vllm 0.4.2

v0.4.2

on Python PyPI

Highlights

Features

• Chunked prefill is ready for testing! It improves inter-token latency in high load scenario by chunking the prompt processing and priortizes decode (#4580)
• Speculative decoding functionalities: logprobs (#4378), ngram (#4237)
• Support FlashInfer as attention backend (#4353)

Models and Enhancements

• Add support for Phi-3-mini (#4298, #4372, #4380)
• Add more histogram metrics (#2764, #4523)
• Full tensor parallelism for LoRA layers (#3524)
• Expanding Marlin kernel to support all GPTQ models (#3922, #4466, #4533)

Dependency Upgrade

• Upgrade to torch==2.3.0 (#4454)
• Upgrade to tensorizer==2.9.0 (#4467)
• Expansion of AMD test suite (#4267)

Progress and Dev Experience

• Centralize and document all environment variables (#4548, #4574)
• Progress towards fully typed codebase (#4337, #4427, #4555, #4450)
• Progress towards pipeline parallelism (#4512, #4444, #4566)
• Progress towards multiprocessing based executors (#4348, #4402, #4419)
• Progress towards FP8 support (#4343, #4332, 4527)

What's Changed

• [Core][Distributed] use existing torch.cuda.device context manager by @youkaichao in #4318
• [Misc] Update ShareGPT Dataset Sampling in Serving Benchmark by @ywang96 in #4279
• [Bugfix] Fix marlin kernel crash on H100 by @alexm-nm in #4218
• [Doc] Add note for docker user by @youkaichao in #4340
• [Misc] Use public API in benchmark_throughput by @zifeitong in #4300
• [Model] Adds Phi-3 support by @caiom in #4298
• [Core] Move ray_utils.py from engine to executor package by @njhill in #4347
• [Bugfix][Model] Refactor OLMo model to support new HF format in transformers 4.40.0 by @Isotr0py in #4324
• [CI/Build] Adding functionality to reset the node's GPUs before processing. by @Alexei-V-Ivanov-AMD in #4213
• [Doc] README Phi-3 name fix. by @caiom in #4372
• [Core]refactor aqlm quant ops by @jikunshang in #4351
• [Mypy] Typing lora folder by @rkooo567 in #4337
• [Misc] Optimize flash attention backend log by @esmeetu in #4368
• [Core] Add shutdown() method to ExecutorBase by @njhill in #4349
• [Core] Move function tracing setup to util function by @njhill in #4352
• [ROCm][Hardware][AMD][Doc] Documentation update for ROCm by @hongxiayang in #4376
• [Bugfix] Fix parameter name in get_tokenizer by @DarkLight1337 in #4107
• [Frontend] Add --log-level option to api server by @normster in #4377
• [CI] Disable non-lazy string operation on logging by @rkooo567 in #4326
• [Core] Refactoring sampler and support prompt logprob for chunked prefill by @rkooo567 in #4309
• [Misc][Refactor] Generalize linear_method to be quant_method by @comaniac in #4373
• [Misc] add RFC issue template by @youkaichao in #4401
• [Core] Introduce DistributedGPUExecutor abstract class by @njhill in #4348
• [Kernel] Optimize FP8 support for MoE kernel / Mixtral via static scales by @pcmoritz in #4343
• [Frontend][Bugfix] Disallow extra fields in OpenAI API by @DarkLight1337 in #4355
• [Misc] Fix logger format typo by @esmeetu in #4396
• [ROCm][Hardware][AMD] Enable group query attention for triton FA by @hongxiayang in #4406
• [Kernel] Full Tensor Parallelism for LoRA Layers by @FurtherAI in #3524
• [Model] Phi-3 4k sliding window temp. fix by @caiom in #4380
• [Bugfix][Core] Fix get decoding config from ray by @esmeetu in #4335
• [Bugfix] Abort requests when the connection to /v1/completions is interrupted by @chestnut-Q in #4363
• [BugFix] Fix min_tokens when eos_token_id is None by @njhill in #4389
• ✨ support local cache for models by @prashantgupta24 in #4374
• [BugFix] Fix return type of executor execute_model methods by @njhill in #4402
• [BugFix] Resolved Issues For LinearMethod --> QuantConfig by @robertgshaw2-neuralmagic in #4418
• [Misc] fix typo in llm_engine init logging by @DefTruth in #4428
• Add more Prometheus metrics by @ronensc in #2764
• [CI] clean docker cache for neuron by @simon-mo in #4441
• [mypy][5/N] Support all typing on model executor by @rkooo567 in #4427
• [Kernel] Marlin Expansion: Support AutoGPTQ Models with Marlin by @robertgshaw2-neuralmagic in #3922
• [CI] hotfix: soft fail neuron test by @simon-mo in #4458
• [Core][Distributed] use cpu group to broadcast metadata in cpu by @youkaichao in #4444
• [Misc] Upgrade to torch==2.3.0 by @mgoin in #4454
• [Bugfix][Kernel] Fix compute_type for MoE kernel by @WoosukKwon in #4463
• [Core]Refactor gptq_marlin ops by @jikunshang in #4466
• [BugFix] fix num_lookahead_slots missing in async executor by @leiwen83 in #4165
• [Doc] add visualization for multi-stage dockerfile by @prashantgupta24 in #4456
• [Kernel] Support Fp8 Checkpoints (Dynamic + Static) by @robertgshaw2-neuralmagic in #4332
• [Frontend] Support complex message content for chat completions endpoint by @fgreinacher in #3467
• [Frontend] [Core] Tensorizer: support dynamic num_readers, update version by @alpayariyak in #4467
• [Bugfix][Minor] Make ignore_eos effective by @bigPYJ1151 in #4468
• fix_tokenizer_snapshot_download_bug by @kingljl in #4493
