• Highlights
• Features
• Improvements
• Bug Fixes
• Validated Hardware
• Validated Configurations

Highlights

• Aligned Gaudi SW Release 1.21 with the improvements on FP8 and INT4 quantization for Intel® Gaudi® AI accelerator
• INT4 quantization enhancements for Intel CPU/GPU

Features

• Support expert parallelism for Mixtral model on Gaudi
• Enhance multi-cards FP8 model save and load on Gaudi
• Enable static FP8 quantization of DeepSeek V3/R1 model on Gaudi
• Support W4A8 mixed precision on Gaudi (experimental)
• Improve compile time on Gaudi when using FP8 (experimental)

Improvements

• Remove numpy version limit for 3.x PyTorch package

Bug Fixes

• Fix graph compile error when quantizing Llama3.2 11B/90B vision model on Gaudi
• Fix segmentation fault issue in LLama2-70B INT4 model on Intel GPU
• Fix accuracy issue caused by duplicated g_idx update for INT4 model on Intel GPU

Validated Hardware 

• Intel Gaudi Al Accelerators (Gaudi 2 and 3)
• Intel Xeon Scalable processor (4th, 5th, 6th Gen)
• Intel Core Ultra Processors (Series 1 and 2)
• Intel Data Center GPU Max Series (1100)
• Intel® Arc™ B-Series Graphics GPU (B580)

Validated Configurations

• Centos 8.4 & Ubuntu 24.04 & Win 11
• Python 3.9, 3.10, 3.11, 3.12
• PyTorch/IPEX 2.4, 2.5, 2.6

• Highlights
• Features
• Improvements
• Bug Fixes
• Validated Hardware
• Validated Configurations

Highlights

• Aligned Gaudi SW Release 1.20 with the improvements on FP8 and INT4 quantization for Intel® Gaudi® AI accelerator
• VLM INT4 weight-only quantization support in transformers-like API on Intel CPU/GPU

Features

• Saving vLLM compatible FP8 model on Gaudi
• FP8 Per-channel Q/DQ and GC integration on Gaudi
• FP8 quantization for mixture of experts (MoE) module on Gaudi
• Saving Hugging Face compatible weight-only INT4 format on Gaudi
• VLM quantization with AutoRound in transformers-like API on Intel CPU/GPU
• Accuracy-aware tuning on PT2E including mixed precision support

Improvements

• FP8 multi-device (Gaudi & GPU) infrastructure support
• Support scaler scale save on Gaudi

Bug Fixes

• Fix incorrect hf_device_map setting for Transformers-like API
• Fix missing IPEX CPU dependency in Transformers-like API example
• Fix device mapping issue found in GPTQ on Llama model
• Fix saving issue in weight-only per-channel quantization

Validated Hardware 

• Intel Gaudi Al Accelerators (Gaudi 2 and 3)
• Intel Xeon Scalable processor (4th, 5th, 6th Gen)
• Intel Core Ultra Processors (Series 1 and 2)
• Intel Data Center GPU Max Series (1100)
• Intel® Arc™ B-Series Graphics GPU (B580)

Validated Configurations

• Centos 8.4 & Ubuntu 24.04 & Win 11
• Python 3.9, 3.10, 3.11, 3.12
• PyTorch/IPEX 2.3, 2.4, 2.5

• Highlights
• Features
• Improvements
• Bug Fixes
• Validated Hardware
• Validated Configurations

Highlights

• Aligned with Habana 1.19 release with the improvements on FP8 and INT4 quantization for Intel® Gaudi® AI accelerator
• INT4 weight-only quantization on Intel® Arc™ B-Series Graphics GPU (code-named BattleMage)

Features

• Saving and loading FP8 checkpoint on Gaudi
• Loading vLLM/llm-compressor compatible FP8 checkpoint on Gaudi
• Arbitrary scale method support on Gaudi
• AutoRound INT4 weight-only quantization on Gaudi
• Block-wise calibration for LLM on Gaudi
• INT4 weight-only quantization on BattleMage

Improvements

• Improve FP8 performance by setting scale as scalar tensor on Gaudi
• Integrate AutoRound 0.4.2 with VLM quantization improvements
• Improve safetensors loading for layer-wise quantization in Transformers-like API
• Improve non-contiguous weight saving in Transformers-like API

