Unrolling tensor subclasses in fwd/bwd split #1489
Merged
+71
−27
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
Signed-off-by: Masaki Kozuki <[email protected]>
Signed-off-by: Masaki Kozuki <[email protected]>
Signed-off-by: Masaki Kozuki <[email protected]>
Signed-off-by: Masaki Kozuki <[email protected]>
Signed-off-by: Masaki Kozuki <[email protected]>
Signed-off-by: Masaki Kozuki <[email protected]>
crcrpar
commented
Nov 28, 2024
| @@ -637,7 +637,7 @@ def _convert_pytorchfunc_to_thundertrace( | |||
| trace = TraceCtx() | |||
| trace.bound_symbols.extend(active_jit_ctx.computation_trace.pop_scope()) | |||
| func_result = unwrap(wrapped_func_result) | |||
| if shallow_copy_output: | |||
| if shallow_copy_output and not trace.bound_symbols: | |||
Comment on lines
+1317
to
+1325
| def _transpose_grad(a: TensorLike, /, dim0: int, dim1: int) -> TensorLike: | ||
| fwd = transpose(a, dim0, dim1) | ||
| g = get_grad(fwd) | ||
| a_grad = transpose(g, dim0, dim1) | ||
| put_grad(a, a_grad) | ||
| return fwd | ||
|
|
||
|
|
||
| register_grad(transpose, _transpose_grad) |
There was a problem hiding this comment.
rel: #1487
needed to avoid prims.permute
Signed-off-by: Masaki Kozuki <[email protected]>
Signed-off-by: Masaki Kozuki <[email protected]>
for more information, see https://pre-commit.ci
crcrpar
commented
Nov 28, 2024
| @@ -269,3 +266,5 @@ def test_torchao_float8_linear(executor, device, _): | |||
|
|
|||
| jitted = executor.make_callable(fp8_model) | |||
| actual = jitted(x) | |||
|
|
|||
| torch.testing.assert_close(actual, expected) | |||
There was a problem hiding this comment.
Signed-off-by: Masaki Kozuki <[email protected]>
for more information, see https://pre-commit.ci
crcrpar
merged commit 06ee30e
into
crpa/subclass-torchao_float8tensor
29 of 36 checks passed
crcrpar
deleted the
crpa/torchao-fp8tensor-flattening-in-fwdbwd-split
branch
crcrpar
added a commit
that referenced
this pull request
Nov 28, 2024
Signed-off-by: Masaki Kozuki <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
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?
In #1415 and #1394, tensor subclasses and their
__torch_dispatch__are unrolled before forward-backward split.It turned out that we want to postpone it in the split as the unrolling seems to be harmful to backward generation.
TODO
note: pytorch/ao#1339 is used