MEGA Hub
New
Docs/Runtime reference

Examples#

Example scripts live under megatensors/examples in the MEGA source tree. They are not installed with the Python wheel, so run this page from a source checkout after installing the package and example dependencies:

BASH
python -m pip install -e "./megatensors[test]"

The commands below assume the current directory is the repository root and python resolves to that environment.

Generate text from a MEGA artifact#

BASH
python megatensors/examples/causal_lm_generate.py \  Qwen3.5-0.8B \  --mega Qwen3.5-0.8B/mega/model.mega.index.json \  --device cuda:0 \  --trust-remote-code \  --prompt "Hello, my name is"

Use --nogds if the host does not have GPUDirect Storage configured:

BASH
python megatensors/examples/causal_lm_generate.py \  Qwen3.5-0.8B \  --mega Qwen3.5-0.8B/mega/model.mega.index.json \  --device cuda:0 \  --nogds \  --prompt "Explain tensor loading in one sentence."

Verify Hugging Face equivalence#

Run a short comparison against the source Hugging Face directory:

BASH
python megatensors/examples/hf_inference_verify.py \  Qwen3.5-0.8B \  --mega Qwen3.5-0.8B/mega/model.mega.index.json \  --device cuda:0 \  --trust-remote-code

Use this before publishing a converted artifact.

Inspect tensor keys#

BASH
python megatensors/examples/extract_keys.py Qwen3.5-0.8B/mega

This is useful when load_model fails because a model class expects different key names.

Sign and verify#

BASH
python megatensors/examples/sign_and_verify.py \  Qwen3.5-0.8B/mega/model.mega.index.json \  --bundle ./signing/release-bundle.pem \  --trusted-roots ./signing/root-ca.pem \  --model-id qwen3.5-0.8b

Minimal Python smoke test#

PYTHON
from megatensors import load_state_dict, load_tokenizer path = "Qwen3.5-0.8B/mega/model.mega.index.json"state = load_state_dict(path, device="cpu")tokenizer = load_tokenizer(path) print(len(state))print(tokenizer("hello").input_ids[:8])