Bug Fixes

• Fix layer-wise quantization issue in GPTQ on client GPU
• Fix glm-4-9b model out-of-memory issue on BattleMage

Validated Hardware 

• Intel Gaudi Al Accelerators (Gaudi 2 and 3)
• Intel Xeon Scalable processor (4th, 5th, 6th Gen)
• Intel Core Ultra Processors (Series 1 and 2)
• Intel Data Center GPU Max Series (1100)
• Intel Arc B-Series Graphics GPU (B580)

Validated Configurations

• Centos 8.4 & Ubuntu 22.04 & Win 11
• Python 3.9, 3.10, 3.11, 3.12
• PyTorch/IPEX 2.3, 2.4, 2.5

• Highlights
• Features
• Improvements
• Validated Hardware
• Validated Configurations

Highlights

• Aligned with Habana 1.18 release with the improvements on FP8 and INT4 quantization for Intel® Gaudi® AI accelerator
• Provided Transformer-like quantization API for weight-only quantization on LLM, which offers transformer-based user one-stop experience for quantization & inference with IPEX on Intel GPU and CPU.

Features

• Add Transformer-like quantization API for weight-only quantization on LLM
• Support fast quantization with light weight recipe and layer-wise approach on Intel AI PC
• Support INT4 quantization of Visual Language Model (VLM), like Llava, Phi-3-vision, Qwen-VL with AutoRound algorithm

Improvements

• Support AWQ format INT4 model loading and converting for IPEX inference in Transformer-like API
• Enable auto-round format export for INT4 model
• Support per-channel INT8 Post Training Quantization for PT2E

Validated Hardware 

• Intel Gaudi Al Accelerators (Gaudi 2 and 3)
• Intel Xeon Scalable processor (4th, 5th, 6th Gen)
• Intel Core Ultra Processors (Series 1 and 2)
• Intel Data Center GPU Max Series (1100)

Validated Configurations

• Centos 8.4 & Ubuntu 22.04 & Win 11
• Python 3.9, 3.10, 3.11, 3.12
• PyTorch/IPEX 2.2, 2.3, 2.4

• Highlights
• Features
• Improvements
• Examples
• Bug Fixes
• Documentations
• Validated Configurations

Highlights

• FP8 quantization and INT4 model loading support on Intel® Gaudi® AI accelerator
• Framework extension API for quantization, mixed-precision and benchmarking
• Accuracy-aware FP16 mixed precision support on Intel® Xeon® 6 Processors
• Performance optimizations and usability improvements on client-side quantization

Features

• [Quantization] Support FP8 quantization on Gaudi (95197d)
• [Quantization] Support INC and Hugging Face model loading on framework extension API for PyTorch (0eced1, bacc16)
• [Quantization] Support Weight-only Quantization on framework extension API for PyTorch (34f0a9, de43d8, 4a4509, 1a4509, a3a065, 1386ac, a0dee9, 503d9e, 84d705, 099b7a, e3c736, e87c95, 2694bb, ec49a2, e7b4b6, a9bf79, ac717b, 915018, 8447d7, dc9328)
• [Quantization] Support static and dynamic quantization in PT2E path (7a4715, 43c358, 30b36b, 1f58f0, 02958d)
• [Quantization] Support SmoothQuant and static quantization in IPEX path with framework extension API (53e6ee, 72fbce, eaa3a5, 95e67e, 855c10, 9c6102, 5dafe5, a5e5f5, 191383, 776645)
• [Quantization] Support Layer-wise Quantization for RTN/GPTQ on framework extension API for PyTorch (649e6b)
• [Quantization] Support Post Training Quantization on framework extension API for Tensorflow (6c27c1, e22c61, f21afb, 3882e9, 2627d3)
• [Quantization] Support Post Training Quantization on Keras3 (f67e86, 047560)
• [Quantization] Support Weight-only Quantization on Gaudi2 (4b9b44, 14868c, 0a3d4b)
• [Quantization] Improve performance and usability of quantization procedure on client side (16a7b1)
• [Quantization] Support auto-device detection on framework extension API for PyTorch (368ba5, 4b9b44, e81a2d, 0a3d4b, 534300, 2a86ae)
• [Quantization] Support Microscaling(MX) Quant for PyTorch (4a24a6, 455f1e)
• [Quantization] Enable cross-devices Half-Quadratic Quantization(HQQ) for LLMs support (db6164, 07f940)
• [Quantization] Support FP8 cast Weight-only Quantization (57ed61)
• [Mixed-Precision] Support FP16 mixed-precision on framework extension autotune API for PyTorch (2e1cdc)
• [Mixed-Precision] Support mixed INT8 with FP16 in PT2E path (fa961e)
• [AutoTune] Support accuracy-aware tuning on framework extension API (e97659, 7b8aec, 5a0374, a4675c, 3a254e, ac47d9, b8d98e, fb6142, fa8e66, d22df5, 09eb5d, c6a8fa)
• [Benchmarking] Implement incbench command for ease-of-use benchmark (2fc725)

Improvements

• [Quantization] Integrate AutoRound v0.3 (bfa27e, [fd9...

• Highlights
• Features
• Improvements
• Examples
• Bug Fixes
• External Contributions
• Validated Configurations

Highlights

• Integrated recent AutoRound with lm-head quantization support and calibration process optimizations
• Migrated ONNX model quantization capability into ONNX project Neural Compressor

Features

• [Quantization] Integrate recent AutoRound with lm-head quantization support and calibration process optimizations (4728fd)
• [Quantization] Support true sequential options in GPTQ (92c942)

Improvements

• [Quantization] Improve WOQ Linear pack/unpack speed with numpy implementation (daa143)
• [Quantization] Auto detect available device when exporting (7be355)
• [Quantization] Refine AutoRound export to support Intel GPU (409231)
• [Benchmarking] Detect the number of sockets when needed (e54b93)

Examples

• Upgrade lm_eval to 0.4.2 in PT and ORT LLM example (fdb509) (54f039)
• Add diffusers/dreambooth example with IPEX (ba4798)

Bug Fixes

• Fix incorrect dtype of unpacked tensor issue in PT (29fdec)
• Fix TF LLM SQ legacy Keras environment variable issue (276449)
• Fix TF estimator issue by adding version check on TF2.16 (855b98)
• Fix missing tokenizer issue in run_clm_no_trainer.py after using lm-eval 0.4.2 (d64029)
• Fix AWQ padding issue in ORT (903da4)
• Fix recover function issue in ORT (ee24db)
• Update model ckpt download url in prepare_model.py (0ba573)
• Fix case where pad_max_length set to None (960bd2)
• Fix a failure for GPU backend (71a9f3)
• Fix numpy versions for rnnt and 3d-unet examples (12b8f4)
• Fix CVEs (5b5579) (25c71a) (47d73b) (41da74)

External Contributions

• Update model ckpt download url in prepare_model.py (0ba573)
• Fix case where pad_max_length set to None (960bd2)
• Add diffusers/dreambooth example with IPEX (ba4798)

Validated Configurations

• Centos 8.4 & Ubuntu 22.04 & Win 11 & MacOS Ventura 13.5
• Python 3.8, 3.9, 3.10, 3.11
• PyTorch/IPEX 2.1, 2.2, 2.3
• TensorFlow 2.14, 2.15, 2.16
• ITEX 2.13.0, 2.14.0, 2.15.0
• ONNX Runtime 1.16, 1.17, 1.18

• Improvement
• Bug Fixes
• Validated Configurations

Improvement

• Improve WOQ AutoRound export (409231, 7ee721)
• Adapt ITREX v1.4 release for example evaluate (9d7a05)
• Update more supported LLM recipes (ce9b16)

Bug Fixes

• Fix WOQ RTN supported layer checking condition (079177)
• Fix in-place processing error in quant_weight function (92533a)

Validated Configurations

• Centos 8.4 & Ubuntu 22.04
• Python 3.10
• TensorFlow 2.15
• ITEX 2.14.0
• PyTorch/IPEX 2.2
• ONNX Runtime 1.17

• Highlights
• Features
• Improvement
• Productivity
• Bug Fixes
• External Contributes
• Validated Configurations

Highlights

• Integrated Weight-Only Quantization algorithm AutoRound and verified on Gaudi2, Intel CPU, NV GPU
• Applied SmoothQuant & Weight-Only Quantization algorithms with 15+ popular LLMs for INT8 & INT4 quantization and published the recipes

