nvidia-nat 1.3.0a20250909__py3-none-any.whl → 1.3.0a20250917__py3-none-any.whl

nat/data_models/optimizer.py

@@ -0,0 +1,149 @@
+# SPDX-FileCopyrightText: Copyright (c) 2021-2025, 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.
+
+from pathlib import Path
+
+from pydantic import BaseModel
+from pydantic import Field
+
+
+class OptimizerMetric(BaseModel):
+    """
+    Parameters used by the workflow optimizer to define a metric to optimize.
+    """
+    evaluator_name: str = Field(description="Name of the metric to optimize.")
+    direction: str = Field(description="Direction of the optimization. Can be 'maximize' or 'minimize'.")
+    weight: float = Field(description="Weight of the metric in the optimization process.", default=1.0)
+
+
+class NumericOptimizationConfig(BaseModel):
+    """
+    Configuration for numeric/enum optimization (Optuna).
+    """
+    enabled: bool = Field(default=True, description="Enable numeric optimization")
+    n_trials: int = Field(description="Number of trials for numeric optimization.", default=20)
+
+
+class PromptGAOptimizationConfig(BaseModel):
+    """
+    Configuration for prompt optimization using a Genetic Algorithm.
+    """
+    enabled: bool = Field(default=False, description="Enable GA-based prompt optimization")
+
+    # Prompt optimization function hooks
+    prompt_population_init_function: str | None = Field(
+        default=None,
+        description="Optional function name to initialize/mutate candidate prompts.",
+    )
+    prompt_recombination_function: str | None = Field(
+        default=None,
+        description="Optional function name to recombine two parent prompts into a child.",
+    )
+
+    # Genetic algorithm configuration
+    ga_population_size: int = Field(
+        description="Population size for genetic algorithm prompt optimization.",
+        default=24,
+    )
+    ga_generations: int = Field(
+        description="Number of generations to evolve in GA prompt optimization.",
+        default=15,
+    )
+    ga_offspring_size: int | None = Field(
+        description="Number of offspring to produce per generation. Defaults to population_size - elitism.",
+        default=None,
+    )
+    ga_crossover_rate: float = Field(
+        description="Probability of applying crossover during reproduction.",
+        default=0.8,
+        ge=0.0,
+        le=1.0,
+    )
+    ga_mutation_rate: float = Field(
+        description="Probability of mutating a child after crossover.",
+        default=0.3,
+        ge=0.0,
+        le=1.0,
+    )
+    ga_elitism: int = Field(
+        description="Number of top individuals carried over unchanged each generation.",
+        default=2,
+    )
+    ga_selection_method: str = Field(
+        description="Parent selection strategy: 'tournament' or 'roulette'.",
+        default="tournament",
+    )
+    ga_tournament_size: int = Field(
+        description="Tournament size when using tournament selection.",
+        default=3,
+    )
+    ga_parallel_evaluations: int = Field(
+        description="Max number of individuals to evaluate concurrently per generation.",
+        default=8,
+    )
+    ga_diversity_lambda: float = Field(
+        description="Strength of diversity penalty (0 disables). Penalizes identical/near-identical prompts.",
+        default=0.0,
+        ge=0.0,
+    )
+
+
+class OptimizerConfig(BaseModel):
+    """
+    Parameters used by the workflow optimizer.
+    """
+    output_path: Path | None = Field(
+        default=None,
+        description="Path to the output directory where the results will be saved.",
+    )
+
+    eval_metrics: dict[str, OptimizerMetric] | None = Field(
+        description="List of evaluation metrics to optimize.",
+        default=None,
+    )
+
+    reps_per_param_set: int = Field(
+        default=3,
+        description="Number of repetitions per parameter set for the optimization.",
+    )
+
+    target: float | None = Field(
+        description=(
+            "Target value for the optimization. If set, the optimization will stop when this value is reached."),
+        default=None,
+    )
+
+    multi_objective_combination_mode: str = Field(
+        description="Method to combine multiple objectives into a single score.",
+        default="harmonic",
+    )
+
+    # Nested configs
+    numeric: NumericOptimizationConfig = NumericOptimizationConfig()
+    prompt: PromptGAOptimizationConfig = PromptGAOptimizationConfig()
+
+
+class OptimizerRunConfig(BaseModel):
+    """
+    Parameters used for an Optimizer R=run
+    """
+    # Eval parameters
+
+    config_file: Path | BaseModel  # allow for instantiated configs to be passed in
+    dataset: str | Path | None  # dataset file path can be specified in the config file
+    result_json_path: str = "$"
+    endpoint: str | None = None  # only used when running the workflow remotely
+    endpoint_timeout: int = 300
+    override: tuple[tuple[str, str], ...] = ()

