minor changes

pyproject.toml

@@ -21,7 +21,7 @@ classifiers = [
     "Programming Language :: Python :: 3.13",
 ]
 dependencies = [
-    "hs2p[asap,cucim,openslide,vips]>=3.2.0",
+    "hs2p[asap,cucim,openslide,vips]>=3.2.1",
     "omegaconf",
     "matplotlib",
     "numpy<2",
@@ -88,7 +88,7 @@ fm = [
     "pandas",
     "pillow",
     "rich",
-    "hs2p[asap,cucim,openslide,vips]>=3.2.0",
+    "hs2p[asap,cucim,openslide,vips]>=3.2.1",
     "wandb",
     "torch>=2.3,<2.8",
     "torchvision>=0.18.0",

slide2vec/inference.py

@@ -17,6 +17,7 @@
 import pandas as pd
 import torch
 from hs2p import SlideSpec, FilterConfig, PreviewConfig, SegmentationConfig, TilingConfig, load_tiling_result, tile_slides
+from hs2p.wsi.backend import resolve_backend
 from hs2p.utils.stderr import run_with_filtered_stderr, run_with_filtered_stdio
 import numpy as np
 from transformers.image_processing_utils import BaseImageProcessor
@@ -209,6 +210,25 @@ def _resolve_on_the_fly_num_workers(num_cucim_workers: int) -> tuple[int, str]:
     return effective_num_workers, " // ".join(details)
 
 
+def _log_on_the_fly_worker_override_once(
+    preprocessing: PreprocessingConfig,
+    execution: ExecutionOptions,
+    tiling_results: Sequence[Any],
+) -> None:
+    if not preprocessing.on_the_fly or preprocessing.read_tiles_from is not None:
+        return
+    if not any(_resolve_slide_backend(preprocessing, tiling_result) == "cucim" for tiling_result in tiling_results):
+        return
+    effective_num_workers, worker_context = _resolve_on_the_fly_num_workers(preprocessing.num_cucim_workers)
+    if effective_num_workers == execution.num_workers:
+        return
+    logging.getLogger(__name__).info(
+        f"on-the-fly mode: setting DataLoader num_workers={effective_num_workers} "
+        f"({worker_context}); "
+        f"ignoring speed.num_dataloader_workers={execution.num_workers}"
+    )
+
+
 def _redirect_worker_output() -> None:
     worker_log_path = os.path.join(
         tempfile.gettempdir(),
@@ -358,6 +378,11 @@ def embed_slides(
                 prepared_slides,
                 tiling_results,
             )
+            _log_on_the_fly_worker_override_once(
+                preprocessing,
+                execution,
+                embeddable_tiling_results,
+            )
             _write_zero_tile_embedding_sidecars(
                 zero_tile_pairs,
                 model=model,
@@ -532,6 +557,11 @@ def embed_patients(
                 prepared_slides,
                 tiling_results,
             )
+            _log_on_the_fly_worker_override_once(
+                preprocessing,
+                execution,
+                embeddable_tiling_results,
+            )
             emit_progress("embedding.started", slide_count=len(embeddable_slides))
             loaded = model._load_backend()
 
@@ -651,6 +681,11 @@ def embed_tiles(
     resolved_tiling_results = _normalize_tiling_results(tiling_results, slide_records)
     resolved_preprocessing = _resolve_model_preprocessing(model, preprocessing)
     hierarchical_mode = _is_hierarchical_preprocessing(resolved_preprocessing)
+    _log_on_the_fly_worker_override_once(
+        resolved_preprocessing,
+        execution,
+        resolved_tiling_results,
+    )
     artifacts: list[TileEmbeddingArtifact] | list[HierarchicalEmbeddingArtifact] = []
     for slide, tiling_result in zip(slide_records, resolved_tiling_results):
         if hierarchical_mode:
@@ -811,6 +846,11 @@ def run_pipeline(
             successful_slides,
             tiling_results,
         )
+        _log_on_the_fly_worker_override_once(
+            resolved_preprocessing,
+            execution,
+            embeddable_tiling_results,
+        )
 
