nvidia-nat 1.2.0rc5__py3-none-any.whl

aiq/profiler/decorators/function_tracking.py

@@ -0,0 +1,254 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import functools
+import inspect
+import uuid
+from typing import Any
+
+from pydantic import BaseModel
+
+from aiq.builder.context import AIQContext
+from aiq.builder.intermediate_step_manager import IntermediateStepManager
+from aiq.data_models.intermediate_step import IntermediateStepPayload
+from aiq.data_models.intermediate_step import IntermediateStepType
+from aiq.data_models.intermediate_step import TraceMetadata
+
+
+# --- Helper function to recursively serialize any object into JSON-friendly data ---
+def _serialize_data(obj: Any) -> Any:
+    """Convert `obj` into a structure that can be passed to `json.dumps(...)`."""
+    if isinstance(obj, BaseModel):
+        # Convert Pydantic model to dict
+        return obj.model_dump()
+
+    if isinstance(obj, dict):
+        return {str(k): _serialize_data(v) for k, v in obj.items()}
+    if isinstance(obj, (list, tuple, set)):
+        return [_serialize_data(item) for item in obj]
+
+    if isinstance(obj, (str, int, float, bool, type(None))):
+        return obj
+
+    # Fallback
+    return str(obj)
+
+
+def _prepare_serialized_args_kwargs(*args, **kwargs) -> tuple[list[Any], dict[str, Any]]:
+    """Serialize args and kwargs before calling the wrapped function."""
+    serialized_args = [_serialize_data(a) for a in args]
+    serialized_kwargs = {k: _serialize_data(v) for k, v in kwargs.items()}
+    return serialized_args, serialized_kwargs
+
+
+def push_intermediate_step(step_manager: IntermediateStepManager,
+                           identifier: str,
+                           function_name: str,
+                           event_type: IntermediateStepType,
+                           args: Any = None,
+                           kwargs: Any = None,
+                           output: Any = None,
+                           metadata: dict[str, Any] | None = None) -> None:
+    """Push an intermediate step to the AIQ Toolkit Event Stream."""
+
+    payload = IntermediateStepPayload(UUID=identifier,
+                                      event_type=event_type,
+                                      name=function_name,
+                                      metadata=TraceMetadata(
+                                          span_inputs=[args, kwargs],
+                                          span_outputs=output,
+                                          provided_metadata=metadata,
+                                      ))
+
+    step_manager.push_intermediate_step(payload)
+
+
+def track_function(func: Any = None, *, metadata: dict[str, Any] | None = None):
+    """
+    Decorator that can wrap any type of function (sync, async, generator,
+    async generator) and executes "tracking logic" around it.
+
+    - If the function is async, it will be wrapped in an async function.
+    - If the function is a generator, it will be wrapped in a generator function.
+    - If the function is an async generator, it will be wrapped in an async generator function.
+    - If the function is sync, it will be wrapped in a sync function.
+    """
+    function_name: str = func.__name__ if func else "<unknown_function>"
+
+    # If called as @track_function(...) but not immediately passed a function
+    if func is None:
+
+        def decorator_wrapper(actual_func):
+            return track_function(actual_func, metadata=metadata)
+
+        return decorator_wrapper
+
+    # --- Validate metadata ---
+    if metadata is not None:
+        if not isinstance(metadata, dict):
+            raise TypeError("metadata must be a dict[str, Any].")
+        if any(not isinstance(k, str) for k in metadata.keys()):
+            raise TypeError("All metadata keys must be strings.")
+
+    # --- Now detect the function type and wrap accordingly ---
+    if inspect.isasyncgenfunction(func):
+        # ---------------------
+        # ASYNC GENERATOR
+        # ---------------------
+
+        @functools.wraps(func)
+        async def async_gen_wrapper(*args, **kwargs):
+            step_manager: IntermediateStepManager = AIQContext.get().intermediate_step_manager
+            # 1) Serialize input
+            serialized_args, serialized_kwargs = _prepare_serialized_args_kwargs(*args, **kwargs)
+
+            invocation_id = str(uuid.uuid4())
+            push_intermediate_step(step_manager,
+                                   invocation_id,
+                                   function_name,
+                                   IntermediateStepType.SPAN_START,
+                                   args=serialized_args,
+                                   kwargs=serialized_kwargs,
+                                   metadata=metadata)
+
+            # 2) Call the original async generator
+            async for item in func(*args, **kwargs):
+                # 3) Serialize the yielded item before yielding it
+                serialized_item = _serialize_data(item)
+                push_intermediate_step(step_manager,
+                                       invocation_id,
+                                       function_name,
+                                       IntermediateStepType.SPAN_CHUNK,
+                                       args=serialized_args,
+                                       kwargs=serialized_kwargs,
+                                       output=serialized_item,
+                                       metadata=metadata)
+                yield item  # yield the original item
+
+            push_intermediate_step(step_manager,
+                                   invocation_id,
+                                   function_name,
+                                   IntermediateStepType.SPAN_END,
+                                   args=serialized_args,
+                                   kwargs=serialized_kwargs,
+                                   output=None,
+                                   metadata=metadata)
+
+            # 4) Post-yield logic if any
+
+        return async_gen_wrapper
+
+    if inspect.iscoroutinefunction(func):
+        # ---------------------
+        # ASYNC FUNCTION
+        # ---------------------
+        @functools.wraps(func)
+        async def async_wrapper(*args, **kwargs):
+            step_manager: IntermediateStepManager = AIQContext.get().intermediate_step_manager
+            serialized_args, serialized_kwargs = _prepare_serialized_args_kwargs(*args, **kwargs)
+            invocation_id = str(uuid.uuid4())
+            push_intermediate_step(step_manager,
+                                   invocation_id,
+                                   function_name,
+                                   IntermediateStepType.SPAN_START,
+                                   args=serialized_args,
+                                   kwargs=serialized_kwargs,
+                                   metadata=metadata)
+
+            result = await func(*args, **kwargs)
+
+            serialized_result = _serialize_data(result)
+            push_intermediate_step(step_manager,
+                                   invocation_id,
+                                   function_name,
+                                   IntermediateStepType.SPAN_END,
+                                   args=serialized_args,
+                                   kwargs=serialized_kwargs,
+                                   output=serialized_result,
+                                   metadata=metadata)
+
+            return result
+
+        return async_wrapper
+
+    if inspect.isgeneratorfunction(func):
+        # ---------------------
+        # SYNC GENERATOR
+        # ---------------------
+        @functools.wraps(func)
+        def sync_gen_wrapper(*args, **kwargs):
+            step_manager: IntermediateStepManager = AIQContext.get().intermediate_step_manager
+            serialized_args, serialized_kwargs = _prepare_serialized_args_kwargs(*args, **kwargs)
+            invocation_id = str(uuid.uuid4())
+            push_intermediate_step(step_manager,
+                                   invocation_id,
+                                   function_name,
+                                   IntermediateStepType.SPAN_START,
+                                   args=serialized_args,
+                                   kwargs=serialized_kwargs,
+                                   metadata=metadata)
+
+            for item in func(*args, **kwargs):
+                serialized_item = _serialize_data(item)
+                push_intermediate_step(step_manager,
+                                       invocation_id,
+                                       function_name,
+                                       IntermediateStepType.SPAN_CHUNK,
+                                       args=serialized_args,
+                                       kwargs=serialized_kwargs,
+                                       output=serialized_item,
+                                       metadata=metadata)
+
+                yield item  # yield the original item
+
+            push_intermediate_step(step_manager,
+                                   invocation_id,
+                                   function_name,
+                                   IntermediateStepType.SPAN_END,
+                                   args=serialized_args,
+                                   kwargs=serialized_kwargs,
+                                   output=None,
+                                   metadata=metadata)
+
+        return sync_gen_wrapper
+
+    @functools.wraps(func)
+    def sync_wrapper(*args, **kwargs):
+        step_manager: IntermediateStepManager = AIQContext.get().intermediate_step_manager
+        serialized_args, serialized_kwargs = _prepare_serialized_args_kwargs(*args, **kwargs)
+        invocation_id = str(uuid.uuid4())
+        push_intermediate_step(step_manager,
+                               invocation_id,
+                               function_name,
+                               IntermediateStepType.SPAN_START,
+                               args=serialized_args,
+                               kwargs=serialized_kwargs,
+                               metadata=metadata)
+
+        result = func(*args, **kwargs)
+
+        serialized_result = _serialize_data(result)
+        push_intermediate_step(step_manager,
+                               invocation_id,
+                               function_name,
+                               IntermediateStepType.SPAN_END,
+                               args=serialized_args,
+                               kwargs=serialized_kwargs,
+                               output=serialized_result,
+                               metadata=metadata)
+
+        return result
+
+    return sync_wrapper

