v0.2.7

What's Changed

• fix: pass rope_theta argument when initializing LlamaLikeBlock for models like qwen2, mistral, etc. by @Shuai-Xie in #568
• Add Gemma2 support. by @radi-cho in #562
• ignore onnx in ignore_patterns by @casper-hansen in #570
• Add Internlm2 support by @Crystalcareai in #576
• quantization fails with old datasets by @stas00 in #593
• doc: replace a broken example with a working one by @stas00 in #595
• Implement NO_KERNELS flag and update torch requirement by @devin-ai-integration in #582
• AWQ Triton kernels. Make autoawq-kernels optional. by @casper-hansen in #608
• device_map defaults to auto by @casper-hansen in #607
• Let installed PyTorch decide required version number by @wasertech in #573
• Replace itrex qbits to ipex woq linear by @jiqing-feng in #549
• enable awq ipex linear in transformers by @jiqing-feng in #610
• fix for "two devices" issue due to RoPE changes by @davedgd in #630
• add qwen2vl support by @kq-chen in #599
• Add support for Phi-3-vision series model by @Isotr0py in #596
• support minicpm3.0 by @LDLINGLINGLING in #605
• Enable Intel GPU path and lora finetune and change examples to support different devices by @jiqing-feng in #631
• Replace custom sharding with save_torch_state_dict from huggingface_hub by @casper-hansen in #644
• New release (0.2.7) + Fix build by @casper-hansen in #647
• Only build once by @casper-hansen in #648

New Contributors

• @Shuai-Xie made their first contribution in #568
• @radi-cho made their first contribution in #562
• @Crystalcareai made their first contribution in #576
• @stas00 made their first contribution in #593
• @wasertech made their first contribution in #573
• @jiqing-feng made their first contribution in #549
• @davedgd made their first contribution in #630
• @kq-chen made their first contribution in #599

Full Changelog: v0.2.6...v0.2.7