code-loader 1.0.191.dev0__tar.gz → 1.0.192__tar.gz

{code_loader-1.0.191.dev0 → code_loader-1.0.192}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: code-loader
-Version: 1.0.191.dev0
+Version: 1.0.192
 Summary: 
 Home-page: https://github.com/tensorleap/code-loader
 License: MIT

{code_loader-1.0.191.dev0 → code_loader-1.0.192}/code_loader/contract/datasetclasses.py

@@ -92,6 +92,16 @@ SectionCallableInterface = Callable[[Union[int, str], PreprocessResponse], npt.N
 InstanceCallableInterface = Callable[[Union[int, str], PreprocessResponse, int], Optional[ElementInstance]]
 InstanceLengthCallableInterface = Callable[[Union[int, str], PreprocessResponse], int]
 
+# (sample_id, prev_inputs, prev_outputs, preprocess) -> next model inputs, or None to end the chain.
+# First call per chain receives prev_inputs=None, prev_outputs=None and returns the initial inputs.
+# Keys starting with '_' are passthrough state: never fed to the model, handed back verbatim in
+# prev_inputs on the next call. The hook must be stateless and deterministically seeded (the engine
+# may replay a step after crash recovery and relies on identical results).
+AutoregressiveStepCallableInterface = Callable[
+    [Union[int, str], Optional[Dict[str, npt.NDArray[np.float32]]], Optional[Dict[str, npt.NDArray[np.float32]]],
+     PreprocessResponse],
+    Optional[Dict[str, npt.NDArray[np.float32]]]]
+
 MetadataSectionCallableInterface = Union[
     Callable[[Union[int, str], PreprocessResponse], int],
     Callable[[Union[int, str], PreprocessResponse], Dict[str, int]],
@@ -239,6 +249,23 @@ class CustomLatentSpaceHandler:
     function: SectionCallableInterface
     name: str = 'custom_latent_space'
 
+
+# How a chain's latent-space vectors are derived from its per-step forward passes.
+# 'last_step': every latent space comes from the final step's forward pass, except input-kind
+# latent spaces which come from the first step (the original sample, before generated content
+# dominates the model inputs). 'mean': every latent space is the elementwise mean over all steps.
+AUTOREGRESSIVE_LATENT_SPACE_AGGREGATIONS = ('last_step', 'mean')
+
+
+@dataclass
+class AutoregressiveStepHandler:
+    function: AutoregressiveStepCallableInterface
+    name: str = 'autoregressive_step'
+    # Per-model-input shapes (unbatched, underscore passthrough keys excluded), discovered by running
+    # the hook's first call at parse time. Fills the role InputHandler.shape plays for input encoders.
+    input_shapes: Optional[Dict[str, List[int]]] = None
+    latent_space_aggregation: str = 'last_step'
+
 @dataclass
 class PredictionTypeHandler:
     name: str
@@ -277,8 +304,9 @@ class DatasetIntegrationSetup:
     metrics: List[MetricHandler] = field(default_factory=list)
     instance_metrics: List[InstanceMetricHandler] = field(default_factory=list)
     custom_layers: Dict[str, CustomLayerHandler] = field(default_factory=dict)
-    custom_latent_spaces: Dict[str, CustomLatentSpaceHandler] = field(default_factory=dict)
+    custom_latent_space: Optional[CustomLatentSpaceHandler] = None
     simulations: List[SimulationHandler] = field(default_factory=list)
+    autoregressive_step: Optional[AutoregressiveStepHandler] = None
 
 
 @dataclass
@@ -289,6 +317,6 @@ class DatasetSample:
     metadata_is_none: Dict[str, bool]
     index: Union[int, str]
     state: DataStateEnum
-    custom_latent_spaces: Optional[Dict[str, npt.NDArray[np.float32]]] = None
+    custom_latent_space: Optional[npt.NDArray[np.float32]] = None
     instance_masks: Optional[Dict[str, ElementInstance]] = None
 

{code_loader-1.0.191.dev0 → code_loader-1.0.192}/code_loader/inner_leap_binder/leapbinder.py

@@ -13,7 +13,8 @@ from code_loader.contract.datasetclasses import SectionCallableInterface, InputH
     CustomMultipleReturnCallableInterfaceMultiArgs, DatasetBaseHandler, custom_latent_space_attribute, \
     RawInputsForHeatmap, VisualizerHandlerData, MetricHandlerData, CustomLossHandlerData, SamplePreprocessResponse, \
     ElementInstanceMasksHandler, InstanceCallableInterface, CustomLatentSpaceHandler, InstanceMetricHandler, \
-    SimulationHandler, _simulation_context
+    SimulationHandler, _simulation_context, AutoregressiveStepHandler, AutoregressiveStepCallableInterface, \
+    AUTOREGRESSIVE_LATENT_SPACE_AGGREGATIONS
 from code_loader.contract.enums import LeapDataType, DataStateEnum, DataStateType, MetricDirection, DatasetMetadataType
 from code_loader.contract.mapping import NodeConnection, NodeMapping, NodeMappingType
 from code_loader.contract.responsedataclasses import DatasetTestResultPayload, LeapAnalysisConfiguration
@@ -511,34 +512,46 @@ class LeapBinder:
         """
         self.setup_container.metadata.append(MetadataHandler(name, function, metadata_type))
 
-    def set_custom_latent_space(self, function: SectionCallableInterface,
-                                name: Optional[str] = None) -> None:
+    def set_custom_latent_space(self, function: SectionCallableInterface) -> None:
         """
-        Register a custom latent space function.
-
-        Multiple custom latent spaces may be registered as long as each has a
-        unique name. They are stored name-keyed (insertion order preserved), so a
-        later registration no longer overrides an earlier one.
+        Set a custom latent space function.
 
