Commits · 3bff7958eb6d717cfbdcf69d01317800083e729a · OpenDAS / vllm_cscc

18 Mar, 2026 1 commit
- x接入mla_cat算子仅在nmz和kvcache-fp8情况下生效，默认关闭，开启需要export VLLM_USE_CAT_MLA=1 · 3bff7958
  yangql authored Mar 18, 2026
  
  3bff7958
17 Mar, 2026 1 commit
- invoke flash_attn in the Qwen2AudioEncoder (transformers) · 3a45ab97
  caihl authored Mar 17, 2026
  
  3a45ab97
16 Mar, 2026 3 commits
- fix: resolve block_shape conflicts between DeepEP MoE and non-DeepEP quantization · f9a04c97
  chenhw5 authored Mar 16, 2026
  
  f9a04c97
- perf: GLM4.7增加MOE调用rmsQuant, fix: 修掉fused_moe向后传递None导致的报错 · 0f6b9a19
  wujl5 authored Mar 16, 2026
  
  0f6b9a19
- [feat]deepseek mtp支持pp模式 · 2ce72b9c
  王敏 authored Mar 16, 2026
  
  2ce72b9c
15 Mar, 2026 1 commit

fanwl authored Mar 15, 2026

- Add VLLM_V1_USE_FA_UNIFIED_ATTN_2D 环境变量
- 0: Triton attention, 1: FA unified attention

eb35ba1b

13 Mar, 2026 3 commits
- rms_norm_opt精度问题解决 · 9404668a
  guanyu1 authored Mar 13, 2026
  
  9404668a
- fix: 修复MOE量化tensor对于其他模型的影响 · 8e726b3f
  wujl5 authored Mar 13, 2026
  
  8e726b3f
- 修改sparse_attn hip后端 · 7cec75a7
  liuchy5 authored Mar 13, 2026
  
  7cec75a7
12 Mar, 2026 8 commits
- Fix：GLM-5量化模型mla_attention layout修复&&sparse_attn fp8支持 · 5b9ad722
  lixh6 authored Mar 12, 2026
  
  5b9ad722
- feat(deepseek-mla):: 精简 fused RMS-RoPE concat 可用性判断 · 9aabf7e7
  laibao authored Mar 12, 2026
  
  9aabf7e7
- feat(deepseek-mla): 接入 VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT 融合链路 · cae53e46
  laibao authored Mar 10, 2026
```
新增环境变量与 MLA 融合接线（wrapper -> attention -> impl） 接入 lightop fused_rms_norm_rope_contiguous，保留回退路径
```
  cae53e46
- moe: 补齐 fill+moe_align 融合开关语义 · 706c031c
  laibao authored Mar 09, 2026
  
  706c031c
- perf: DS V2模型MOE部分增加rmsQuant · 168ceef7
  wujl5 authored Mar 12, 2026
  
  168ceef7
- fix: fix bug http://hpczentao.sugon.com/bug-view-118388.html · 8bf99b0b
  wanglong3 authored Mar 12, 2026
  
  8bf99b0b
- perf: DS V2模型MLA中增加rmsQuant · 2350c778
  wujl5 authored Mar 12, 2026
  
  2350c778
- perf: DS v2增加DTBMM融合,默认关闭 · 6ca1362b
  wujl5 authored Mar 12, 2026
  
  6ca1362b
11 Mar, 2026 3 commits
- dpsk_v32的mtp层的dense加载适配 · 96abd293
  yangql authored Mar 11, 2026
  
  96abd293
- feat:修复dsa的mqa接口兼容glm5 · 66979358
  liuchy5 authored Mar 11, 2026
  
  66979358
- mla: 恢复 opt-cat 在 prefill 和 decode 的拼接路由 · ee355d86
  laibao authored Mar 07, 2026
  
  ee355d86
10 Mar, 2026 2 commits
- clean code for Deepseek V2. · 38b14e8b
  wujl5 authored Mar 10, 2026
  
  38b14e8b
- Fix：修复调用Triton MoE gemm时缺失的参数，对齐接口 · 4ad7a1fe
  lixh6 authored Mar 10, 2026
  
  4ad7a1fe
09 Mar, 2026 1 commit
- 适配在bmz上的mla的kvcache_e5m2和e4m3量化的支持 · 0daa00fb
  yangql authored Mar 09, 2026
  
  0daa00fb
07 Mar, 2026 1 commit
- feat: support shared expert fusion. · 83c871fb
  wanglong3 authored Mar 07, 2026
  
  83c871fb
06 Mar, 2026 3 commits
- perf:Deepseek v2模型增加rmsQuant和siluMulQuant融合 · 58a36508
  wujl5 authored Mar 06, 2026
  
  58a36508
- [perf]glm4_moe模型适配rmsquant和silu_quant融合算子 · 110bbdd5
  王敏 authored Mar 06, 2026
  
  110bbdd5
- [perf]添加Module支持split qkv+rmsnorm+rope+kvcache融合算子，GLM4_MOE完成适配 · f1a7696f
  王敏 authored Mar 06, 2026
  
  f1a7696f
05 Mar, 2026 3 commits
- 修复dsa的workspace的bug，以及添加环境变量关闭DSAVLLM_DISABLE_DSA=1 · cb1a27d2
  yangql authored Mar 05, 2026
  
  cb1a27d2
- add qwen3-asr · c50e6dee
  weishb authored Mar 05, 2026
  
  c50e6dee
- 修复channel-int8 的config读取bug · cfd6a543
  SAC_fanth authored Mar 05, 2026
  
  cfd6a543
04 Mar, 2026 1 commit
- feat:fix dsa · 6e7c8326
  liuchy5 authored Mar 04, 2026
  
  6e7c8326
03 Mar, 2026 1 commit
- support dsa · d71496bf
  zhuwenwen authored Mar 03, 2026
  
  d71496bf
02 Mar, 2026 2 commits
- fix(qwen3): 修复VL场景fused RoPE分支条件，并更正qwen3_moe中rms_mrope参数顺序 · 02b0d6ba
  laibao authored Mar 02, 2026
  
  02b0d6ba
- [Model] GLM adaptation · afe3ea1e
  zhuwenwen authored Mar 02, 2026
  
  afe3ea1e
28 Feb, 2026 1 commit
- 修复qwen3.5的使用dtype为fp16的picecwise的推理模式 · b5e8d01e
  yangql1 authored Feb 28, 2026
  
  b5e8d01e
25 Feb, 2026 1 commit
- 修复block-wise 用triton报错的bug · 359c16db
  SAC_fanth authored Feb 25, 2026
  
  359c16db
24 Feb, 2026 2 commits

修复channel-wise 用triton报错的bug · 9d44744c
SAC_fanth authored Feb 24, 2026

9d44744c

feat(moe): 增加 Qwen3 路由 logits 的环境变量可控采集能力 · 9ae319b3

laibao authored Feb 21, 2026

新增 router_capture 工具，用于按 num_tokens/rank 过滤并落盘 MoE router logits
在 Qwen3MoeSparseMoeBlock 中接入采集调用，并在 torch.compile 场景下自动跳过
补充 VLLM_MOE_ROUTER_CAPTURE* 环境变量

9ae319b3

19 Feb, 2026 1 commit
- 支持非deepseek模型使用moe_fused_gate · 93548076
  王敏 authored Feb 19, 2026
  
  93548076
16 Feb, 2026 1 commit
- add qwen3_5 · 4dc838d3
  Rayyyyy authored Feb 16, 2026
  
  4dc838d3