Custom/gradients dispatch #632
+475
−11
There are no files selected for viewing
This file contains hidden or 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
This file contains hidden or 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
This file contains hidden or 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
This file contains hidden or 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
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,145 @@ | ||
| /* Copyright 2026 The TensorFlow Authors. All Rights Reserved. | ||
|
|
||
| Licensed under the Apache License, Version 2.0 (the "License"); | ||
| you may not use this file except in compliance with the License. | ||
| You may obtain a copy of the License at | ||
|
|
||
| http://www.apache.org/licenses/LICENSE-2.0 | ||
|
|
||
| Unless required by applicable law or agreed to in writing, software | ||
| distributed under the License is distributed on an "AS IS" BASIS, | ||
| WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| See the License for the specific language governing permissions and | ||
| limitations under the License. | ||
| ======================================================================= | ||
| */ | ||
| package org.tensorflow.op; | ||
|
|
||
| import java.lang.reflect.Constructor; | ||
| import java.util.List; | ||
| import java.util.concurrent.ConcurrentHashMap; | ||
| import java.util.concurrent.ConcurrentMap; | ||
| import org.tensorflow.AbstractGradientAdapter; | ||
| import org.tensorflow.Graph; | ||
| import org.tensorflow.GraphOperation; | ||
| import org.tensorflow.Operand; | ||
| import org.tensorflow.Output; | ||
| import org.tensorflow.internal.c_api.TFJ_Scope; | ||
|
|
||
| /** | ||
| * Dispatching adapter for Java-side custom gradient registration. | ||
| * | ||
| * <p>This class mirrors the behavior of TensorFlow Python's {@code tf.RegisterGradient} mechanism | ||
| * by providing a centralized dispatch layer for custom gradients in the Java API. | ||
| * | ||
| * <p>Gradients may be registered in one of two forms for a given op type: | ||
| * | ||
| * <ul> | ||
| * <li>A raw gradient ({@link RawCustomGradient}) operating directly on {@link GraphOperation} and | ||
| * {@link Output} objects. | ||
| * <li>A typed gradient ({@link CustomGradient}) operating on generated {@link RawOpInputs} | ||
| * subclasses. | ||
| * </ul> | ||
| * | ||
| * <p>For any given op type, exactly one gradient definition is permitted: either raw or typed. | ||
| * Duplicate registrations, or attempts to mix raw and typed gradients for the same op type, are | ||
| * rejected to prevent ambiguous dispatch behavior. | ||
| * | ||
| * <p>At runtime, {@link #apply(Graph, TFJ_Scope, GraphOperation, List)} determines the operation | ||
| * type and dispatches to the corresponding registered gradient implementation. | ||
| */ | ||
| final class DispatchingGradientAdapter extends AbstractGradientAdapter { | ||
|
|
||
| private final ConcurrentMap<String, RawCustomGradient> raw = new ConcurrentHashMap<>(); | ||
| private final ConcurrentMap<String, TypedEntry<?>> typed = new ConcurrentHashMap<>(); | ||
|
|
||
| private static String dupMsg(String opType, String existingKind, String newKind) { | ||
| return "A " | ||
| + existingKind | ||
| + " gradient is already registered for op type '" | ||
| + opType | ||
| + "'. Raw and typed registrations are mutually exclusive; cannot register " | ||
| + newKind | ||
| + "."; | ||
| } | ||
|
|
||
| static final class TypedEntry<T extends RawOpInputs<?>> { | ||
| final CustomGradient<T> grad; | ||
| final Class<T> inputClass; | ||
| final Constructor<T> ctor; | ||
|
|
||
| TypedEntry(CustomGradient<T> grad, Class<T> inputClass) { | ||
| this.grad = grad; | ||
| this.inputClass = inputClass; | ||
| try { | ||
| this.ctor = inputClass.getConstructor(org.tensorflow.GraphOperation.class); | ||
| } catch (NoSuchMethodException e) { | ||
| throw new IllegalArgumentException( | ||
| "Inputs class " + inputClass.getName() + " must have a public ctor(GraphOperation).", | ||
| e); | ||
| } | ||
| } | ||
| } | ||
|
|
||
| void putRaw(String opType, RawCustomGradient g) { | ||
| if (typed.containsKey(opType)) { | ||
| throw new IllegalStateException(dupMsg(opType, "typed", "raw")); | ||
| } | ||
| RawCustomGradient prev = raw.putIfAbsent(opType, g); | ||
| if (prev != null) { | ||
| throw new IllegalStateException( | ||
| "A raw gradient is already registered for op type '" + opType + "'."); | ||
| } | ||
| } | ||
|
|
||
| <T extends RawOpInputs<?>> void putTyped( | ||
| String opType, CustomGradient<T> g, Class<T> inputClass) { | ||
| if (raw.containsKey(opType)) { | ||
| throw new IllegalStateException(dupMsg(opType, "raw", "typed")); | ||
| } | ||
| TypedEntry<?> prev = typed.putIfAbsent(opType, new TypedEntry<>(g, inputClass)); | ||
| if (prev != null) { | ||
| throw new IllegalStateException( | ||
| "A typed gradient is already registered for op type '" + opType + "'."); | ||
| } | ||
| } | ||
|
|
||
| @Override | ||
| protected List<Operand<?>> apply( | ||
| Graph graph, TFJ_Scope scope, GraphOperation operation, List<Output<?>> gradInputs) { | ||
|
|
||
| final String opType = operation.type(); | ||
|
|
||
| RawCustomGradient rg = raw.get(opType); | ||
|
Collaborator
There was a problem hiding this comment. This logic prefers raw gradients over typed ones, but there isn't anything documented about why it prefers them or if it makes sense to add both raw and typed gradients for the same op. It would be good to clarify this, and if it doesn't make sense to have both kinds of gradients the adapter should reject them in the puts.
