- 12 Mar, 2026 1 commit
-
-
wanghl6 authored
-
- 09 Mar, 2026 1 commit
-
-
wanghl6 authored
-
- 11 Feb, 2026 1 commit
-
-
zhuwenwen authored
-
- 09 Feb, 2026 1 commit
-
-
zhuwenwen authored
-
- 06 Feb, 2026 1 commit
-
-
wujl5 authored
-
- 05 Feb, 2026 1 commit
-
-
wanglong3 authored
feat: support moe sum when topk==9 bugfix: Fix mtp model load error when eable shared experts fusion.
-
- 28 Jan, 2026 1 commit
-
-
laibao authored
增加环境变量开关,禁用 Marlin W16A16 MoE 路径 强制 Triton 且权重已是 Marlin packed 时给出明确报错 Marlin 支持探测改为 best-effort(不再依赖 VLLM_USE_LIGHTOP)
-
- 23 Jan, 2026 1 commit
-
-
zhuwenwen authored
-
- 21 Jan, 2026 2 commits
- 17 Jan, 2026 2 commits
- 16 Jan, 2026 4 commits
-
-
zhuwenwen authored
add VLLM_USE_FUSED_CACHE_QUANT_BMM_MLA to use fused rmsnorm + contiguous + rope(for dpsk-v3) + concat_and_cache_mla + q quant, control bmm(todo) + cat +mla (fp8)
-
laibao authored
新增环境变量 VLLM_MOE_ROUTER_CAPTURE / DIR / RANK / MAX_LAYERS / NUM_TOKENS_* 用于开关与过滤控制 新增 router_capture.py,支持按 num_tokens 分桶抓取 router logits 并落盘 在 qwen3_moe 中接入抓取逻辑,默认关闭,仅在开启时记录 固定 skip_profile / skip_stack_funcs 为默认启用,避免抓到 warmup/profile 形状 统一配置入口到 vllm.envs,作为运行时基准
-
zhuwenwen authored
-
zhuwenwen authored
-
- 15 Jan, 2026 1 commit
-
-
wanglong3 authored
-
- 14 Jan, 2026 1 commit
-
-
zhuwenwen authored
-
- 05 Jan, 2026 1 commit
-
-
wanglong3 authored
-
- 04 Jan, 2026 4 commits
- 30 Dec, 2025 1 commit
-
-
jujl1 authored
-
- 26 Dec, 2025 1 commit
-
-
zhuwenwen authored
-
- 24 Dec, 2025 1 commit
-
-
laibao authored
新增环境变量 VLLM_V1_USE_REDUCED_TOPK_TOPP_SAMPLER 用于开关控制 扩展 SamplingMetadata,增加 max_top_k 与 has_any_no_top_k 在 InputBatch 侧计算 top-k 的主机端汇总信息,避免 device 同步 更新 Sampler/TopKTopPSampler 传递并使用新参数以启用优化采样
-
- 23 Dec, 2025 2 commits
- 21 Dec, 2025 1 commit
-
-
yangql authored
-
- 18 Dec, 2025 3 commits
- 16 Dec, 2025 1 commit
-
-
zhuwenwen authored
set VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT=1
-
- 15 Dec, 2025 1 commit
-
-
王敏 authored
-
- 11 Dec, 2025 1 commit
-
-
laibao authored
- 添加环境变量开关 VLLM_USE_MARLIN_W16A16_MOE(及 fused RMS+RoPE 标志位) - 增加 W16A16 Marlin 权重重排/缓存逻辑 - 集成 lightop 的 W16A16 MoE fast path,复用打包缓存避免重复 pack 并在 GPU 卸载原布局
-
- 05 Dec, 2025 1 commit
-
-
yangql authored
-
- 02 Dec, 2025 1 commit
-
-
zhuwenwen authored
-
- 01 Dec, 2025 2 commits
- 28 Nov, 2025 1 commit
-
-
jujl1 authored
-