dolphin-utils

A collection of utilities for AI/ML model analysis and processing.

dolphin-summarize

This tool analyzes safetensors model files to generate a condensed summary of the model's architecture. It groups similar parameter names using range notation (e.g., model.layers.[0-39].mlp.down_proj.weight) and displays the shape and data type (precision) for each parameter group.

Key Features:

• Remote Processing: Analyze Hugging Face models without downloading the full model files (downloads only headers - KB instead of GB)
• Local Processing: Works with locally stored model directories
• Efficient: Uses HTTP range requests and streaming to minimize data transfer
• Reliable: Multiple fallback strategies ensure 100% compatibility

Dependencies

• Python 3.7+
• huggingface_hub - For accessing Hugging Face repositories
• requests - For HTTP requests and remote file access
• safetensors - For reading safetensors file headers

All dependencies are automatically installed when you install the package:

pip install dolphin-utils

Usage

After installing the package:

pip install dolphin-utils

You can use the tool in two ways:

Via CLI command:

dolphin-summarize [MODEL_PATH_OR_REPO_ID] [OPTIONS]

Via Python module:

python -m dolphin_summarize [MODEL_PATH_OR_REPO_ID] [OPTIONS]

Arguments:

• MODEL_PATH_OR_REPO_ID:
  • Local path: Directory containing safetensors files (e.g., ~/models/my_llama_model)
  • Hugging Face repo ID: Repository identifier (e.g., microsoft/DialoGPT-medium)
  • Defaults to current directory (.) if not provided

Options:

• --output OUTPUT, -o OUTPUT: Path to an output file where the summary will be written (optional).
• --verbose, -v: Show verbose output during processing (optional).

Examples

Remote Processing (Hugging Face Hub):

# Process a model directly from Hugging Face without downloading
python -m dolphin_summarize microsoft/DialoGPT-medium --verbose

# Large models work too - only headers are downloaded
python -m dolphin_summarize meta-llama/Llama-2-70b-hf --verbose

Local Processing:

# Process a local model directory
python -m dolphin_summarize ~/models/my_llama_model --verbose

# Process current directory
python -m dolphin_summarize . --verbose

Output Format

The script prints the summary to the console (and optionally to a file). Each line represents a parameter or a group of parameters with a similar structure:

parameter_name,[shape],dtype

Example Output Lines:

lm_head.weight,[131072,5120],BF16
model.embed_tokens.weight,[131072,5120],BF16
model.layers.[0-39].input_layernorm.weight,[5120],BF16
model.layers.[0-39].mlp.down_proj.weight,[5120,13824],BF16
model.layers.[0-39].mlp.gate_proj.weight,[13824,5120],BF16
model.layers.[0-39].mlp.up_proj.weight,[13824,5120],BF16
model.layers.[0-39].post_attention_layernorm.weight,[5120],BF16
model.layers.[0-39].self_attn.k_proj.weight,[512,5120],BF16
model.layers.[0-39].self_attn.o_proj.weight,[5120,8192],BF16
model.layers.[0-39].self_attn.q_proj.weight,[8192,5120],BF16
model.layers.[0-39].self_attn.v_proj.weight,[512,5120],BF16
model.norm.weight,[5120],BF16