Collaborator
There was a problem hiding this comment. Can you add javadoc to the top of this class noting the overall purpose of it (to provide Java side dispatching for gradients mirroring TF-Python), that it only accepts either raw or typed gradients for a given op, and that it rejects duplicate assignments. |
||
| if (rg != null) { | ||
| // NativeScope & Ops constructors are package-private => must be in org.tensorflow.op | ||
| Scope nativeScope = | ||
| new NativeScope(scope, graph, operation.name()).withSubScope(operation.name()); | ||
| return rg.call(new Ops(nativeScope), operation, gradInputs); | ||
| } | ||
|
|
||
| @SuppressWarnings("rawtypes") | ||
| TypedEntry te = typed.get(opType); | ||
| if (te != null) { | ||
| return applyTyped(graph, scope, operation, gradInputs, te); | ||
| } | ||
|
|
||
| throw new IllegalStateException("No Java custom gradient registered for op type: " + opType); | ||
| } | ||
|
|
||
| private <T extends RawOpInputs<?>> List<Operand<?>> applyTyped( | ||
| Graph graph, | ||
| TFJ_Scope scope, | ||
| GraphOperation operation, | ||
| List<Output<?>> gradInputs, | ||
| TypedEntry<T> te) { | ||
| try { | ||
| T inputs = te.ctor.newInstance(operation); | ||
| Scope nativeScope = | ||
| new NativeScope(scope, graph, operation.name()).withSubScope(operation.name()); | ||
| return te.grad.call(new Ops(nativeScope), inputs, gradInputs); | ||
| } catch (ReflectiveOperationException e) { | ||
| throw new RuntimeException("Failed to instantiate inputs for " + te.inputClass.getName(), e); | ||
| } | ||
| } | ||
| } | ||
This file contains hidden or 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
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,40 @@ | ||
| /* Copyright 2026 The TensorFlow Authors. All Rights Reserved. | ||
|
|
||
| Licensed under the Apache License, Version 2.0 (the "License"); | ||
| you may not use this file except in compliance with the License. | ||
| You may obtain a copy of the License at | ||
|
|
||
| http://www.apache.org/licenses/LICENSE-2.0 | ||
|
|
||
| Unless required by applicable law or agreed to in writing, software | ||
| distributed under the License is distributed on an "AS IS" BASIS, | ||
| WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| See the License for the specific language governing permissions and | ||
| limitations under the License. | ||
| ======================================================================= | ||
| */ | ||
| package org.tensorflow.op; | ||
|
|
||
| import org.tensorflow.internal.c_api.TFJ_GradFuncAdapter; | ||
|
|
||
| /** Public bridge to a single native gradient adapter. */ | ||
| public final class GradientDispatch { | ||
|
|
||
| // package-private adapter that can access NativeScope/Ops constructors | ||
| static final DispatchingGradientAdapter ADAPTER = new DispatchingGradientAdapter(); | ||
|
|
||
| private GradientDispatch() {} | ||
|
|
||
| public static TFJ_GradFuncAdapter adapter() { | ||
| return ADAPTER; | ||
| } | ||
|
|
||
| public static void putRaw(String opType, RawCustomGradient gradient) { | ||
| ADAPTER.putRaw(opType, gradient); | ||
| } | ||
|
|
||
| public static <T extends RawOpInputs<?>> void putTyped( | ||
| String opType, CustomGradient<T> gradient, Class<T> inputClass) { | ||
| ADAPTER.putTyped(opType, gradient, inputClass); | ||
| } | ||
| } |
This file contains hidden or 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
Oops, something went wrong.
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.
Uh oh!
There was an error while loading. Please reload this page.