API Reference

Value

Scalar-valued autograd node.

from strands_microgpt import Value

a = Value(2.0)

Operations

Operation	Usage	Description
+	a + b	Addition
*	a * b	Multiplication
-	a - b	Subtraction
/	a / b	Division
**	a ** n	Power
.relu()	a.relu()	ReLU activation
.exp()	a.exp()	Exponential
.log()	a.log()	Natural log
.backward()	a.backward()	Compute gradients

Attributes

Attribute	Type	Description
.data	float	The scalar value
.grad	float	The gradient (after .backward())

---

Tokenizer

Character-level tokenizer with BOS token.

Constructor

Tokenizer(chars: List[str])

Class Methods

Method	Description
Tokenizer.from_docs(docs)	Build from list of strings
Tokenizer.from_dict(data)	Deserialize from dict

Instance Methods

Method	Returns	Description
.encode(text)	List[int]	Text → token IDs
.decode(ids)	str	Token IDs → text
.to_dict()	Dict	Serialize for saving

Attributes

Attribute	Type	Description
.vocab_size	int	Number of tokens (chars + BOS)
.bos	int	BOS token ID
.chars	List[str]	Character vocabulary

---

MicroGPT

Pure-Python GPT transformer.

Constructor

MicroGPT(
    vocab_size: int,
    n_layer: int = 1,
    n_embd: int = 16,
    block_size: int = 16,
    n_head: int = 4,
    seed: int = 42,
)

Methods

Method	Returns	Description
.train_on_docs(docs, tokenizer, ...)	List[float]	Train on documents, return losses
.generate(tokenizer, ...)	List[str]	Generate text samples
.forward(token_id, pos_id, keys, values)	List[Value]	Single-token forward pass
.save_checkpoint(path, tokenizer, metadata)	None	Save to JSON

Class Methods

Method	Returns	Description
.load_checkpoint(path)	(MicroGPT, Tokenizer, Dict)	Load from JSON
.from_dataset(url, path, **kwargs)	(MicroGPT, Tokenizer, docs)	Load data + create model

Attributes

Attribute	Type	Description
.num_params	int	Total parameter count
.n_layer	int	Number of layers
.n_embd	int	Embedding dimension

---

MicroGPTModel

Strands Model provider interface.

from strands_microgpt import MicroGPTModel

model = MicroGPTModel(
    dataset_url="...",
    num_steps=1000,
    n_layer=1,
    n_embd=16,
    temperature=0.5,
    num_samples=20,
)

Implements the full Strands Model interface: stream(), format_request(), format_chunk(), etc.

---

Tools

microgpt_train

microgpt_train(
    dataset_url: str = "...",
    dataset_path: str = "",
    num_steps: int = 1000,
    n_layer: int = 1,
    n_embd: int = 16,
    block_size: int = 16,
    n_head: int = 4,
    learning_rate: float = 0.01,
    seed: int = 42,
    checkpoint_path: str = "",
) -> Dict[str, Any]

microgpt_generate

microgpt_generate(
    checkpoint_path: str = "/tmp/microgpt_checkpoint.json",
    num_samples: int = 20,
    temperature: float = 0.5,
    seed: int = 0,
) -> Dict[str, Any]