v0.2.6

Major changes

• Fast model execution with CUDA/HIP graph
• W4A16 GPTQ support (thanks to @chu-tianxiang)
• Fix memory profiling with tensor parallelism
• Fix *.bin weight loading for Mixtral models

What's Changed

• Fix typing in generate function for AsyncLLMEngine & add toml to requirements-dev by @mezuzza in #2100
• Fix Dockerfile.rocm by @tjtanaa in #2101
• avoid multiple redefinition by @MitchellX in #1817
• Add a flag to include stop string in output text by @yunfeng-scale in #1976
• Add GPTQ support by @chu-tianxiang in #916
• [Docs] Add quantization support to docs by @WoosukKwon in #2135
• [ROCm] Temporarily remove GPTQ ROCm support by @WoosukKwon in #2138
• simplify loading weights logic by @esmeetu in #2133
• Optimize model execution with CUDA graph by @WoosukKwon in #1926
• [Minor] Delete Llama tokenizer warnings by @WoosukKwon in #2146
• Fix all-reduce memory usage by @WoosukKwon in #2151
• Pin PyTorch & xformers versions by @WoosukKwon in #2155
• Remove dependency on CuPy by @WoosukKwon in #2152
• [Docs] Add CUDA graph support to docs by @WoosukKwon in #2148
• Temporarily enforce eager mode for GPTQ models by @WoosukKwon in #2154
• [Minor] Add more detailed explanation on quantization argument by @WoosukKwon in #2145
• [Minor] Fix xformers version by @WoosukKwon in #2158
• [Minor] Add Phi 2 to supported models by @WoosukKwon in #2159
• Make sampler less blocking by @Yard1 in #1889
• [Minor] Fix a typo in .pt weight support by @WoosukKwon in #2160
• Disable CUDA graph for SqueezeLLM by @WoosukKwon in #2161
• Bump up to v0.2.6 by @WoosukKwon in #2157

New Contributors

• @mezuzza made their first contribution in #2100
• @MitchellX made their first contribution in #1817

Full Changelog: v0.2.5...v0.2.6