[FRONTEND][TFLITE] TFLite frontend ported from Relay #18685

srkreddy1238 · 2026-01-23T11:04:41Z

Verified for entire range of classification nets
Quantization is disabled at the moment
There exists few unspoorted ops in convertion maps which is need to be mapped in future when relax op inventory grows.

Verified for entire range of classification nets Quantization is disabled at the moment There exists few unspoorted ops in convertion maps which is need to be mapped in future when relax op inventory grows.

gemini-code-assist · 2026-01-23T11:04:58Z

Summary of Changes

Hello @srkreddy1238, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request integrates a TensorFlow Lite (TFLite) frontend into the TVM Relax framework, allowing for the conversion and execution of TFLite models. The implementation includes a specialized FlexBuffer parser to handle TFLite's custom options and is validated by a broad range of tests, ensuring compatibility with common TensorFlow operations and popular Keras classification models. This lays the groundwork for broader TFLite model support in Relax.

Highlights

New TFLite Frontend: Introduced a new frontend for TensorFlow Lite (TFLite) within the TVM Relax framework, enabling the conversion of TFLite models into Relax graphs.
FlexBuffer Decoder: Implemented a partial FlexBuffer decoder to correctly parse custom options embedded within TFLite models.
Comprehensive Test Coverage: Added an extensive test suite covering various TensorFlow operations (e.g., binary, element-wise, logical, reduce, conv2d, pooling) and end-to-end validation with several Keras classification networks.
Current Limitations: Quantization support is currently disabled, and a few TFLite operations are not yet mapped, pending future expansion of the Relax op inventory.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

gemini-code-assist

Code Review

This pull request introduces the TFLite frontend for Relax, enabling conversion of TFLite graphs into Relax graphs. It includes a Flexbuffer decoder for custom options and a comprehensive set of tests for various TensorFlow operations and Keras applications. While the overall structure is good, there are several critical issues related to missing imports, incorrect handling of data types, and potential bugs in the Flexbuffer deserialization and quantized operations that need to be addressed before merging. Additionally, some security and maintainability concerns were identified.

gemini-code-assist · 2026-01-23T11:07:26Z

python/tvm/relax/frontend/tflite/tflite_flexbuffer.py

+            unpack_str = "<B"
+        elif byte_width == 4:
+            unpack_str = "<i"
+        assert unpack_str != ""


The indirect_jump function currently only handles byte_width values of 1 and 4. However, the BitWidth enum (lines 25-31) indicates that byte_width can also be 2 or 8 (derived from 1 << BitWidth(root_packed_type & 3)). This will lead to an assert unpack_str != "" failure or incorrect parsing for byte_width values of 2 or 8.

Suggested change

assert unpack_str != ""

unpack_str = "<B"

elif byte_width == 2:

unpack_str = "<H"

elif byte_width == 4:

unpack_str = "<i"

elif byte_width == 8:

unpack_str = "<q"

gemini-code-assist · 2026-01-23T11:07:26Z

python/tvm/relax/frontend/tflite/tflite_flexbuffer.py

+        root_end -= root_byte_width
+
+        root_type = FlexBufferType(root_packed_type >> 2)
+        byte_width = 1 << BitWidth(root_packed_type & 3)


The byte_width calculation 1 << BitWidth(root_packed_type & 3) can result in values 1, 2, 4, or 8. However, the indirect_jump function (lines 82-86) only handles byte_width values of 1 and 4. This discrepancy will cause incorrect behavior or errors when processing Flexbuffers with 2-byte or 8-byte widths.

gemini-code-assist · 2026-01-23T11:07:26Z

tests/python/relax/test_frontend_tflite.py

+
+import ssl
+
+ssl._create_default_https_context = ssl._create_unverified_context


Disabling SSL certificate verification globally using ssl._create_unverified_context is a security risk and generally discouraged. For testing purposes, consider using a more targeted approach, such as a context manager, or ensuring the test environment has proper certificates.

gemini-code-assist · 2026-01-23T11:07:26Z

python/tvm/relax/frontend/tflite/tflite_flexbuffer.py

+            elif value_type == FlexBufferType.FBT_FLOAT:
+                value = struct.unpack("<f", value_bytes)[0]
+            else:
+                raise Exception


Raising a generic Exception is not ideal. It's better to raise a more specific exception type (e.g., NotImplementedError or ValueError) with a descriptive message to help in debugging.

Suggested change

raise Exception

raise NotImplementedError(f"FlexBufferType {value_type} not supported for vector decoding.")

gemini-code-assist · 2026-01-23T11:07:26Z

tests/python/relax/test_frontend_tflite.py

+        # Limiting the tests for CI
+        (keras_app.Xception, (1, 299, 299, 3)),
+        # (keras_app.VGG16, (1, 224, 224, 3)),
+        # (keras_app.VGG19, (1, 224, 224, 3)),
+        (keras_app.ResNet50, (1, 224, 224, 3)),
+        # (keras_app.ResNet50V2, (1, 224, 224, 3)),
+        # (keras_app.ResNet101, (1, 224, 224, 3)),
+        # (keras_app.ResNet101V2, (1, 224, 224, 3)),
+        # (keras_app.ResNet152, (1, 224, 224, 3)),
+        # (keras_app.ResNet152V2, (1, 224, 224, 3)),
+        (keras_app.InceptionResNetV2, (1, 299, 299, 3)),
+        # (keras_app.MobileNet, (1, 224, 224, 3)),
+        (keras_app.MobileNetV2, (1, 224, 224, 3)),
+        (keras_app.DenseNet121, (1, 224, 224, 3)),
+        # (keras_app.DenseNet169, (1, 224, 224, 3)),
+        # (keras_app.DenseNet201, (1, 224, 224, 3)),
+        (keras_app.NASNetMobile, (1, 224, 224, 3)),
+        # (keras_app.NASNetLarge, (1, 331, 331, 3)),
+        (keras_app.EfficientNetB0, (1, 224, 224, 3)),
+        # (keras_app.EfficientNetB1, (1, 240, 240, 3)),
+        # (keras_app.EfficientNetB2, (1, 260, 260, 3)),
+        # (keras_app.EfficientNetB3, (1, 300, 300, 3)),
+        # (keras_app.EfficientNetB4, (1, 380, 380, 3)),
+        # (keras_app.EfficientNetB5, (1, 456, 456, 3)),
+        # (keras_app.EfficientNetB6, (1, 528, 528, 3)),
+        # (keras_app.EfficientNetB7, (1, 600, 600, 3)),
+        (keras_app.EfficientNetV2B0, (1, 224, 224, 3)),
+        # (keras_app.EfficientNetV2B1, (1, 240, 240, 3)),
+        # (keras_app.EfficientNetV2B2, (1, 260, 260, 3)),
+        # (keras_app.EfficientNetV2B3, (1, 300, 300, 3)),
+        # (keras_app.EfficientNetV2S, (1, 384, 384, 3)),
+        # (keras_app.EfficientNetV2M, (1, 480, 480, 3)),
+        # (keras_app.EfficientNetV2L, (1, 480, 480, 3)),
+        (keras_app.ConvNeXtTiny, (1, 224, 224, 3)),
+        # (keras_app.ConvNeXtSmall, (1, 224, 224, 3)),
+        # (keras_app.ConvNeXtBase, (1, 224, 224, 3)),
+        # (keras_app.ConvNeXtLarge, (1, 224, 224, 3)),
+        # (keras_app.ConvNeXtXLarge, (1, 224, 224, 3)),
+    ],


Many network tests are commented out with the note "Limiting the tests for CI". While this might be necessary for CI resource constraints, it means a significant portion of the TFLite frontend's functionality for these Keras models is not being tested. This could lead to regressions or undetected issues. Consider enabling these tests for local development or in a less constrained CI environment, or adding a tracking issue to re-enable them.

[FRONTEND][TFLITE] TFLite frontend ported from Relax

c30dd7f

Verified for entire range of classification nets Quantization is disabled at the moment There exists few unspoorted ops in convertion maps which is need to be mapped in future when relax op inventory grows.

gemini-code-assist bot reviewed Jan 23, 2026

View reviewed changes

srkreddy1238 added 3 commits January 23, 2026 19:11

import

f315cea

GPU tests commented

ce3d5dd

CI tests

03918e8

srkreddy1238 changed the title ~~[FRONTEND][TFLITE] TFLite frontend ported from Relax~~ [FRONTEND][TFLITE] TFLite frontend ported from Relay Jan 27, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[FRONTEND][TFLITE] TFLite frontend ported from Relay #18685

[FRONTEND][TFLITE] TFLite frontend ported from Relay #18685

srkreddy1238 commented Jan 23, 2026

Uh oh!

gemini-code-assist bot commented Jan 23, 2026

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

gemini-code-assist bot Jan 23, 2026

Uh oh!

gemini-code-assist bot Jan 23, 2026

Uh oh!

gemini-code-assist bot Jan 23, 2026

Uh oh!

gemini-code-assist bot Jan 23, 2026

Uh oh!

gemini-code-assist bot Jan 23, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant


		import ssl

		ssl._create_default_https_context = ssl._create_unverified_context

	raise Exception
	raise NotImplementedError(f"FlexBufferType {value_type} not supported for vector decoding.")

[FRONTEND][TFLITE] TFLite frontend ported from Relay #18685

Are you sure you want to change the base?

[FRONTEND][TFLITE] TFLite frontend ported from Relay #18685

Conversation

srkreddy1238 commented Jan 23, 2026

Uh oh!

gemini-code-assist bot commented Jan 23, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Jan 23, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 23, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 23, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 23, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 23, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant