tracebloc 4.0.2__tar.gz → 4.0.3__tar.gz

{tracebloc-4.0.2 → tracebloc-4.0.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tracebloc
-Version: 4.0.2
+Version: 4.0.3
 Summary: Build AI with anyone. On data that can't move. SDK for the tracebloc collaborative AI workspace.
 Author-email: tracebloc <lukas@tracebloc.io>
 License-Expression: MIT

{tracebloc-4.0.2 → tracebloc-4.0.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "tracebloc"
-version = "4.0.2"
+version = "4.0.3"
 description = "Build AI with anyone. On data that can't move. SDK for the tracebloc collaborative AI workspace."
 license = "MIT"
 license-files = ["license.txt"]

{tracebloc-4.0.2 → tracebloc-4.0.3}/tracebloc/linking/model_dataset.py

@@ -120,7 +120,6 @@ class LinkModelDataset(DataPlanMethods, TrainingPlanMethods, AugmentationMethods
                 "image_attributes",
                 "class_order",
                 "conversion",
-                "dataset_sequence_length",
                 "max_batch_size",
             },
         )
@@ -183,7 +182,6 @@ class LinkModelDataset(DataPlanMethods, TrainingPlanMethods, AugmentationMethods
             category=self._category,
             batch_size=self._batch_size,
             resources=self._resources,
-            dataset_length=kwargs.get("dataset_sequence_length"),
         )
         if self._converted.sequence_length is not None:
             kwargs["data_shape"] = self._converted.sequence_length.value

{tracebloc-4.0.2 → tracebloc-4.0.3}/tracebloc/linking/regeneration.py

@@ -73,9 +73,9 @@ def dataset_counts(
 ) -> dict[str, int]:
     """The counts the link answer gives for the dataset, by :data:`COUNT_NAMES`.
 
-    Classes for a task that has them, the feature count from the match point
-    the model was compared on, and the sequence length and vocabulary when the
-    answer carries them.
+    Classes for a task that has them, and the feature count from the match point
+    the model was compared on. The answer carries no sequence length or
+    vocabulary size, so a package keeps its own.
     """
     counts: dict[str, int] = {}
     if category not in NO_OUTPUT_CLASSES:
@@ -89,36 +89,18 @@ def dataset_counts(
             expected = point.get("expected") if isinstance(point, Mapping) else None
             if _positive_int(expected) is not None:
                 counts["num_feature_points"] = expected
-    for name, key in (
-        ("sequence_length", "sequenceLength"),
-        ("vocab_size", "vocabSize"),
-    ):
-        if _positive_int(content.get(key)) is not None:
-            counts[name] = content[key]
     return counts
 
 
 def changed_counts(
     header: Mapping[str, SourcedValue], dataset: Mapping[str, int]
 ) -> dict[str, tuple[Any, int]]:
-    """``{name: (package's, dataset's)}`` for each count the package must take again.
-
-    A sequence length is taken again only when the package's is longer than the
-    dataset's sequences: a shorter one is the model's own limit, which a
-    rebuild keeps.
-    """
+    """``{name: (package's, dataset's)}`` for each count the package must take again."""
     changed = {}
     for name, wanted in dataset.items():
         declared = header.get(name)
-        if declared is None:
-            continue
-        have = declared.value
-        if name == "sequence_length":
-            differs = isinstance(have, int) and have > wanted
-        else:
-            differs = have != wanted
-        if differs:
-            changed[name] = (have, wanted)
+        if declared is not None and declared.value != wanted:
+            changed[name] = (declared.value, wanted)
     return changed
 
 

{tracebloc-4.0.2 → tracebloc-4.0.3}/tracebloc/linking/response.py

@@ -241,12 +241,6 @@ def _render_mismatch_diagnostics(
         # A one-output head is 1-wide whatever the dataset's bucket count.
         width = 1 if category in SINGLE_OUTPUT else content.get("dataset_classes")
         expected_msg += f"  output_classes : {width}\n"
-    if "datasetShape" in content:
-        expected_msg += f"  shape          : {content['datasetShape']}\n"
-    if "sequenceLength" in content:
-        expected_msg += f"  sequence_length: {content['sequenceLength']}\n"
-    if "vocabSize" in content:
-        expected_msg += f"  vocab_size     : {content['vocabSize']}\n"
     keys_to_match = list(features.keys()) if features is not None else None
     if "match_points" in content and keys_to_match:
         match_points = content["match_points"]
@@ -269,8 +263,6 @@ def _render_mismatch_diagnostics(
     model_msg = ""
     if category not in NO_OUTPUT_CLASSES:
         model_msg += f"  output_classes : {content.get('output_class')}\n"
-    if "inputShape" in content:
-        model_msg += f"  shape          : {content['inputShape']}\n"
     for key_to_match in compared_keys(content, keys_to_match):
         if category in [
             TABULAR_CLASSIFICATION,
@@ -317,9 +309,6 @@ class LinkSuccessState(NamedTuple):
     #: output ``i``), or ``None`` when the backend sent none. The training plan
     #: prints it in place of ``allClasses`` (backend#5686).
     class_order: object = None
-    #: The dataset's sequence length (``sequenceLength``), or ``None`` when the
-    #: backend sent none. A converted text model's plan trains at most this long.
-    sequence_length: object = None
 