         Args:
-        function (SectionCallableInterface): The latent-space handler function.
+        function (SectionCallableInterface): The metadata handler function.
             This function receives:
                 subset (PreprocessResponse): The subset of the data.
                 index (int): The index of the sample within the subset.
-            This function should return a numpy float32 array containing the latent
-            space vec of the sample.
-        name (Optional[str]): Unique name for this custom latent space. Defaults to
-            the reserved single-LS name for backward compatibility.
+            This function should numpy float32 array contains the latent space vec of the sample.
         """
-        if name is None:
-            name = custom_latent_space_attribute
-        if name in self.setup_container.custom_latent_spaces:
-            raise Exception(
-                f"A custom latent space named '{name}' is already registered. Each "
-                f"@tensorleap_custom_latent_space must have a unique name "
-                f"(pass name='...' to distinguish them)."
-            )
-        self.setup_container.custom_latent_spaces[name] = CustomLatentSpaceHandler(function, name)
+        self.setup_container.custom_latent_space = CustomLatentSpaceHandler(function)
+
+    def set_autoregressive_step(self, function: AutoregressiveStepCallableInterface,
+                                latent_space_aggregation: str = 'last_step') -> None:
+        """
+        Set the autoregressive step hook — the feedback function that drives a chain:
+        it supplies the model's initial inputs on its first call (prev_inputs=None, prev_outputs=None)
+        and turns each step's inputs/outputs into the next step's inputs until it returns None.
+        An autoregressive integration has no input encoders; the hook is the sole input source.
+        latent_space_aggregation declares how the chain's latent-space vectors are derived from
+        its steps (see AUTOREGRESSIVE_LATENT_SPACE_AGGREGATIONS).
+        """
+        if self.setup_container.autoregressive_step is not None:
+            raise Exception('tensorleap_autoregressive_step is already defined. '
+                            'Only one autoregressive step hook is allowed per integration.')
+        if latent_space_aggregation not in AUTOREGRESSIVE_LATENT_SPACE_AGGREGATIONS:
+            raise Exception(f'tensorleap_autoregressive_step: unknown latent_space_aggregation '
+                            f'{latent_space_aggregation!r}. Supported values: '
+                            f'{", ".join(AUTOREGRESSIVE_LATENT_SPACE_AGGREGATIONS)}.')
+        self.setup_container.autoregressive_step = AutoregressiveStepHandler(
+            function, latent_space_aggregation=latent_space_aggregation)
+
+        # Builtin chain metadata, declared at parse time so it survives the reporter's
+        # metadata type mapping; the placeholder values are overwritten by the engine when a
+        # chain finalizes (realized length, truncated-by-safety-cap flag).
+        def builtin_chain_metadata(idx: Any, preprocess: PreprocessResponse) -> Dict[str, Any]:
+            return {'length': -1, 'truncated': False}
+
+        self.set_metadata(builtin_chain_metadata, 'builtin_chain')
 
     def set_custom_layer(self, custom_layer: Type[Any], name: str, inspect_layer: bool = False,
                          kernel_index: Optional[int] = None, use_custom_latent_space: bool = False) -> None:
@@ -750,6 +763,7 @@ class LeapBinder:
                     )
 
     def check(self) -> None:
+        self.validate_autoregressive_setup()
         preprocess_result = self.get_preprocess_result()
         self.check_preprocess(preprocess_result)
         self.check_handlers(preprocess_result)
@@ -757,6 +771,43 @@ class LeapBinder:
         self.validate_ignore_latent_spaces()
         print("Successful!")
 
+    def validate_autoregressive_setup(self) -> None:
+        """Structural rules for autoregressive integrations (no-op when the hook is absent).
+
+        Order-independent: runs after the whole script registered, because decorators may
+        appear in any order in the integration file.
+        """
+        if self.setup_container.autoregressive_step is None:
+            return
+        if self.setup_container.inputs:
+            raise Exception(
+                "An autoregressive integration must not declare input encoders — "
+                "tensorleap_autoregressive_step is the sole source of model inputs "
+                "(its first call returns the initial inputs). Found input encoders: "
+                f"{[handler.name for handler in self.setup_container.inputs]}.")
+        if self.setup_container.simulations:
+            raise Exception(
+                "Simulations are not supported together with tensorleap_autoregressive_step yet: "
+                "the autoregressive entry points do not resolve synthetic sample ids. Remove the "
+                "simulation decorators or the autoregressive hook.")
+        if self.setup_container.instance_masks:
+            raise Exception(
+                "Element instances are not supported together with tensorleap_autoregressive_step: "
+                "instance generation masks the sample's encoded inputs, which do not exist in an "
+                "autoregressive integration. Remove the instance encoders or the autoregressive hook.")
+        if self.setup_container.custom_latent_space is not None:
+            raise Exception(
+                "tensorleap_custom_latent_space is not supported together with "
+                "tensorleap_autoregressive_step: the custom latent space function only sees "
+                "(sample_id, preprocess) — it runs before generation and cannot observe the chain. "
+                "The chain's latent is the final step's model latent.")
+        if not self.setup_container.prediction_types:
+            raise Exception(
+                "An autoregressive integration must declare prediction types "
+                "(tensorleap_load_model(prediction_types=[...])): the engine keys each step's "
+                "model outputs by the declared names to build the hook's prev_outputs dict. "
+                "Without them, prev_outputs would be empty at runtime.")
+
     def validate_ignore_latent_spaces(self) -> None:
         """Validate leap_analysis_configuration.ignore_latent_spaces against this binder.
 

{code_loader-1.0.191.dev0 → code_loader-1.0.192}/code_loader/inner_leap_binder/leapbinder_decorators.py

@@ -21,12 +21,12 @@ from code_loader.contract.datasetclasses import CustomCallableInterfaceMultiArgs
     CustomMultipleReturnCallableInterfaceMultiArgs, ConfusionMatrixCallableInterfaceMultiArgs, CustomCallableInterface, \
     VisualizerCallableInterface, MetadataSectionCallableInterface, PreprocessResponse, SectionCallableInterface, \
     ConfusionMatrixElement, SamplePreprocessResponse, PredictionTypeHandler, InstanceCallableInterface, ElementInstance, \
-    InstanceLengthCallableInterface
+    InstanceLengthCallableInterface, AutoregressiveStepCallableInterface
 from code_loader.contract.enums import MetricDirection, LeapDataType, DatasetMetadataType, DataStateType
 from code_loader import leap_binder, LeapLoader
 from code_loader.contract.mapping import NodeMapping, NodeMappingType, NodeConnection
 from code_loader.contract.visualizer_classes import LeapImage, LeapImageMask, LeapTextMask, LeapText, LeapGraph, \
-    LeapHorizontalBar, LeapImageWithBBox, LeapImageWithHeatmap, LeapVideo
+    LeapHorizontalBar, LeapImageWithBBox, LeapImageWithHeatmap, LeapVideo, LeapValidationError
 from code_loader.inner_leap_binder.leapbinder import mapping_runtime_mode_env_var_mame
 from code_loader.mixpanel_tracker import clear_integration_events, AnalyticsEvent, emit_integration_event_once
 
