`198cd55`

ruff auto-format after S23 edits (line-length-driven, no logic)

Authored by mfwolffe <wolffemf@dukes.jmu.edu> 2 weeks ago

Status	File	+	-
M	`src/dlm_sway/backends/hf.py`	4	4
M	`src/dlm_sway/core/scoring.py`	1	3
M	`src/dlm_sway/probes/cluster_kl.py`	2	1
M	`src/dlm_sway/probes/delta_kl.py`	2	1

src/dlm_sway/backends/hf.pymodified

          log_probs = F.log_softmax(logits.float(), dim=-1).squeeze(0)
          return _topk_to_token_dist(log_probs, top_k=top_k)
 -    def next_token_dist_batch(
 -        self, prompts: Sequence[str], *, top_k: int = 256
 -    ) -> list[TokenDist]:
 +    def next_token_dist_batch(self, prompts: Sequence[str], *, top_k: int = 256) -> list[TokenDist]:
          """Batched forward via tokenizer left-padding.
          Decoder-only LMs need left-padding because the last-token
                      attention_mask=tokens.get("attention_mask"),
                  ).logits[:, -1, :]  # (B, V) — left-pad makes "last" always the real last token
              log_probs = F.log_softmax(logits.float(), dim=-1)  # (B, V)
 -            return [_topk_to_token_dist(log_probs[row], top_k=top_k) for row in range(len(miss_prompts))]
 +            return [
 +                _topk_to_token_dist(log_probs[row], top_k=top_k) for row in range(len(miss_prompts))
 +            ]
          return self._inst.cached_batch(
              "next_token_dist", self.id, list(prompts), top_k, compute_misses

src/dlm_sway/core/scoring.pymodified

          """
          ...
 -    def next_token_dist_batch(
 -        self, prompts: Sequence[str], *, top_k: int = 256
 -    ) -> list[TokenDist]:
 +    def next_token_dist_batch(self, prompts: Sequence[str], *, top_k: int = 256) -> list[TokenDist]:
          """Batched variant of :meth:`next_token_dist`.
          Returns one :class:`TokenDist` per entry in ``prompts``, in the

src/dlm_sway/probes/cluster_kl.pymodified

          with ctx.backend.as_finetuned() as ft_view:
              ft_dists = ft_view.next_token_dist_batch(list(spec.prompts), top_k=top_k)
          divergences: list[float] = [
 -            divergence(b, f, kind=spec.divergence) for b, f in zip(base_dists, ft_dists, strict=True)
 +            divergence(b, f, kind=spec.divergence)
 +            for b, f in zip(base_dists, ft_dists, strict=True)
+         ]
          # Aggregate per-cluster means + variances. A cluster that

src/dlm_sway/probes/delta_kl.pymodified

          with ctx.backend.as_finetuned() as ft_view:
              ft_dists = ft_view.next_token_dist_batch(list(spec.prompts), top_k=top_k)
          divergences: list[float] = [
 -            divergence(b, f, kind=spec.divergence) for b, f in zip(base_dists, ft_dists, strict=True)
 +            divergence(b, f, kind=spec.divergence)
 +            for b, f in zip(base_dists, ft_dists, strict=True)
+         ]
          raw_mean = statistics.fmean(divergences)