feat(tools): HuggingFace config parser for transformer spec

[pradhyum6144]

Summary

• Adds tools/hf_parser.py that converts HuggingFace config.json into ModelPack transformer spec format using PR [WIP] feat: add model architecture configuration #111's vocabulary (issue Transformer specification and auto-generation method for the existing models #164)
• Supports 6 model families: Mistral (GQA), Mixtral (MoE), Qwen2 (attention bias), GPT-2 (MHA/LayerNorm), DeepSeek-V2 (MLA + mixed layers), and unknown models with NEEDS_REVIEW fallback
• Includes tools/hf_parser_test.py with 26 unit tests covering all supported architectures
• Builds on PR docs: Add HuggingFace field mapping for architecture spec #185's field mapping research with executable tooling — parses MLA attention fields (kv_lora_rank, q_lora_rank, qk_nope_head_dim), DeepSeek MoE routing params, and mixed layer detection

Features

Feature	Details
Attention types	MHA, GQA, MLA
FFN types	MLP (gated/ungated), MoE (with shared experts)
Normalization	RMSNorm, LayerNorm
Position embeddings	RoPE (with theta/scaling), learned
MLA fields	kv_lora_rank, q_lora_rank, qk_nope_head_dim, qk_rope_head_dim, v_head_dim
MoE routing	routed_scaling_factor, topk_method, norm_topk_prob
Mixed layers	first_k_dense_replace, moe_layer_freq
Unknown models	Fields marked NEEDS_REVIEW for human verification

Usage

# From HuggingFace model ID
python tools/hf_parser.py meta-llama/Meta-Llama-3-8B

# From local config.json
python tools/hf_parser.py --file path/to/config.json

# JSON output
python tools/hf_parser.py mistralai/Mistral-7B-v0.3 --format json

Test plan

• 26 unit tests pass (pytest tools/hf_parser_test.py -v)
• Tested against real HF configs: Mistral-7B, Mixtral-8x7B, Qwen2-7B, GPT-2, DeepSeek-V2-Lite
• Unknown model types produce NEEDS_REVIEW markers instead of crashing
• Head dim correctly derived when not explicit in config

Relates to #164

[gemini-code-assist]

Summary of Changes

Hello, 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 delivers a crucial utility for standardizing the representation of HuggingFace transformer model configurations. By providing a robust parser that translates diverse config.json structures into a unified ModelPack specification, it significantly streamlines the process of integrating and analyzing various large language models, ensuring consistency and reducing manual effort in model definition.

Highlights

• HuggingFace Config Parser: Introduced tools/hf_parser.py to convert HuggingFace config.json files into the ModelPack transformer specification format, aligning with PR [WIP] feat: add model architecture configuration #111's vocabulary.
• Broad Model Support: Implemented support for six distinct model families, including Mistral (GQA), Mixtral (MoE), Qwen2 (attention bias), GPT-2 (MHA/LayerNorm), and DeepSeek-V2 (MLA + mixed layers), with a NEEDS_REVIEW fallback for unknown models.
• Comprehensive Testing: Included tools/hf_parser_test.py with 26 unit tests to ensure accurate parsing across all supported architectures and edge cases.
• Advanced Field Parsing: Enhanced parsing capabilities to correctly extract MLA attention fields (kv_lora_rank, q_lora_rank, qk_nope_head_dim), DeepSeek MoE routing parameters, and mixed layer detection, building on prior research.

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.

Footnotes

1. 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 a valuable tool for converting HuggingFace model configurations into the ModelPack transformer specification format. The implementation is well-structured and covers a good range of model architectures and their specific parameters. The inclusion of a comprehensive test suite is also a great practice. I've identified a bug in the YAML output generation that could lead to invalid output, a minor documentation inconsistency, and an opportunity for a small refactoring to improve maintainability. Overall, this is a solid contribution.

[gemini-code-assist]

The current string formatting for YAML output does not escape special characters like double quotes within the string value. This can result in invalid YAML. For example, a value of a "b" c would be rendered as key: "a "b" c", which is not valid. You can fix this by using json.dumps() which correctly handles string escaping for JSON, and the output is also valid for YAML.

Suggested change

	lines.append(f'{prefix}{key}: "{value}"')
	lines.append(f'{prefix}{key}: {json.dumps(value)}')

[gemini-code-assist]

The usage example in the docstring mentions a --file flag, but the script's argument parser is implemented to accept a positional argument for both model IDs and file paths. This example should be updated to reflect the actual implementation for clarity.

Suggested change

	python tools/hf_parser.py --file path/to/config.json
	python tools/hf_parser.py path/to/config.json

[gemini-code-assist]

For checking membership, using a set is more idiomatic and performant than a tuple, especially as the list of models grows. I suggest converting this tuple to a set. For even better maintainability, you could define this as a module-level constant.

Suggested change

	use_gated = model_type in (
	"llama", "mistral", "mixtral", "qwen2", "qwen2_moe", "phi3",
	"gemma", "gemma2", "deepseek_v2", "deepseek_v3",
	)
	use_gated = model_type in {
	"llama", "mistral", "mixtral", "qwen2", "qwen2_moe", "phi3",
	"gemma", "gemma2", "deepseek_v2", "deepseek_v3",
	}