Features

• [Quantization] Integrate Weight-Only Quantization algorithm AutoRound (5c7f33, dfd083, 9a7ddd, cf1de7)
• [Quantization] Quantize weight with in-place mode in Weight-Only Quantization (deb1ed)
• [Pruning] Enable SNIP on multiple cards using DeepSpeed ZeRO-3 (49ab28)
• [Pruning] Support new pruning approach Wanda and DSNOT for PyTorch LLM (7a3671)

Improvement

• [Quantization] SmoothQuant code structure refactor (a8d81c)
• [Quantization] Optimize the workflow of parsing Keras model (b816d7)
• [Quantization] Support static_groups options in GPTQ API (1c426a)
• [Quantization] Update TEQ train dataloader (d1e994)
• [Quantization] WeightOnlyLinear keeps self.weight after recover (2835bd)
• [Quantization] Add version condition for IPEX prepare init (d96e14)
• [Quantization] Enhance the ORT node name checking (f1597a)
• [Pruning] Stop the tuning process early when enabling smooth quant (844a03)

Productivity

• ORT LLM examples support latest optimum version (26b260)
• Add coding style docs and recommended VS Code setting (c1f23c)
• Adapt transformers 4.37 loading (6133f4)
• Upgrade pre-commit checker for black/blacken-docs/ruff (7763ed)
• Support CI summary in PR comments (d4bcdd))
• Notebook example update to install latest INC & TF, add metric in fit (4239d3)

Bug Fixes

• Fix QA IPEX example fp32 input issue (c4de19)
• Update Conditions of Getting min-max during TF MatMul Requantize (d07175)
• Fix TF saved_model issues (d8e60b)
• Fix comparison of module_type and MulLinear (ba3aba)
• Fix ORT calibration issue (cd6d24)
• Fix ORT example bart export failure (b0dc0d)
• Fix TF example accuracy diff during benchmark and quantization (5943ea)
• Fix bugs for GPTQ exporting with static_groups (b4e37b)
• Fix ORT quant issue caused by tensors having same name (0a20f3)
• Fix Neural Solution SQL/CMD injection (14b7b0)
• Fix the best qmodel recovery issue (f2d9b7)
• Fix logger issue (83bc77)
• Store token in protected file (c6f9cc)
• Define the default SSL context (b08725)
• Fix IPEX stats bug (5af383)
• Fix ORT calibration for Dml EP (c58aea)
• Fix wrong socket number retrieval for non-english system (5b2a88)
• Fix trust remote for llm examples (2f2c9a)

External Contributes

• Intel Mac support (21cfeb)
• Add PTQ example for PyTorch CV Segment Anything Model (bd5e69)

Validated Configurations

• Centos 8.4 & Ubuntu 22.04 & Win 11 & MacOS Ventura 13.5
• Python 3.8, 3.9, 3.10, 3.11
• TensorFlow 2.13, 2.14, 2.15
• ITEX 2.13.0, 2.14.0
• PyTorch/IPEX 2.0, 2.1, 2.2
• ONNX Runtime 1.15, 1.16, 1.17

• Improvement
• Bug Fixes
• Examples
• Validated Configurations

Improvement

• Narrow down the tuning space of SmoothQuant auto-tune (9600e1)
• Support ONNXRT Weight-Only Quantization with different dtypes (5119fc)
• Add progress bar for ONNXRT Weight-Only Quantization and SmoothQuant (4d26e3)

Bug Fixes

• Fix SmoothQuant alpha-space generation (33ece9)
• Fix inputs error for SmoothQuant example_inputs (39f63a)
• Fix LLMs accuracy regression with IPEX 2.1.100 (3cb6d3)
• Fix quantizable add ops detection on IPEX backend (4c004d)
• Fix range step bug in ORTSmoothQuant (40275c)
• Fix unit test bugs and update CI versions (6c78df, 835805)
• Fix notebook issues (08221e)

Examples

• Add verified LLMs list and recipes for SmoothQuant and Weight-Only Quantization (f19cc9)
• Add code-generaion evaluation for Weight-Only Quantization GPTQ (763440)

Validated Configurations

• Centos 8.4 & Ubuntu 22.04
• Python 3.10
• TensorFlow 2.14
• ITEX 2.14.0.1
• PyTorch/IPEX 2.1.0
• ONNX Runtime 1.16.3

• Highlights
• Features
• Improvement
• Productivity
• Bug Fixes
• Examples
• Validated Configurations

Highlights

• Supported layer-wise quantization for PyTorch RTN/GPTQ Weight-Only Quantization and ONNX Runtime W8A8 quantization.
• Supported Weight-Only Quantization tuning for ONNX Runtime backend.
• Supported GGML double quant on RTN/GPTQ Weight-Only Quantization with FW extension API
• Supported SmoothQuant of Big Saved Model for TensorFlow Backend.

