This PR batch the fix for a few accuracy failures issues during training by raising tolerance. I do that only for models that I think it fails not due to real issue.
## sebotnet33ts_256
The accuracy test for this model start to fail around June 05 [link](https://hud.pytorch.org/benchmark/timm_models/inductor_with_cudagraphs?dashboard=torchinductor&startTime=Sun%2C%2002%20Jun%202024%2007%3A19%3A38%20GMT&stopTime=Tue%2C%2002%20Jul%202024%2007%3A19%3A38%20GMT&granularity=day&mode=training&dtype=amp&lBranch=main&lCommit=04a0d856207d83c2031e4b9cb6825ba3e0092850&rBranch=main&rCommit=e62925930f6a62f6aeeb1fe1a661a9bd3352b53d&model=sebotnet33ts_256).
I can not repro locally, but from the log from the dashboard:
```
RMSE (res-fp64): 0.09441, (ref-fp64): 0.02971 and shape=torch.Size([1536]). res.dtype: torch.float32, multiplier: 3.000000, tol: 0.040000
```
raising the tolerance should fix it.
## DebertaForQuestionAnswering
This model fails accuracy test on the dashboard only in max-autotune mode. I can not repro locally by command:
```
TORCHINDUCTOR_MAX_AUTOTUNE=1 time python benchmarks/dynamo/huggingface.py --accuracy --no-translation-validation --training --amp --backend inductor --device cuda --only DebertaForQuestionAnswering
```
From error message on the dashboard:
```
RMSE (res-fp64): 0.01803, (ref-fp64): 0.00537 and shape=torch.Size([2]). res.dtype: torch.float32, multiplier: 3.000000, tol: 0.010000
```
0.02 tolerance should suppress this error.
## gluon_inception_v3
This model fail on the dashboard in max-autotune mode. I can not repro locally by command
```
TORCHINDUCTOR_MAX_AUTOTUNE=1 time python benchmarks/dynamo/timm_models.py --accuracy --training --amp --backend inductor --disable-cudagraphs --device cuda --only gluon_inception_v3
```
From error message on the dashboard
```
RMSE (res-fp64): 0.02798, (ref-fp64): 0.00730 and shape=torch.Size([384]). res.dtype: torch.float32, multiplier: 3.000000, tol: 0.010000
Accuracy failed for key name Mixed_7c.branch3x3dbl_3a.bn.running_var
```
raising tolerance should suppress this error.
# mobilenetv3_large_100
Fail in MA model. I can not repro locally by command
```
TORCHINDUCTOR_MAX_AUTOTUNE=1 time python benchmarks/dynamo/timm_models.py --accuracy --training --amp --backend inductor --disable-cudagraphs --device cuda --only
```
The error message on the dashboard is
```
RMSE (res-fp64): 0.29754, (ref-fp64): 0.05205 and shape=torch.Size([]). res.dtype: torch.float32, multiplier: 3.000000, tol: 0.040000
```
The tensor is so small that the noise can be high. I use larger multiplier for smaller tensor in torch._dynamo.utils.same.
# yolov3
Fail on dashboard with error
```
Error on the dashboard: RMSE (res-fp64): 0.01278, (ref-fp64): 0.00246 and shape=torch.Size([256]). res.dtype: torch.float32, multiplier: 3.000000, tol: 0.001000
```
Fix it by using a larger multiplier for smaller tensors and raising the tolereance.
# timm_efficientdet
Fail on the dashboard with error
```
E0623 18:37:43.638000 139924418725056 torch/_dynamo/utils.py:1468] RMSE (res-fp64): 0.00096, (ref-fp64): 0.00009 and shape=torch.Size([2]). res.dtype: torch.float32, multiplier: 3.000000, tol: 0.001000
```
But I can not repro locally with command
```
time python benchmarks/dynamo/torchbench.py --backend inductor --amp --performance --only timm_efficientdet --training
```
Raise the tolerance should fix.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/129941
Approved by: https://github.com/jansel
ghstack dependencies: #129996
The `usort` config in `pyproject.toml` has no effect due to a typo. Fixing the typo make `usort` do more and generate the changes in the PR. Except `pyproject.toml`, all changes are generated by `lintrunner -a --take UFMT --all-files`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/127126
Approved by: https://github.com/kit1980
The `usort` config in `pyproject.toml` has no effect due to a typo. Fixing the typo make `usort` do more and generate the changes in the PR. Except `pyproject.toml`, all changes are generated by `lintrunner -a --take UFMT --all-files`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/127126
Approved by: https://github.com/kit1980
ghstack dependencies: #127122, #127123, #127124, #127125
We encountered some model accuracy failures as the tolerance is critical. In general, we align with CUDA practice. This PR intends to adjust the tolerance for Torchbench models for training mode on Intel GPU devices and aligns with CUDA.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125213
Approved by: https://github.com/desertfire
Biggest movement is 4% HF inference, 9% TIMM inference. Note, this is max-autotune mode so we are more tolerant of compilation increases. We could improve compilation time by limiting:
```
# Take how many of the top triton kernels to benchmark epilogue
max_epilogue_benchmarked_choices = 3
```
There is a hf_Whisper failure which you can repro on main without this stack with `TORCHINDUCTOR_MAX_AUTOTUNE_GEMM_BACKENDS=TRITON TORCHINDUCTOR_MAX_AUTOTUNE=1 python benchmarks/dynamo/torchbench.py --backend inductor --amp --accuracy --training --only hf_Whisper`. When you turn off epilogue fusion, it fixes the accuracy. I bisected the failure to an epilogue, however when you compare the results of that epilogue with the corresponding separate kernels the results of the output are equivalent.
Inference:
<img width="1686" alt="image" src="https://github.com/pytorch/pytorch/assets/11477974/0b240080-cd33-4c08-89d3-583103b1fb0c">
Training:
<img width="1329" alt="Screenshot 2024-04-16 at 6 16 30 PM" src="https://github.com/pytorch/pytorch/assets/11477974/db0afcc9-7288-4c27-84ce-4fc1a5690788">
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124031
Approved by: https://github.com/Chillee, https://github.com/shunting314
ghstack dependencies: #124030, #122642, #123229, #122825
torchrec_dlrm training fail the accuracy check when max-autotune is enabled.
I found there is no real issue in PT2. We fail to get fp64 reference results for the accuracy check. In max-autotune mode numerical may change a bit and cause the cosine similarity check fail. Using fp64 baseline is more reliable and make the test pass.
The reason why we are not using a fp64 baseline earlier is because torchrec uses a dataclass [Batch](99e6e669b5/torchrec/datasets/utils.py (L28)) to represent the input. We use pytree to cast model and inputs to fp64. pytree can not look into a dataclass. My fix is to convert the dataclass to namedtuple to be more pytree friendly
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122012
Approved by: https://github.com/jansel, https://github.com/eellison
After this, the sam_fast benchmark can now be run in the pytorch repo:
```
SEGMENT_ANYTHING_FAST_USE_FLASH_4=0 benchmarks/dynamo/torchbench.py --inference --amp --performance --backend=inductor --explain --only sam_fast
```
sam_fast is designed for inference only, with cuda and amp on. The code adds these restrictions to the benchmark.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121420
Approved by: https://github.com/oulgen, https://github.com/msaroufim
Fix https://github.com/pytorch/pytorch/issues/120545 . The reason why these models fail accuracy test with freezing is due to the conv-batchnorm fusion. Conv-batchnorm fusion causes relative big numerical churn.
For the failed TIMM models, raising the tolerance to `8 * 1e-2` can make the test pass.
For the failed TB models, the numerical difference is too large. Having a discussion with @eellison , we decided to skip them with freezing for now.
One the other hand, we probably should dig more why the conv-bn fusion cause such large numerical difference.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121054
Approved by: https://github.com/eellison
This PR moves other aspects of torchbench's model configuration (e.g. batch size,
tolerance requirements, etc.) into a new YAML file: `torchbench.yaml`. It also merges the
recently added `torchbench_skip_models.yaml` file inside the `skip` key.
This is an effort so that external consumers are able to easily replicate the performance
results and coverage results from the PyTorch HUD.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120299
Approved by: https://github.com/jansel
# Summary
## PR Dependencies
I don't use ghstack :( this is a PR where it would have been helpful. That beings said I am going to peel off some PRs to make reviewing this easier:
- [x] Separate build flags for Flash and MemEff: #107985
### Description
This pull request updates the version of _scaled_dot_product_flash_attention from version 1 to version 2. The changes are based on the flash attention code originally authored by @tridao
### Changes Made
The majority of the changes in this pull request involve:
- Copying over the flash_attention sources.
- Updating header files.
- Removing padding and slicing code from within the flash_attention kernel and relocating it to the composite implicit region of the SDPA. This was need to make the kernel functional and appease autograd.
- Introducing a simple kernel generator to generate different instantiations of the forward and backward flash templates.
- Adding conditional compilation (ifdef) to prevent building when nvcc is invoked with gencode < sm80.
- Introducing a separate dependent option for mem_eff_attention, as flash_attention v2 lacks support for Windows and cannot be built for sm50 generation codes.
- Modifying build.sh to reduce parallelization on sm86 runners and to lower the maximum parallelization on the manywheel builds. This adjustment was made to address out-of-memory issues during the compilation of FlashAttentionV2 sources.
- Adding/Updating tests.
### Notes for Reviewers
This is not a fun review, and I apologize in advance.
Most of the files-changed are in the flash_attn/ folder. The only files of interest here IMO:
- aten/src/ATen/native/transformers/cuda/flash_attn/flash_api.cpp
- aten/src/ATen/native/transformers/cuda/flash_attn/kernels/generate_kernels.py ( this has been incorporated upstream to flash-attention github)
There are a number of files all related to avoiding OOMs in CI/CD. These are typically shell scripts.
### Follow up items
- Include the updates from e07aa036db and 9e5e8bc91e | https://github.com/pytorch/pytorch/issues/108108
### Work Items
- [x] I don't think Windows will be supported for 3.1.0 - Need to update cmakee
- [x] Let multi_query/attention pass through and test | UPDATE: I have the fast path implemented here: https://github.com/pytorch/pytorch/pull/106730 but since this will require changes to semantics of math to call repeat_interleave, I think this should be done as a followup.
- [x] Had to drop cutlass back to 3.0.0 to get it to compile. Need to figure out how to upgrade to 3.1.0 and later. Spoke with Tri and he is going to be taking a look. Note: compiling with clang currently errors for the cute headers.
- [x] Update test exercise above codepath
- [x] Still need to disable on seq_len % 128 != 0 for backward( Tri beat me to it a4f148b6ab)
- [x] Add determinism warning to BWD, Tri got to this one as well: 1c41d2b
- [x] Update dispatcher to universally prefer FlashV2
- [x] Update tests to exercise new head_dims
- [x] Move the head_dim padding from kernel to top level composite implicit function in order to make it purely functional
- [x] Create template generator script
- [x] Initial cmake support for building kernels/ folder
- [x] Replay CudaGraph changes
### Results
#### Forward only
The TFlops are reported here are on a100 that is underclocked.
#### Forward+Backward
Ran a sweep and for large compute bound sizes we do see a ~2x performance increase for forw+back.
<img width="1684" alt="Screenshot 2023-07-20 at 3 47 47 PM" src="https://github.com/pytorch/pytorch/assets/32754868/fdd26e07-0077-4878-a417-f3a418b6fb3b">
Pull Request resolved: https://github.com/pytorch/pytorch/pull/105602
Approved by: https://github.com/huydhn, https://github.com/cpuhrsch
Move detectron2_fcos_r_50_fpn to amp. The minifier showed the following snippet as causing the divergence, where inductor has better numerics than eager:
```
import torch
def foo(x):
return x > .2
inp = torch.tensor([.2002], device="cuda", dtype=torch.bfloat16)
print(foo(inp))
print(torch.compile(foo)(inp))
```
doctr_reco_predictor had very minimal divergence (.002 vs .001 required), bumping tolerance here.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108598
Approved by: https://github.com/shunting314
# Summary
## PR Dependencies
I don't use ghstack :( this is a PR where it would have been helpful. That beings said I am going to peel off some PRs to make reviewing this easier:
- [x] Separate build flags for Flash and MemEff: #107985
### Description
This pull request updates the version of _scaled_dot_product_flash_attention from version 1 to version 2. The changes are based on the flash attention code originally authored by @tridao
### Changes Made
The majority of the changes in this pull request involve:
- Copying over the flash_attention sources.
- Updating header files.
- Removing padding and slicing code from within the flash_attention kernel and relocating it to the composite implicit region of the SDPA. This was need to make the kernel functional and appease autograd.
- Introducing a simple kernel generator to generate different instantiations of the forward and backward flash templates.
- Adding conditional compilation (ifdef) to prevent building when nvcc is invoked with gencode < sm80.
- Introducing a separate dependent option for mem_eff_attention, as flash_attention v2 lacks support for Windows and cannot be built for sm50 generation codes.
- Modifying build.sh to reduce parallelization on sm86 runners and to lower the maximum parallelization on the manywheel builds. This adjustment was made to address out-of-memory issues during the compilation of FlashAttentionV2 sources.
- Adding/Updating tests.
### Notes for Reviewers
This is not a fun review, and I apologize in advance.
Most of the files-changed are in the flash_attn/ folder. The only files of interest here IMO:
- aten/src/ATen/native/transformers/cuda/flash_attn/flash_api.cpp
- aten/src/ATen/native/transformers/cuda/flash_attn/kernels/generate_kernels.py ( this has been incorporated upstream to flash-attention github)
There are a number of files all related to avoiding OOMs in CI/CD. These are typically shell scripts.
### Follow up items
- Include the updates from e07aa036db and 9e5e8bc91e | https://github.com/pytorch/pytorch/issues/108108
### Work Items
- [x] I don't think Windows will be supported for 3.1.0 - Need to update cmakee
- [x] Let multi_query/attention pass through and test | UPDATE: I have the fast path implemented here: https://github.com/pytorch/pytorch/pull/106730 but since this will require changes to semantics of math to call repeat_interleave, I think this should be done as a followup.
- [x] Had to drop cutlass back to 3.0.0 to get it to compile. Need to figure out how to upgrade to 3.1.0 and later. Spoke with Tri and he is going to be taking a look. Note: compiling with clang currently errors for the cute headers.
- [x] Update test exercise above codepath
- [x] Still need to disable on seq_len % 128 != 0 for backward( Tri beat me to it a4f148b6ab)
- [x] Add determinism warning to BWD, Tri got to this one as well: 1c41d2b
- [x] Update dispatcher to universally prefer FlashV2
- [x] Update tests to exercise new head_dims
- [x] Move the head_dim padding from kernel to top level composite implicit function in order to make it purely functional
- [x] Create template generator script
- [x] Initial cmake support for building kernels/ folder
- [x] Replay CudaGraph changes
### Results
#### Forward only
The TFlops are reported here are on a100 that is underclocked.
#### Forward+Backward
Ran a sweep and for large compute bound sizes we do see a ~2x performance increase for forw+back.
<img width="1684" alt="Screenshot 2023-07-20 at 3 47 47 PM" src="https://github.com/pytorch/pytorch/assets/32754868/fdd26e07-0077-4878-a417-f3a418b6fb3b">
Pull Request resolved: https://github.com/pytorch/pytorch/pull/105602
Approved by: https://github.com/huydhn, https://github.com/cpuhrsch
Fixes inference accuracy for `doctr_reco_predictor` and `pyhpc_turbulent_kinetic_energy`.
For the `same(float, float)` comparison we weren't going through the more rigorous tensor comparison path which takes into account the fp64 base results.
Also return True when fp64 base result are not well formed (nan).
I debugged these models and the source of divergence were innocuous:
`doctr_reco_predictor` - can be fixed by turning off layout optimization, decomp for batch norm
`pyhpc_turbulent_kinetic_energy` - divergence caused because fused kernel keeps precision in fp32 instead of casting back and forth from/to fp32/bf16. Fused kernel is better precision, anyway.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108202
Approved by: https://github.com/jansel
