Lookaside for torch.ops.higher_order.autograd_function_apply
#1256
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
t-vi
reviewed
Oct 3, 2024
crcrpar
force-pushed
the
crpa/lookaside_autograd-function-apply
branch
from
b4647ed
to
71db6cd
Compare
IvanYashchuk
reviewed
Oct 4, 2024
thunder/torch/__init__.py
Outdated
There was a problem hiding this comment.
Can we keep this code? I think it does a good job at separation of concerns. The job if "jit_ext" is to make things Thunder friendly.
There was a problem hiding this comment.
Then it feels like we should have a better utility to make a callable friendly to Thunder before doing what this PR is trying to do
There was a problem hiding this comment.
You have already done the complicated part of making callables friendly to Thunder. What I mean is that it's now possible to remove a bit of complexity of registering a grad rule by using the functions that were removed:
diff --git a/thunder/core/jit_ext.py b/thunder/core/jit_ext.py
index d98836a2..9d390e7d 100644
--- a/thunder/core/jit_ext.py
+++ b/thunder/core/jit_ext.py
@@ -864,9 +864,18 @@ def _general_jit_torch_ops_higher_order_autograd_function_apply(fwd, bwd, *fwd_a
bwd_trace._siginfo = SigInfo.from_name_and_args(f"bwd_{sym_id}", saved_values + grads)
@wraps(bwd_trace.python_callable())
- def bwd_impl_callable(*args, **kwargs):
+ def bwd_impl_callable(ctx, *args, **kwargs):
return thunder.core.trace_interpreter.interpret_trace(bwd_trace, *args, **kwargs)
+ @wraps(augmented_fwd_trace.python_callable())
+ def fwd_impl_callable(ctx, *args, **kwargs):
+ return thunder.core.trace_interpreter.interpret_trace(augmented_fwd_trace, *args, **kwargs)
+
+ from thunder.torch import autograd_function_apply
+ wrapped_fwd = wrap_const(fwd_impl_callable)
+ wrapped_bwd = wrap_const(bwd_impl_callable)
+ return interpreter_needs_wrap(autograd_function_apply)(wrapped_fwd, wrapped_bwd, *fwd_args, **fwd_kwargs)
+
@wraps(core_of_fwd)
def grad_transform(*args, **kwargs):
from thunder.core.transforms import get_grad, put_grads
crcrpar
added a commit
that referenced
this pull request
Nov 13, 2024
Signed-off-by: Masaki Kozuki <[email protected]>
crcrpar
force-pushed
the
crpa/lookaside_autograd-function-apply
branch
from
627845d
to
f349308
Compare
crcrpar
added a commit
that referenced
this pull request
Nov 13, 2024
Signed-off-by: Masaki Kozuki <[email protected]>
crcrpar
force-pushed
the
crpa/lookaside_autograd-function-apply
branch
2 times, most recently
from
9c51ae2
to
94c3409
Compare
crcrpar
added a commit
that referenced
this pull request
Nov 13, 2024
Signed-off-by: Masaki Kozuki <[email protected]>
crcrpar
force-pushed
the
crpa/lookaside_autograd-function-apply
branch
from
c5621cd
to
dd702f5
Compare
crcrpar
added a commit
that referenced
this pull request
Nov 14, 2024
Signed-off-by: Masaki Kozuki <[email protected]>
crcrpar
force-pushed
the
crpa/lookaside_autograd-function-apply
branch
from
dd702f5
to
1b85a21
Compare
thunder/core/jit_ext.py
Outdated
| from thunder.core import utils | ||
| from thunder.core.baseutils import sequencify | ||
| from thunder.core.pytree import tree_flatten, tree_map | ||
| from thunder.core.transforms import VJPDual, augmented_forward_impls, backward_impls |
There was a problem hiding this comment.
These imports are unused.
thunder/core/jit_ext.py
Outdated
| if p in producer_map: | ||
| prod_bsym = producer_map[p] | ||
| tensor_to_prod_bsym[variableify(p)] = prod_bsym | ||
| prod_bsym_to_tensor = {v: unvariableify(k) for k, v in tensor_to_prod_bsym.items()} |
There was a problem hiding this comment.
This variable is unused.
thunder/core/jit_ext.py
Outdated
| return aug_fwd_trace | ||
| aug_fwd_result = aug_fwd_trace.output | ||
| output, saved_values = unwrap(aug_fwd_result) | ||
| wrapped_output = wrap(output, provenance=aug_fwd_provenance) |
There was a problem hiding this comment.
This variable is unused.
thunder/torch/__init__.py
Outdated
There was a problem hiding this comment.
You have already done the complicated part of making callables friendly to Thunder. What I mean is that it's now possible to remove a bit of complexity of registering a grad rule by using the functions that were removed:
diff --git a/thunder/core/jit_ext.py b/thunder/core/jit_ext.py
index d98836a2..9d390e7d 100644
--- a/thunder/core/jit_ext.py
+++ b/thunder/core/jit_ext.py
@@ -864,9 +864,18 @@ def _general_jit_torch_ops_higher_order_autograd_function_apply(fwd, bwd, *fwd_a
bwd_trace._siginfo = SigInfo.from_name_and_args(f"bwd_{sym_id}", saved_values + grads)
@wraps(bwd_trace.python_callable())
- def bwd_impl_callable(*args, **kwargs):
+ def bwd_impl_callable(ctx, *args, **kwargs):
return thunder.core.trace_interpreter.interpret_trace(bwd_trace, *args, **kwargs)
+ @wraps(augmented_fwd_trace.python_callable())
+ def fwd_impl_callable(ctx, *args, **kwargs):
+ return thunder.core.trace_interpreter.interpret_trace(augmented_fwd_trace, *args, **kwargs)
+
+ from thunder.torch import autograd_function_apply
+ wrapped_fwd = wrap_const(fwd_impl_callable)
+ wrapped_bwd = wrap_const(bwd_impl_callable)
+ return interpreter_needs_wrap(autograd_function_apply)(wrapped_fwd, wrapped_bwd, *fwd_args, **fwd_kwargs)
+
@wraps(core_of_fwd)
def grad_transform(*args, **kwargs):
from thunder.core.transforms import get_grad, put_grads
thunder/core/jit_ext.py
Outdated
Comment on lines
870
to
871
| @wraps(core_of_fwd) | ||
| def grad_transform(*args, **kwargs): |
There was a problem hiding this comment.
This part can be removed if we reuse existing functions and use interpreter_needs_wrap(autograd_function_apply) here.
thunder/torch/__init__.py
Outdated
There was a problem hiding this comment.
Let's not remove this code and use it inside the lookaside.
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
force-pushed
the
crpa/lookaside_autograd-function-apply
branch
from
1b85a21
to
7729af1
Compare
4 tasks
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?
As per #1248, the support of
torch.ops.higher_order.autograd_function_applywould be a bit more flexible by tracing into bothfwdandbwd.