Notion LLM Config Table Management

Database Information

• Database ID: ddfd95bd-109a-4ac6-955c-90541cc53d5e
• Data Source ID: 0a9fafd6-2cc2-4d6b-b6f0-3797ea777421
• Location: Family Notes workspace → "LLM config table"
• Purpose: Track LLM model configurations across multiple providers (Llama, GPT-2, Qwen, others)

Data Sources

Llama Models

Primary source: ~/projects/github/llama-models/models/sku_list.py

• Contains all registered Llama models with architecture parameters
• Constants: LLAMA2_VOCAB_SIZE = 32000, LLAMA3_VOCAB_SIZE = 128256
• WARNING: sku_list.py has incorrect ffn_dim_multiplier values for Llama 2 7B and 13B models
  • sku_list.py shows 1.3 for all Llama 2 models, but actual checkpoints differ
  • Always verify d_ff values against HuggingFace configs, not just sku_list.py

GPT-2 Models

Primary source: HuggingFace config.json files

• Base repos: openai-community/gpt2, gpt2-medium, gpt2-large, gpt2-xl
• All variants: vocab_size=50257, max_context=1024, uses BPE tokenizer
• Uses learned positional embeddings (not RoPE)

User Preferences

1. No emojis in any updates or content
2. Batch operations preferred over sequential updates
3. Focus on practical utility - add "Main" column for filtering to representative models
4. Verify calculations - user will catch errors, so double-check formulas

Database Schema Key Columns

Identity

• Model Name (title), Model Family, Model Type, Core Model ID
• HuggingFace Repo, Variant

Architecture Parameters

• d_hidden = model hidden dimension (dim from Llama arch_args)
• d_ff = feed-forward hidden dimension (see calculation formulas below)
• d_ff / d_hidden Ratio = d_ff / d_hidden, rounded to 3 decimals
  • Llama 2: 2.688 (7B), 2.7 (13B), 3.5 (70B)
  • Llama 3+: 2.667 (3B), 3.125 (Guard INT4), 3.25 (405B), 3.5 (most common), 4.0 (1B)
  • GPT-2: 4.0 (all variants - standard transformer)
  • Null for models without d_ff (Llama 4 has empty arch_args)
• Gated MLP (checkbox) = gated MLP/activation (SwiGLU for Llama, standard GELU for GPT-2)
• n_layers, n_heads, n_kv_heads, head_dim
• Multiple Of, Norm Eps

Tokenization

• Tokenizer: Llama 2 Tokenizer (32k vocab) or Llama 3 Tokenizer (128k vocab)
• Vocab Size: 32000 (Llama 2), 128256 (Llama 3+)

MoE Architecture

• Is MoE (checkbox)
• Num Experts: Llama 4: 16 (Scout), 128 (Maverick)
• Top K Experts: Number of routed experts per token
• MoE Routing: "Token Choice" (Llama 4) vs "Expert Choice"
• Has Shared Expert: Llama 4 has 1 shared + 1 routed per token
• Activated Params (B): Llama 4: 17B
• Total Params (B): Llama 4: 109B (Scout), 400B (Maverick)

RoPE Configuration

• RoPE Theta, RoPE Freq Base
• Use Scaled RoPE (checkbox)

Other

• Quantization Format, PTH File Count, Max Context Size
• Main (checkbox) - marks representative models for filtering

Update Patterns

Property Updates (Preferred)

mcp__notion__notion-update-page({
    "page_id": "page-id-here",
    "command": "update_properties",
    "properties": {
        "Tokenizer": "Llama 3 Tokenizer",
        "Vocab Size": 128256,
        "Is MoE": "__YES__",  # Checkboxes use __YES__/__NO__
        "Num Experts": 16
    }
})

Batch Updates

• Use parallel tool calls when updating multiple independent pages
• Group by model family for logical batching (e.g., all Scout models together)

Search & Fetch

# Search within database
mcp__notion__notion-search({
    "query": "Llama 4",
    "query_type": "internal",
    "data_source_url": "collection://0a9fafd6-2cc2-4d6b-b6f0-3797ea777421"
})

# Fetch page details
mcp__notion__notion-fetch({"id": "page-id-or-url"})

Common Tasks

Adding New Models

1. Parse ~/projects/github/llama-models/models/sku_list.py to get model data
2. Extract arch_args and calculate derived fields (head_dim, d_ff using Llama formula, d_ff/d_hidden ratio)
3. Determine tokenizer based on model family (see Llama Model Family Mappings)
4. Set MoE fields for Llama 4 models
5. Set "Gated MLP" to Yes (all Llama models use SwiGLU)
6. Create pages in batches using mcp__notion__notion-create-pages

Marking Representative Models

Representative "Main" models (user preference):

• Llama 2: 7b chat, 70b chat
• Llama 3.1: 8b instruct, 70b instruct, 405b instruct (FP8)
• Llama 3.2: 1b instruct, 3b instruct, 11b vision instruct (user trains small models)
• Llama 3.3: 70b instruct
• Llama 4: Scout instruct, Maverick instruct
• GPT-2: All variants (117M, 345M, 774M, 1.5B) marked as Main

Schema Updates

mcp__notion__notion-update-database({
    "database_id": "ddfd95bd-109a-4ac6-955c-90541cc53d5e",
    "properties": {
        "New Column": {"type": "number", "number": {}}
    }
})

