|
When should you *not* use custom_{fwd/bwd}?
|
|
0
|
321
|
February 16, 2024
|
|
Casting Inputs Using custom_fwd Disables Gradient Tracking
|
|
2
|
630
|
February 8, 2024
|
|
Wrong Tensor type when using Flash Attention 1.0.9
|
|
0
|
367
|
February 1, 2024
|
|
Autocast with BCELoss() on CPU
|
|
2
|
917
|
January 18, 2024
|
|
Torch.nan not supported in int16
|
|
1
|
507
|
January 9, 2024
|
|
How to use float16 for all tensor operations?
|
|
4
|
2612
|
January 1, 2024
|
|
How to switch mixed-precision mode in training
|
|
2
|
505
|
December 26, 2023
|
|
Gradient with Automatic Mixed Precision
|
|
2
|
794
|
November 23, 2023
|
|
Changing dtype drastically affects training time
|
|
1
|
459
|
November 15, 2023
|
|
AMP on cpu: No Gradscaler necessary / available?
|
|
1
|
1173
|
November 14, 2023
|
|
Does torch.cuda.amp support O2 almost FP16 training now?
|
|
1
|
744
|
November 2, 2023
|
|
Why would GradientScaler work
|
|
3
|
510
|
October 28, 2023
|
|
Training loss behaves strangely in mixed-precision training
|
|
5
|
865
|
October 20, 2023
|
|
Gradients'dtype is not fp16 when using torch.cuda.amp
|
|
3
|
615
|
October 20, 2023
|
|
Model distillation with mixed-precision training
|
|
4
|
586
|
October 9, 2023
|
|
Unexpected execution time difference for identical operations on GPU
|
|
8
|
675
|
September 25, 2023
|
|
Performance regression in torch 2.0 with deterministic algorithms
|
|
2
|
805
|
September 22, 2023
|
|
Is autocast expected to reflect changes to weights?
|
|
1
|
546
|
September 20, 2023
|
|
How to handle the value outside the fp16 range when casting?
|
|
6
|
1954
|
September 11, 2023
|
|
Gradients type in torch.cuda.amp
|
|
3
|
918
|
August 22, 2023
|
|
Torch autocast's gradient
|
|
3
|
749
|
August 21, 2023
|
|
Scaler.step(optimizer) in FP16 or FP32?
|
|
1
|
818
|
August 2, 2023
|
|
Amp on cpu 50x slower and high memory allocation
|
|
0
|
575
|
August 1, 2023
|
|
Why the loss_scale getting smaller and smaller?
|
|
1
|
662
|
July 17, 2023
|
|
Dataset half precision
|
|
1
|
592
|
July 11, 2023
|
|
Cudnn.allow_tf32 makes my network slower
|
|
5
|
900
|
July 4, 2023
|
|
What is the correct way to use mixed-precision training with OneCycleLR
|
|
3
|
842
|
June 20, 2023
|
|
How to replace apex.amp by pytorch amp?
|
|
1
|
2693
|
June 14, 2023
|
|
Fused mixed precision updates with PyTorch amp
|
|
4
|
1472
|
June 7, 2023
|
|
Fp16 matmul - CUDA kernel output differs from torch
|
|
2
|
1142
|
May 31, 2023
|