Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

fix torch compile (#2459)

Wenhua Cheng799ebee79aea
diff --git a/auto_round/compressors/model_free.py b/auto_round/compressors/model_free.py
index 9d5c54f4b..c9dbdd5c8 100644
--- a/auto_round/compressors/model_free.py
+++ b/auto_round/compressors/model_free.py
@@ -1832,6 +1832,10 @@ class ModelFreeCompressor(_ModelFreeCompressorCore):
model=model_name_or_path,
iters=0,
disable_opt_rtn=disable_opt_rtn,
+ # Forward the user's torch.compile choice so the fallback compressor
+ # honors an explicit --enable_torch_compile request. ``None`` keeps the
+ # fallback's default situational policy (auto-disable for RTN/OPT-RTN).
+ enable_torch_compile=enable_torch_compile,
tokenizer=tokenizer,
scheme=copy.deepcopy(scheme),
layer_config=copy.deepcopy(layer_config),