Replace https://github.com/pytorch/pytorch/pull/138947 for re-import.
Replaces https://github.com/ROCm/pytorch/pull/1592
This PR contains the initial implementation of SDPA with composable_kernel backend. The CK path can be forced by simply calling torch.backends.cuda.preferred_rocm_fa_library("ck"). Similarly, you can force the incumbent aotriton implementation by passing in "aotriton" or "default". As you'd expect, not setting this option will result in aotriton to be used as the backend. In the case of CK, if pytorch deems flash attention usable, then it will use the CK path in all the same places aotriton would have been used. This PR makes no changes to the heuristics which select which attention scheme to use (i.e. flash attention vs memory efficient attention vs math etc etc). It only gets called when flash attention is both enabled (via USE_FLASH_ATTENTION) and is selected at runtime by the existing heuristics.
Files located in pytorch/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha* have been pulled from https://github.com/Dao-AILab/flash-attention courtesy of @tridao's hard work who is the co-author
NOTE: In order to use this backend, the user MUST set USE_CK_FLASH_ATTENTION=1 in their environment when they build PyTorch.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143695
Approved by: https://github.com/malfet
Co-authored-by: Andy Lugo <Andy.LugoReyes@amd.com>
Co-authored-by: Jithun Nair <jithun.nair@amd.com>
Replaces https://github.com/ROCm/pytorch/pull/1592
This PR contains the initial implementation of SDPA with composable_kernel backend. The CK path can be forced by simply calling `torch.backends.cuda.preferred_rocm_fa_library("ck")`. Similarly, you can force the incumbent aotriton implementation by passing in "aotriton" or "default". As you'd expect, not setting this option will result in aotriton to be used as the backend. In the case of CK, if pytorch deems flash attention usable, then it will use the CK path in all the same places aotriton would have been used. This PR makes no changes to the heuristics which select which attention scheme to use (i.e. flash attention vs memory efficient attention vs math etc etc). It only gets called when flash attention is both enabled (via `USE_FLASH_ATTENTION`) and is selected at runtime by the existing heuristics.
Files located in pytorch/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha* have been pulled from https://github.com/Dao-AILab/flash-attention courtesy of @tridao's hard work who is the co-author
NOTE: In order to use this backend, the user MUST set USE_CK_FLASH_ATTENTION=1 in their environment when they build PyTorch.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/138947
Approved by: https://github.com/pruthvistony, https://github.com/xw285cornell, https://github.com/leitian
Co-authored-by: Xiaodong Wang <xw285@cornell.edu>
Notable new features for SDPA operators on AMD systems from AOTriton 0.8b:
1. Nestedtensor support;
2. MQA/GQA support;
3. Restore Efficient attention support for causal=True and seqlen_q != seqlen_k cases;
+ The kernel should use top-left alignment, bottom right alignment will be added later
4. Move gfx1100 (RX7900/W7800/W7900) out of experimental support status.
However, users are strongly recommended to update to ROCM 6.2.4, notably for
its firmware updates.
Related unit tests are enabled as well.
Notable related changes from AOTriton 0.8b:
1. AOTriton 0.8b moves the GPU kernel out of libaotriton.so to a separate directory `aotriton.images`;
2. LZMA replaces ZSTD as GPU kernel compression algorithm for better compression ratio: aotriton0.8b (.so + aotriton.images take 350MB) compared to aotriton0.7b .so: 800MB
3. The compression cannot be disabled now, and `liblzma` is hard run-time dependency.
+ Should not be a problem, since `lzma` is part of Python Standard Library
Pull Request resolved: https://github.com/pytorch/pytorch/pull/140172
Approved by: https://github.com/jithunnair-amd, https://github.com/jeffdaily
Co-authored-by: Jithun Nair <37884920+jithunnair-amd@users.noreply.github.com>
Notable changes:
1. Enable CudaGraph related tests
2. Fix UT problems
3. EXPERIMENTAL Navi31 support. User should enable Navi31 support with Env Var `TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1`
Know Problem:
1. `test/test_transformers.py` will massive failures and/or NaN outputs with `--use-pytest`
+ Update: Confirmed skip `class TestSDPAPrivateUse1Only` can fix the problem with `--use-pytest`
Note:
AOTriton 0.7b adds support to nestedtenosrs+SDPA but need more work (and consequently a separate PR) to enable it.
Fixes#133540
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134498
Approved by: https://github.com/pruthvistony, https://github.com/jeffdaily, https://github.com/malfet
Use oneDNN BRGEMM on packed data to get better performance on the 5th generation of Xeon where Intel® Advanced Matrix Extensions (AMX) will have fp16 support, e.g. amx-fp16.
Multiple models have achieved acceleration, for instance, FP16 stable diffusion v2.1 has achieved over 50% improvement.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/131879
Approved by: https://github.com/jgong5, https://github.com/peterbell10
ghstack dependencies: #131878
# UPDATE:
This is take 3 of https://github.com/pytorch/pytorch/pull/131863 which was landed via co dev but not applying correclty
# Summary
Changes the stance of SDPA on what to do for fully masked out rows
## Current Behavior
Several PyTorch users have expressed frustration over this issue:
- https://github.com/pytorch/pytorch/issues/41508
- https://github.com/pytorch/pytorch/issues/103749
- https://github.com/pytorch/pytorch/issues/103963
These are significant issues with extensive discussion but no satisfactory resolution. The PyTorch team's consensus, as stated here:
https://github.com/pytorch/pytorch/issues/24816#issuecomment-524415617
Can be paraphrased as follows:
When passing in fully masked out rows, attention becomes ambiguous. We have two main options:
1. Uniformly attend to all values:
```python
scores[masked_out_rows] = 1 / len(row)
out[masked_out_rows] = 1 / len(row) * value
```
2. Decide that attention between no queries (masked) and no keys (masked) is meaningless:
```python
output[fully_masked_rows] = NaN
```
We went with option 2. Partially because it was easier to implement, but also people argued that users can slice the output to remove the NaNs:
``` Python
>fill_value = -float("inf")
>row0 = torch.randn(4)
>row1 = torch.tensor([(fill_value for _ in range(4)])
>matrix = torch.stack([row0, row1]).requires_grad_(True)
>out = torch.softmax(matrix, 1)
>out = out[0]
>print(out)
tensor([0.5377, 0.2729, 0.0692, 0.1201])
```
Cool, problem solved. But what happends when you call backwards..
```Python
>out.backward(torch.ones_like(out))
>print(matrix.grad)
tensor([[3.0957e-08, 1.4157e-08, 7.7802e-10, 1.3713e-08],
[ nan, nan, nan, nan]])
```
Those pesky NaNs are back!
## Why do we see NaNs today?
The core of the problem revolves around using softmax function in sdpa:
```python
> row = torch.tensor([(-float("inf")) for _ in range(4)])
> torch.softmax(row, 0)
tensor([nan, nan, nan, nan])
```
## Quick Aside: Masking in Attention
Attention itself doesn't have a concept of masking. The `sdpa` function has an argument called `attn_mask`, which would be more accurately named `attn_bias`. This is because we don't actually "mask" entries when computing attention. Instead, due to implementation details([performance](https://github.com/pytorch/pytorch/issues/25110#issuecomment-524519087)), we add a value to the masked-out query/key pairs.
We use a large negative number (typically -inf) to decrease the attention weight, as softmax assigns more weight to larger values.
## Alternative Approaches
If we use a very large negative number instead of -inf:
```python
> row = torch.tensor([(-1e6) for _ in range(4)])
> torch.softmax(row, 0)
tensor([0.2500, 0.2500, 0.2500, 0.2500])
```
However if users always remembered to "slice" out their outputs i.e.:
```Python
>fill_value = -1e6
>...
>out.backward(torch.ones_like(out))
>print(matrix.grad)
tensor([[-0.0563, -0.0564, 0.1613, -0.0486],
[ 0.0000, 0.0000, 0.0000, 0.0000]])
```
This would bring us back into a better state.
## A Third Option
We don't necessarily need to alter the behavior of softmax for -inf or very large negative numbers. The fundamental goal is to exclude certain query/key pairs from attention, regardless of the underlying implementation.
This PR implements the new semantic for masking w/ attention in fully masked-out rows:
```python
out[masked_out_rows] = 0
```
**Important Note**: This idea isn't entirely new. The [MaskedTensor](https://pytorch.org/tutorials/prototype/maskedtensor_overview#safe-softmax) prototype, a tensor subclass, was designed to handle such cases. However, it remains a prototype feature and hasn't gained widespread adoption.
## Details
This PR stack does 3 things:
1. Adds a PRIVATE _safe_softmax op
2. Updates semantic for flash_cpu fused kernel
3. Updates semantic for efficient_cuda fused kernel
_safe_softmax is not supposed to be used generically and is only meant to be used within the context of SDPA. Due to this fact instead of decomposing softmax and checking for -inf rows we instead "cheat" and use nan_to_num.
Why I think this is okay? (please find a counter point if avail)
There are multiple ways NaNs can emerge. For the fully masked out rows case nan_to_num works. But what if there were other NaNs, wouldn't this silently remove them?
The only case that this can happen is if the input itself had a NaN or an Inf
For example:
```Python
a = torch.ones([4], requires_grad=False, dtype=torch.float16)
a[1] = torch.finfo(torch.float16).max
print(a.softmax(-1))
```
Will return
`tensor([0., 1., 0., 0.], dtype=torch.float16)`
Where
```Python
a = torch.ones([4], requires_grad=False, dtype=torch.float16)
a[1] = float("inf")
a.softmax(-1)
```
returns:
`tensor([nan, nan, nan, nan], dtype=torch.float16)`
If we dont want to even allow for the possibility of "inf" or "NaN" attention scores to be converted to 0 then we can implemented it something like this
```Python
max = torch.max(a, dim=-1, keepdim=True)
exp = torch.exp(a - max.values)
denom = torch.sum(exp, dim=-1, keepdim=True)
softmax = exp / denom
softmax = torch.where(max.values == float('-inf'), 0.0, softmax)
```
however we would be paying for this in math performance.
## Why Now
I think one point that has substantially changed where PyTorch should lie on this argument is the fact that we have fused implementations for SDPA now. And these fused implementations allow us to easily and performantly support this new semantic.
Differential Revision: [D61418679](https://our.internmc.facebook.com/intern/diff/D61418679)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/133882
Approved by: https://github.com/soulitzer
- Reduced number of skipped test cases
- Merged redundant test cases
**Benchmark:**
| | Original | New |
| ----- | ----- | ----- |
| Run time | 60 mins | 35 mins |
| Total tests | 75k | 18k |
| Skipped tests | 20k | 4k |
_These are approximate numbers from running test_transformers.py on a single H100, and can change based on the device._
Pull Request resolved: https://github.com/pytorch/pytorch/pull/133049
Approved by: https://github.com/drisspg
Summary:
feikou observed the big numerical gaps when using math backend on AMD and NV GPUs. It's mainly because we are not using higher precision FP32 for the intermediate accumulated/materialized parts.
Since math backend is expected to be slower anyways, and we expect math backend to generate the correct reference result, I think it should be worth to upcast FP16/BF16 input to FP32, and do FP32/TF32 computations, and then downcast FP32 output back to FP16/BF16.
Differential Revision: D58710805
Pull Request resolved: https://github.com/pytorch/pytorch/pull/128922
Approved by: https://github.com/xw285cornell, https://github.com/drisspg
Summary:
feikou observed the big numerical gaps when using math backend on AMD and NV GPUs. It's mainly because we are not using higher precision FP32 for the intermediate accumulated/materialized parts.
Since math backend is expected to be slower anyways, and we expect math backend to generate the correct reference result, I think it should be worth to upcast FP16/BF16 input to FP32, and do FP32/TF32 computations, and then downcast FP32 output back to FP16/BF16.
Differential Revision: D58710805
Pull Request resolved: https://github.com/pytorch/pytorch/pull/128922
Approved by: https://github.com/xw285cornell, https://github.com/drisspg