@@ -197,11 +197,11 @@ def validate_output_structure(result, func_name: str, expected_type_name="np.nda
 def batch_warning(result, func_name):
     if len(result.shape) > 0 and result.shape[0] == 1:
         warnings.warn(
-            f"{func_name} warning: Tensorleap will add a batch dimension at axis 0 to the output of {func_name}, "
-            f"although the detected size of axis 0 is already 1. "
-            f"This may lead to an extra batch dimension (e.g., shape (1, 1, ...)). "
-            f"Please ensure that the output of '{func_name}' is not already batched "
-            f"to avoid computation errors."
+            f"{func_name} warning: Tensorleap will add a batch dimension at axis 0 of "
+            f"{func_name}'s output, but axis 0 is already size 1 — this can produce a "
+            f"doubly-batched shape (e.g. (1, 1, ...)). If axis 0 is a real dimension "
+            f"(e.g. a single channel) and not an accidental batch, you can ignore this; "
+            f"only fix it if you pre-batched the output."
         )
 
 
@@ -316,16 +316,16 @@ def tensorleap_integration_test():
                     first_tb = traceback.extract_tb(e.__traceback__)[-1]
                     file_name = Path(first_tb.filename).name
                     line_number = first_tb.lineno
-                    if isinstance(e, TypeError) and 'is not subscriptable' in str(e):
-                        update_env_params_func("code_mapping", "x")
-                        raise (f'Invalid integration code. File {file_name}, line {line_number}: '
-                               f"indexing is supported only on the model's predictions inside the integration test. Please remove this indexing operation usage from the integration test code.")
+                    update_env_params_func("code_mapping", "x")
+                    if isinstance(e, LeapValidationError):
+                        raise Exception(f'Invalid integration code. File {file_name}, line {line_number}: {e}')
+                    elif isinstance(e, TypeError) and 'is not subscriptable' in str(e):
+                        raise Exception(f'Invalid integration code. File {file_name}, line {line_number}: '
+                                        f"indexing is supported only on the model's predictions inside the integration test. Please remove this indexing operation usage from the integration test code.")
                     else:
-                        update_env_params_func("code_mapping", "x")
-
-                        raise (f'Invalid integration code. File {file_name}, line {line_number}: '
-                               f'Integration test is only allowed to call Tensorleap decorators. '
-                               f'Ensure any arithmetics, external library use, Python logic is placed within Tensorleap decoders')
+                        raise Exception(f'Invalid integration code. File {file_name}, line {line_number}: '
+                                        f'Integration test is only allowed to call Tensorleap decorators. '
+                                        f'Ensure any arithmetics, external library use, Python logic is placed within Tensorleap decoders')
                 finally:
                     if mapping_runtime_mode_env_var_mame in os.environ:
                         del os.environ[mapping_runtime_mode_env_var_mame]
@@ -349,6 +349,14 @@ def _safe_get_item(key):
 
 
 def tensorleap_load_model(prediction_types: Optional[List[PredictionTypeHandler]] = _UNSET):
+    """Register the model-loading function.
+
+    ``prediction_types``: declare one ``PredictionTypeHandler`` per model output, in
+    output order. Multi-output models must declare an entry for EVERY output — the
+    count is validated against the model's outputs — so for an output you don't consume
+    as a prediction, pass a throwaway handler to satisfy the count. Omit the argument
+    only when the model has a single prediction output.
+    """
     prediction_types_was_provided = prediction_types is not _UNSET
 
     if not prediction_types_was_provided:
@@ -603,7 +611,9 @@ def tensorleap_load_model(prediction_types: Optional[List[PredictionTypeHandler]
 
                 # onnx runtime interface
                 def run(self, output_names, input_dict):
-                    assert output_names is None
+                    if output_names is not None:
+                        raise LeapValidationError("Tensorleap doesn't support selecting outputs by name — "
+                                                  "use model.run(None, inputs) to return all outputs.")
                     assert isinstance(input_dict, dict), \
                         f'Expected input_dict to be a dict, got {type(input_dict)} instead.'
                     for i, (input_key, elem) in enumerate(input_dict.items()):
@@ -1568,9 +1578,8 @@ def tensorleap_metadata(
     return decorating_function
 
 
-def tensorleap_custom_latent_space(name: Optional[str] = None):
+def tensorleap_custom_latent_space():
     def decorating_function(user_function: SectionCallableInterface):
-        ls_name = name if name is not None else user_function.__name__
         def _validate_input_args(sample_id: Union[int, str], preprocess_response: PreprocessResponse):
             assert isinstance(sample_id, (int, str)), \
                 (f'tensorleap_custom_latent_space validation failed: '
@@ -1590,9 +1599,9 @@ def tensorleap_custom_latent_space(name: Optional[str] = None):
             if result.ndim > 1:
                 flat_dim = int(np.prod(result.shape))
                 store_general_warning(
-                    key=("tensorleap_custom_latent_space_flatten", ls_name, tuple(result.shape)),
+                    key=("tensorleap_custom_latent_space_flatten", tuple(result.shape)),
                     message=(
-                        f"tensorleap_custom_latent_space '{ls_name}' returned per-sample shape {tuple(result.shape)} "
+                        f"tensorleap_custom_latent_space returned per-sample shape {tuple(result.shape)} "
                         f"(ndim={result.ndim}). Tensorleap assumes per-sample shape (d, ...) and will "
                         f"flatten to ({flat_dim},) before downstream visualization and clustering. "
                         f"If you want a different aggregation (e.g. global average pooling), do it "
@@ -1611,7 +1620,7 @@ def tensorleap_custom_latent_space(name: Optional[str] = None):
 
             return result
 
-        leap_binder.set_custom_latent_space(inner_without_validate, ls_name)
+        leap_binder.set_custom_latent_space(inner_without_validate)
 
         def inner(sample_id, preprocess_response):
             if os.environ.get(mapping_runtime_mode_env_var_mame):
@@ -1629,6 +1638,246 @@ def tensorleap_custom_latent_space(name: Optional[str] = None):
     return decorating_function
 
 
+def tensorleap_autoregressive_step(latent_space_aggregation: str = 'last_step'):
+    """The feedback hook that drives an autoregressive chain.
+
+    Signature of the decorated function:
+        (sample_id, prev_inputs, prev_outputs, preprocess) -> Optional[Dict[str, np.ndarray]]
+
+    First call per chain: prev_inputs=None, prev_outputs=None — returns the initial model inputs.
+    Later calls receive the previous step's model inputs and outputs (unbatched, per-sample) and
+    return the next step's full input dict; returning None ends the chain. Keys starting with '_'
+    are passthrough state: never fed to the model, handed back verbatim in prev_inputs next call.
+    The hook must be stateless and deterministically seeded (seed stochasticity from sample_id) —
+    the engine replays steps after crash recovery and relies on identical results.
+
+    latent_space_aggregation ('last_step' | 'mean') declares how the chain's latent-space
+    vectors are derived from its per-step forward passes. 'last_step' (default): every latent
+    space comes from the final step's forward pass, except input-kind latent spaces which come
+    from the first step — the original sample, before generated content dominates the model
+    inputs. 'mean': every latent space is the elementwise mean over all steps of the chain.
+    """
+    assert isinstance(latent_space_aggregation, str), \
+        ('tensorleap_autoregressive_step must be called with parentheses: '
+         '@tensorleap_autoregressive_step() or '
+         "@tensorleap_autoregressive_step(latent_space_aggregation='mean').")
+
+    def decorating_function(user_function: AutoregressiveStepCallableInterface):
+        # Contract: shapes and key set are fixed across all calls AND all chains (compiled graphs
+        # need static shapes), so one record shared across samples is the correct scope.
+        first_call_record: Dict[str, Any] = {'keys': None, 'shapes': None}
+
+        def _normalize_args(args, kwargs):
+            expected_names = ["sample_id", "prev_inputs", "prev_outputs", "preprocess"]
+            validate_args_structure(
+                *args,
+                types_order=[Union[int, str], Union[dict, type(None)], Union[dict, type(None)],
+                             PreprocessResponse],
+                func_name=user_function.__name__, expected_names=expected_names, **kwargs)
+            normalized = []
+            for i, name in enumerate(expected_names):
+                normalized.append(args[i] if i < len(args) else kwargs[name])
+            return normalized
+
+        def _validate_input_args(sample_id, prev_inputs, prev_outputs, preprocess_response):
+            assert (prev_inputs is None) == (prev_outputs is None), \
+                (f'{user_function.__name__}() validation failed: '
+                 f'prev_inputs and prev_outputs must both be None (first call) or both be dicts '
+                 f'(subsequent calls). Got prev_inputs={type(prev_inputs).__name__}, '
+                 f'prev_outputs={type(prev_outputs).__name__}.')
+            assert type(sample_id) == preprocess_response.sample_id_type, \
+                (f'{user_function.__name__}() validation failed: '
+                 f'Argument sample_id should be as the same type as defined in the preprocess response '
+                 f'{preprocess_response.sample_id_type}. Got {type(sample_id)}.')
+
+        def _validate_result(result, is_first_call):
+            if result is None:
+                assert not is_first_call, \
+                    (f'{user_function.__name__}() validation failed: '
+                     f'the first call (prev_inputs=None, prev_outputs=None) must return the initial '
+                     f'model inputs dict — returning None on the first call would produce an empty chain.')
+                return
+            assert isinstance(result, dict), \
+                (f'{user_function.__name__}() validation failed: '
+                 f'the return value must be a dict of named model input tensors, or None to end the '
+                 f'chain. Got {type(result)}.')
+            model_input_keys = [key for key in result if not key.startswith('_')]
+            assert model_input_keys, \
+                (f'{user_function.__name__}() validation failed: '
+                 f'the returned dict contains no model inputs — every key starts with "_" '
+                 f'(underscore keys are passthrough state, never fed to the model).')
+            for key, value in result.items():
+                assert isinstance(key, str), \
+                    (f'{user_function.__name__}() validation failed: '
+                     f'all keys in the returned dict must be strings. Got {type(key)}.')
+                if key.startswith('_'):
+                    continue
+                assert isinstance(value, np.ndarray), \
+                    (f'{user_function.__name__}() validation failed: '
+                     f'value for model input "{key}" must be a numpy array. Got {type(value)}.')
+
+        def _validate_consistency(result):
+            # Rules 2-3 of the autoregressive contract: identical key set and identical
+            # (non-passthrough) shapes on every call — compiled graphs need static shapes.
+            if result is None:
+                return
+            keys = set(result.keys())
+            shapes = {key: tuple(value.shape) for key, value in result.items()
+                      if not key.startswith('_') and isinstance(value, np.ndarray)}
+            if first_call_record['keys'] is None:
+                first_call_record['keys'] = keys
+                first_call_record['shapes'] = shapes
+                return
+            assert keys == first_call_record['keys'], \
+                (f'{user_function.__name__}() validation failed: '
+                 f'every call must return the same key set as the first call. '
+                 f'First call keys: {sorted(first_call_record["keys"])}, this call: {sorted(keys)}.')
+            for key, shape in shapes.items():
+                assert shape == first_call_record['shapes'][key], \
+                    (f'{user_function.__name__}() validation failed: '
+                     f'tensor shapes must be identical on every call (fixed-length padding + mask for '
+                     f'growing sequences). Input "{key}" was {first_call_record["shapes"][key]} on the '
+                     f'first call and {shape} now.')
+
+        def _assert_deterministic(first_result, second_result):
+            error_message = (
+                f'{user_function.__name__}() validation failed: '
+                f'the hook is not deterministic — two calls with identical arguments returned different '
