Files
miles/docker/patch/cu13/te_dequantized_backward_override.patch
T

23 lines
772 B
Diff

--- a/pytorch/module/grouped_linear.py
+++ b/pytorch/module/grouped_linear.py
@@ -431,6 +431,8 @@
backward_override = None
if backward_override == "high_precision":
save_original_input = True
+ elif backward_override == "dequantized":
+ save_original_input = False
num_gemms = len(m_splits)
weights = weights_and_biases[:num_gemms]
--- a/pytorch/module/linear.py
+++ b/pytorch/module/linear.py
@@ -285,6 +285,8 @@
is_fsdp2 = args.is_fsdp2
if backward_override == "high_precision":
save_original_input = True
+ elif backward_override == "dequantized":
+ save_original_input = False
# NVTX label for profiling
nvtx_label = "transformer_engine._Linear.forward"