         if tiling_only:
             emit_progress(
@@ -1011,6 +1051,11 @@ def run_pipeline_with_coordinates(
             slide_records,
             tiling_results,
         )
+        _log_on_the_fly_worker_override_once(
+            resolved_preprocessing,
+            execution,
+            embeddable_tiling_results,
+        )
         _write_zero_tile_embedding_sidecars(
             zero_tile_pairs,
             model=model,
@@ -1535,13 +1580,7 @@ def _compute_tile_embeddings_for_slide(
     loader_kwargs = _embedding_dataloader_kwargs(loaded, execution)
     resolved_backend = _resolve_slide_backend(preprocessing, tiling_result)
     if preprocessing.on_the_fly and preprocessing.read_tiles_from is None and resolved_backend == "cucim":
-        effective_num_workers, worker_context = _resolve_on_the_fly_num_workers(preprocessing.num_cucim_workers)
-        if effective_num_workers != execution.num_workers:
-            logging.getLogger(__name__).info(
-                f"on-the-fly mode: setting DataLoader num_workers={effective_num_workers} "
-                f"({worker_context}); "
-                f"ignoring speed.num_dataloader_workers={execution.num_workers}"
-            )
+        effective_num_workers, _ = _resolve_on_the_fly_num_workers(preprocessing.num_cucim_workers)
         loader_kwargs["num_workers"] = effective_num_workers
         if effective_num_workers == 0:
             loader_kwargs.pop("persistent_workers", None)
@@ -1620,13 +1659,7 @@ def _compute_hierarchical_embeddings_for_slide(
     loader_kwargs = _embedding_dataloader_kwargs(loaded, execution)
     resolved_backend = _resolve_slide_backend(preprocessing, tiling_result)
     if resolved_backend == "cucim":
-        effective_num_workers, worker_context = _resolve_on_the_fly_num_workers(preprocessing.num_cucim_workers)
-        if effective_num_workers != execution.num_workers:
-            logging.getLogger(__name__).info(
-                f"on-the-fly hierarchical mode: setting DataLoader num_workers={effective_num_workers} "
-                f"({worker_context}); "
-                f"ignoring speed.num_dataloader_workers={execution.num_workers}"
-            )
+        effective_num_workers, _ = _resolve_on_the_fly_num_workers(preprocessing.num_cucim_workers)
         loader_kwargs["num_workers"] = effective_num_workers
         if effective_num_workers == 0:
             loader_kwargs.pop("persistent_workers", None)
@@ -2425,7 +2458,8 @@ def _resolve_device(device: str, default_device):
 
 
 def _describe_device_mode(model, execution: ExecutionOptions) -> str:
-    if model._requested_device == "cpu":
+    requested_device = getattr(model, "_requested_device", None)
+    if requested_device == "cpu":
         return "cpu"
     if execution.num_gpus and execution.num_gpus > 1:
         return f"{execution.num_gpus} gpus"
@@ -2722,6 +2756,19 @@ def _tile_slides(
 ) -> list[Any]:
     _preload_asap_wholeslidedata(preprocessing)
     tiling_cfg, segmentation_cfg, filtering_cfg, preview_cfg, read_coordinates_from, resume = _build_hs2p_configs(preprocessing)
+    for slide in slides:
+        backend_selection = resolve_backend(
+            tiling_cfg.requested_backend,
+            wsi_path=slide.image_path,
+            mask_path=slide.mask_path,
+        )
+        if backend_selection.reason is not None:
+            emit_progress(
+                "backend.selected",
+                sample_id=slide.sample_id,
+                backend=backend_selection.backend,
+                reason=backend_selection.reason,
+            )
 
     def _run_tile_slides():
         return tile_slides(
@@ -2932,7 +2979,8 @@ def ensure_defaults() -> tuple[int, float]:
 
 
 def _validate_multi_gpu_execution(model, execution: ExecutionOptions) -> None:
-    if model._requested_device == "cpu":
+    requested_device = getattr(model, "_requested_device", None)
+    if requested_device == "cpu":
         raise ValueError("ExecutionOptions.num_gpus > 1 is incompatible with device='cpu'")
     if not torch.cuda.is_available():
         raise RuntimeError("ExecutionOptions.num_gpus > 1 requires CUDA")

slide2vec/progress.py

@@ -68,6 +68,13 @@ def write_log(self, message: str, *, stream=None) -> None:
         print(message, file=target, flush=True)
 