+                f'outputs. Seed any stochasticity (e.g. initial noise) from sample_id: the engine '
+                f'replays steps after crash recovery and relies on identical results.')
+            if first_result is None or second_result is None:
+                assert first_result is None and second_result is None, error_message
+                return
+            assert set(first_result.keys()) == set(second_result.keys()), error_message
+            for key, value in first_result.items():
+                if key.startswith('_') or not isinstance(value, np.ndarray):
+                    continue
+                assert np.array_equal(value, second_result[key]), error_message
+
+        def _squeeze_test_batch_dim(tensors_dict):
+            # Inside the integration test the model works on batch-1 tensors (this wrapper adds the
+            # batch axis to returned inputs, and model outputs come back batched). The user's hook
+            # body must see the same unbatched per-sample tensors it will see at engine runtime, so
+            # strip the leading batch-1 axis before calling it.
+            if tensors_dict is None:
+                return None
+            squeezed = {}
+            for key, value in tensors_dict.items():
+                if not key.startswith('_') and isinstance(value, np.ndarray) \
+                        and value.ndim > 0 and value.shape[0] == 1:
+                    squeezed[key] = value[0]
+                else:
+                    squeezed[key] = value
+            return squeezed
+
+        def inner_without_validate(sample_id, prev_inputs, prev_outputs, preprocess_response):
+            global _called_from_inside_tl_decorator
+            _called_from_inside_tl_decorator += 1
+            try:
+                result = user_function(sample_id, prev_inputs, prev_outputs, preprocess_response)
+            finally:
+                _called_from_inside_tl_decorator -= 1
+            return result
+
+        leap_binder.set_autoregressive_step(inner_without_validate,
+                                            latent_space_aggregation=latent_space_aggregation)
+
+        def inner(*args, **kwargs):
+            if not _call_from_tl_platform:
+                set_current('tensorleap_autoregressive_step')
+            sample_id, prev_inputs, prev_outputs, preprocess_response = _normalize_args(args, kwargs)
+
+            is_top_level_in_test = (_called_from_inside_tl_decorator == 0
+                                    and _called_from_inside_tl_integration_test_decorator)
+            if is_top_level_in_test:
+                prev_inputs = _squeeze_test_batch_dim(prev_inputs)
+                prev_outputs = _squeeze_test_batch_dim(prev_outputs)
+
+            _validate_input_args(sample_id, prev_inputs, prev_outputs, preprocess_response)
+
+            result = inner_without_validate(sample_id, prev_inputs, prev_outputs, preprocess_response)
+
+            if is_top_level_in_test:
+                # Rule 7: determinism double-call. An unseeded sampler fails here, on the user's
+                # machine, instead of silently forking a chain on engine crash-replay.
+                second_result = inner_without_validate(sample_id, prev_inputs, prev_outputs,
+                                                       preprocess_response)
+                _assert_deterministic(result, second_result)
+
+            _validate_result(result, is_first_call=prev_inputs is None)
+            _validate_consistency(result)
+
+            if is_top_level_in_test and result is not None:
+                result = {key: (np.expand_dims(value, axis=0)
+                                if not key.startswith('_') and isinstance(value, np.ndarray) else value)
+                          for key, value in result.items()}
+
+            if not _call_from_tl_platform:
+                update_env_params_func("tensorleap_autoregressive_step", "v")
+            return result
+
+        class _MappingInputsPlaceholder:
+            """Mapping-mode stand-in for the hook's returned dict.
+
+            Indexing it by a model-input name yields a placeholder whose node_mapping the model
+            placeholder tags Input0/Input1/... — exactly how input encoders wire the graph. One
+            NodeConnection is recorded per non-passthrough key, once.
+            """
+            def __init__(self):
+                self._items: Dict[str, Any] = {}
+
+            def __getitem__(self, key):
+                assert isinstance(key, str), \
+                    f'Expected a string model input name, got {type(key)} instead.'
+                if key not in self._items:
+                    class TempMapping:
+                        pass
+                    ret = TempMapping()
+                    ret.node_mapping = NodeMapping(key, NodeMappingType.Input)
+                    if not key.startswith('_'):
+                        leap_binder.mapping_connections.append(NodeConnection(ret.node_mapping, None))
+                    self._items[key] = ret
+                return self._items[key]
+
+            def get(self, key, default=None):
+                # Deliberate alias of [] — the fixed-key-set contract makes a defaulted lookup
+                # meaningless in mapping mode, and every accessed key must wire into the graph.
+                return self[key]
+
+            def _unsupported_iteration(self):
+                raise Exception(
+                    'Iterating over the autoregressive hook result is not supported inside the '
+                    'integration test. Pass the model its inputs with explicit keys, e.g. '
+                    "model.run(None, {'x': inputs['x'], 't': inputs['t']}).")
+
+            def items(self):
+                self._unsupported_iteration()
+
+            def keys(self):
+                self._unsupported_iteration()
+
+            def values(self):
+                self._unsupported_iteration()
+
+        mapping_placeholder = _MappingInputsPlaceholder()
+
+        def mapping_inner(*args, **kwargs):
+            return mapping_placeholder
+
+        def final_inner(*args, **kwargs):
+            if os.environ.get(mapping_runtime_mode_env_var_mame):
+                return mapping_inner(*args, **kwargs)
+            else:
+                return inner(*args, **kwargs)
+
+        if not _call_from_tl_platform:
+            register_decorator_row("tensorleap_autoregressive_step")
+
+        return final_inner
+
+    return decorating_function
+
+
 def tensorleap_preprocess():
     def decorating_function(user_function: Callable[[], List[PreprocessResponse]]):
         leap_binder.set_preprocess(user_function)
@@ -2428,12 +2677,12 @@ def tensorleap_status_table():
                     ready = False
                     next_step = row["name"]
 
