Lightning Habana 1.7.0

New Features and Enhancements

• Added generalized hpu parallel strategy #237
• Added Intel neural compressor and deprecated Habana Quantization toolkit #235
• Added utility to parse device name from hl-smi #232
• Added more tests for torch.compile #213
• Fixed getting device name without hlsmi #240
• Fixed torch device specification for FSDP on HPU #222
• Fixed fsdp strategy compatibility with latest release #226
• Fixed fsdp example #227
• Fixed multi card tests for all tenants and use default broadcast #234
• Fixed distributed initialization order #239
• Upgraded to Synapse AI Release 1.18.0 #245, #221, #207
• Updated infrastructure #205, #208, #209, #206, #210, #211, #117, #212,
  #214, #217, #220, #223, #224, #229, #231, #236, #238, #241, #242, #243, #247, #250

For more information on Intel Habana Synapse AI release, refer to Gaudi release notes

Contributors

@ankitgola005 @jyothisambolu @Borda @jerome-habana

Full Changelog: 1.6.0...1.7.0

Lightning Habana 1.7.0 Release Candidate

Preview release for 1.7.0

Lightning Habana 1.6.0

New Features and Enhancements

• Added FSDP strategy with fabric on HPU #198
• Added additional dtype support #194
• Added HPUPrecisionPlugin support for fp8 training with Transformer engine #195
• Fixed label name issue with torch.compile #185
• Fixed deepspeed configs & tests based on synapse AI release 1.15.1 and latest PTL fabric #184
• Fixed incompatibility issue for PyTorch>=2.3.0 #193
• Upgraded to Synapse AI Release 1.16.0 #191
• Updated hpu compile backend #183
• Updated HQT APIs #192
• Updated infrastructure #186, #188, #189, #187, #196, #199, #202, #200, #201, #197, #204, #203, #190

For more information on Intel Gaudi Synapse AI release, refer to Gaudi release notes

Contributors

@ankitgola005 @jyothisambolu @Borda @jerome-habana

Full Changelog: 1.5.0...1.6.0

Lightning Habana 1.6.0 Release Candidate

Preview release for 1.6.0

Lightning Habana 1.5.0

New Features and Enhancements

• Added fp8 inference support #162
• Added experimental FSDP support #174, #182
• Added support for Intel Gaudi Profiler #158, #172
• Added fp8 training and inference support to deepspeed strategy #176
• Fixed tests failing on Gaudi2 #154
• Fixed contextmanager to be compatible with Python 3.8 #155
• Replaced HPUParallelStrategy with HPUDDPStrategy #160
• Upgraded to Intel Gaudi Release 1.15.1 #171, #166
• Updated dynamic shape tests to use metric APIs #163
• Code restructure and added more profiler tests #159
• Updated document with transformer engine support #165
• Updated infrastructure #156, #157, #161, #167, #169, #168, #170, #173, #177, #180, #179, #178

For more information on Intel Habana Synapse AI release, refer to Gaudi release notes

Contributors

@ankitgola005 @jyothisambolu @Borda @jerome-habana

Full Changelog: 1.4.0...1.5.0

Lightning Habana 1.4.0

New Features and Enhancements

• Added fp8 training support with HPU #149
• Added deepspeed precision plugin for HPU #147
• Added Multi-tenancy examples with HPU #111
• Added hpu graphs & examples #141
• Fixed fabric support for HPU #126, #139
• Fixed API break due to non-strict loading in Trainer #150
• Fixed profiler tests #129
• Deprecated aot_hpu_training_backend #148
• Updated examples & unit tests #131, #130, #142
• Upgraded to Synapse AI Release 1.14.0 #140
• Updated infrastructure #128, #132, #134, #133, #135, #138, #137, #145, #143, , #144, #146, #153

For more information on Intel Habana Synapse AI release, refer to Gaudi release notes

Contributors

@ankitgola005 @jyothisambolu @Borda @jerome-habana

Full Changelog: 1.3.0...1.4.0

Lightning Habana 1.3.0

New Features and Enhancements

• Upgraded to Synapse AI Release 1.13.0 #114
• Updated infrastructure #108, #109, #122, #124, #125, #123,
• Added DeepSpeed inference strategy with HPU #110
• Showcased dynamic shapes with disk caching #107
• Showcased multi-tenancy with parallel agents #99, #112, #113, #116, #118
• Showcased torch compile with HPU #119
• Showcased HPU graph #115

For more information on Intel Habana Synapse AI release, refer to Gaudi release notes

Contributors

@assassin1991 @jyothisambolu @Borda @jerome-habana

Full Changelog: 1.2.0...1.3.0

Lightning Habana 1.3.0 Release Candidate

pre-release of LH 1.3.0

Lightning Habana 1.2.0

New Features and Enhancements

• Upgraded to lightning 2.1.0 #79, #100, #105
• Upgraded to latest Synapse AI Release 1.12.1 #90, #106
• Updated infrastructure #83, #84, #86, #85, #82, #80, #89, #87, #88, #97, #96, #101
• Added notes on HW upgrade process #92, #102
• Upgraded ci workflows #97, #96, #103, #104
• Added environment variables cleanup with HPU distributed backend #91
• Added tests, example and documentation for HPUPrecisionPlugin with autocast #94
• Added tests to validate checkpoint resuming with deepspeed #95

For more information on Intel Habana Synapse AI release, refer to Gaudi release notes

Contributors

@assassin1991 @jyothisambolu @Borda @jerome-habana

Full Changelog: 1.1.0...1.2.0

Lightning Habana 1.1.0

New Features and Enhancements

• Infrastructure updated #57, #58, #59, #60, #70, #71, #73, #74, #72
• Upgrade to HPU Release 1.11.0 #61
• Added documentation with examples for DeepSpeed with HPU #64
• Deprecate HMP precision plugin #65
• Added support to get Gaudi synapse release version #75
• Added support for autocast through HPUPrecision plugin #66
• Fixed optimizer priority based on deepspeed specification #69
• Fixed right synapse AI version check for autocast #76
• Added tests to demonstrate HPU graphs #67
• Enhanced test coverage of HPUDeepSpeed strategy #68
• Implemented reduce with HPUParallel plugin #77

For more information on Intel Habana synapse AI release, please refer Gaudi release notes

Contributors

@assassin1991 @Borda @jyothisambolu @jerome-habana

Full Changelog: 1.0.1...1.1.0