nat/data_models/temperature_mixin.py

@@ -16,9 +16,10 @@
 import re
 
 from pydantic import BaseModel
-from pydantic import Field
 
 from nat.data_models.gated_field_mixin import GatedFieldMixin
+from nat.data_models.optimizable import OptimizableField
+from nat.data_models.optimizable import SearchSpace
 
 
 class TemperatureMixin(
@@ -35,9 +36,9 @@ class TemperatureMixin(
     Attributes:
         temperature: Sampling temperature in [0, 1]. Defaults to 0.0 when supported on the model.
     """
-    temperature: float | None = Field(
+    temperature: float | None = OptimizableField(
         default=None,
         ge=0.0,
         le=1.0,
         description="Sampling temperature in [0, 1]. Defaults to 0.0 when supported on the model.",
-    )
+        space=SearchSpace(high=0.9, low=0.1, step=0.2))

nat/data_models/top_p_mixin.py

@@ -16,9 +16,10 @@
 import re
 
 from pydantic import BaseModel
-from pydantic import Field
 
 from nat.data_models.gated_field_mixin import GatedFieldMixin
+from nat.data_models.optimizable import OptimizableField
+from nat.data_models.optimizable import SearchSpace
 
 
 class TopPMixin(
@@ -35,9 +36,9 @@ class TopPMixin(
     Attributes:
         top_p: Top-p for distribution sampling. Defaults to 1.0 when supported on the model.
     """
-    top_p: float | None = Field(
+    top_p: float | None = OptimizableField(
         default=None,
         ge=0.0,
         le=1.0,
         description="Top-p for distribution sampling. Defaults to 1.0 when supported on the model.",
-    )
+        space=SearchSpace(high=1.0, low=0.5, step=0.1))

nat/embedder/nim_embedder.py

@@ -50,7 +50,7 @@ class NIMEmbedderModelConfig(EmbedderBaseConfig, RetryMixin, name="nim"):
                                        description=("The truncation strategy if the input on the "
                                                     "server side if it's too large."))
 
-    model_config = ConfigDict(protected_namespaces=())
+    model_config = ConfigDict(protected_namespaces=(), extra="allow")
 
 
 @register_embedder_provider(config_type=NIMEmbedderModelConfig)

nat/embedder/openai_embedder.py

@@ -27,7 +27,7 @@ from nat.data_models.retry_mixin import RetryMixin
 class OpenAIEmbedderModelConfig(EmbedderBaseConfig, RetryMixin, name="openai"):
     """An OpenAI LLM provider to be used with an LLM client."""
 
-    model_config = ConfigDict(protected_namespaces=())
+    model_config = ConfigDict(protected_namespaces=(), extra="allow")
 
     api_key: str | None = Field(default=None, description="OpenAI API key to interact with hosted model.")
     base_url: str | None = Field(default=None, description="Base url to the hosted model.")

nat/eval/config.py

@@ -27,7 +27,7 @@ class EvaluationRunConfig(BaseModel):
     """
     Parameters used for a single evaluation run.
     """
-    config_file: Path
+    config_file: Path | BaseModel
     dataset: str | None = None  # dataset file path can be specified in the config file
     result_json_path: str = "$"
     skip_workflow: bool = False

nat/eval/evaluate.py

@@ -449,10 +449,14 @@ class EvaluationRun:
         from nat.runtime.loader import load_config
 
         # Load and override the config
-        if self.config.override:
+        config = None
+        if isinstance(self.config.config_file, BaseModel):
+            config = self.config.config_file
+        elif self.config.override:
             config = self.apply_overrides()
         else:
             config = load_config(self.config.config_file)
+
         self.eval_config = config.eval
         workflow_alias = self._get_workflow_alias(config.workflow.type)
         logger.debug("Loaded %s evaluation configuration: %s", workflow_alias, self.eval_config)

nat/eval/register.py

@@ -17,6 +17,10 @@
 
 # Import evaluators which need to be automatically registered here
 from .rag_evaluator.register import register_ragas_evaluator
+from .runtime_evaluator.register import register_avg_llm_latency_evaluator
+from .runtime_evaluator.register import register_avg_num_llm_calls_evaluator
+from .runtime_evaluator.register import register_avg_tokens_per_llm_end_evaluator
+from .runtime_evaluator.register import register_avg_workflow_runtime_evaluator
 from .swe_bench_evaluator.register import register_swe_bench_evaluator
 from .trajectory_evaluator.register import register_trajectory_evaluator
 from .tunable_rag_evaluator.register import register_tunable_rag_evaluator

nat/eval/runtime_evaluator/__init__.py

@@ -0,0 +1,14 @@
+# 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.