• Unable to find Punica extension issue during source code installation by @kingljl in #4494
• [Core] Centralize GPU Worker construction by @njhill in #4419
• [Misc][Typo] type annotation fix by @HarryWu99 in #4495
• [Misc] fix typo in block manager by @Juelianqvq in #4453
• Allow user to define whitespace pattern for outlines by @robcaulk in #4305
• [Misc]Add customized information for models by @jeejeelee in #4132
• [Test] Add ignore_eos test by @rkooo567 in #4519
• [Bugfix] Fix the fp8 kv_cache check error that occurs when failing to obtain the CUDA version. by @AnyISalIn in #4173
• [Bugfix] Fix 307 Redirect for /metrics by @robertgshaw2-neuralmagic in #4523
• [Doc] update(example model): for OpenAI compatible serving by @fpaupier in #4503
• [Bugfix] Use random seed if seed is -1 by @sasha0552 in #4531
• [CI/Build][Bugfix] VLLM_USE_PRECOMPILED should skip compilation by @tjohnson31415 in #4534
• [Speculative decoding] Add ngram prompt lookup decoding by @leiwen83 in #4237
• [Core] Enable prefix caching with block manager v2 enabled by @leiwen83 in #4142
• [Core] Add multiproc_worker_utils for multiprocessing-based workers by @njhill in #4357
• [Kernel] Update fused_moe tuning script for FP8 by @pcmoritz in #4457
• [Bugfix] Add validation for seed by @sasha0552 in #4529
• [Bugfix][Core] Fix and refactor logging stats by @esmeetu in #4336
• [Core][Distributed] fix pynccl del error by @youkaichao in #4508
• [Misc] Remove Mixtral device="cuda" declarations by @pcmoritz in #4543
• [Misc] Fix expert_ids shape in MoE by @WoosukKwon in #4517
• [MISC] Rework logger to enable pythonic custom logging configuration to be provided by @tdg5 in #4273
• [Bug fix][Core] assert num_new_tokens == 1 fails when SamplingParams.n is not 1 and max_tokens is large & Add tests for preemption by @rkooo567 in #4451
• [CI]Add regression tests to ensure the async engine generates metrics by @ronensc in #4524
• [mypy][6/N] Fix all the core subdirectory typing by @rkooo567 in #4450
• [Core][Distributed] enable multiple tp group by @youkaichao in #4512
• [Kernel] Support running GPTQ 8-bit models in Marlin by @alexm-nm in #4533
• [mypy][7/N] Cover all directories by @rkooo567 in #4555
• [Misc] Exclude the tests directory from being packaged by @itechbear in #4552
• [BugFix] Include target-device specific requirements.txt in sdist by @markmc in #4559
• [Misc] centralize all usage of environment variables by @youkaichao in #4548
• [kernel] fix sliding window in prefix prefill Triton kernel by @mmoskal in #4405
• [CI/Build] AMD CI pipeline with extended set of tests. by @Alexei-V-Ivanov-AMD in #4267
• [Core] Ignore infeasible swap requests. by @rkooo567 in #4557
• [Core][Distributed] enable allreduce for multiple tp groups by @youkaichao in #4566
• [BugFix] Prevent the task of _force_log from being garbage collected by @Atry in #4567
• [Misc] remove chunk detected debug logs by @DefTruth in #4571
• [Doc] add env vars to the doc by @youkaichao in #4572
• [Core][Model runner refactoring 1/N] Refactor attn metadata term by @rkooo567 in #4518
• [Bugfix] Allow "None" or "" to be passed to CLI for string args that default to None by @mgoin in #4586
• Fix/async chat serving by @schoennenbeck in #2727
• [Kernel] Use flashinfer for decoding by @LiuXiaoxuanPKU in #4353
• [Speculative decoding] Support target-model logprobs by @cadedaniel in #4378
• [Misc] add installation time env vars by @youkaichao in #4574
• [Misc][Refactor] Introduce ExecuteModelData by @comaniac in #4540
• [Doc] Chunked Prefill Documentation by @rkooo567 in #4580
• [Kernel] Support MoE Fp8 Checkpoints for Mixtral (Static Weights with Dynamic/Static Activations) by @mgoin in #4527
• [CI] check size of the wheels by @simon-mo in #4319
• [Bugfix] Fix inappropriate content of model_name tag in Prometheus metrics by @DearPlanet in #3937
• bump version to v0.4.2 by @simon-mo in #4600
• [CI] Reduce wheel size by not shipping debug symbols by @simon-mo in #4602

New Contributors

• @zifeitong made their first contribution in #4300
• @caiom made their first contribution in #4298
• @Alexei-V-Ivanov-AMD made their first contribution in #4213
• @normster made their first contribution in #4377
• @FurtherAI made their first contribution in #3524
• @chestnut-Q made their first contribution in #4363
• @prashantgupta24 made their first contribution in #4374
• @fgreinacher made their first contribution in #3467
• @alpayariyak made their first contribution in #4467
• @HarryWu99 made their first contribution in #4495
• @Juelianqvq made their first contribution in #4453
• @robcaulk made their first contribution in #4305
• @AnyISalIn made their first contribution in #4173
• @sasha0552 made their first contribution in #4531
• @tdg5 made their first contribution in #4273
• @itechbear made their first contribution in #4552
• @markmc made their first contribution in #4559
• @Atry made their first contribution in #4567
• @schoennenbeck made their first contribution in #2727
• @DearPlanet made their first contribution in #3937

Full Changelog: v0.4.1...v0.4.2