GitLab

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

moe: 补齐 fill+moe_align 融合开关语义

See merge request dcutoolkit/deeplearing/vllm!484

增加max_cudagraph_capture_size，细化capture的范围

See merge request dcutoolkit/deeplearing/vllm!494

perf: DS V2模型MOE部分增加rmsQuant

See merge request dcutoolkit/deeplearing/vllm!492

fix: fix bug http://hpczentao.sugon.com/bug-view-118388.html

See merge request dcutoolkit/deeplearing/vllm!491

mrope_1d修改

See merge request dcutoolkit/deeplearing/vllm!490

perf: DS V2模型MLA中增加rmsQuant

See merge request dcutoolkit/deeplearing/vllm!487

mla:  opt-cat 在 prefill 和 decode 的拼接路由

See merge request dcutoolkit/deeplearing/vllm!483

perf: DS v2增加DTBMM融合,默认关闭

See merge request dcutoolkit/deeplearing/vllm!480

dpsk_v32的mtp层的dense加载适配

See merge request dcutoolkit/deeplearing/vllm!479

feat:修复dsa的mqa接口兼容glm5

See merge request dcutoolkit/deeplearing/vllm!478

Fix：修复调用Triton MoE gemm时缺失的参数，对齐接口

See merge request dcutoolkit/deeplearing/vllm!476

# Conflicts:
#   vllm/model_executor/layers/fused_moe/fused_moe.py

feat: support shared expert fusion.

See merge request dcutoolkit/deeplearing/vllm!469

Fix：Extend MAX_VPT from 32 to 256 to accommodate large-scale MoE models (e.g., GLM-5-quantized model).

适配在bmz上的mla的kvcache_e5m2和e4m3量化的支持

See merge request dcutoolkit/deeplearing/vllm!472

clean code for Deepseek V2.

See merge request dcutoolkit/deeplearing/vllm!471

优化015 fp8 kvscale

See merge request dcutoolkit/deeplearing/vllm!473

perf:Deepseek v2模型增加rmsQuant和siluMulQuant融合

See merge request dcutoolkit/deeplearing/vllm!468

support qwen3-asr

See merge request dcutoolkit/deeplearing/vllm!466

[perf]glm4_moe模型适配rmsquant和silu_quant融合算子

See merge request dcutoolkit/deeplearing/vllm!467