huihui-ai/

grok-2

huihui-ai/grok-2

1. Collecting safetensors files: Locates all pytorch_model-*.safetensors files in the specified model directory.
2. Loading files into cache: Loads all safetensors files into memory and builds a key-to-file mapping.
3. Merging Tensor Parallel (TP) shards: Merges shards for tensor parallelism (TP=8) along specific dimensions and verifies the merged tensor shapes.
4. Grouping weights by layer: Organizes weights by model layer, with special weights (e.g., lm_head.weight, model.embed_tokens.weight, and model.norm.weight) handled separately.
5. Saving merged weights: Saves the grouped weights as new safetensors files and generates a new index file pytorch_model.bin.index.json.

Features

• Input: Safetensors files in the xai-org/grok-2 model directory.
• Output: Layer-organized safetensors files and an index file in the huihui-ai/grok-2 directory.
• Tensor Parallelism Support: Handles TP=8 shards, merging tensors along specific dimensions (w1.weight and w3.weight along dim=0, w2.weight along dim=1).
• Error Handling: Includes warnings and handling for missing files, shape mismatches, and other exceptions.
• Shape Validation: Verifies shapes for specific weights (e.g., MoE layer weights), ensuring merged tensors match expected shapes (e.g., (16384, 8192) or (8192, 16384)).

Usage

1. Install the required Python libraries:
   pip install torch safetensors
2. Place the script in an environment with the xai-org/grok-2 model directory.
3. Run the script:
   python convert_safetensors.py
4. Output files will be saved in the huihui-ai/grok-2 directory, including layer-organized safetensors files and an index file.

Notes

• Ensure the input directory xai-org/grok-2 contains valid pytorch_model-*.safetensors files.
• The script assumes a tensor parallelism degree of 8 (tp_count = 8). Modify the tp_count value in the script if needed.
• Memory requirements may be high; run on a machine with sufficient memory.
• If shards are missing or shapes mismatch, the script will print warnings and attempt to proceed.