auto-round
[Accuracy improvement] Re-implement opt-rtn for both imatrix and model-free, covering INT & NVFP
#2362
Closed
Overview
Commits
15
Changes
View On GitHub
Commits
  • [EXP] support NVFP4 in model free
    xin3he committed 39 days ago
  • feat: enhance NVFP4 support with fake quantization and global scale handling
    xin3he committed 39 days ago
  • feat: implement optimized RTN for NVFP4 E5M3 with scale search enhancements; support search scale in model_free for NVFP4 dtype
    xin3he committed 39 days ago
  • feat: implement fused scale updates for NVFP4 projections across output shards; enhance global scale handling in quantization
    xin3he committed 39 days ago
  • feat: add disable_opt_rtn support and enhance NVFP4 scale handling; implement neighbor scale search logic
    xin3he committed 39 days ago
  • feat: add symmetric search offsets function for scale search; remove unused logging for non-default scale selection in NVFP4
    xin3he committed 39 days ago
  • feat: enhance NVFP4 support with disable_opt_rtn functionality; add neighbor search steps for scale refinement
    xin3he committed 38 days ago
  • Merge remote-tracking branch 'origin/main' into xinhe/9-15
    Copilot committed 38 days ago
  • fix: resolve autoround merge conflict with main
    Copilot committed 38 days ago
  • fix: normalize legacy scheme aliases in merge resolution
    Copilot committed 38 days ago
  • feat: enhance NVFP4 support with new validation and configuration functions
    xin3he committed 38 days ago
  • Merge remote-tracking branch 'origin/main' into xinhe/9-15
    Copilot committed 38 days ago
  • fix: restore wrapper nvfp compatibility
    Copilot committed 38 days ago
  • feat: enhance NVFP4 support with validation functions and improved search scale handling
    xin3he committed 38 days ago
  • Merge branch 'main' into xinhe/9-15
    xin3he committed 38 days ago
FAQ Terms Privacy Refunds Impressum
Loading