 
 def _apply_link_success(
@@ -495,9 +484,6 @@ def _apply_link_success(
     print("Please set a training plan.")
     resources = content.get("resources")
     class_order = content.get("class_order")
-    sequence_length = content.get("sequenceLength")
-    if isinstance(sequence_length, bool) or not isinstance(sequence_length, int):
-        sequence_length = None
     return LinkSuccessState(
         total_images=total_images,
         num_classes=num_classes,
@@ -508,5 +494,4 @@ def _apply_link_success(
         resources=resources if isinstance(resources, dict) else None,
         image_attributes=image_attributes,
         class_order=class_order if class_order_text(class_order) is not None else None,
-        sequence_length=sequence_length,
     )

{tracebloc-4.0.2 → tracebloc-4.0.3}/tracebloc/linking/user_methods.py

@@ -156,7 +156,6 @@ class UserLinkMethods:
             image_attributes=self._User__image_attributes,
             class_order=self._User__class_order,
             conversion=self._User__conversion,
-            dataset_sequence_length=self._User__dataset_sequence_length,
         )
 
     def _User__prepare_stored_model(  # noqa: N802 — User's mangled name, kept (RFC-0348 D.5)
@@ -267,7 +266,6 @@ class UserLinkMethods:
             self._User__resources = state.resources
             self._User__image_attributes = state.image_attributes
             self._User__class_order = state.class_order
-            self._User__dataset_sequence_length = state.sequence_length
             return
 
         raise DatasetLinkingError(

{tracebloc-4.0.2 → tracebloc-4.0.3}/tracebloc/training/converted.py

@@ -100,14 +100,13 @@ def text_sequence_length(
     category: str,
     batch_size: Any,
     resources: Any,
-    dataset_length: Any,
     lora_enabled: bool,
     tokenizer_path: Any = None,
 ) -> Any:
     """The length a converted text model trains at (``SequenceLength``).
 
     The model's own limit, bounded by the memory estimate within the client's
-    per-run memory, or the dataset's length when that is shorter.
+    per-run memory.
     """
     from tracebloc.upload.sequence_length import (
         choose_sequence_length,
@@ -123,7 +122,7 @@ def text_sequence_length(
         memory_bytes=limit[0] if limit else None,
         lora_enabled=lora_enabled,
     )
-    return choose_sequence_length(resolved, dataset_length)
+    return choose_sequence_length(resolved)
 
 
 @dataclass(frozen=True)
@@ -155,7 +154,6 @@ def converted_plan(
     category: str,
     batch_size: Any,
     resources: Any,
-    dataset_length: Any,
 ) -> ConvertedPlan:
     """The :class:`ConvertedPlan` for the upload ``conversion`` produced."""
     from tracebloc.training.lora import LORA_CATEGORIES, lora_by_default
@@ -174,7 +172,6 @@ def converted_plan(
                 category=category,
                 batch_size=batch_size,
                 resources=resources,
-                dataset_length=dataset_length,
                 lora_enabled=lora,
                 tokenizer_path=getattr(written, "tokenizer", None),
             )

{tracebloc-4.0.2 → tracebloc-4.0.3}/tracebloc/user.py

@@ -148,7 +148,6 @@ class User(UserUploadMethods, UserLinkMethods):
         ``tracebloc.checks.trains`` can resolve a model file exactly the way
         ``upload_model`` does without a session (``_for_local_checks``)."""
         self.__ext = ".py"
-        self.__dataset_sequence_length = None  # the link answer's sequenceLength
         self.__data_shape = 224
         self.__batch_size = 16
         self.__max_batch_size = None  # the model file's, for memory sizing

{tracebloc-4.0.2 → tracebloc-4.0.3}/tracebloc.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tracebloc
-Version: 4.0.2
+Version: 4.0.3
 Summary: Build AI with anyone. On data that can't move. SDK for the tracebloc collaborative AI workspace.
 Author-email: tracebloc <lukas@tracebloc.io>
 License-Expression: MIT