Features

• [Quantization] Support GGML double quant in Weight-Only Quantization for RTN and GPTQ (05c15a)
• [Quantization] Support Weight-Only Quantization tuning for ONNX Runtime backend (6d4ea5, 934ba0, 4fcfdf)
• [Quantization] Support SmoothQuant block-wise alpha-tuning (ee6bc2)
• [Quantization] Support SmoothQuant of Big Saved Model for TensorFlow Backend (3b2925, 4f2c35)
• [Quantization] Support PyTorch layer-wise quantization for GPTQ (ee5450)
• [Quantization] support PyTorch layer-wise quantization for RTN (ebd1e2)
• [Quantization] Support ONNX Runtime layer-wise W8A8 quantization (6142e4, 5d33a5)
• [Common] [Experimental] FW extension API implement (76b8b3, 8447d7, 258236)
• [Quantization] [Experimental] FW extension API for PT backend support Weight-Only Quantization (915018, dc9328)
• [Quantization] [Experimental] FW extension API for TF backend support Keras Quantization (2627d3)
• [Quantization] IPEX 2.1 XPU (CPU+GPU) support (af0b50, cf847c)

Improvement

• [Quantization] Add use_optimum_format for export_compressed_model in Weight-Only Quantization (5179da, 0a0644)
• [Quantization] Enhance ONNX Runtime quantization with DirectML EP (db0fef, d13183, 098401, 6cad50)
• [Quantization] Support restore ipex model from json (c3214c)
• [Quantization] ONNX Runtime add attr to MatMulNBits (7057e3)
• [Quantization] Increase SmoothQuant auto alpha running speed (173c18)
• [Quantization] Add SmoothQuant alpha search space as a config argument (f9663d)
• [Quantization] Add SmoothQuant weight_clipping as a default_on option (1f4aec)
• [Quantization] Support SmoothQuant with MinMaxObserver (45b496)
• [Quantization] Support Weight-Only Quantization with fp16 for PyTorch backend (d5cb56)
• [Quantization] Support trace with dictionary type example_inputs (afe315)
• [Quantization] Support falcon Weight-Only Quantization (595d3a)
• [Common] Add deprecation decorator in experimental fold (aeb3ed)
• [Common] Remove 1.x API dependency (ee617a)
• [Mixed Precision] Support PyTorch eager mode BF16 MixedPrecision (3bfb76)

Productivity

• Support quantization and benchmark on macOS (16d6a0)
• Support ONNX Runtime 1.16.0 (d81732, 299af9, 753783)
• Support TensorFlow new API for gnr-base (8160c7)

Bug Fixes

• Fix GraphModule object has no attribute bias (7f53d1)
• Fix ONNX model export issue (af0aea, eaa57f)
• Add clip for ONNX Runtime SmoothQuant (cbb69b)
• Fix SmoothQuant minmax observer init (b1db1c)
• Fix SmoothQuant issue in get/set_module (dffcfe)
• Align sparsity with block-wise masks in progressive pruning (fcdc29)

Examples

• Support peft model with SmoothQuant (5e21b7)
• Enable two ONNX Runtime examples table-transformer-detection (550cee), BEiT (7265df)

Validated Configurations

• Centos 8.4 & Ubuntu 22.04 & Win10 & MacOS Ventura 13.5
• Python 3.8, 3.9, 3.10, 3.11
• TensorFlow 2.13, 2.14, 2.15
• ITEX 1.2.0, 2.13.0.0, 2.14.0.1
• PyTorch/IPEX 1.13.0+cpu, 2.0.1+cpu, 2.1.0
• ONNX Runtime 1.14.1, 1.15.1, 1.16.3
• MXNet 1.9.1