-        if _crashed["value"]:
-            print(f"\nScript crashed before completing all steps. crashed at function '{_current_func['name']}'.")
+        if code_mapping_failure[0]:
+            print(f"\n{CROSS} {code_mapping_failure_mes}")
             return
 
-        if code_mapping_failure[0]:
-            print(f"\n{CROSS + code_mapping_failure_mes}.")
+        if _crashed["value"]:
+            print(f"\nScript crashed before completing all steps. crashed at function '{_current_func['name']}'.")
             return
 
         print(ready_mess) if ready else print(
@@ -2448,6 +2697,10 @@ def tensorleap_status_table():
             code_mapping_failure[0] = 1
         if status == "v":
             _set_status(name, CHECK)
+            # A decorator that reports success is no longer the "currently running"
+            # one, so clear the pointer. This keeps a later body-level raise from
+            # being blamed on the last decorator that actually succeeded.
+            _current_func["name"] = None
         else:
             _set_status(name, CROSS)
 
@@ -2459,6 +2712,12 @@ def tensorleap_status_table():
             table.append({"name": "tensorleap simulation (optional)", "Added to integration": UNKNOWN})
         _sim_tracking[sim_name] = "registered"
 
+    def register_decorator_row(name: str):
+        # Rows for decorators that only apply to some integration kinds (e.g. the autoregressive
+        # step hook) are appended on registration, so unrelated integrations never show them.
+        if _find_row(_remove_suffix(name, " (optional)")) is None:
+            table.append({"name": name, "Added to integration": UNKNOWN})
+
     def mark_sim_result(sim_name: str, passed: bool):
         _sim_tracking[sim_name] = "passed" if passed else "failed"
 
@@ -2494,10 +2753,14 @@ def tensorleap_status_table():
     def handle_exception(exc_type, exc_value, exc_traceback):
         _crashed["value"] = True
         crashed_name = _current_func["name"]
-        if crashed_name:
+        if crashed_name and not code_mapping_failure[0]:
             row = _find_row(crashed_name)
             if row:
                 row["Added to integration"] = CROSS
+        elif not crashed_name:
+            # Crash with no decorator currently running = the integration-test body /
+            # code flow itself failed; report that rather than blaming a decorator.
+            code_mapping_failure[0] = 1
 
         traceback.print_exception(exc_type, exc_value, exc_traceback)
         run_on_exit()
@@ -2505,11 +2768,12 @@ def tensorleap_status_table():
     atexit.register(run_on_exit)
     sys.excepthook = handle_exception
 
-    return set_current, update_env_params, register_sim, mark_sim_result
+    return set_current, update_env_params, register_sim, mark_sim_result, register_decorator_row
 
 
 if not _call_from_tl_platform:
-    set_current, update_env_params_func, register_sim, mark_sim_result = tensorleap_status_table()
+    set_current, update_env_params_func, register_sim, mark_sim_result, register_decorator_row = \
+        tensorleap_status_table()
 
 
 

{code_loader-1.0.191.dev0 → code_loader-1.0.192}/code_loader/leaploader.py

@@ -251,7 +251,11 @@ class LeapLoader(LeapLoaderBase):
 
         metadata, metadata_is_none = self.get_metadata(state, sample_id)
 
-        custom_latent_spaces = self._get_custom_latent_spaces(sample_id, preprocess_result[state])
+        custom_latent_space = None
+        if global_leap_binder.setup_container.custom_latent_space is not None:
+            custom_latent_space = global_leap_binder.setup_container.custom_latent_space.function(sample_id,
+                                                                                                  preprocess_result[
+                                                                                                      state])
         instance_mask = self._get_instances_masks(state, sample_id, instance_id)
         sample = DatasetSample(inputs=self._get_inputs(state, sample_id),
                                gt=None if state == DataStateEnum.unlabeled else self._get_gt(state, sample_id),
@@ -259,7 +263,7 @@ class LeapLoader(LeapLoaderBase):
                                metadata_is_none=metadata_is_none,
                                index=sample_id,
                                state=state,
-                               custom_latent_spaces=custom_latent_spaces,
+                               custom_latent_space=custom_latent_space,
                                instance_masks=instance_mask)
         return sample
 
@@ -290,7 +294,11 @@ class LeapLoader(LeapLoaderBase):
                     handler.name, handler_result
                 )
 
-        custom_latent_spaces = self._get_custom_latent_spaces(original_sample_id, preprocess)
+        custom_latent_space = None
+        if global_leap_binder.setup_container.custom_latent_space is not None:
+            custom_latent_space = global_leap_binder.setup_container.custom_latent_space.function(
+                original_sample_id, preprocess
+            )
 
         return DatasetSample(
             inputs=inputs,
@@ -299,7 +307,7 @@ class LeapLoader(LeapLoaderBase):
             metadata_is_none=metadata_is_none,
             index=synthetic_index,
             state=DataStateEnum.additional,
-            custom_latent_spaces=custom_latent_spaces,
+            custom_latent_space=custom_latent_space,
             instance_masks=None,
         )
 
@@ -314,12 +322,16 @@ class LeapLoader(LeapLoaderBase):
 
                 integration_test_test_payload = self._check_integration_test_exists()
                 test_payloads.append(integration_test_test_payload)
+                global_leap_binder.validate_autoregressive_setup()
                 preprocess_test_payload = self._check_preprocess()
                 test_payloads.append(preprocess_test_payload)
                 handlers_test_payloads = self._check_handlers()
                 test_payloads.extend(handlers_test_payloads)
                 simulation_test_payloads = self._check_simulations()
                 test_payloads.extend(simulation_test_payloads)
+                autoregressive_test_payload = self._check_autoregressive_step()
+                if autoregressive_test_payload is not None:
+                    test_payloads.append(autoregressive_test_payload)
                 is_valid = all([payload.is_passed for payload in test_payloads])
                 setup_response = self.get_dataset_setup_response(handlers_test_payloads)
 
@@ -456,6 +468,76 @@ class LeapLoader(LeapLoaderBase):
             result_payloads.append(test_result)
         return result_payloads
 
