Skip to content

SoumilB7/unfold

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

234 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

MODEL UNFOLDER

your one click model unfolder

PyPI

from model_unfolder import unfold
unfold("meta-llama/Meta-Llama-3-8B")

Meta-Llama-3-8B architecture diagram


Install

pip install model-unfolder

# for local development
pip install -e .
pip install transformers   # only required to load by model ID

Three ways to call it

from model_unfolder import unfold

# 1) by HuggingFace model ID — only config.json is downloaded, never weights
unfold("meta-llama/Meta-Llama-3-8B")
unfold("deepseek-ai/DeepSeek-V3")

# 2) from a transformers AutoConfig
from transformers import AutoConfig
unfold(AutoConfig.from_pretrained("Qwen/Qwen2.5-7B", trust_remote_code=True))

# 3) from a raw config.json dict — no transformers install needed
import json
unfold(json.load(open("config.json")))

Built on transformers

Pass a model ID and unfold calls transformers.AutoConfig.from_pretrained(model_id) under the hood (parser.py). It only retries with trust_remote_code=True when Transformers says the config requires remote code.

Auth-token from your environment

Gated models (Llama-3, Mistral, Gemma, …) need a HuggingFace token. unfold reuses whatever transformers / huggingface_hub already see:

# Either set an env var
export HF_TOKEN="hf_xxxxxxxx"            # also accepted: HUGGING_FACE_HUB_TOKEN

# or use the CLI cache (persists across sessions)
huggingface-cli login

# or load a .env in your notebook
# >>> from dotenv import load_dotenv; load_dotenv()

No extra config in model_unfolder itself.

Save / export

diagram = unfold(cfg)
diagram.save("model.html")   # standalone interactive HTML
diagram.save("model.json")   # expanded architecture JSON (no rendering)
diagram.param_count()        # {"total": ..., "active": ..., "per_layer": [...]}
diagram.to_ir()              # full IR dict

Param estimates are close to published numbers — DeepSeek-V3 reports ~675B (~41B active), Llama-3-8B reports 8.03B.

Models supported

Transformers

Family Models
DeepSeek DeepSeek-V2, DeepSeek-V3 (+ MTP head), Kimi K2
Llama Llama 3 / 3.1 / 3.2 / 3.3, OLMo-2, Llama 4 Scout / Maverick (MoE + iRoPE NoPE layers)
Mistral Mistral 7B, Mixtral 8x7B / 8x22B, Mistral Medium 3.5
Qwen Qwen2 / 2.5, Qwen2-MoE, Qwen3, Qwen3-MoE, Qwen3.5 / 3.6 (+ MTP)
Gemma Gemma 2 9B / 27B (interleaved local+global), Gemma 3 / 3n (+ PLE), Gemma 4 31B / E2B / E4B (+ PLE), RecurrentGemma 2B / 9B (LRU + local attention)
Cohere Command R, Command R+, Command R7B (QK-Norm attention)
Jamba Jamba (SSM + attention hybrid, MoE)
Zamba Zamba 7B, Zamba2 2.7B / 7B (Mamba SSM + weight-shared attention)
Mamba Mamba 130M–2.8B, Mamba-2 (pure SSM, no attention)
Falcon Falcon 7B / 40B (parallel attn+FFN), Falcon-H1 (Mamba-2 SSM)
MiniMax MiniMax-Text-01 (lightning + softmax hybrid, MoE)
RWKV RWKV-4 / 5 / 6 (pure recurrent, no attention)

Diffusors

Coming soon.

Custom

Drop a request in issues.

License

Apache 2.0.

About

Visualize any model

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages