release: 0.3.3

Bump version 0.3.0 -> 0.3.3 (0.3.1/0.3.2 were published from PyPI without a
matching repo bump; 0.3.3 is the next free version). Correct requires-python to
>=3.10 (codebase uses dataclass(slots=True)) and add the 0.3.3 changelog entry
covering the PruneCallback dict fix (#39) and faithful group-sensitivity (#40).

Author: nathanhubens

CHANGELOG.md

@@ -2,6 +2,18 @@
 
 <!-- do not remove -->
 
+## 0.3.3
+
+### Bug Fixes
+- `PruneCallback` now accepts a per-layer `dict` for `pruning_ratio` (matching `Pruner`); previously raised `TypeError: '>' not supported between instances of 'dict' and 'int'` (#39)
+- Sensitivity analysis pruning mode now measures each layer by applying its exact per-layer target `{name: level}` — the same operation `Pruner`/`PruneCallback` perform — so the reported Δ faithfully predicts a real prune. Fixes residual/skip-coupled layers (ResNet stem conv, block `conv2`, `downsample`) being falsely ranked "Most Robust" with Δ=0 because they could not be pruned in isolation (#40)
+
+### Enhancements
+- Sensitivity analysis: coupled layers now share a `group_id` (and Δ); layers that cannot be pruned independently (output `Linear`, attention) are marked `prunable=False` and surfaced separately rather than ranked; `analyze(layer_types=...)` restricts the analysis to chosen module types (accepts a single type or a tuple); internal `Pruner` notices are silenced during analysis
+
+### Metadata
+- Corrected `requires-python` to `>=3.10` (the codebase uses `dataclass(slots=True)`)
+
 ## 0.3.0
 
 ### New Features

fasterai/__init__.py

@@ -1 +1 @@
-__version__ = "0.3.0"
+__version__ = "0.3.3"

pyproject.toml

@@ -7,7 +7,7 @@ name = "fasterai"
 dynamic = ["version"]
 description = "A library to make neural networks lighter and faster with fastai"
 readme = "README.md"
-requires-python = ">=3.6"
+requires-python = ">=3.10"
 license = "Apache-2.0"
 authors = [{name = "Nathan Hubens", email = "nathan.hubens@gmail.com"}]
 keywords = ['pruning', 'sparsification', 'knowledge-distillation', 'quantization', 'compression', 'deep-learning']