Important Notes

Pitfalls to Avoid

1. Don't clear existing fields - Only specify properties you're updating
2. Column ordering - Cannot be changed via API (view-level setting in UI)
3. Checkbox format - Must use "__YES__" or "__NO__", not boolean
4. Llama 4 models - Have empty arch_args={} in sku_list.py, no d_hidden/d_ff available

Model Family Mappings

Llama

• llama2: Llama 2 Tokenizer, 32k vocab
  • Actual d_ff values (verified from HuggingFace):
    • 7B: d_ff=11008 (NOT 14336 from sku_list.py formula)
    • 13B: d_ff=13824 (NOT 17920 from sku_list.py formula)
    • 70B: d_ff=28672 (correct in sku_list.py)
• llama3, llama3_1, llama3_2, llama3_3, llama4, safety: Llama 3 Tokenizer, 128k vocab

GPT-2

• gpt2: BPE tokenizer, 50257 vocab, 1024 max context
• Naming: "GPT-2", "GPT-2 Medium", "GPT-2 Large", "GPT-2 XL"
• Model Type: "base" (all are base models)
• Parameter counts: 117M (base), 345M (medium), 774M (large), 1.5B (XL)
• Config parameter mappings: n_embd → d_hidden, n_layer → n_layers

Llama MoE Architecture (Llama 4 only)

• Only Llama 4 models are MoE
• Architecture: 1 shared expert (always active) + 1 routed expert (top_k=1)
• Effective: 2 experts per token (shared + routed)
• Routing: Token Choice (each token selects expert via router scores)

Llama Gated MLP / Activation Function

• All Llama models use gated MLP with SwiGLU activation
• Implementation: w2(F.silu(w1(x)) * w3(x)) (from models/llama3/model.py)
• Uses 3 weight matrices (w1, w2, w3) instead of standard 2-matrix FFN
• SwiGLU = Swish-Gated Linear Unit (Swish is same as SiLU)

Calculation Formulas

General Derived Fields

head_dim = d_hidden / n_heads
d_ff_d_hidden_ratio = round(d_ff / d_hidden, 3)

Llama d_ff Calculation

From actual Llama code (models/llama3/model.py):

# Initial: hidden_dim = 4 * dim
hidden_dim = int(2 * hidden_dim / 3)  # = int(8 * dim / 3)
if ffn_dim_multiplier is not None:
    hidden_dim = int(ffn_dim_multiplier * hidden_dim)
# Round up to multiple_of
hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
# Result is d_ff

IMPORTANT: Llama 2 Model-Specific Behavior

Llama 2 models use different ffn_dim_multiplier values than what's in sku_list.py:

• 7B & 13B: Use ffn_dim_multiplier=None (pure 8d/3 formula)
  • 7B: d_ff=11008, ratio=2.688
  • 13B: d_ff=13824, ratio=2.7
• 70B: Uses ffn_dim_multiplier=1.3 (as specified in sku_list.py)
  • 70B: d_ff=28672, ratio=3.5

This was verified from actual HuggingFace checkpoint configs. The sku_list.py incorrectly shows ffn_dim_multiplier=1.3 for all Llama 2 models.

Alternative formula from HuggingFace transformers (equivalent for models with multiple_of=256):

# For Llama 2 7B, 13B (no multiplier)
def compute_intermediate_size(n, multiple_of=256):
    return int(math.ceil(n * 8 / 3) + multiple_of - 1) // multiple_of * multiple_of

GPT-2 d_ff Calculation

Standard transformer architecture:

d_ff = 4 * d_hidden  # Always 4x for all GPT-2 variants

• Not explicitly in config.json, but defined in model architecture
• All GPT-2 models have d_ff/d_hidden ratio of exactly 4.0

Helpful Scripts & References

Llama

Working directory: ~/projects/github/llama-models/

Scripts:

• Parse models: parse_llama_models.py
• Prepare Notion data: prepare_notion_data.py
• Update tokenizers: update_tokenizers.py
• Bulk updates: bulk_update_llama3.py
• Calculate d_ff: fix_d_ff.py (correct d_ff calculation using actual Llama formula)
• Calculate ratios: calculate_ratios.py (d_ff/d_hidden ratios and gated MLP status)
• Data files: d_ff_corrections.json, complete_notion_updates.json, ratio_updates.json
• Documentation: COMPLETION_SUMMARY.md (d_ff update history), STATUS.md (current status)

References:

• Llama models repo: ~/projects/github/llama-models/
• Model definitions: models/sku_list.py
• Architecture files: models/llama{2,3,4}/args.py
• MoE implementation: models/llama4/moe.py

GPT-2

HuggingFace References:

• Base: https://huggingface.co/openai-community/gpt2
• Medium: https://huggingface.co/gpt2-medium
• Large: https://huggingface.co/gpt2-large
• XL: https://huggingface.co/gpt2-xl
• Config location: /raw/main/config.json (append to repo URL)

Architecture Notes:

• Uses learned positional embeddings (not RoPE) - add to Notes field
• Activation: GELU (gelu_new variant) - not gated MLP
• Standard attention (n_kv_heads = n_heads, no GQA)
• All variants have head_dim = 64
• Norm epsilon: 1e-05 (layer_norm_epsilon in config)