 
+def _format_backend_selected_message(payload: dict[str, Any]) -> str:
+    reason = payload.get("reason")
+    if reason:
+        return f"[backend] {payload['sample_id']}: {reason}"
+    return f"[backend] {payload['sample_id']}: using {payload['backend']}"
+
+
 class PlainTextCliProgressReporter:
     def __init__(self, *, stream=None) -> None:
         self.stream = stream or sys.stdout
@@ -133,6 +140,8 @@ def _format_line(self, kind: str, payload: dict[str, Any]) -> str | None:
                 f"Embedding finished: {payload['slides_completed']}/{payload['slide_count']} slides, "
                 f"{payload['tile_artifacts']} tile artifacts, {payload['slide_artifacts']} slide artifacts"
             )
+        if kind == "backend.selected":
+            return _format_backend_selected_message(payload)
         if kind == "run.finished":
             return f"Run finished successfully. Logs: {payload['logs_dir']}"
         if kind == "run.failed":
@@ -313,6 +322,9 @@ def emit(self, event: ProgressEvent) -> None:
                 _embedding_summary_rows(payload),
             )
             return
+        if kind == "backend.selected":
+            self.console.print(_format_backend_selected_message(payload))
+            return
         if kind == "run.finished":
             self._print_summary(
                 "Run Complete",

slide2vec/utils/config.py

@@ -151,14 +151,19 @@ def setup(args):
 def hf_login():
     from huggingface_hub import login
 
-    if "HF_TOKEN" not in os.environ and distributed.is_main_process():
-        hf_token = getpass.getpass(
+    token = os.environ.get("HF_TOKEN")
+    prompted = False
+    if token is None and distributed.is_main_process():
+        token = getpass.getpass(
             "Enter your Hugging Face API token (input will not be visible): "
         )
-        os.environ["HF_TOKEN"] = hf_token
+        os.environ["HF_TOKEN"] = token
+        prompted = True
+    if token is None:
+        return
     if distributed.is_enabled_and_multiple_gpus():
         import torch.distributed as dist
 
         dist.barrier()
-    if distributed.is_main_process():
-        login(os.environ["HF_TOKEN"])
+    if distributed.is_main_process() and prompted:
+        login(token)

tasks/lessons.md

@@ -3,6 +3,11 @@
 ## 2026-04-12
 
 - When refactoring CLI parsing to support `parse_known_args()`, prefer updating the test double to match the real parser API instead of adding a production fallback for mocks. Keep the runtime code clean unless the fallback is genuinely needed by real callers.
+- When a regression test stubs a helper that normally returns a structured object, keep the production code strict and make the stub return the expected attributes instead of broadening the runtime API to accept a placeholder value.
+- When stdout noise comes from both a local wrapper and an upstream dependency, suppress the common-case message at the upstream boundary and keep the per-item local note at debug level instead of stacking multiple INFO lines.
+- When a diagnostic line is still useful, move it to the run boundary and log it once per run instead of per slide or per worker.
+- When a backend-selection reason is user-relevant, emit it as a structured progress event and let the reporter choose how to render it instead of filtering it out in the API.
+- When a CLI depends on both slide2vec and hs2p progress systems, activate them with a shared reporter bridge so upstream events do not disappear into the default null reporter.
 
 ## 2026-04-10
 

tests/test_progress.py

@@ -46,6 +46,9 @@ def __init__(self, file=None, **kwargs):
         def print(self, message, **kwargs):
             self.lines.append((message, kwargs))
 
+        def log(self, message, **kwargs):
+            self.lines.append((message, kwargs))
+
     class FakeProgress:
         def __init__(self, *args, **kwargs):
             self.tasks = {}
@@ -131,6 +134,7 @@ def parse_known_args(self, argv=None):
     assert isinstance(progress.get_progress_reporter(), progress.NullProgressReporter)
 
 
+
 def test_cli_entrypoint_returns_zero(monkeypatch):
     import slide2vec.cli as cli
 

tests/test_regression_core.py

@@ -349,6 +349,26 @@ def test_execution_options_logs_resolved_auto_num_workers(monkeypatch, caplog):
     assert "ExecutionOptions: num_workers=18 (requested=auto)" in caplog.text
     assert "num_workers=auto" not in caplog.text
 
+
+def test_hf_login_skips_hub_login_when_token_is_already_set(monkeypatch):
+    import slide2vec.utils.config as config
+
+    called = False
+
+    def _fake_login(*args, **kwargs):
+        del args, kwargs
+        nonlocal called
+        called = True
+
+    monkeypatch.setenv("HF_TOKEN", "token-from-env")
+    monkeypatch.setattr(config.distributed, "is_main_process", lambda: True)
+    monkeypatch.setattr(config.distributed, "is_enabled_and_multiple_gpus", lambda: False)
+    monkeypatch.setattr("huggingface_hub.login", _fake_login)
+
+    config.hf_login()
+
+    assert called is False
+
 def test_execution_options_from_config_maps_cli_fields(tmp_path: Path):
     cfg = SimpleNamespace(
         output_dir=str(tmp_path),