Skip to content

Latest commit

 

History

History
251 lines (147 loc) · 10.6 KB

CHANGELOG.md

File metadata and controls

251 lines (147 loc) · 10.6 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[1.7.0] - 2024-10-DD

Added

  • Added utility to get device name from hl-smi (#232)
  • Integrated Intel Neural Compressor for FP8 inference (#235)

Changed

  • Updated to Intel Gaudi software Release 1.16.2 (#207)
  • Updated to Intel Gaudi software Release 1.17.0 (#221)
  • Modified torch device specification for FSDP on HPU (#222)
  • Updated strategy to use default fork (#234)
  • Updated hpu parallel strategy as base class (#237)
  • Updated to Intel Gaudi software Release 1.18.0 (#245)

Fixed

  • Fixed device name retrieval without hlsmi (#240)

Removed

Deprecated

  • Deprecated support for Habana Quantization Toolkit. (#235)

[1.6.0] - 2024-06-28

Added

  • Added support for additional dtypes (#194)
  • Added more tests of FSDP with HPU (#197)
  • Added FSDP strategy with fabric on HPU (#198)

Changed

  • Updated to common hpu_backend interface for compile support. (#183)
  • Updated to Intel Gaudi software Release 1.16.0 (#191)
  • Updated HQT APIs to be in accordance with Intel Gaudi software Release 1.16.0 (#192)
  • Updated HPUPrecisionPlugin for fp8 based on Intel Gaudi software Release 1.16.0. (#195)

Fixed

  • Fixed deepspeed documentation & tests based on synapse AI release 1.15.1 and latest PTL fabric. (#184)
  • Workaround to resolve label name issue in HPUProfiler with torch.compile. (#185)
  • Fixed incompatibility issue for PyTorch>=2.3.0 (#193)

Removed

Deprecated

[1.5.0] - 2024-05-03

Added

  • Added support for Intel Gaudi Profiler. Deprecate HABANA_PROFILE environment variable from HPUProfiler. (#158)
  • Added support for FP8 inference. (#162)
  • Added support for LightningCLI. (#173)
  • Added experimental support for FSDP on HPU. (#174)
  • Added support for FP8 inference with DeepSpeed. (#176)
  • Updated the lightning version check for using FSDP. (#182)

Changed

  • Changed HPUParallelStrategy to HPUDDPStrategy (#160)
  • Changed HPU docker image based on Synapse AI release 1.15.0 (#166)
  • Updated to Intel Gaudi software Release 1.15.1 (#171)

Fixed

  • Fixed "No profiler activity found" error with HPUProfiler. (#172)

Removed

Deprecated

[1.4.0] - 2024-02-16

Added

  • Added DeepSpeed precision plugin for HPU (#147)
  • Added support for fp8 training. (#149)

Changed

  • Decoupled return strings of firmware, synapse version helper (#137)
  • Changed HPU docker image based on Synapse AI release 1.14.0 (#140)

Fixed

  • Fixed fabric imports for HPU strategies (#126)
  • Enabling tests and examples of fabric with HPU (#139)
  • Fixes an API break due to non-strict loading in Trainer (#150)

Removed

Deprecated

  • aot_hpu_training_backend will be deprecated. Use hpu_backend instead for torch compile with hpu (#148)

[1.3.0] - 2023-12-06

Added

  • Added support for Deepspeed inference on HPU with tests and documentation (#110)
  • Added tests, examples, and documentation for dynamic shapes with recipe caching (#107)
  • Added preview of torch compile with tests and documentation (#119)

Changed

  • Changed HPU docker image based on Synapse AI release 1.13.0 (#114)

Fixed

  • Fixed fabric imports for HPU strategies (#126)

Removed

Deprecated

[1.2.0] - 2023-10-26

Added

  • Added tests, examples and documentation for HPUPrecisionPlugin with autocast (#94)
  • Added test to validate checkpoint resuming with HPUDeepSpeedStrategy (#95)
  • Added support for lightning 2.1 (#100, #105)

Changed

  • Changed HPU docker image based on synapse AI release 1.12.0 (#90)
  • Use standard API's and Remove env variable to get HPU distributed backend (#91)
  • Changed HPU docker image based on synapse AI release 1.12.1, updated hooks (#106)

Fixed

Removed

Deprecated

[1.1.0] - 2023-09-26

Added

  • Documentation with examples for using DeepSpeed with HPU (#64)
  • Add autocast using HPUPrecision plugin (#66, #75)
  • Demonstrate HPU Graphs support (#67)
  • Enhance test coverage of DeepSpeed strategy on HPU (#68)
  • Added version check helper to use right release (#75, #76)
  • Implement reduce with parallel plugin (#77)

Changed

  • Changed HPU docker image based on synapse AI release 1.11.0 & upgraded deepspeed plugin to version 0.9.4 (#61)

Fixed

  • Fixed optimizer priority based on deepspeed specification (#36)
  • Fixed missing extras in package (#70)

Deprecated

  • Warn on HMP deprecation from HPUPrecision plugin (#65)

[1.0.1] - 2023-07-26

Added

  • Added tests for mixed precision training (#36)
  • Example to include mixed precision training (#54)

Changed

  • Enabled skipped tests based on registered strategy, accelerator (#46)

Fixed

  • Fixed Attribute Error (#43)
  • Fixed wrong imports (#44)
  • Fixed graph breaks in test/val phases in lazy mode (#45)

[1.0.0] - 2023-06-14

Added

  • Added HPU support for fabric (#11)
  • Added Pytorch HPU profiler support (#15)
  • Added basic HPU infra support for deep speed (#21)
  • Added Pytorch HPU datamodule support (#16)

Changed

  • Changed code hierarchy in compliance with base lightning code for pytorch (#12)
  • Changed default HPU docker image based on HPU release 1.10.0 (#30)

Fixed

  • Fixed mnist example test (#20)
  • Habana's dataloader hang with Lightning 2.0.x (#29)
  • Make #29 applicable only for gaudi devices (#39)
  • Fixed environment initialization for hpus and fixed docs (#40)
  • Fixed docs and added work around to make use hpu media packages without signature issues (#41)

Removed

  • Cleaning up env's ID for HPU parallel plugins based on synapse AI release 1.9 (#28)
  • Remove unnecessary import checks which degrade performance (#38)