github InternLM/lmdeploy v0.16.0

2 hours ago

What's Changed

🚀 Features

💥 Improvements

  • refactor: report cache usage directly by @lvhan028 in #4798
  • SM90 native BF16/FP8 GEMM kernels, fused-SiLU quantization, and linear test harness by @lzhangzz in #4795
  • refactor: split api server endpoints by @lvhan028 in #4797
  • refactor(pytorch): derive CUDA step metadata from selected operators by @grimoire in #4805
  • optimize and modularize SSM prefix caching by @grimoire in #4788
  • perf(guided-decoding): optimize with async D2H copy and xgrammar v0.2.1 by @windreamer in #4605
  • refactor(serve): split chat_completions endpoint into a package by @lvhan028 in #4840
  • feat(chat-completions): add usage.completion_tokens_details by @lvhan028 in #4842
  • feat(pytorch): add optimized Gluon blocked FP8 GEMM for Hopper by @grimoire in #4830
  • perf(pytorch): add opt-in torch.compile for decode CUDA graphs by @grimoire in #4808
  • Ssm prefix cache non aligned by @grimoire in #4799
  • perf: optimize GLM-5.2 serving by @CUHKSZzxy in #4827
  • refactor: separate request preprocessing from generation by @lvhan028 in #4856

🐞 Bug fixes

  • [Bugfix] Fix PyTorch H2D input lifetime across CUDA streams by @grimoire in #4792
  • fix(serve): reject empty/falsy prompt input in format_prompts and AsyncEngine.generate by @SuperMarioYL in #4803
  • Fix ray mp duplicate output by @RunningLeon in #4833
  • fix(turbomind): dispatch cuMemcpyBatchAsync by CUDA runtime version by @lvhan028 in #4838
  • fix(disagg): use JSON instead of pickle for P2P ZMQ requests (#4804) by @Anai-Guo in #4812
  • fix(serve): emit signatures for Anthropic thinking blocks by @matrix72c in #4851
  • Fix int4 KV cache quantization range when the packed head width is not a power of two by @truong-v in #4850
  • fix: harden serving request validation by @lvhan028 in #4872
  • fix: fix allgather/allgather2d for cuda-ipc when byte_width is not multiple of uint by @irexyc in #4873

📚 Documentations

🌐 Other

New Contributors

Full Changelog: v0.15.0...v0.16.0

Don't miss a new lmdeploy release

NewReleases is sending notifications on new releases.