GitLab

[fix]修复mtp eager模式下显存占用增加问题

See merge request dcutoolkit/deeplearing/vllm!180

[feat]优化mtp/eagle的计算逻辑，减少第1层并行解码的计算重复

See merge request dcutoolkit/deeplearing/vllm!178

add v1 engine zero overhead

See merge request dcutoolkit/deeplearing/vllm!177

已修改        vllm/utils/__init__.py

See merge request dcutoolkit/deeplearing/vllm!176

[fix]修复v1 mtp接受率低的问题

See merge request dcutoolkit/deeplearing/vllm!174

This reverts commit 20b6cf64.

V0.9.2 dev w8a8 new

See merge request dcutoolkit/deeplearing/vllm!173

[feat]支持mtp模型full_cuda_graph

See merge request dcutoolkit/deeplearing/vllm!172

已修改        vllm/utils/__init__.py

See merge request dcutoolkit/deeplearing/vllm!171

# Conflicts:
#	vllm/v1/spec_decode/eagle.py
#	vllm/v1/worker/gpu_model_runner.py