Gpt2safetensors #2459

LakshmiKalaKadali · 2025-11-24T12:19:43Z

GPT2 model conversion from keras to hf safetensors format. colab gist

gemini-code-assist · 2025-11-24T12:20:08Z

Summary of Changes

Hello @LakshmiKalaKadali, 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 significantly extends the model export capabilities by enabling the conversion of Keras GPT-2 models to the Hugging Face Safetensors format. This allows for seamless interoperability between Keras-based GPT-2 implementations and the broader Hugging Face ecosystem, facilitating easier model sharing and deployment. The changes involve adding specific conversion logic for GPT-2's architecture and tokenizer, integrating this into the existing export framework, and providing a robust testing mechanism to ensure the fidelity of the converted models.

Highlights

GPT-2 Safetensors Export: Introduced comprehensive functionality to convert Keras GPT-2 models, including their configurations and weights, into the Hugging Face Safetensors format.
Tokenizer Asset Saving: Added a save_assets method to GPT2Tokenizer for proper serialization of vocabulary and merge files, crucial for Hugging Face compatibility.
Hugging Face Exporter Integration: Updated the general hf_exporter.py utility to recognize and correctly handle GPT-2 models and their tokenizers, including specific logic for tied weights and different tokenizer asset formats.
Conversion Script Enhancement: Modified the convert_gpt2_checkpoints.py script to incorporate the new Safetensors export capability, allowing users to generate Hugging Face compatible GPT-2 models directly.
Robust Testing: Included a dedicated test suite (gpt2_test.py) to validate the accuracy of the GPT-2 model conversion, comparing Keras and Hugging Face model configurations and output logits.

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 adds functionality to convert Keras-based GPT-2 models to the Hugging Face safetensors format. This involves a new export script for GPT-2, updates to the generic Hugging Face exporter, and modifications to the GPT-2 tokenizer and checkpoint conversion script. The changes are generally well-structured, but there are some critical issues. I've identified violations of the backend-agnostic principle, a potential runtime error in the exporter logic, and some maintainability concerns. My feedback includes suggestions to use keras.ops for backend-agnostic tensor operations, fix a critical bug, and improve code style and dependency management.

keras_hub/src/utils/transformers/export/gpt2.py

gemini-code-assist · 2025-11-24T12:22:57Z

keras_hub/src/utils/transformers/export/hf_exporter.py

+        if (
+            "lm_head.weight" in weights_dict_torch
+            and "transformer.wte.weight" in weights_dict_torch
+        ):
+            wte = weights_dict_torch["transformer.wte.weight"]
+            lm = weights_dict_torch["lm_head.weight"]
+
+        if wte.data_ptr() == lm.data_ptr():
+            weights_dict_torch["lm_head.weight"] = lm.clone().contiguous()


There is a potential NameError here. The variables wte and lm are defined within an if block but used in a subsequent, un-nested if block on line 106. If the condition on line 99 is false, wte and lm will not be defined, causing a runtime error. The second if statement should be nested inside the first.

Suggested change

if (

"lm_head.weight" in weights_dict_torch

and "transformer.wte.weight" in weights_dict_torch

):

wte = weights_dict_torch["transformer.wte.weight"]

lm = weights_dict_torch["lm_head.weight"]

if wte.data_ptr() == lm.data_ptr():

weights_dict_torch["lm_head.weight"] = lm.clone().contiguous()

if (

"lm_head.weight" in weights_dict_torch

and "transformer.wte.weight" in weights_dict_torch

):

wte = weights_dict_torch["transformer.wte.weight"]

lm = weights_dict_torch["lm_head.weight"]

if wte.data_ptr() == lm.data_ptr():

weights_dict_torch["lm_head.weight"] = lm.clone().contiguous()

keras_hub/src/utils/transformers/export/gpt2_test.py

tools/checkpoint_conversion/convert_gpt2_checkpoints.py

gemini-code-assist · 2025-11-24T12:22:58Z

tools/checkpoint_conversion/convert_gpt2_checkpoints.py

    num_params = PRESET_MAP[FLAGS.preset][0]
    hf_model_name = PRESET_MAP[FLAGS.preset][1]

+    os.system("pip install requests")


Running pip install from within a script is not a recommended practice for managing dependencies. It can lead to an unpredictable environment and conflicts. Project dependencies should be managed declaratively, for example, in a requirements.txt file.