aiq/profiler/forecasting/config.py

@@ -0,0 +1,18 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# If you have any global constants or defaults
+DEFAULT_MODEL_TYPE = "randomforest"
+DEFAULT_MATRIX_LENGTH = 10

aiq/profiler/forecasting/model_trainer.py

@@ -0,0 +1,75 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# forecasting/model_trainer.py
+
+import logging
+
+from aiq.profiler.forecasting.config import DEFAULT_MODEL_TYPE
+from aiq.profiler.forecasting.models import ForecastingBaseModel
+from aiq.profiler.forecasting.models import LinearModel
+from aiq.profiler.forecasting.models import RandomForestModel
+from aiq.profiler.intermediate_property_adapter import IntermediatePropertyAdaptor
+
+logger = logging.getLogger(__name__)
+
+
+def create_model(model_type: str) -> ForecastingBaseModel:
+    """
+    A simple factory method that returns a model instance
+    based on the input string. Extend this with more model
+    classes (e.g., PolynomialModel, RandomForestModel, etc.).
+    """
+    if model_type == "linear":
+        return LinearModel()
+    if model_type == "randomforest":
+        return RandomForestModel()
+
+    raise ValueError(f"Unsupported model_type: {model_type}")
+
+
+class ModelTrainer:
+    """
+    Orchestrates data preprocessing, training, and returning
+    a fitted model.
+
+    Parameters
+    ----------
+    model_type: str, default = "randomforest"
+        The type of model to train. Options include "linear" and "randomforest".
+    """
+
+    def __init__(self, model_type: str = DEFAULT_MODEL_TYPE):
+        self.model_type = model_type
+        self._model = create_model(self.model_type)
+
+    def train(self, raw_stats: list[list[IntermediatePropertyAdaptor]]) -> ForecastingBaseModel:
+        """
+        Train the model using the `raw_stats` training data.
+
+        Parameters
+        ----------
+        raw_stats: list[list[IntermediatePropertyAdaptor]]
+            Stats collected by the profiler.
+
+        Returns
+        -------
+        ForecastingBaseModel
+            A fitted model.
+        """
+
+        self._model.fit(raw_stats)
+
+        return self._model

aiq/profiler/forecasting/models/__init__.py

@@ -0,0 +1,22 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# forecasting/models/__init__.py
+
+from .forecasting_base_model import ForecastingBaseModel
+from .linear_model import LinearModel
+from .random_forest_regressor import RandomForestModel
+
+__all__ = ["ForecastingBaseModel", "LinearModel", "RandomForestModel"]

aiq/profiler/forecasting/models/forecasting_base_model.py

@@ -0,0 +1,40 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# forecasting/models/base_model.py
+
+from abc import ABC, abstractmethod
+import numpy as np
+
+
+class ForecastingBaseModel(ABC):
+    """
+    Abstract base class for all models in this package.
+    """
+
+    @abstractmethod
+    def fit(self, raw_stats):
+        """
+        Train/fine-tune the model on the provided dataset.
+        """
+        pass
+
+    @abstractmethod
+    def predict(self, raw_stats) -> np.ndarray:
+        """
+        Predict using the trained model.
+        Returns a np.ndarray, shape = (N, 4).
+        """
+        pass

