transformers
bf74a954 - Give quantizers the model, checkpoint renames, skip-list resolution and dequantized saves

Commit
28 days ago
Give quantizers the model, checkpoint renames, skip-list resolution and dequantized saves - HfQuantizer.get_weight_conversions(model): converters may depend on the modules that were built. - HfQuantizer.get_checkpoint_renamings(model): renames of the checkpoint's own keys, applied before the model's conversion mapping (a producer's naming). - get_modules_to_not_convert also takes the modules a checkpoint skip-list names under the checkpoint's names (checkpoint_module_namer); the legacy FP8 quantizer's own forward-renaming pass goes. - Core dequantized save: each transform the quantizer derives from a model transform remembers it (_origin); a quantizer with dequantized_saves_unquantized=True saves a dequantized load through the model's own conversions that fired. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Parents
Loading