Commits · 9d4330d2980ced3f321bc4960687318bf6ac9054 · OpenDAS / vllm_cscc · GitLab

27 Jan, 2026 4 commits
- refactor: 抽离 runner 侧 KV compression 逻辑并统一 slot mapping · 9d4330d2
  laibao authored Jan 27, 2026
  
  9d4330d2
- refactor: 抽离 flash_attn 的 KV compression 逻辑到 vllm/v1/kv_compression · 3adc766e
  laibao authored Jan 27, 2026
  
  3adc766e
- refactor: 将 kv_compression 的 Triton 内核迁移到 vllm/v1/kv_compression · 9db5ff3b
  laibao authored Jan 27, 2026
  
  9db5ff3b
- refactor: 抽取调度器 KV 长度账本逻辑；修复 num_kv_tokens 初始化 · 3bc7eb74
  laibao authored Jan 27, 2026
  
  3bc7eb74
23 Jan, 2026 2 commits
- fix: 修复 MTP runner 缺失 _extract_layer_index 导致的 KV 压缩崩溃 · 4634cbcf
  laibao authored Jan 23, 2026
  
  4634cbcf
- feat: kvpress flash_attn 实现非 chunked Top‑K compaction · 863f93e6
  laibao authored Jan 23, 2026
  
  863f93e6
22 Jan, 2026 5 commits
- feat: kvpress flash_attn（scheme 3）生成 prompt-end payload · a9ebf337
  laibao authored Jan 22, 2026
  
  a9ebf337
- feat: kvpress flash_attn 透传 KV 压缩元数据 · b6a27380
  laibao authored Jan 22, 2026
  
  b6a27380
- feat: kvpress runner 支持 chunked prefill prompt-end 一次性 KV compaction · 54b1bf44
  laibao authored Jan 22, 2026
  
  54b1bf44
- feat: kvpress runner 侧生成 Top-K 压缩元数据 · faf55520
  laibao authored Jan 22, 2026
  
  faf55520
- feat: kvpress runner 侧按 num_kv_tokens 计算 KV 写入位置 · 2df94aa9
  laibao authored Jan 22, 2026
  
  2df94aa9
21 Jan, 2026 5 commits
- feat: kvpress新增 runner 侧 KV 压缩状态/位置打通 · ad069e33
  laibao authored Jan 21, 2026
  
  ad069e33
- feat: kvpress新增调度层 KV 压缩逻辑 · 2fde0fa2
  laibao authored Jan 21, 2026
  
  2fde0fa2
- feat: kvpress新增 KV cache 申请/截断支持 · eef99f73
  laibao authored Jan 21, 2026
  
  eef99f73
- feat: kvpress新增 KV 压缩状态与元数据打通 · 8d3d07fc
  laibao authored Jan 21, 2026
  
  8d3d07fc
- feat: kvpress新增 SnapKV 打分与 KV compaction Triton 内核 · 5036e878
  laibao authored Jan 21, 2026
  
  5036e878
20 Jan, 2026 2 commits
- feat: kvpress新增 KV 压缩预算计算模块 · d3acd4a5
  laibao authored Jan 20, 2026
  
  d3acd4a5
- feat: kvpress新增 KV 压缩配置开关（默认关闭） · ade2749c
  laibao authored Jan 20, 2026
  
  ade2749c
06 Jan, 2026 1 commit
- fix: 修复deepseek在PP>1开chunkprefill精度问题 · 4b752578
  jujl1 authored Jan 06, 2026
  
  4b752578
05 Jan, 2026 3 commits
- feat: enable shared expert overlap. · ee19dca6
  wanglong3 authored Jan 03, 2026
  
  ee19dca6
- feat：为 GLM4 和 Llama 模型新增 MultiModalConfigProxy，以支持扁平配置与嵌套的多模态配置（text_config） · 952f0347
  laibao authored Jan 05, 2026
  
  952f0347
- 删除调试信息 · c773cc66
  yangql authored Jan 05, 2026
  
  c773cc66
04 Jan, 2026 5 commits
- 解决pd分离非对称切分通信组过多问题 · 294cc23a
  xiabo authored Jan 04, 2026
  
  294cc23a
- 区分pcie和hglink custom allreduce的使用 · 1b5aa25e
  xiabo authored Jan 04, 2026
  
  1b5aa25e
- fix: 只有当kv block中不含有MTP的假数据时才会被cached，以修复cache_full_blocks同一个kv block保存两次的bug · f7cb8c7b
  jujl1 authored Jan 04, 2026
  
  f7cb8c7b
- perf: 加速 v1 InputBatch.add_request 的 token_ids 拷贝 · fbe8587a
  laibao authored Jan 04, 2026
```
新增环境变量开关 VLLM_V1_FAST_TOKEN_ID_COPY（默认关闭）。开启后在 CachedRequestState 中缓存 prompt_token_ids 的 np.int32，并在 add_request 里用 np.copyto 写入 token_ids_cpu，避免长 prompt 场景反复 list->NumPy 转换开销（尤其是抢占/反复进出 batch 时）
```
  fbe8587a
- perf: DS-量化模型融合qa和kva的gemm · 577eb49f
  wujl5 authored Jan 04, 2026
  
  577eb49f
31 Dec, 2025 5 commits
- fix: chunk-prefill生成的token被错误返回 · 2c1de3fa
  jujl1 authored Dec 31, 2025
  
  2c1de3fa
- Revert "增加mtp>1的pad" · d8c36dbf
  yangql authored Dec 31, 2025
```
This reverts commit 059b1961.
```
  d8c36dbf
- [feat]解决deepep auto模式 mtp >1 时卡住 · 37d2b35d
  王敏 authored Dec 31, 2025
  
  37d2b35d
- 增加mtp>1的pad · 059b1961
  yangql authored Dec 31, 2025
  
  059b1961
- feat: 兼容MTP零消耗和主模型+MTP零消耗(VLLM_ZERO_OVERHEAD_ENHANCE=1)开启 · 76e22965
  jujl1 authored Dec 31, 2025
  
  76e22965
30 Dec, 2025 3 commits
- feat: 兼容MTP零消耗和主模型+MTP零消耗(VLLM_ZERO_OVERHEAD_ENHANCE=1)开启 · 8a413453
  jujl1 authored Dec 30, 2025
  
  8a413453
- 解决pd分离非对称切分通信组过多问题 · 84e5aba2
  xiabo authored Dec 30, 2025
  
  84e5aba2
- fix: 解决列表访问越界 · 5208b291
  jujl1 authored Dec 30, 2025
  
  5208b291
29 Dec, 2025 5 commits
- [feat]解决deepep auto模式卡住问题 · cb64c6bc
  王敏 authored Dec 29, 2025
  
  cb64c6bc
- fix: 解决同时处理prefill和decode时的prefill请求token计数错误 · bc945a5a
  jujl1 authored Dec 24, 2025
  
  bc945a5a
- fix: support chunk-prefill and fix bug in check_stop · 96197e48
  jujl1 authored Dec 23, 2025
  
  96197e48
- feat: triton kernel 实现 update_input · 89639c96
  jujl1 authored Dec 23, 2025
  
  89639c96
- feat: 主模型+mtp提前返回 · 0936ee97
  jujl1 authored Nov 28, 2025
  
  0936ee97