HuggingFace Model Analysis Guide

This guide explains how to analyze a HuggingFace model to extract the information needed to build a TransformerLens Architecture Adapter.

Read the model’s config.json

Every HF model has a config.json that contains architecture details. You can access it via:

from transformers import AutoConfig
config = AutoConfig.from_pretrained("model-name-or-path")
print(config)

Or via the HuggingFace API:

curl -s "https://huggingface.co/model-name/resolve/main/config.json" | python -m json.tool

Key config fields to extract

HF Config Field	TL Config Field	Description
hidden_size	d_model	Model dimension
num_attention_heads	n_heads	Number of attention heads
num_key_value_heads	n_key_value_heads	KV heads (for GQA; if absent or equal to n_heads, not GQA)
intermediate_size	d_mlp	MLP intermediate dimension
num_hidden_layers	n_layers	Number of transformer blocks
vocab_size	d_vocab	Vocabulary size
max_position_embeddings	n_ctx	Maximum sequence length
rms_norm_eps	eps	Normalization epsilon
model_type	—	Architecture family (e.g., “llama”, “gpt2”, “mistral”)
architectures	architecture	HF class name (e.g., ["LlamaForCausalLM"])

Determine architecture characteristics

Normalization type

Check the model code or config:

• RMSNorm → normalization_type = "RMS" — Look for RMSNorm in the model code, or rms_norm_eps in config

• LayerNorm → normalization_type = "LN" — Look for LayerNorm, or layer_norm_eps / layer_norm_epsilon in config

Also identify the epsilon attribute name:

• "variance_epsilon" (Llama)

• "rms_norm_eps" (some models expose this directly)

• "layer_norm_eps" (GPT-2, BERT)

• "eps" (generic)

Positional embedding type

• Rotary (RoPE) → positional_embedding_type = "rotary" — Most modern models (Llama, Mistral, Qwen, Gemma)

• Learned/Standard → positional_embedding_type = "standard" — GPT-2, OPT

• Check for RotaryEmbedding class in the model code

Attention type

• Multi-Head Attention (MHA) — n_key_value_heads == n_heads or field absent

• Grouped Query Attention (GQA) — n_key_value_heads < n_heads (e.g., Llama 3, Mistral)

• Multi-Query Attention (MQA) — n_key_value_heads == 1 (e.g., Falcon)

MLP type

• Gated MLP (SwiGLU) → gated_mlp = True — Has gate/up/down projections (Llama, Qwen, Gemma)

• Standard MLP → gated_mlp = False — Has fc1/fc2 or c_fc/c_proj (GPT-2)

QKV layout

• Separate Q/K/V — Most models: q_proj, k_proj, v_proj

• Combined QKV — GPT-2 style: single c_attn or query_key_value linear layer

Inspect module names

To find the exact HuggingFace module paths for the component mapping:

from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("model-name", torch_dtype="auto")

# Print all named modules
for name, module in model.named_modules():
    print(f"{name}: {type(module).__name__}")

What to look for

Map these HF module paths to TL component mapping entries:

TL Name	Look for in HF	Common HF Paths
embed	Token embedding	model.embed_tokens, transformer.wte
pos_embed	Position embedding (if standard)	transformer.wpe
rotary_emb	Rotary embedding (if RoPE)	model.rotary_emb, model.layers.0.self_attn.rotary_emb
blocks	Layer list	model.layers, transformer.h, model.decoder.layers
ln1	Pre-attention norm	input_layernorm, ln_1
ln2	Post-attention norm	post_attention_layernorm, ln_2
attn	Self-attention module	self_attn, attn
attn.q	Query projection	q_proj, query
attn.k	Key projection	k_proj, key
attn.v	Value projection	v_proj, value
attn.o	Output projection	o_proj, out_proj, dense, c_proj
attn.qkv	Combined QKV (if used)	c_attn, query_key_value
mlp	MLP module	mlp, feed_forward
mlp.gate	Gate projection (if gated)	gate_proj, w1
mlp.in	Up/input projection	up_proj, c_fc, fc1, w3
mlp.out	Down/output projection	down_proj, c_proj, fc2, w2
ln_final	Final layer norm	model.norm, transformer.ln_f, model.final_layernorm
unembed	LM head	lm_head, embed_out

Check for biases

# Check if a specific layer has bias
layer = model.model.layers[0]
print(f"Q bias: {layer.self_attn.q_proj.bias is not None}")
print(f"MLP in bias: {layer.mlp.up_proj.bias is not None}")

Document which layers lack biases — this affects the “Optional Parameters” section of the adapter docstring.

Examine state dict keys

# Print all parameter names and shapes
for key, param in model.state_dict().items():
    print(f"{key}: {param.shape}")

This helps verify:

• Weight naming patterns match your component mapping

• Tensor shapes match expected dimensions

• No unexpected parameters that need special handling

Find an existing similar adapter

Check if a similar architecture already has an adapter. Most new models are variants of existing patterns:

If your model is like…	Start from adapter…
Llama, Mistral, Qwen2, Gemma	llama.py
GPT-2, GPT-J	gpt2.py
BLOOM, Falcon	bloom.py or falcon.py
T5, encoder-decoder	t5.py
MoE model	mixtral.py or granite_moe.py
Multimodal (vision+text)	llava.py or gemma3_multimodal.py

Quick reference: decision tree

1. Does the model use RMSNorm or LayerNorm?
   → RMSNorm: normalization_type="RMS", use RMSNormalizationBridge
   → LayerNorm: normalization_type="LN", use NormalizationBridge

2. Does the model use RoPE or learned positional embeddings?
   → RoPE: positional_embedding_type="rotary", add RotaryEmbeddingBridge, use PositionEmbeddingsAttentionBridge
   → Learned: positional_embedding_type="standard", add PosEmbedBridge

3. Are Q/K/V separate or combined?
   → Separate: use PositionEmbeddingsAttentionBridge with q/k/v/o submodules
   → Combined: use JointQKVAttentionBridge with qkv/o submodules

4. Does the MLP have a gate projection?
   → Yes (gate+up+down): gated_mlp=True, use GatedMLPBridge
   → No (in+out): gated_mlp=False, use MLPBridge

5. Is n_key_value_heads < n_heads?
   → Yes: GQA — set n_key_value_heads on cfg
   → No: standard MHA — no special handling needed