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 Inference tensors do not track version counter error (#1502)

Wang, Mengnie2b88a5cc6c5
diff --git a/auto_round/compressors/base.py b/auto_round/compressors/base.py
index d56cba25d..5472c373d 100644
--- a/auto_round/compressors/base.py
+++ b/auto_round/compressors/base.py
@@ -1367,10 +1367,10 @@ class BaseCompressor(object):
not any(m.children())
and len(m.state_dict()) > 0
and m.global_name not in tied_weights_layers
+ and self.is_immediate_saving
):
set_module(self.model, m.global_name, copy.deepcopy(m))
- if self.is_immediate_saving:
- shard_writer(self, name=m.global_name)
+ shard_writer(self, name=m.global_name)
m.to("meta")
clear_memory(device_list=self.device_list)
memory_monitor.log_summary()
@@ -1401,10 +1401,14 @@ class BaseCompressor(object):
clear_memory(device_list=self.device_list)
memory_monitor.log_summary()
- elif not any(m.children()) and len(m.state_dict()) > 0 and n not in tied_weights_layers:
+ elif (
+ not any(m.children())
+ and len(m.state_dict()) > 0
+ and n not in tied_weights_layers
+ and self.is_immediate_saving
+ ):
set_module(self.model, n, copy.deepcopy(m))
- if self.is_immediate_saving:
- shard_writer(self, name=n)
+ shard_writer(self, name=n)
m.to("meta")
# Convert remaining fp8