[Bug Fix] Remove eager re-exports from inference __init__ to avoid he…

[yubofredwang]

These imports are useless and block vllm to run without SGLang deps, removing

[Copilot]

Pull request overview

This PR removes eager package-level re-exports in torchspec.inference (and torchspec.inference.engine) to prevent import-time loading of optional inference backends (notably SGLang), and updates the training entrypoint to import the inference factory function directly.

Changes:

• Update torchspec/train_entry.py to import prepare_inference_engines from torchspec.inference.factory.
• Remove re-exported symbols and __all__ from torchspec/inference/__init__.py.
• Remove re-exported engine symbols and __all__ from torchspec/inference/engine/__init__.py.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
torchspec/train_entry.py	Switches to explicit import from torchspec.inference.factory after removing torchspec.inference re-exports.
torchspec/inference/init.py	Removes inference factory re-exports to avoid eager imports and optional dependency loading.
torchspec/inference/engine/init.py	Removes engine class re-exports to avoid eager imports and optional dependency loading.

Comments suppressed due to low confidence (2)

torchspec/inference/engine/init.py:20

• Dropping all re-exports from torchspec.inference.engine breaks prior imports like from torchspec.inference.engine import VllmEngine/InferenceEngine. If you still want to avoid importing optional dependencies at module import time, a lazy __getattr__-based re-export (optionally guarded with try/except and a targeted ImportError message when sglang/vllm are missing) would preserve the API without the eager dependency loading.

# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.

torchspec/inference/init.py:20

• Removing the re-exports from torchspec.inference is a breaking API change (e.g., from torchspec.inference import prepare_inference_engines will now fail). To avoid eager imports while preserving the public entrypoints, consider using a lazy-export pattern (e.g., module-level __getattr__/__all__, like torchspec.transfer.mooncake.__init__) that imports factory only when the attribute is accessed and can surface a clearer error for missing optional deps.

# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.