analysis
analysis
¶
ForkAnalysis(spec: ExecutionSpec, base: Optional[Node] = None)
¶
Bases: AnalysisBase[ForkAnalysisConfig, GPT]
Export tokenwise fork diversity and a per-fork logit lens.
capture_fork_blocks(module: Any, method: str) -> bool
staticmethod
¶
Capture fork outputs before attention/MLP, and boundary block states.
measure_tokenwise_diversity(states: np.ndarray, token_index: np.ndarray, sequence_length: int) -> list[dict[str, int | float | None]]
staticmethod
¶
Measure non-rightmost sibling cosine distance for each input token.
decode_token(token_id: int) -> str
¶
Keep padded vocabulary IDs visible even without a tokenizer spelling.
measure_stage(*, stage: str, block_output: tuple[jax.Array, jax.Array, jax.Array], predictions: tuple[np.ndarray, np.ndarray, np.ndarray], token_ids: list[int], token_text: list[str]) -> tuple[list[dict[str, Any]], list[dict[str, Any]]]
¶
Turn one captured fork stage into diversity and lens records.