InternLM/lmdeploy v0.18.0 on GitHub

What's Changed

[Feat]: Support output input logprobs by @RunningLeon in #4793
Expand SM90 quantized GEMMs and unify TurboMind linear execution by @lzhangzz in #4943
[ascend] support glm52 by @wanfengcxz in #4928
Support dflash for qwen3.5 by @RunningLeon in #4789
Move shared expert into the MoE FFN layer; shard dense FFN node-locally in MoE models by @lzhangzz in #4973

fix: shard MLA attention under hybrid DP and TP by @CUHKSZzxy in #4908
refactor(serving): share chat serving runner by @lvhan028 in #4876
feat(pytorch): prototype TurboMind W4A16 backend for AWQ linear by @qescccczmr in #4897
refactor(pytorch): redesign CacheEngine around plans and allocations by @grimoire in #4862
fix: reject multimodal requests for text models by @CUHKSZzxy in #4926
refactor(pytorch): replace backend operator builders with typed build specs by @grimoire in #4890
Migrate TurboMind to C++20 by @lzhangzz in #4946
feat(pytorch): add piecewise CUDA graph prefill by @grimoire in #4895
feat: add XTuner TileLang sparse MLA backend by @CUHKSZzxy in #4904
feat: add request-only cache usage metric by @grimoire in #4940
refactor: stream tool parameters incrementally by @lvhan028 in #4802
feat: support checkpoint-engine weight updates by @lvhan028 in #4913
feat: add symmetric-memory LM-head all-gather by @qescccczmr in #4915
Refactor PyTorch scheduler ownership and API boundaries by @grimoire in #4921
feat(gpt-oss): combine xgrammar structural_tag grammar with Harmony parser by @windreamer in #4907
Decouple GEMM workspace from Linear handle by @lzhangzz in #4978
feat(kv_connector): mooncake store support mtp by @caikun-pjlab in #4948
fix(pytorch): avoid health RPC scheduling delay by @lvhan028 in #4966
refactor: remove legacy OpenAI API client by @lvhan028 in #4993

fix: fall back to model generation config by @CUHKSZzxy in #4929
fix: initialize MLA KV-B after online FP8 quantization by @CUHKSZzxy in #4925
fix tool_choice='required' by @lvhan028 in #4909
fix: exclude glm-5.2 mtp projection from online fp8 by @CUHKSZzxy in #4945
Fix TurboMind vision workspace overflows by @irexyc in #4924
fix(proxy): interpolate node_url in the terminate_node failure response by @Anai-Guo in #4941
fix(disagg): make zmq_disconnect synchronous so p2p_drop_connect actually closes the sockets by @Anai-Guo in #4916
Remove cuBLAS grouped GEMM and give cuBLAS dense priority over native BF16/FP16 kernels by @lzhangzz in #4975
fix(internvit): allocate pinned input buffers on demand by @irexyc in #4982
fix(serve): tolerate messages without a content key in multimodal preprocessing by @matrix72c in #4994
fix: enforce weights-only loading for MemDecode router checkpoints by @lvhan028 in #4988
fix(dlinfer): remove CUDA hardcode in NTK rotary embedding for non-CUDA accelerators by @li-lizhe in #4986
fix: build NCCL stub with public device API header by @lvhan028 in #4998
fix: correct tool choice handling for Intern-S and GPT-OSS by @lvhan028 in #4997
fix: restore getenv after environment parsing errors by @adenzhou1350 in #5000

[Docs] Fix A100 FP16 benchmark labels by @BingH225 in #4919
fix(deps): pin xgrammar below 0.2.5.post1 to unbreak unit-test CI by @windreamer in #4931
Assert the logprobs count in the restful return checks by @David-Wu1119 in #4934
docs: align docstring parameter names with signatures by @simpleqt in #4938
[Docs] Fix syntax in the Chinese output logits example by @BingH225 in #4959
docs: fix dead heading anchors in CONTRIBUTING and get_started pages by @simpleqt in #4937
improve(autotest): replace APIClient with OpenAI SDK and strict param checks by @littlegy in #4887
ci: add swap for CUDA release builds by @lvhan028 in #4974
[ci] refactor(autotest): consolidate layout-based pytest cases by @zhulinJulia24 in #4881
Test: update sleep/wakeup and abort scenarios by @littlegy in #4528
docs: sync ja README quickstart python version with en/zh (3.10 -> 3.12) by @simpleqt in #4939
bump version to v0.18.0 by @lvhan028 in #4990

Full Changelog: v0.17.0...v0.18.0