Skip to content

CHANGELOG.md

Latest commit

 

History

History
74 lines (55 loc) · 2.55 KB

CHANGELOG.md

File metadata and controls

74 lines (55 loc) · 2.55 KB

HF-OSS Demo changelog

Uses changelog conventions. Uses semantic versioning.

Guiding Principles

  • Changelogs are for humans, not machines.
  • There should be an entry for every single version.
  • The same types of changes should be grouped.
  • Versions and sections should be linkable.
  • The latest version comes first.
  • The release date of each version is displayed.
  • Mention whether you follow Semantic Versioning.

Types of changes

  • Added for new features.
  • Changed for changes in existing functionality.
  • Deprecated for soon-to-be removed features.
  • Removed for now removed features.
  • Fixed for any bug fixes.
  • Security in case of vulnerabilities.

[1.3.3] - 2023-01-04

  • Deprecated max workspace size flag to memory pool limits for TensorRT
  • Added t5-11b support
  • Changed T5 demo kv cache TRT memory organization to avoid D2D copy

[1.3.2] - 2022-11-17

  • Added beam search support for GPT2 demo
  • Added KV cache support for GPT2 demo
  • Fixed perplexity calculation array size out of max_length
  • Fixed trt KV cache engine profile to only accept input_length = 1
  • Fixed external onnx weight file name overwrite issue

[1.3.1] - 2022-11-04

  • Added beam search support for T5 demo
  • Added KV cache support for T5 demo

[1.3.0] - 2022-11-03

  • Added perplexity calculation for all samples
  • Added precision override to checkpoints.
  • Fixed TensorRT BART checkpoint not working.

[1.2.5] - 2022-10-08

  • Added beam search support for BART

[1.2.4] - 2022-09-30

  • Added notebooks for BART demo
  • Enabled flexible control on (a) percentile latency reports (b) engine building profile other than standard maximum input/output length config

[1.2.3] - 2022-06-30

  • Added KV cache support for BART demo

[1.2.2] - 2022-06-14

  • Added BART demo

[1.2.1] - 2022-05-20

  • Added benchmark action to T5 frameworks/onnxrt and GPT2 frameworks/trt for performance benchmarking. It uses random inputs with fixed lengths and disables early stopping such that we can compare the performance with other frameworks.
  • Added batch_size > 1 support to GPT2 trt sample.

[1.2.0] - 2022-03-29

  • Added benchmark action to T5 trt for performance benchmarking. It uses random inputs with fixed lengths and disables early stopping such that we can compare the performance with other frameworks.

[1.1.0] - 2022-02-09

  • Added -o or --save-output-fpath which saves a pickled version of the NetworkResult object. Useful for testing.

[1.0.0] - 2022

  • Added initial working example of HF samples and notebooks.