|
Half precision training time same as full precision
|
|
4
|
346
|
September 19, 2024
|
|
Why bf16 do not need loss scaling?
|
|
4
|
5606
|
September 5, 2024
|
|
What does the `use_fast_accum` option do in `torch._scaled_mm`
|
|
1
|
1183
|
August 27, 2024
|
|
Weight parameters with 8, 14 bit precisions?
|
|
3
|
77
|
August 20, 2024
|
|
How to perform mixed precsion on single F.linear?
|
|
1
|
86
|
August 11, 2024
|
|
Model forward pass in AMP gives NaN
|
|
0
|
180
|
August 5, 2024
|
|
Fp16 inference time cost
|
|
2
|
329
|
August 1, 2024
|
|
Is it a good idea to use float16/bfloat16 for inference?
|
|
2
|
490
|
August 1, 2024
|
|
Why nn.LSTM still use float16 in hidden_state, even if set to bfloat16 or float32 already?
|
|
0
|
134
|
July 29, 2024
|
|
Prediction is different with or without padding: the model is sensiitive to floating point precision?
|
|
0
|
79
|
July 26, 2024
|
|
Question about bfloat16 operations in AMP and cuda
|
|
3
|
798
|
July 11, 2024
|
|
Torch.matmul launch different CUDA kernel from cublas
|
|
2
|
466
|
July 6, 2024
|
|
Fp16 overflow when computing matmul in autocast context
|
|
5
|
2045
|
July 5, 2024
|
|
Mixed precision training with transformer embeddings stored in fp16
|
|
0
|
261
|
June 13, 2024
|
|
Autocast keep cache across multiple forward pass
|
|
0
|
202
|
June 5, 2024
|
|
Precision 16 run problem
|
|
2
|
422
|
June 4, 2024
|
|
Torch.save numerical differences
|
|
6
|
1936
|
May 31, 2024
|
|
AMP during inference
|
|
1
|
947
|
May 31, 2024
|
|
GradScaler for CPU with AMP
|
|
8
|
2565
|
May 28, 2024
|
|
Alternative to torch.inverse for 16 bit
|
|
2
|
1230
|
May 6, 2024
|
|
Current CUDA Device does not support bfloat16. Please switch dtype to float16
|
|
1
|
3735
|
April 26, 2024
|
|
Cuda half2 support
|
|
0
|
201
|
April 25, 2024
|
|
How much does TORCH.AMP improve performance
|
|
1
|
421
|
April 22, 2024
|
|
Why bfloat16 matmul is significantly slower than float32?
|
|
0
|
495
|
April 16, 2024
|
|
No gradient received in mixed precision training
|
|
2
|
796
|
April 12, 2024
|
|
What's the use of `scaled_grad_params` in this example of gradient penalty with scaled gradients?
|
|
4
|
325
|
April 9, 2024
|
|
Bfloat16 from float16 issues
|
|
0
|
673
|
April 1, 2024
|
|
FP8 support on H100
|
|
8
|
6353
|
March 8, 2024
|
|
Converting float16 tensor to numpy causes rounding
|
|
2
|
1144
|
February 26, 2024
|
|
Is Autocast Failing to Cast Gradients?
|
|
1
|
428
|
February 19, 2024
|