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 param missing bug (#1008)

Wang, Mengni7f9fa0eef4c0
diff --git a/auto_round/compressors/diffusion/compressor.py b/auto_round/compressors/diffusion/compressor.py
index b57e927b6..e17692532 100644
--- a/auto_round/compressors/diffusion/compressor.py
+++ b/auto_round/compressors/diffusion/compressor.py
@@ -191,6 +191,7 @@ class DiffusionCompressor(BaseCompressor):
input_others: dict,
indices: list[int],
device: str,
+ cache_device: str = "cpu",
) -> torch.Tensor:
output_config = output_configs.get(block.__class__.__name__, [])
idx = None if "hidden_states" not in output_config else output_config.index("hidden_states")
@@ -207,7 +208,7 @@ class DiffusionCompressor(BaseCompressor):
current_input_others.update(current_input_ids)
current_input_ids = hidden_states
output_q = block_forward(block, current_input_ids, current_input_others, self.amp, self.amp_dtype, device, idx)
- return output_q
+ return output_q.to(cache_device)
@torch.no_grad()
def _get_block_outputs(