GitLab

set VLLM_USE_FUSED_QA_KVA_GEMM=1
feat:w4a8Linear调用apply_int8_linear，以支持blaslt

fuse_moe_fp8接入marlin算子
fix(v1)：修复抢占恢复时 BlockTable 溢出
feat(moe)：新增 VLLM_USE_MOE_W16A16_TRTION 强制 Triton MoE
fix: 解决原版0消耗chunk-prefill崩溃问题
fp8增加fused_moe_gate参数

[opt] 优化epsp代码, 零消耗添加epsp
update VLLM_USE_FUSED_RMS_ROPE=0 (default). for qwen3, VLLM_USE_FUSED_RMS_ROPE=1 (default)
feat(moe/marlin): Marlin W16A16 MoE 自动探测并预打包（去掉手动开关）
perf(qwen3): 融合 q/k RMSNorm + RoPE
fused_moe_fp8接入lmslim

[qwen3-480b] MoE(TN) configs for nmz TP=8
[opt] 优化deepep相关代码
[fix] 修复deepseek moe模型的awq量化推理bug和精度问题, 修复awq模型的VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD设置位置, update_state,优化性能，去除冗余操作
pcie 解决custom cudagraph模式需要拷贝的问题，需要配合dtk进行使用
[feat] Switch default w8a8 gemm impl to blaslt. Support w8a8-fp8 GEMM backend.MoE 路由抓取：新增 router_capture 工具链与 envs 统一配置
[envs] set VLLM_CUSTOM_CACHE=1、VLLM_USE_FUSED_RMS_ROPE=1、VLLM_USE_FUSED_FILL_RMS_CAT=1、VLLM_USE_FLASH_ATTN_FP8=1、VLLM_USE_FLASH_MLA_FP8=1、update VLLM_USE_TOPK_RENORM

Merge branch 'v0.9.2-dev-for-yuanbao' of http://10.16.6.30/dcutoolkit/deeplearing/vllm into v0.9.2-dev-for-yuanbao

[pd] 修改非堆成切分的判断, pd分离使用默认调度，set VLLM_USE_PD_SPLIT=1
[perf](fused-moe): 预打包 Marlin W16A16 MoE 权重，降低 warmup 显存峰值
[fix]修复丢弃MTP代码报错，pp+chunkprefill多并发input ids更新bug, 修复不开启融合图的断言错误, PP 场景 decode 阶段 token 被误丢弃导致卡住

feat: support moe sum when topk==9

bugfix: Fix mtp model load error when eable shared experts fusion.

feat: Support enable rms quant and shared expert overlap at same time.

See merge request dcutoolkit/deeplearing/vllm!352

fix：更新 Marlin W16A16 MoE 中支持的断言消息

See merge request dcutoolkit/deeplearing/vllm!349

解决custom allreduce在dp情况下的其服务错误问题

See merge request dcutoolkit/deeplearing/vllm!348

DS量化模型重构atten和moe调用rmsquant融合逻辑。

See merge request dcutoolkit/deeplearing/vllm!345

bugfix: Fix the startup crash issue of the service when USE_FUSED_RMS_QUANT=1

See merge request dcutoolkit/deeplearing/vllm!344

fix: 修复deepseek在PP>1开chunkprefill精度问题

See merge request dcutoolkit/deeplearing/vllm!342

feat: enable shared expert overlap.

See merge request dcutoolkit/deeplearing/vllm!339

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）

See merge request dcutoolkit/deeplearing/vllm!336

合并deepep的auto模式分支"

See merge request dcutoolkit/deeplearing/vllm!335

V0.9.2 dev pdncclinit

See merge request dcutoolkit/deeplearing/vllm!333

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

See merge request dcutoolkit/deeplearing/vllm!331

perf: 加速 v1 InputBatch.add_request 的 token_ids 拷贝

See merge request dcutoolkit/deeplearing/vllm!330

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）

perf: DS-量化模型融合qa和kva的gemm

See merge request dcutoolkit/deeplearing/vllm!329