Llama-4 mapping

[danielhanchen]

Make sure to read the contributing guidelines before submitting a PR

Summary by CodeRabbit

• New Features
  • Enhanced the mapping configuration to support a new generation model architecture (llama4), expanding overall system versatility.
  • Revised mapping definitions now improve compatibility with a broader range of model components.
  • These updates contribute to smoother performance and more robust integration for advanced model variations.
  • The improvements lay a solid foundation for future enhancements and increased adaptability.

[coderabbitai]

Walkthrough

The pull request extends the tensor mapping configuration in the gguf-py/gguf/tensor_mapping.py file. It adds several new tensor mappings to the TensorNameMap class, updating both the general mappings (mappings_cfg) and block-specific mappings (block_mappings_cfg). These changes are introduced to support additional tensor names for the "llama4" model architecture.

Changes

File

Change Summary

gguf-py/gguf/tensor_mapping.py

- Added new tensor mappings ("language_model.model.embed_tokens", "language_model.lm_head", "language_model.model.norm") in TensorNameMap.mappings_cfg. - Added multiple new block mappings (e.g., input_layernorm, self_attn.q_proj, self_attn.k_proj, self_attn.v_proj, self_attn.o_proj, post_attention_layernorm, and various feed_forward configurations) in TensorNameMap.block_mappings_cfg for supporting the "llama4" model architecture.

Sequence Diagram(s)

sequenceDiagram
  participant U as User
  participant TM as TensorNameMap

  U->>TM: Request tensor mapping for a given key
  alt Lookup in mappings_cfg
    TM->>TM: Retrieve mapping (embed_tokens, lm_head, norm)
  end
  alt Lookup in block_mappings_cfg
    TM->>TM: Retrieve mapping (layers: input_layernorm, self_attn, feed_forward, etc.)
  end
  TM-->>U: Return the mapped configuration

Loading

Poem

I'm a rabbit in the code so bright,
Hopping through mappings by day and night.
New tensor names now dance in the air,
With llama4 support, they shine with flair!
Crunching carrots of code with every bite,
Happy hops as the changes take flight.
🥕✨

---

📜 Recent review details

Configuration used: CodeRabbit UI
Review profile: CHILL
Plan: Pro

📥 Commits

Reviewing files that changed from the base of the PR and between 656babd and b14a449.

📒 Files selected for processing (1)

• gguf-py/gguf/tensor_mapping.py (16 hunks)

🧰 Additional context used

🧬 Code Definitions (1)

gguf-py/gguf/tensor_mapping.py (1)

gguf-py/gguf/constants.py (1)

• MODEL_TENSOR (295-431)

🔇 Additional comments (5)

gguf-py/gguf/tensor_mapping.py (5)

33-33: Tensor mapping for Llama4 token embeddings looks good

The tensor mapping added for Llama4's token embeddings follows the appropriate pattern with the "language_model.model.embed_tokens" path, which is distinct from but similar to other Llama variants.

---

71-71: Tensor mapping for Llama4 output layer looks good

The "language_model.lm_head" tensor mapping for the output layer is properly added and maintains the consistent naming pattern used by other model architectures.

---

94-94: Tensor mapping for Llama4 output normalization looks good

The added "language_model.model.norm" tensor mapping for output normalization follows the consistent pattern with other model architectures.

---

136-136: Tensor mappings for Llama4 attention components look good

The tensor mappings for attention components (input layernorm, query/key/value/output projections, and post-attention layernorm) follow the expected naming pattern for Llama4 architecture. These mappings maintain consistency with other model architectures while accounting for Llama4's specific naming structure.

Also applies to: 176-176, 191-191, 205-205, 232-232, 270-270

---

290-290: Tensor mappings for Llama4 feed-forward and MoE components look good

The added mappings properly support Mixture of Experts (MoE) architecture components including:

• Router tensor for expert routing
• Standard feed-forward projections (up/gate/down)
• Expert-specific projections
• Shared expert projections

These mappings follow the consistent pattern with other MoE model architectures like Mixtral while maintaining Llama4's specific naming structure.

Also applies to: 328-328, 337-337, 343-343, 364-364, 373-373, 379-379, 408-408, 418-418, 424-424

✨ Finishing Touches

• 📝 Generate Docstrings

---

🪧 Tips

Chat

There are 3 ways to chat with CodeRabbit:

• Review comments: Directly reply to a review comment made by CodeRabbit. Example:
  • I pushed a fix in commit <commit_id>, please review it.
  • Generate unit testing code for this file.
  • Open a follow-up GitHub issue for this discussion.
• Files and specific lines of code (under the "Files changed" tab): Tag @coderabbitai in a new review comment at the desired location with your query. Examples:
  • @coderabbitai generate unit testing code for this file.
  • @coderabbitai modularize this function.
• PR comments: Tag @coderabbitai in a new PR comment to ask questions about the PR branch. For the best results, please provide a very specific query, as very limited context is provided in this mode. Examples:
  • @coderabbitai gather interesting stats about this repository and render them as a table. Additionally, render a pie chart showing the language distribution in the codebase.
  • @coderabbitai read src/utils.ts and generate unit testing code.
  • @coderabbitai read the files in the src/scheduler package and generate a class diagram using mermaid and a README in the markdown format.
  • @coderabbitai help me debug CodeRabbit configuration file.

Note: Be mindful of the bot's finite context window. It's strongly recommended to break down tasks such as reading entire modules into smaller chunks. For a focused discussion, use review comments to chat about specific files and their changes, instead of using the PR comments.

CodeRabbit Commands (Invoked using PR comments)

• @coderabbitai pause to pause the reviews on a PR.
• @coderabbitai resume to resume the paused reviews.
• @coderabbitai review to trigger an incremental review. This is useful when automatic reviews are disabled for the repository.
• @coderabbitai full review to do a full review from scratch and review all the files again.
• @coderabbitai summary to regenerate the summary of the PR.
• @coderabbitai generate docstrings to generate docstrings for this PR.
• @coderabbitai resolve resolve all the CodeRabbit review comments.
• @coderabbitai plan to trigger planning for file edits and PR creation.
• @coderabbitai configuration to show the current CodeRabbit configuration for the repository.
• @coderabbitai help to get help.

Other keywords and placeholders

• Add @coderabbitai ignore anywhere in the PR description to prevent this PR from being reviewed.
• Add @coderabbitai summary to generate the high-level summary at a specific location in the PR description.
• Add @coderabbitai anywhere in the PR title to generate the title automatically.

CodeRabbit Configuration File (.coderabbit.yaml)

• You can programmatically configure CodeRabbit by adding a .coderabbit.yaml file to the root of your repository.
• Please see the configuration documentation for more information.
• If your editor has YAML language server enabled, you can add the path at the top of this file to enable auto-completion and validation: # yaml-language-server: $schema=https://coderabbit.ai/integrations/schema.v2.json

Documentation and Community

• Visit our Documentation for detailed information on how to use CodeRabbit.
• Join our Discord Community to get help, request features, and share feedback.
• Follow us on X/Twitter for updates and announcements.

[ngxson]

@danielhanchen oh sorry I didn't notice this PR. Thanks a lot!!

I'll "forward" this PR to upstream llama.cpp (doing a git cherry-pick)