lint:
- test/test_fake_tensor.py
- test/test_flop_counter.py
- torch/_export/verifier.py
with same rules as other files, it was a night mare for me to update tests in one of the skipped files
with not being able to lint them locally like other files with lintrunner -a.
note that those file do have active dev and not old not touched files.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154261
Approved by: https://github.com/angelayi, https://github.com/Skylion007
Fixes#126268
I've basically followed @ezyang suggestion (I think) to use `func.decompose(...)`. Since `__torch_dispatch__` won't be called a second time for the same op, I've added a second `TorchDispatchMode` (`_DecomposedCounterMode`) that simpy dispatches to the parent flop counter. Using `self` as the inner context manager is not possible, since the second call to `__enter__` would re-initialize the counter's tracking state.
Let me know if there's something wrong with this implementation, since I'm quite unsure how the decomposition thing actually works :D
Pull Request resolved: https://github.com/pytorch/pytorch/pull/138508
Approved by: https://github.com/ezyang
Co-authored-by: Edward Z. Yang <ezyang@meta.com>
Fixes#126268
I've basically followed @ezyang suggestion (I think) to use `func.decompose(...)`. Since `__torch_dispatch__` won't be called a second time for the same op, I've added a second `TorchDispatchMode` (`_DecomposedCounterMode`) that simpy dispatches to the parent flop counter. Using `self` as the inner context manager is not possible, since the second call to `__enter__` would re-initialize the counter's tracking state.
Let me know if there's something wrong with this implementation, since I'm quite unsure how the decomposition thing actually works :D
Pull Request resolved: https://github.com/pytorch/pytorch/pull/138508
Approved by: https://github.com/ezyang
Previously, FlopCounterMode would ignore any custom ops registered
through `register_flop_formula`. The problem was:
- register_flop_formula(target) requires target to be an OpOverloadPacket.
- register_flop_formula used register_decomposition to populate its registry
- register_decomposition decomposes the OpOverloadPacket into OpOverload before
putting it into the registry
- FlopCounterMode ignores OpOverloads in its registry (it assumes the
registry is a dictionary mapping OpOverloadPacket to flop formula).
register_decomposition is too heavy of a hammer, plus this isn't a
decomposition, so I changed the registration mechanism.
Test Plan:
- new tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/131777
Approved by: https://github.com/Chillee
Looks like one of the first failures seen is `test_causal_variants_compile_causal_variant_CausalVariant_LOWER_RIGHT_shape0_cuda` when `test_causal_variants_causal_variant_CausalVariant_LOWER_RIGHT_shape0_cuda` passes.
What seems interesting here is that the `torch.compile` version fails while the eager version passes. Not sure what the difference would be here...
Nevertheless, is there a recommended mechanism to skip cuDNN SDPA as a backend for this test? CC @drisspg
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125343
Approved by: https://github.com/Skylion007
Looks like one of the first failures seen is `test_causal_variants_compile_causal_variant_CausalVariant_LOWER_RIGHT_shape0_cuda` when `test_causal_variants_causal_variant_CausalVariant_LOWER_RIGHT_shape0_cuda` passes.
What seems interesting here is that the `torch.compile` version fails while the eager version passes. Not sure what the difference would be here...
Nevertheless, is there a recommended mechanism to skip cuDNN SDPA as a backend for this test? CC @drisspg
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125343
Approved by: https://github.com/Skylion007
Looks like one of the first failures seen is `test_causal_variants_compile_causal_variant_CausalVariant_LOWER_RIGHT_shape0_cuda` when `test_causal_variants_causal_variant_CausalVariant_LOWER_RIGHT_shape0_cuda` passes.
What seems interesting here is that the `torch.compile` version fails while the eager version passes. Not sure what the difference would be here...
Nevertheless, is there a recommended mechanism to skip cuDNN SDPA as a backend for this test? CC @drisspg
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125343
Approved by: https://github.com/Skylion007
This patch implements `with sdpa_kernel(SDPBackend.EFFICIENT_ATTENTION):` by reusing AOTriton's accelerated SDPA implementation
Known limitations:
- Only supports MI200/MI300X GPUs
- Does not support varlen
- Does not support `CausalVariant`
- Optional arguments `causal_diagonal` and `seqlen_k` in `_efficient_attention_forward/backward` must be null
- Does not work well with inductor's SDPA rewriter. The rewriter has been updated to only use math and flash attention on ROCM.
This PR also uses a different approach of installing AOTriton binary instead of building it from source in the base docker image. More details on motivation: https://github.com/pytorch/pytorch/pull/124885#issuecomment-2153229129
`PYTORCH_TEST_WITH_ROCM=1 PYTORCH_TESTING_DEVICE_ONLY_FOR="cuda" python test/test_transformers.py` yields "55028 passed, 20784 skipped" results with this change. [Previous result](https://hud.pytorch.org/pr/127528) of `test_transformers.py` was 0 error, 0 failure, 55229 skipped out of 75517 tests in total (the XML report does not contain total number of passed tests).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124885
Approved by: https://github.com/malfet
This adds implementations for:
* _flash_attention_forward
* _efficient_attention_forward
* _flash_attention_backward
* _efficient_attention_backward
These flop counts are implemented as follows:
* Unbind the batch elements
* Calculate flops individually for each element in the batch
* Sum the final result
This means that we are accessing the concrete sequence lengths (which could be slow, and may trigger a GPU/CPU sync); but, the FLOP numbers will vary with the sparsity of the NestedTensor - more accurate than if we just assumed we padded everything.
Differential Revision: [D57120139](https://our.internmc.facebook.com/intern/diff/D57120139)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125776
Approved by: https://github.com/Chillee
This does a few things that were originally a few PRs but I am on a new machine and don't have ghstack.
If it is too problematic to review, I can re-split, just let me know.
This does:
- Cleanup context manager use in test_flop_counter
- Remove need for mod argument in FlopCounterMode, warning about it
- Re-implement a Module tracker from scratch using global forward Module use and multi_grad_hook (we cannot use global backward Module hook because they don't look for nested Tensor and they're custom Function based instead of multi_grad_hook).
- Update FlopCouterMode to use the new ModuleTracker. All the existing test suite passes as-is (only changes there are new tests and refactoring mentioned above)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125352
Approved by: https://github.com/mikaylagawarecki
Note about the Updates:
This PR:
1. skips more flash attention related UTs on MI200
2. Fix additional ATen compiling errors after hipification
3. Fix the author "root" of a specific commit
4. Includes the patch from Nikita in favor of block level static initialization.
CAVEAT: This revised PR has a commit that modifies the CI to force its running on MI200 nodes. That specific commit must be reverted before merge.
Original PR (https://github.com/pytorch/pytorch/pull/114309) Note:
This pull requests add initial Flash Attention support for AMD/ROCM platform. It added a specialized Triton repository/branch as a compile-time dependency for Flash Attention math library on AMD/ROCM. This triton submodule is not used at runtime and will not be shipped to the final pytorch package. We have the plan to release this specialized Triton as a separate project.
Know limitations:
- Only supports MI200 series GPU (i.e., `gcnArchName == gfx90a:sramecc+:xnack-`.
- Only supports power of two sequence lengths.
- No support for varlen APIs.
- Only support head dimension 16,32,64,128.
- Performance is still being optimized.
Fixes#112997
Pull Request resolved: https://github.com/pytorch/pytorch/pull/115981
Approved by: https://github.com/malfet
# Summary:
This pull request **removes** support for non-square sequence lengths in causal attention when using FlashAttention V2.
### Why are doing this
// FlashAttention 2 updated the default mask meaning for causal in this PR:
// 9e5e8bc91e it is now aligned to lower_right which would be a BC break
// for non-square masks. We will not support non-square masks for causal w/ FAV2
For more context see:
https://github.com/pytorch/pytorch/issues/108108
### Followup
A large number of people will likely want to use FAV2 with lower_right causal attention for non equal sequence lengths. See this RFC : https://github.com/pytorch/pytorch/issues/110681
Pull Request resolved: https://github.com/pytorch/pytorch/pull/111007
Approved by: https://github.com/cpuhrsch
# 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
# 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
This PR fixes https://github.com/pytorch/pytorch/issues/103684.
- Instead of registering forward hooks in `__init__()`, do it upon `__enter__()`.
- De-register those forward hooks upon `__exit__()`.
- Achieve this by saving an additional mapping `_module_to_forward_hook_handles: Dict[nn.Module, _ForwardHookHandles]`. Only the values in the mapping (i.e. not the keys) are useful for this change. (A `List[_ForwardHookHandles]` would suffice.)
- The unit test accesses private attributes `_forward_hooks` and `_forward_pre_hooks` :/
Note that this PR is technically not backward compatible since it does not register the hooks upon `__init__()`, which means that you will not get the flops counting without the context manager.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/103744
Approved by: https://github.com/Chillee
High level approach:
1. I generated a bunch of data comparing FlashAttention and Cutlass implementations (https://pastebin.com/pe0j3YeK)
2. I trained a decision tree using standard train/val split methodology and hyperparameter sweeps (https://pastebin.com/fjYX1HjR).
2a. I did a bunch of feature augmentation to capture interactions between features.
The heuristic I ended up with is:
```
use_flash = seq_len / (num_heads * batch_size) > 6
```
TL;DR: On my dataset, where FlashAttention and Cutlass differ by more than 10%, the existing heuristic achieves 69% accuracy. My new heuristic achieves 94% accuracy.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/99644
Approved by: https://github.com/ngimel, https://github.com/drisspg
Overall, an example usage. Note that this *also* captures backwards FLOPs.
```
import torchvision.models as models
import torch
from torch.utils.flop_counter import FlopCounterMode
inp = torch.randn(1, 3, 224, 224, device='cpu')
mod = models.resnet18()
flop_counter = FlopCounterMode(mod, depth=1)
with flop_counter:
mod(inp).sum().backward()
```
<img width="326" alt="image" src="https://user-images.githubusercontent.com/6355099/222023068-3491e405-f195-4e11-b679-36b19a1380c7.png">
You can control the depth of the module hierarchy with the `depth` attribute (which defaults to 2). For example, if I don't limit it, this is what it outputs.
<img width="366" alt="image" src="https://user-images.githubusercontent.com/6355099/222023306-3d880bb6-f534-4f98-bf10-83c4353acefc.png">
## Other APIs
FlopCounterMode(custom_mapping=...): Allows for custom flop counting functions
FlopCounterMode.get_table(depth=...): Explicitly get the table as a string
FlopCounterMode.flop_counts: Contains the flop information as a Dict[hierarchy: str, Dict[Op, int]]
FlopCounterMode.register_hierarchy(f, name): Allows you to register additional "hierarchies" for a function.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/95751
Approved by: https://github.com/ngimel, https://github.com/albanD
