GPTQModel v1.0.6

What's Changed

Patch release to fix loading of quantized Llama 3.2 Vision model.

• [FIX] mllama loader by @LRL-ModelCloud in #404

Full Changelog: v1.0.5...v1.0.6

GPTQModel v1.0.5

What's Changed

Added partial quantization support Llama 3.2 Vision model. v1.0.5 allows quantization of text-layers (layers responsible for text-generation) only. We will add vision layer support shortly. A Llama 3.2 11B Vision Instruct models will quantize to 50% of the size in 4bit mode. Once vision layer support is added, the size will reduce to expected ~1/4.

• [MODEL] Add Llama 3.2 Vision (mllama)* support by @LRL-ModelCloud in #401

Full Changelog: v1.0.4...v1.0.5

GPTQModel v1.0.4

What's Changed

Liger Kernel support added for ~50% vram reduction in quantization stage for some models. Added toggle to disable parallel packing to avoid oom larger models. Transformers depend updated to 4.45.0 for Llama 3.2 support.

• [FEATURE] add a parallel_packing toggle by @LRL-ModelCloud in #393
• [FEATURE] add liger_kernel support by @LRL-ModelCloud in #394

Full Changelog: v1.0.3...v1.0.4

GPTQModel v1.0.3

What's Changed

• [MODEL] Add minicpm3 by @LDLINGLINGLING in #385
• [FIX] fix minicpm3 support by @LRL-ModelCloud in #387
• [MODEL] Added GRIN-MoE support by @LRL-ModelCloud in #388

New Contributors

• @LDLINGLINGLING made their first contribution in #385
• @mrT23 made their first contribution in #386

Full Changelog: v1.0.2...v1.0.3

GPTQModel v1.0.2

What's Changed

Upgrade the AutoRound package to v0.3.0. Pre-built WHL and PyPI source releases are now available. Installation can be done by downloading our pre-built WHL or using pip install gptqmodel --no-build-isolation.

• [CORE] Autoround v0.3 by @LRL-ModelCloud in #368
• [CI] Lots of CI fixups by @CSY-ModelCloud

Full Changelog: v1.0.0...v1.0.2

v1.0.0

What's Changed

40% faster multi-threaded packing, new lm_eval api, fixed python 3.9 compat.

• Add lm_eval api by @PZS-ModelCloud in #338
• Multi-threaded packing in quantization by PZS-ModelCloud in #354
• [CI] Add TGI unit test by @PZS-ModelCloud in #348
• [CI] Updates by @CSY-ModelCloud in #347, #352, #353, #355, @CSY-ModelCloud in #357
• Fix python 3.9 compat by @PZS-ModelCloud in #358

Full Changelog: v0.9.11...v1.0.0

GPTQModel v0.9.11

What's Changed

Added LG EXAONE 3.0 model support. New dynamic per layer/module flexible quantization where each layer/module may have different bits/params. Added proper sharding support to backend.BITBLAS. Auto-heal quantization errors due to small damp values.

• [CORE] add support for pack and shard to bitblas by @LRL-ModelCloud in #316
• Add dynamic bits by @PZS-ModelCloud in #311, #319, #321, #323, #327
• [MISC] Adjust the validate order of QuantLinear when BACKEND is AUTO by @ZX-ModelCloud in #318
• add save_quantized log model total size by @PZS-ModelCloud in #320
• Auto damp recovery by @CSY-ModelCloud in #326
• [FIX] add missing original_infeatures by @CSY-ModelCloud in #337
• Update Transformers to 4.44.0 by @Qubitium in #336
• [MODEL] add exaone model support by @LRL-ModelCloud in #340
• [CI] Upload wheel to local server by @CSY-ModelCloud in #339
• [MISC] Fix assert by @CSY-ModelCloud in #342

Full Changelog: v0.9.10...v0.9.11

GPTQModel v0.9.10

What's Changed

Ported vllm/nm gptq_marlin inference kernel with expanded bits (8bits), group_size (64,32), and desc_act support for all GPTQ models with format = FORMAT.GPTQ. Auto calculate auto-round nsamples/seglen parameters based on calibration dataset. Fixed save_quantized() called on pre-quantized models with non-supported backends. HF transformers depend updated to ensure Llama 3.1 fixes are correctly applied to both quant and inference stage.

• [CORE] add marlin inference kernel by @ZX-ModelCloud in #310
• [CI] Increase timeout to 40m by @CSY-ModelCloud in #295, #299
• [FIX] save_quantized() by @ZX-ModelCloud in #296
• [FIX] autoround nsample/seqlen to be actual size of calibration_dataset. by @LRL-ModelCloud in #297, @LRL-ModelCloud in #298
• Update HF transformers to 4.43.3 by @Qubitium in #305
• [CI] remove test_marlin_hf_cache_serialization() by @ZX-ModelCloud in #314

Full Changelog: v0.9.9...v0.9.10

GPTQModel v0.9.9

What's Changed

Added Llama-3.1 support, Gemma2 27B quant inference support via vLLM, auto pad_token normalization, fixed auto-round quant compat for vLLM/SGLang.

• [CI] by @CSY-ModelCloud in #238, #236, #237, #241, #242, #243, #246, #247, #250
• [FIX] explicitly call torch.no_grad() by @LRL-ModelCloud in #239
• Bitblas update by @Qubitium in #249
• [FIX] calib avg for calib dataset arg passed as tensors by @Qubitium, @LRL-ModelCloud in #254, #258
• [MODEL] gemma2 27b can load with vLLM now by @LRL-ModelCloud in #257
• [OPTIMIZE] to optimize vllm inference, set an environment variable 'VLLM_ATTENTI… by @LRL-ModelCloud in #260
• [FIX] hard set batch_size to 1 for 4.43.0 transformer due to compat/regression by @LRL-ModelCloud in #279
• FIX vllm llama 3.1 support by @Qubitium in #280
• Use better defaults values for quantization config by @Qubitium in #281
• [REFRACTOR] Cleanup backend and model_type usage by @LRL-ModelCloud in #276
• [FIX] allow auto_round lm_head quantization by @LRL-ModelCloud in #282
• [FIX] [MODEL] Llama-3.1-8B-Instruct's eos_token_id is a list by @CSY-ModelCloud in #284
• [FIX] add release_vllm_model, and import destroy_model_parallel in release_vllm_model by @LRL-ModelCloud in #288
• [FIX] autoround quants compat with vllm/sglang by @Qubitium in #287

Full Changelog: v0.9.8...v0.9.9

GPTQModel v0.9.8

What's Changed

1. Marlin end-to-end in/out feature padding for max model support
2. Run quantized models (FORMAT.GPTQ) directly using fast vLLM backend!
3. Run quantized models (FORMAT.GPTQ) directly using fast SGLang backend!

• 🚀 🚀 [CORE] Marlin end-to-end in/out feature padding by @LRL-ModelCloud in #183 #192
• 🚀 🚀 [CORE] Add vLLM Backend for FORMAT.GPTQ by @PZS-ModelCloud in #190
• 🚀 🚀 [CORE] Add SGLang Backend by @PZS-ModelCloud in #191
• 🚀 [CORE] Use Triton v2 to pack gptq/gptqv2 formats by @LRL-ModelCloud in #202
• ✨ [CLEANUP] remove triton warmup by @Qubitium in #200
• 👾 [FIX] 8bit choosing wrong packer by @Qubitium in #199
• ✨ [CI] [CLEANUP] Improve Unit Tests by CSY, PSY, and ZYC
• ✨ [DOC] Consolidate Examples by ZYC in #225

Full Changelog: v0.9.7...v0.9.8