dustylm-sdk 0.1.1

The official Python SDK for DustyLM: an 8M-parameter model that talks like a robot vacuum.

dustylm

Run DustyLM, the 8M parameter robot vacuum language model, in a few lines.

pip install dustylm-sdk

from dustylm import DustyLM
model = DustyLM.from_pretrained("mkhordoo/dusty-8m-sft")
response = model.chat([{"role": "user", "content": "who are you?"}])
print(response["choices"][0]["message"]["content"])

# beep. i am a little robot. i clean floors and find crumbs.

Backends

Backend	Install	Model file	Best for
torch (default)	pip install dustylm-sdk	~32 MB (FP32)	Inspection and experimentation
onnx	pip install dustylm-sdk[onnx]	~8–10 MB (int8)	A compact model artifact

These sizes describe the model files, not the complete Python environment. Installation size varies by platform. Both backends currently include PyTorch as an SDK dependency.

# ONNX backend
model = DustyLM.from_pretrained("mkhordoo/dusty-8m-sft", backend="onnx")

Loading Local Checkpoints

Pass a Hugging Face repository ID to download its model and tokenizer automatically, as shown in the quick-start example:

model = DustyLM.from_pretrained("mkhordoo/dusty-8m-sft")

For files already downloaded to your machine, pass their local directory instead. With the default filenames, that directory must contain tokenizer.json and either model.pt (PyTorch) or model_int8.onnx (ONNX):

model = DustyLM.from_pretrained("./my-checkpoint/")

You can use any local directory as the artifact root. model_file and tokenizer_file are resolved relative to that directory, so they may use different filenames or relative paths.

If you trained a model with the main dusty-lm repository, its checkpoint and tokenizer are stored in separate artifact directories. Load the promoted SFT checkpoint with:

from dustylm import DustyLM

model = DustyLM.from_pretrained(
    "artifacts/checkpoints",
    model_file="dusty8m_sft.pt",
    tokenizer_file="../tokenizers/dusty_tokenizer.json",
)

response = model.chat([{"role": "user", "content": "who are you?"}])

To inspect an intermediate checkpoint instead, change model_file to an existing step file such as dusty8m_sft_step_100.pt.

The SDK infers several architecture dimensions from PyTorch checkpoint shapes and currently expects DustyLM-compatible configurations.

API

DustyLM.from_pretrained

DustyLM.from_pretrained(
    repo_id_or_path: str = "mkhordoo/dusty-8m-sft",
    *,
    model_file: str | None = None,
    tokenizer_file: str | None = None,
    backend: str = "torch",
)

Argument	Default	Description
repo_id_or_path	"mkhordoo/dusty-8m-sft"	HF Hub repo ID or local directory path
model_file	"model.pt" (torch) / "model_int8.onnx" (onnx)	Override the model filename in the directory
tokenizer_file	"tokenizer.json"	Override the tokenizer filename
backend	"torch"	"torch" or "onnx"

DustyLM.chat

model.chat(
    messages: list[dict],
    temperature: float = 0.7,
    max_tokens: int = 64,
    top_p: float = 0.9,
) -> dict

Returns an OpenAI-style chat completion dict.

---

Built from the dusty-lm training repository.