v0.2.8.post1

Minor toml file patch

v0.2.8

Bug fixing:
-Fix static cache init with new transformers version
-Add mxfp vllm patching utils
-Improve cuda graphs/compile settings for transformer models

v0.2.7.post1

Bug fixing:
-HIP graph fix in generation: bc8f4c7
-Fix HQQLinear with None linear inputs: 3b86ac9

v0.2.7

• Fix nan bug when max - min is very small: 373cbea
• Add DISABLE_CUDA=1 env variable to disable building cuda kernels for then aten backend. This allows faster pip build. 861f690
• Improve memory usage a566c78
• Fix vLLM torch fallback logic: d3f14b4

v0.2.6

• Fix cuda build
• torchcompile() support for hqq_aten
• bfloat16 support for vllm/hqq
• Update vllm utils to support hqq_gemlite and hqq_torch aliases
• FIx vLLM v1 issues
• Extend save_to_safetensors to VLMs

Full Changelog: v0.2.5...0.2.6

v0.2.5

-Fix .name in backends
-Skip gemlite invalid in/out feature sizes in VLLM patching
-Faster VLLM packing via GemLite

v.0.2.3.post1

Bug fixes:

• Check W_q in state dict to fix peft issue #151
• Fix bugs related to AutoHQQHFModel.save_to_safetensors

v0.2.3

• VLLM support via patching - GemLite backend + on-the-fly quantization
• Add support for Aria
• Add support to load quantized SequenceClassification
• Faster decoding via (custom cudagraphs, sdpa math backend, etc.)
• Fix bugs related torch compile and hf_generator related to the newer transformers versions
• Fix bugs related to saving quantized models with no grouping
• Fix bugs related to saving large quantized models
• Update examples
• Add support for HQQLinear .to(device)

v0.2.2

HQQ v0.2.2

• Support static cache compilation without using HFGenerator
• Fixing various issues related to torch.compile

v.0.2.1

HQQ v0.2.1

• HQQLinear.state_dict() for non-initialized layers. Mainly used in for huggingface/transformers#33141