off-by: mgoin <mgoin64@gmail.com>
(cherry picked from commit e346e2d0

)
Signed-off-by: Robert Shaw <rshaw@neuralmagic.com>
```
  daa2784b
03 Feb, 2026 6 commits
- skip aiter · e5f2ff72
  zhuwenwen authored Feb 03, 2026
  
  e5f2ff72
- add moe_fused_gate · 0386844b
  zhuwenwen authored Feb 03, 2026
  
  0386844b
- [torch.compile] Speed up MOE handling in forward_context (#33184) · 611b1875
  Richard Zou authored Jan 27, 2026
```
Signed-off-by: Richard Zou <zou3519@gmail.com>
(cherry picked from commit d9aa39a3)
```
  611b1875
- [Misc][Build] Lazy load cv2 in nemotron_parse.py (#33189) · eec3546b
  Kiersten Stokes authored Jan 29, 2026
```
Signed-off-by: kiersten-stokes <kierstenstokes@gmail.com>
(cherry picked from commit 9e138cb0)
```
  eec3546b
- fix run error · b31c7251
  zhuwenwen authored Feb 03, 2026
  
  b31c7251
- [Release] Fix format and cherry-pick (#33618) · 31a64c63
  Zhewen Li authored Feb 02, 2026
```
Signed-off-by: zhewenli <zhewen@inferact.ai>
Co-authored-by: zhewenli <zhewen@inferact.ai>
```
  31a64c63
02 Feb, 2026 1 commit
- [Release] patch step3p5 attention class in v0.15.1 release (#33602) · 57eae2f8
  Zhewen Li authored Feb 02, 2026
```
Signed-off-by: zhewenli <zhewen@inferact.ai>
Co-authored-by: zhewenli <zhewen@inferact.ai>
```
  57eae2f8