+    def _check_autoregressive_step(self) -> Optional[DatasetTestResultPayload]:
+        handler = global_leap_binder.setup_container.autoregressive_step
+        if handler is None:
+            return None
+        test_result = DatasetTestResultPayload(handler.name)
+        try:
+            preprocess_result = self._preprocess_result()
+            input_shapes: Dict[str, List[int]] = {}
+            for state, preprocess_response in preprocess_result.items():
+                if preprocess_response.sample_ids_to_instance_mappings:
+                    raise Exception('Element instances are not supported together with '
+                                    'tensorleap_autoregressive_step.')
+                sample_id = preprocess_response.sample_ids[0]
+                first_result = handler.function(sample_id, None, None, preprocess_response)
+                second_result = handler.function(sample_id, None, None, preprocess_response)
+                if not isinstance(first_result, dict):
+                    raise Exception('The autoregressive step hook must return a dict of initial model '
+                                    f'inputs on its first call, got {type(first_result)} '
+                                    f'(state: {state.name}).')
+                if not isinstance(second_result, dict) or \
+                        set(second_result.keys()) != set(first_result.keys()):
+                    raise Exception('The autoregressive step hook is not deterministic — two calls '
+                                    'with identical arguments returned different results '
+                                    f'(state: {state.name}). Seed any stochasticity from '
+                                    'sample_id: the engine replays steps after crash recovery '
+                                    'and relies on identical results.')
+                model_input_items = {key: value for key, value in first_result.items()
+                                     if not key.startswith('_')}
+                if not model_input_items:
+                    raise Exception('The autoregressive step hook returned no model inputs on its first '
+                                    'call — every key starts with "_" (underscore keys are passthrough '
+                                    f'state, never fed to the model). State: {state.name}.')
+                for key, value in model_input_items.items():
+                    if not isinstance(value, np.ndarray):
+                        raise Exception(f'The autoregressive step hook returned a non-numpy value for '
+                                        f'model input "{key}" ({type(value)}). State: {state.name}.')
+                    if key in input_shapes and input_shapes[key] != list(value.shape):
+                        raise Exception(f'The autoregressive step hook returned different shapes for '
+                                        f'model input "{key}" across states: {input_shapes[key]} vs '
+                                        f'{list(value.shape)}. Shapes must be fixed.')
+                    input_shapes[key] = list(value.shape)
+                    if not np.array_equal(value, second_result[key]):
+                        raise Exception('The autoregressive step hook is not deterministic — two calls '
+                                        'with identical arguments returned different outputs for model '
+                                        f'input "{key}". Seed any stochasticity from sample_id: the '
+                                        'engine replays steps after crash recovery and relies on '
+                                        'identical results.')
+            handler.input_shapes = input_shapes
+        except Exception as e:
+            line_number, file_name, stacktrace = get_root_exception_file_and_line_number()
+            test_result.display[TestingSectionEnum.Errors.name] = \
+                f"{repr(e)} in file {file_name}, line_number:  {line_number}\nStacktrace:\n{stacktrace}"
+            test_result.is_passed = False
+        return test_result
+
+    def _ensure_autoregressive_input_shapes(self) -> Dict[str, List[int]]:
+        handler = global_leap_binder.setup_container.autoregressive_step
+        assert handler is not None
+        if handler.input_shapes is None:
+            preprocess_result = self._preprocess_result()
+            first_state_response = next(iter(preprocess_result.values()))
+            first_result = handler.function(first_state_response.sample_ids[0], None, None,
+                                            first_state_response)
+            if not isinstance(first_result, dict):
+                raise Exception('The autoregressive step hook must return a dict of initial model '
+                                f'inputs on its first call, got {type(first_result)}.')
+            handler.input_shapes = {key: list(value.shape) for key, value in first_result.items()
+                                    if not key.startswith('_') and isinstance(value, np.ndarray)}
+        return handler.input_shapes
+
     def run_simulation(self, sim_name, params=None, n_samples=1, seed=0,
                        sample_ids=None, extend_preprocess=True):
         # type: (str, Optional[Dict[str, Any]], int, int, Optional[List[str]], bool) -> Dict[str, Any]
@@ -646,6 +728,14 @@ class LeapLoader(LeapLoaderBase):
                 raise Exception(f"cant calculate shape for input, input name:{inp.name}")
             inputs.append(DatasetInputInstance(name=inp.name, shape=inp.shape, channel_dim=inp.channel_dim))
 
+        if setup.autoregressive_step is not None:
+            # The hook is the sole input source; report its step-0 inputs as the dataset inputs so
+            # graph building, input-coverage validation and type dictionaries see real specs.
+            # channel_dim=-1 is intentional for AR inputs — there is no InputHandler to carry a
+            # user-declared value; expose it on the decorator if a use case ever needs otherwise.
+            for input_name, input_shape in self._ensure_autoregressive_input_shapes().items():
+                inputs.append(DatasetInputInstance(name=input_name, shape=input_shape, channel_dim=-1))
+
         ground_truths = []
         for gt in setup.ground_truths:
             if gt.shape is None:
@@ -753,7 +843,21 @@ class LeapLoader(LeapLoaderBase):
         return result_agg
 
     def _get_inputs(self, state: DataStateEnum, sample_id: Union[int, str]) -> Dict[str, npt.NDArray[np.float32]]:
-        return self._get_dataset_handlers(global_leap_binder.setup_container.inputs, state, sample_id)
+        inputs = self._get_dataset_handlers(global_leap_binder.setup_container.inputs, state, sample_id)
+        autoregressive_handler = global_leap_binder.setup_container.autoregressive_step
+        if autoregressive_handler is not None:
+            # Autoregressive integrations have no input encoders — the hook's first call supplies
+            # the sample's (step-0) model inputs, so every sample-fetching flow (graph validation,
+            # analysis, import checks) sees a fully-populated inputs dict.
+            preprocess_response = self._preprocess_result()[state]
+            step_zero_inputs = autoregressive_handler.function(sample_id, None, None, preprocess_response)
+            if not isinstance(step_zero_inputs, dict):
+                raise Exception(
+                    'The autoregressive step hook must return a dict of initial model inputs on its '
+                    f'first call, got {type(step_zero_inputs)} for sample {sample_id}.')
+            inputs.update({key: value for key, value in step_zero_inputs.items()
+                           if not key.startswith('_')})
+        return inputs
 
     def _get_instances_masks(self, state: DataStateEnum, sample_id: Union[int, str], instance_id: int) -> Optional[Dict[str, ElementInstance]]:
         if instance_id is None:
@@ -776,6 +880,12 @@ class LeapLoader(LeapLoaderBase):
     def _get_gt(self, state: DataStateEnum, sample_id: Union[int, str]) -> Dict[str, npt.NDArray[np.float32]]:
         return self._get_dataset_handlers(global_leap_binder.setup_container.ground_truths, state, sample_id)
 
