chore: drop fastcluster to allow numpy 2 (#147)

Author: mplatzer

mostlyai/qa/_accuracy.py

@@ -19,7 +19,6 @@
 from typing import Any, Literal
 from collections.abc import Callable, Iterable
 
-import fastcluster
 import numpy as np
 import pandas as pd
 import phik.phik
@@ -1234,7 +1233,7 @@ def calculate_correlations(binned: pd.DataFrame, corr_cols: Iterable[str] | None
             # determine column order via hierarchical linkage based on data; this mimics the logic used within
             # `seaborn.clustermap`. We wrap this in try/except to be on the safe side, as `dendogram` and `linkage` can
             # raise errors for edge cases. And because the improved sort order is not critical to have in place.
-            corr_link = fastcluster.linkage(corr, method="complete", metric="euclidean")
+            corr_link = scipy.cluster.hierarchy.linkage(corr, method="complete", metric="euclidean")
             corr_idx = scipy.cluster.hierarchy.dendrogram(corr_link, no_plot=True)["leaves"]
             corr_cols = corr.columns[corr_idx]
         except Exception:

mostlyai/qa/assets/__init__.py

@@ -29,7 +29,7 @@
 
 
 def read_html_assets() -> dict[str, str]:
-    return {fn: Path(HTML_ASSETS_PATH / fn).read_text(encoding='utf-8') for fn in _HTML_ASSET_FILES}
+    return {fn: Path(HTML_ASSETS_PATH / fn).read_text(encoding="utf-8") for fn in _HTML_ASSET_FILES}
 
 
 def load_tokenizer():

pyproject.toml

@@ -25,17 +25,17 @@ classifiers = [
 ]
 dependencies = [
     "pydantic>=2.0.0,<3.0.0",
-    "numpy>=1.26.3,<2.0.0",
+    "numpy>=1.26.3",
     "scipy>=1.11.0",
     "pandas>=2.0.0",
     "pyarrow>=14.0.0",
     "plotly>=6.0.0",
     "phik>=0.12.4",
-    "fastcluster>=1.2.6",
     "joblib>=1.4.2",
     "Jinja2>=3.1.2",
     "scikit-learn>=1.4.0",
     "sentence-transformers>=3.1.0",
+    "accelerate>=1.5.0",
     "rich>=13.9.4,<14",
     "skops>=0.11.0",
     "torch>=2.6.0",