aiq/profiler/forecasting/models/linear_model.py

@@ -0,0 +1,196 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+
+import numpy as np
+
+from aiq.profiler.forecasting.models.forecasting_base_model import ForecastingBaseModel
+from aiq.profiler.intermediate_property_adapter import IntermediatePropertyAdaptor
+
+logger = logging.getLogger(__name__)
+
+
+class LinearModel(ForecastingBaseModel):
+    """
+    A linear regression model that conforms to the BaseModel interface.
+    """
+
+    def __init__(self):
+        super().__init__()
+
+        try:
+            from sklearn.linear_model import LinearRegression
+        except ImportError:
+            logger.error("scikit-learn is not installed. Please install scikit-learn to use the LinearModel "
+                         "profiling model or install `aiq[profiler]` to install all necessary profiling packages.")
+
+            raise
+
+        self.model = LinearRegression()
+        self.matrix_length = None
+
+    def fit(self, raw_stats: list[list[IntermediatePropertyAdaptor]]):
+        """
+        X: shape (N, M)  # M = matrix_length * 4
+        y: shape (N, 4)
+        """
+        x_flat, y_flat = self._prep_for_model_training(raw_stats)
+
+        logger.info("Training dataset size: X=%s, y=%s", x_flat.shape, y_flat.shape)
+
+        # 3) Fit
+        self.model.fit(x_flat, y_flat)
+
+    def predict(self, raw_stats: list[list[IntermediatePropertyAdaptor]]) -> np.ndarray:
+        """
+        Predict using the fitted linear model.
+        Returns shape (N, 4)
+        """
+        X = self._prep_single(raw_stats)
+        return self.model.predict(X)
+
+    def _prep_single(self, raw_stats: list[list[IntermediatePropertyAdaptor]]) -> np.ndarray:
+        arr, _ = self._extract_token_usage_meta(raw_stats)
+        arr = arr[0]
+        n_rows = arr.shape[0]
+
+        matrix_length = self.matrix_length
+
+        assert matrix_length is not None, "matrix_length must be set before calling _prep_single"
+
+        if n_rows >= matrix_length:
+            # Keep the latest matrix_length rows
+            x_mat = arr[-matrix_length:, :]
+        else:
+            # Pad with zeros at the top
+            pad_size = matrix_length - n_rows
+            pad_block = np.zeros((pad_size, arr.shape[1]), dtype=arr.dtype)
+            x_mat = np.vstack([pad_block, arr])
+
+        return x_mat
+
+    def _prep_for_model_training(self, raw_stats: list[list[IntermediatePropertyAdaptor]]):
+        raw_matrices, matrix_length = self._extract_token_usage_meta(raw_stats)
+
+        self.matrix_length = matrix_length
+
+        x_list = []
+        y_list = []
+        for arr in raw_matrices:
+            samples = self._preprocess_for_forecasting(arr, matrix_length)
+            for (x_mat, y_mat) in samples:
+                x_list.append(x_mat)
+                y_list.append(y_mat)
+
+        # 2) Flatten features
+        x_flat, y_flat = self._flatten_features(x_list, y_list)
+
+        return x_flat, y_flat
+
+    def _extract_token_usage_meta(self, all_requests_data: list[list[IntermediatePropertyAdaptor]]):
+        import math
+
+        all_run_data = []
+        call_stack_sizes = []
+
+        for prompt in all_requests_data:
+            run_data = []
+            seconds_between_call_map = {}
+
+            for stat in prompt:
+                if stat.event_type.value == "LLM_START":
+                    seconds_between_call_map[stat.UUID] = stat.seconds_between_calls
+
+                if stat.event_type.value == "LLM_END":
+                    step_data = [
+                        seconds_between_call_map[stat.UUID],
+                        stat.token_usage.prompt_tokens,
+                        stat.token_usage.completion_tokens
+                    ]
+
+                    run_data.append(step_data)
+
+            all_run_data.append(run_data)
+            call_stack_sizes.append(len(run_data))
+
+        all_run_data = [np.array(run) for run in all_run_data]
+        recommended_matrix_length = math.ceil(sum(call_stack_sizes) / len(call_stack_sizes))
+
+        return all_run_data, recommended_matrix_length
+
+    def _preprocess_for_forecasting(self, arr: np.ndarray, matrix_length: int):
+        """
+        Given a 2D NumPy array `arr` of shape (n_rows, 4), generate a list of
+        (input_array, output_array) pairs for forecasting, each of shape:
+
+        - input_array: (matrix_length, 4) after padding/trimming
+        - output_array: (1, 4)
+        """
+        n_rows = arr.shape[0]
+
+        # partial_sums[i] = sum of arr[i:] per column
+        partial_sums = np.flip(np.cumsum(np.flip(arr, axis=0), axis=0), axis=0)
+
+        samples = []
+        for i in range(n_rows):
+            x_untrimmed = arr[:i + 1, :]
+            # Trim or pad
+            current_len = x_untrimmed.shape[0]
+            if current_len > matrix_length:
+                x_mat = x_untrimmed[-matrix_length:, :]
+            elif current_len < matrix_length:
+                pad_size = matrix_length - current_len
+                pad_block = np.zeros((pad_size, x_untrimmed.shape[1]), dtype=arr.dtype)
+                x_mat = np.vstack([pad_block, x_untrimmed])
+            else:
+                x_mat = x_untrimmed
+
+            # Compute output
+            if i == n_rows - 1:
+                y_vec = np.array([0, 0, 0, 0], dtype=arr.dtype)
+            else:
+                n_below = n_rows - (i + 1)
+                sum_below = partial_sums[i + 1]
+                avg_col0 = sum_below[0] / n_below
+                sum_rest = sum_below[1:]
+                y_vec = np.concatenate(([avg_col0], sum_rest))
+
+            samples.append((x_mat, y_vec.reshape(1, 4)))
+
+        return samples
+
+    def _flatten_features(self, x_list, y_list):
+        """
+        x_list: list of arrays, each of shape (matrix_length, 4)
+        y_list: list of arrays, each of shape (1, 4)
+
+        Returns:
+            x_flat: np.array of shape (N, matrix_length*4)
+            y_flat: np.array of shape (N, 4)
+        """
+        flattened_x = []
+        flattened_y = []
+
+        for x_mat, y_mat in zip(x_list, y_list):
+            x_1d = x_mat.flatten()  # shape -> (matrix_length*4,)
+            y_1d = y_mat.flatten()  # shape -> (4,)
+            flattened_x.append(x_1d)
+            flattened_y.append(y_1d)
+
+        x_flat = np.array(flattened_x)
+        y_flat = np.array(flattened_y)
+        logger.debug("Flattened features to shapes: %s (X), %s (y).", x_flat.shape, y_flat.shape)
+        return x_flat, y_flat