+    def get_gt(self, state: DataStateEnum, sample_id: Union[int, str]) -> Dict[str, npt.NDArray[np.float32]]:
+        """Ground-truth tensors only ({} when no GT encoders are declared) — lets the engine's
+        rollout fetch a chain's GT without re-running the autoregressive hook via get_sample."""
+        self.exec_script()
+        return self._get_gt(state, sample_id)
+
     @lru_cache()
     def _metadata_name_to_type(self) -> Dict[str, DatasetMetadataType]:
         global_leap_binder.check_preprocess(self._preprocess_result())
@@ -866,29 +976,72 @@ class LeapLoader(LeapLoaderBase):
 
         return id_type
 
-    @staticmethod
-    def _get_custom_latent_spaces(
-            sample_id: Union[int, str],
-            preprocess: "PreprocessResponse") -> Optional[Dict[str, npt.NDArray[np.float32]]]:
-        handlers = global_leap_binder.setup_container.custom_latent_spaces
-        if not handlers:
-            return None
-        return {name: handler.function(sample_id, preprocess) for name, handler in handlers.items()}
-
     @lru_cache()
     def has_custom_latent_space_decorator(self) -> bool:
         self.exec_script()
-        return len(global_leap_binder.setup_container.custom_latent_spaces) > 0
+        return global_leap_binder.setup_container.custom_latent_space is not None
 
     @lru_cache()
-    def get_custom_latent_space_names(self) -> Tuple[str, ...]:
-        """Ordered names of all registered custom latent spaces.
+    def has_autoregressive_step(self) -> bool:
+        self.exec_script()
+        return global_leap_binder.setup_container.autoregressive_step is not None
 
-        Registration order is the canonical index mapping consumed by the engine
-        (name i -> user_custom_i). Returns a tuple so the lru_cache value is hashable.
+    @lru_cache()
+    def get_autoregressive_latent_space_aggregation(self) -> str:
+        self.exec_script()
+        handler = global_leap_binder.setup_container.autoregressive_step
+        if handler is None:
+            return 'last_step'
+        return handler.latent_space_aggregation
+
+    def run_autoregressive_step(self, sample_id: Union[int, str],
+                                prev_inputs: Optional[Dict[str, npt.NDArray[np.float32]]],
+                                prev_outputs: Optional[Dict[str, npt.NDArray[np.float32]]],
+                                state: DataStateEnum) -> Optional[Dict[str, npt.NDArray[np.float32]]]:
+        """Execute the user's autoregressive step hook against the in-process preprocess result.
+
+        This is the engine rollout's per-step entry point (invoked on the generic pod). The first
+        call per chain passes prev_inputs=None, prev_outputs=None and returns the initial model
+        inputs; a None return ends the chain.
         """
         self.exec_script()
-        return tuple(global_leap_binder.setup_container.custom_latent_spaces.keys())
+        handler = global_leap_binder.setup_container.autoregressive_step
+        if handler is None:
+            raise Exception('run_autoregressive_step was called but the integration has no '
+                            'tensorleap_autoregressive_step hook.')
+        preprocess_response = self._preprocess_result()[state]
+        result = handler.function(sample_id, prev_inputs, prev_outputs, preprocess_response)
+        if result is None:
+            return None
+        if not isinstance(result, dict):
+            raise Exception('The autoregressive step hook must return a dict of model inputs or '
+                            f'None to end the chain, got {type(result)} for sample {sample_id}.')
+        # Parse-time validation only exercises step 0; enforce the fixed key-set/shape contract
+        # on every step here so drift fails with an actionable message instead of a cryptic
+        # engine error on a compiled graph.
+        expected_shapes = self._ensure_autoregressive_input_shapes()
+        model_inputs = {key: value for key, value in result.items() if not key.startswith('_')}
+        if set(model_inputs) != set(expected_shapes):
+            raise Exception('The autoregressive step hook must return the same model input keys '
+                            f'on every call. Expected {sorted(expected_shapes)}, got '
+                            f'{sorted(model_inputs)} for sample {sample_id}.')
+        for key, value in model_inputs.items():
+            if isinstance(value, np.ndarray) and list(value.shape) != expected_shapes[key]:
+                raise Exception('The autoregressive step hook must return fixed tensor shapes on '
+                                f'every call (use fixed-length padding + mask for growing '
+                                f'sequences). Input "{key}" was {expected_shapes[key]} on the '
+                                f'first call and {list(value.shape)} now (sample {sample_id}).')
+        return result
+
+    def get_prediction_names_in_order(self) -> List[str]:
+        """Model output names by declared prediction-type order, used to key prev_outputs for the
+        hook. AR integrations must declare prediction types (validate_autoregressive_setup) and
+        the count is checked against the model's outputs at integration-test time; the declared
+        ORDER matching the model's output order is the user's responsibility — it is not
+        validated."""
+        self.exec_script()
+        return [prediction_type.name for prediction_type in
+                global_leap_binder.setup_container.prediction_types]
 
     def get_instances_data(self, state: DataStateEnum) -> Tuple[Dict[str, List[str]], Dict[str, str]]:
         preprocess_result = self._preprocess_result()

{code_loader-1.0.191.dev0 → code_loader-1.0.192}/code_loader/leaploaderbase.py

@@ -149,9 +149,10 @@ class LeapLoaderBase:
     def has_custom_latent_space_decorator(self) -> bool:
         pass
 
-    @abstractmethod
-    def get_custom_latent_space_names(self) -> Tuple[str, ...]:
-        pass
+    def has_autoregressive_step(self) -> bool:
+        # Default False so existing LeapLoaderBase implementations stay valid; the concrete
+        # loaders that can answer (LeapLoader in-process, LeapLoaderWithRedis via RPC) override.
+        return False
 
     @abstractmethod
     def get_heatmap_visualizer_raw_vis_input_arg_name(self, visualizer_name: str) -> Optional[str]:

{code_loader-1.0.191.dev0 → code_loader-1.0.192}/pyproject.toml

@@ -1,7 +1,7 @@
 [tool.poetry]
 name = "code-loader"
 
-version = "1.0.191.dev0"
+version = "1.0.192"
 description = ""
 authors = ["dorhar <doron.harnoy@tensorleap.ai>"]
 license = "MIT"