Model Loading: Code

Runnable companions to Model Loading: Hub to GPU. The first two are stdlib-only Python (3.11+) and need no install.

File	What it does	Run
safetensors_inspect.py	Parses the safetensors header by hand (8-byte LE length, JSON, byte buffer). Lists tensors, dtypes, shapes, byte sizes, element totals; validates sizes and holes; handles model.safetensors.index.json (shard map vs shard headers); decodes one tensor (F32/F16/BF16/ints) into a Python list; inspects remote files with HTTP Range requests (honors HF_TOKEN)	python safetensors_inspect.py --selftest
config_explain.py	Reads a config.json and prints total and active params, attention variant (MHA/GQA/MQA/MLA), MoE layout, KV bytes per token, max context, RoPE scaling, EOS ids, and a cold-load time estimate per storage path	python config_explain.py samples/*.json
load_with_transformers.py	Reference load path: snapshot_download(allow_patterns=...), AutoTokenizer.apply_chat_template, AutoModelForCausalLM.from_pretrained(dtype="auto", device_map="auto"), generate. Defaults to HuggingFaceTB/SmolLM2-135M-Instruct	uv run --no-project --with transformers --with torch --with accelerate python load_with_transformers.py
samples/	Verbatim public config.json files: Llama 3.1 8B Instruct (dense, GQA, llama3 RoPE), Qwen3-30B-A3B (MoE, 128 experts top-8), DeepSeek-V3 (MLA, 256 experts + shared, FP8, YaRN)

Examples

# Inspect a remote checkpoint without downloading the weights
python safetensors_inspect.py --show 5 --tensor model.norm.weight --limit 6 \
  --url https://huggingface.co/HuggingFaceTB/SmolLM2-135M-Instruct/resolve/main/model.safetensors

# Check a customer's sharded checkpoint: index vs shard headers
python safetensors_inspect.py /mnt/models/customer-ft/model.safetensors.index.json --show 0

# KV cache sizing with an FP8 cache
python config_explain.py samples/llama-3.1-8b-instruct.config.json --kv-dtype fp8

The Llama sample reproduces the official checkpoint exactly: config_explain.py computes 8,030,261,248 parameters, and the Hub index reports total_size = 16,060,522,496 bytes in BF16 (2 bytes each).

Sample sources: Llama 3.1 8B Instruct (gated; values match the public mirror), Qwen3-30B-A3B, DeepSeek-V3. Fetched October 2026.