Reduce torch eager dispatch overhead

[MarcosAsh]

Summary

Reduces Keras[torch] eager dispatch overhead by avoiding unnecessary allocations and indirection in the hot path.

Addresses #22561.

Changes

• any_symbolic_tensors: iterate args directly instead of tree.flatten
• convert_to_tensor: fast path for already-correct torch tensors
• Operation.call: inline error handling, skip wrapper creation per call
• Layer.call: fast path for single-tensor eager inference (bypasses CallSpec, input validation, mask population, autocast scoping)
• _set_mask_metadata/ mask population: skip tree ops for single tensors
• assert_input_compatibility: skip tree.flatten for single-spec case

Benchmark

Colab notebook on T4
GPU:

Metric	PyTorch	master	This PR	Improvement
add (256x256)	9.0us	51.5us (5.71x)	16.3us (1.81x)	3.2x faster
matmul (256x256)	17.9us	90.3us (5.04x)	43.9us (2.45x)	2.1x faster
softmax (64x1000)	8.8us	51.8us (5.91x)	17.8us (1.95x)	2.9x faster
CNN forward	232.1us	3099.7us (13.35x)	1896.7us (8.50x)	1.6x faster
Dense call overhead	--	263.2us	46.8us	5.6x less overhead

The remaining CNN gap includes Conv2D's NHWC/NCHW data format conversion overhead, which is a separate problem (#18457).

[gemini-code-assist]

Code Review

This pull request implements performance optimizations across Keras by introducing fast paths for common execution scenarios, such as single-tensor inputs and eager inference, to reduce overhead from tree.flatten and traceback wrappers. Key changes include optimized symbolic tensor detection, faster tensor conversion in the Torch backend, and streamlined layer call logic. A bug was identified in the new error-handling mechanism in operation.py that would cause the loss of the original traceback during exception injection, and a fix was suggested to ensure proper debugging information is preserved.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 58.78378% with 61 lines in your changes missing coverage. Please review.
✅ Project coverage is 83.19%. Comparing base (9708582) to head (8f00faa).
⚠️ Report is 13 commits behind head on master.

Files with missing lines	Patch %	Lines
keras/src/utils/traceback_utils.py	3.03%	32 Missing ⚠️
keras/src/layers/input_spec.py	54.54%	13 Missing and 7 partials ⚠️
keras/src/ops/operation.py	28.57%	5 Missing ⚠️
keras/src/backend/common/keras_tensor.py	81.25%	1 Missing and 2 partials ⚠️
keras/src/backend/torch/core.py	90.90%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #22616      +/-   ##
==========================================
- Coverage   83.29%   83.19%   -0.10%     
==========================================
  Files         596      596              
  Lines       68138    68311     +173     
  Branches    10613    10694      +81     
==========================================
+ Hits        56754    56834      +80     
- Misses       8638     8683      +45     
- Partials     2746     2794      +48     

Flag	Coverage Δ
keras	83.01% <58.78%> (-0.10%)	⬇️
keras-jax	59.60% <52.02%> (-0.07%)	⬇️
keras-numpy	55.24% <51.35%> (-0.10%)	⬇️
keras-openvino	53.29% <49.32%> (-0.10%)	⬇️
keras-tensorflow	60.95% <52.02%> (-0.09%)	⬇️
keras-torch	59.76% <58.78%> (-0.09%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[MarcosAsh]

/gemini review

[gemini-code-assist]

Code Review

This pull request introduces several performance optimizations across the Keras backend and layer infrastructure, primarily by implementing fast paths for common eager execution scenarios. These changes include optimized tensor checks in any_symbolic_tensors, a fast path for convert_to_tensor in the Torch backend, streamlined input compatibility checks, and a significant reduction in overhead for layer calls and mask metadata handling. Additionally, the pull request refactors traceback injection to avoid creating wrapper functions per call, instead injecting argument information only when an exception occurs. I have provided a high-severity suggestion to remove an unnecessary boolean cast in the layer call fast path to prevent potential device-to-host synchronization in PyTorch.

[MarcosAsh]

I think this was too large a PR so I am doing them in more digestible smaller PRs.