Skip to content

Add DeepSeek-V4.1 model - #1895

Open
Goekdeniz-Guelmez wants to merge 4 commits into
ml-explore:mainfrom
Goekdeniz-Guelmez:add-deepseek-v4.1
Open

Goekdeniz-Guelmez wants to merge 4 commits into
ml-explore:mainfrom
Goekdeniz-Guelmez:add-deepseek-v4.1

Conversation

@Goekdeniz-Guelmez

Copy link
Copy Markdown
Contributor

Addresses #1894

- Implemented tests to verify that the Transformers DeepSeek-V4.1 model loads correctly and matches the output of the MLX implementation.
- Added a test to ensure cached decoding matches the full forward pass output.
- Included a tiny configuration for the DeepSeek-V4.1 model to facilitate testing.
@Goekdeniz-Guelmez

Copy link
Copy Markdown
Contributor Author

Text backbone implementation of DeepSeek-V4.1-Flash. The implementation was checked against the pending
Transformers reference using a random init 172,720-parameter checkpoint.

Not tested against the real weights.

… with nn.RMSNorm, and enhance SwiGLU activation implementation.
… scaling and inverse handling; add unit tests for partial RoPE support with strided positions.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants