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: support list inputs in diffusion tuning cache (#2346)

Wang, Changc53611fbd8ca
diff --git a/auto_round/compressors/diffusion/tuning_cache.py b/auto_round/compressors/diffusion/tuning_cache.py
index a3f9b55f4..57581a1c6 100644
--- a/auto_round/compressors/diffusion/tuning_cache.py
+++ b/auto_round/compressors/diffusion/tuning_cache.py
@@ -74,6 +74,9 @@ class DiffusionTuningCache:
@classmethod
def create(cls, block, runner, inputs, others, outputs, sampler, iters, budget_gib, device):
+ # Single-output blocks (e.g. Wan) pass a tensor list to the next block.
+ if isinstance(inputs, list):
+ inputs = {"hidden_states": inputs}
if runner.batch_size != 1 or not isinstance(inputs, dict) or iters <= 0:
return None
cache = cls()