gemini-code-assist · 2025-11-24T12:22:58Z

keras_hub/src/models/gpt2/gpt2_tokenizer.py

+        import json
+        import os
+        import shutil


For better readability and consistency, imports should be at the top of the file. Please move import json, import os, and import shutil to the top-level of the module to follow standard Python conventions.

gemini-code-assist · 2025-11-24T12:22:58Z

keras_hub/src/utils/transformers/export/gpt2.py

+        q_w = keras_model.get_layer(
+            f"transformer_layer_{i}"
+        )._self_attention_layer._query_dense.kernel
+        k_w = keras_model.get_layer(
+            f"transformer_layer_{i}"
+        )._self_attention_layer._key_dense.kernel
+        v_w = keras_model.get_layer(
+            f"transformer_layer_{i}"
+        )._self_attention_layer._value_dense.kernel


Accessing private layer attributes like _self_attention_layer and its sub-layers makes this code brittle. If the internal structure of TransformerDecoder changes, this export script will break. It would be more robust to expose these weights via a public API on the layer to create a more stable interface.

sachinprasadhs · 2025-11-26T00:06:42Z

Resolve the gemini suggested changes and mark them as resolved for the respective comments, make the changes as backend agnostic, not specific to tf.
Once these comments are addressed, I will go thorough the files in detail.

LakshmiKalaKadali · 2025-11-26T01:23:59Z

Sure Sachin, I am working on it. Thank You

…

On Wed, Nov 26, 2025 at 5:37 AM Sachin Prasad ***@***.***> wrote: *sachinprasadhs* left a comment (keras-team/keras-hub#2459) <#2459 (comment)> Resolve the gemini suggested changes and mark them as resolved for the respective comments, make the changes as backend agnostic, not specific to tf. Once these comments are addressed, I will go thorough the files in detail. — Reply to this email directly, view it on GitHub <#2459 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BDVX3HLEK7VQA2OGD4ODON336TVKRAVCNFSM6AAAAACNAUC4WOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTKNZYGEZTOMJYGQ> . You are receiving this because you were mentioned.Message ID: ***@***.***>

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

LakshmiKalaKadali added 2 commits November 24, 2025 16:40

GPT2 kerashub to hf safetensor conversion

5add1a3

GPT2 kerashub to hf safe tensor conversion

7146a5c

gemini-code-assist bot reviewed Nov 24, 2025

View reviewed changes

sachinprasadhs added the stat:awaiting response from contributor label Nov 26, 2025

LakshmiKalaKadali and others added 5 commits November 26, 2025 09:38

Update keras_hub/src/utils/transformers/export/gpt2.py

a8b6d07

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

Update keras_hub/src/utils/transformers/export/gpt2.py

3c59953

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

Update keras_hub/src/utils/transformers/export/gpt2_test.py

adcc364

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

Update keras_hub/src/utils/transformers/export/gpt2_test.py

d8e71af

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

Update tools/checkpoint_conversion/convert_gpt2_checkpoints.py

45e2e9d

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Gpt2safetensors #2459

Gpt2safetensors #2459

LakshmiKalaKadali commented Nov 24, 2025

Uh oh!

gemini-code-assist bot commented Nov 24, 2025

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

Uh oh!

Uh oh!

gemini-code-assist bot Nov 24, 2025

Uh oh!

Uh oh!

Uh oh!

Uh oh!

gemini-code-assist bot Nov 24, 2025

Uh oh!

gemini-code-assist bot Nov 24, 2025

Uh oh!

gemini-code-assist bot Nov 24, 2025

Uh oh!

sachinprasadhs commented Nov 26, 2025

Uh oh!

LakshmiKalaKadali commented Nov 26, 2025 via email

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Gpt2safetensors #2459

Are you sure you want to change the base?

Gpt2safetensors #2459

Conversation

LakshmiKalaKadali commented Nov 24, 2025

Uh oh!

gemini-code-assist bot commented Nov 24, 2025

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!

Uh oh!

Uh oh!

gemini-code-assist bot Nov 24, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

gemini-code-assist bot Nov 24, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Nov 24, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Nov 24, 2025

Choose a reason for hiding this comment

Uh oh!

sachinprasadhs commented Nov 26, 2025

Uh oh!

LakshmiKalaKadali commented Nov 26, 2025 via email

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants