Skip to content

Conversation

@xin3he
Copy link
Contributor

@xin3he xin3he commented Jan 18, 2024

Type of Change

Migration weight-only implementation to 3x API.

Description

  • Design review
  • Refine weight-only configuration
    • RTN
    • GPTQ
    • AWQ
    • TEQ
  • Separate public utilities from weight_only algos
    • RTN
    • GPTQ
    • AWQ
    • TEQ
  • Unify interface for quantization
    • RTN
    • GPTQ
    • AWQ
    • TEQ
  • Migrate algos.
    • AWQ
    • TEQ
  • Implement double_quant.
    • AWQ
    • TEQ
  • Add UTs for each algos on global configuration.
    • RTN
    • GPTQ
    • AWQ
    • TEQ
  • Add UTs for each algos on per-op configuration quantization.
    • RTN
    • GPTQ
    • AWQ
    • TEQ

Expected Behavior & Potential Risk

UT pass

@xin3he xin3he added the draft label Jan 18, 2024
@xin3he xin3he changed the title Xinhe/woq PyTorch Weight-only 3x API Implementation Jan 18, 2024
@xin3he xin3he marked this pull request as draft January 18, 2024 06:04
@xin3he
Copy link
Contributor Author

xin3he commented Jan 18, 2024

For INC unification, PR: Rename RTNWeightOnlyConfig to RTNConfig merge first.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

draft PyTorch Related to PyTorch F/W WIP

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants