[tensor wrapper subclass] Add support for torchao.float8 mlp #1585
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
crcrpar
force-pushed
the
tensor_subclass_3
branch
from
b4ecee4 to
cafa119
Compare
github-actions
bot
added
the
documentation
crcrpar
force-pushed
the
tensor_subclass_3
branch
from
cafa119 to
5afba28
Compare
crcrpar
force-pushed
the
tensor_subclass_2
branch
from
515d425 to
2e763f1
Compare
crcrpar
force-pushed
the
tensor_subclass_3
branch
from
5afba28 to
6e6b077
Compare
crcrpar
force-pushed
the
tensor_subclass_2
branch
from
2e763f1 to
a534003
Compare
crcrpar
force-pushed
the
tensor_subclass_3
branch
from
6e6b077 to
d9ed305
Compare
crcrpar
force-pushed
the
tensor_subclass_2
branch
from
e22488f to
db3a6c9
Compare
- Add `scaled_mm`
- Change how the lookaside of `torch.autograd.Function.apply` applies dce
taking the failure of apex fused rms norm into consideration.
```python
@torch.no_grad()
@no_autocast
def FusedRMSNormAffineMixedDtypesFunction(t_0, t_1, tup11, f12, b13):
# /usr/local/lib/python3.12/dist-packages/apex/normalization/fused_layer_norm.py:128: weight_ = weight.contiguous()
# t_0: "cuda:0 f32[4, 5, 3, 2]"
# t_1: "cuda:0 f32[3, 2]"
# /usr/local/lib/python3.12/dist-packages/apex/normalization/fused_layer_norm.py:127: input_ = input.contiguous()
t5 = ltorch.contiguous(t_0, memory_format=_torch_memory_format_0) # t5: "cuda:0 f32[4, 5, 3, 2]"
# t5 = prims.stride_order(t_0, (3, 2, 1, 0)) # t5: "cuda:0 f32[4, 5, 3, 2]"
# /usr/local/lib/python3.12/dist-packages/apex/normalization/fused_layer_norm.py:128: weight_ = weight.contiguous()
t6 = ltorch.contiguous(t_1, memory_format=_torch_memory_format_0) # t6: "cuda:0 f32[3, 2]"
# t6 = prims.stride_order(t_1, (1, 0)) # t6: "cuda:0 f32[3, 2]"
(t10, t9) = apex_fused_rms_norm_forward_affine_mixed_dtypes(t5, (3, 2), t6, 1e-05)
return t10
```
For this trace, `thunder.core.transforms.dce` replaces `t9` with `_`
then the augmented forward trace would lose the access to it. So by
reusing the augmented forward trace in the basic forward trace, `dce`
would not do so.
Signed-off-by: Masaki Kozuki <[email protected]>
crcrpar
force-pushed
the
tensor_subclass_3
branch
from
d9ed305 to
8435406
Compare
github-actions
bot
removed
the
documentation
|
needs to fix the backward of torchao.float8 in tests. The cause seems to be the mismatch or row-major or column-major of the inputs to |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Multiple changes for thunder.jit to support a torchao.float8 MLP (see the test):
torch._scaled_mm_general_jit_torch_autograd_function_apply_lookaside