nat/eval/runtime_evaluator/evaluate.py

@@ -0,0 +1,123 @@
+# 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.
+
+from __future__ import annotations
+
+from collections import defaultdict
+from dataclasses import dataclass
+
+from nat.data_models.intermediate_step import IntermediateStepType
+from nat.eval.evaluator.base_evaluator import BaseEvaluator
+from nat.eval.evaluator.evaluator_model import EvalInputItem
+from nat.eval.evaluator.evaluator_model import EvalOutputItem
+from nat.profiler.intermediate_property_adapter import IntermediatePropertyAdaptor
+
+
+@dataclass
+class _CallTiming:
+    start_ts: float | None = None
+    end_ts: float | None = None
+
+    @property
+    def latency(self) -> float | None:
+        if self.start_ts is None or self.end_ts is None:
+            return None
+        return max(0.0, self.end_ts - self.start_ts)
+
+
+class AverageLLMLatencyEvaluator(BaseEvaluator):
+    """
+    Mean difference between connected LLM_START and LLM_END events (same UUID).
+    The score is the average latency in seconds for the item. Reasoning contains per-call latencies.
+    """
+
+    def __init__(self, max_concurrency: int = 8):
+        super().__init__(max_concurrency=max_concurrency, tqdm_desc="Evaluating Avg LLM Latency")
+
+    async def evaluate_item(self, item: EvalInputItem) -> EvalOutputItem:  # noqa: D401
+        calls: dict[str, _CallTiming] = defaultdict(_CallTiming)
+
+        for step in (IntermediatePropertyAdaptor.from_intermediate_step(s) for s in item.trajectory):
+            if step.event_type == IntermediateStepType.LLM_START:
+                calls[step.UUID].start_ts = step.event_timestamp
+            elif step.event_type == IntermediateStepType.LLM_END:
+                calls[step.UUID].end_ts = step.event_timestamp
+
+        latencies = [ct.latency for ct in calls.values() if ct.latency is not None]
+        avg_latency = sum(latencies) / len(latencies) if latencies else 0.0
+
+        reasoning = {
+            "num_llm_calls": len(latencies),
+            "latencies": latencies,
+        }
+        return EvalOutputItem(id=item.id, score=round(avg_latency, 4), reasoning=reasoning)
+
+
+class AverageWorkflowRuntimeEvaluator(BaseEvaluator):
+    """
+    Average workflow runtime per item: max(event_timestamp) - min(event_timestamp) across the trajectory.
+    The score is the runtime in seconds for the item.
+    """
+
+    def __init__(self, max_concurrency: int = 8):
+        super().__init__(max_concurrency=max_concurrency, tqdm_desc="Evaluating Avg Workflow Runtime")
+
+    async def evaluate_item(self, item: EvalInputItem) -> EvalOutputItem:  # noqa: D401
+        if not item.trajectory:
+            return EvalOutputItem(id=item.id, score=0.0, reasoning={"note": "no steps"})
+
+        timestamps = [s.event_timestamp for s in item.trajectory]
+        runtime = max(timestamps) - min(timestamps)
+        return EvalOutputItem(id=item.id, score=round(max(0.0, runtime), 4), reasoning={"steps": len(timestamps)})
+
+
+class AverageNumberOfLLMCallsEvaluator(BaseEvaluator):
+    """
+    Average number of LLM calls per item. The score is the count for the item.
+    """
+
+    def __init__(self, max_concurrency: int = 8):
+        super().__init__(max_concurrency=max_concurrency, tqdm_desc="Evaluating Avg # LLM Calls")
+
+    async def evaluate_item(self, item: EvalInputItem) -> EvalOutputItem:  # noqa: D401
+        num_calls = sum(1 for s in item.trajectory if s.event_type == IntermediateStepType.LLM_END)
+        return EvalOutputItem(id=item.id, score=float(num_calls), reasoning={"num_llm_end": num_calls})
+
+
+class AverageTokensPerLLMEndEvaluator(BaseEvaluator):
+    """
+    Average total tokens per LLM_END event: sum of prompt and completion tokens if available.
+    The score is the average tokens per LLM_END for the item (0 if none).
+    """
+
+    def __init__(self, max_concurrency: int = 8):
+        super().__init__(max_concurrency=max_concurrency, tqdm_desc="Evaluating Avg Tokens/LLM_END")
+
+    async def evaluate_item(self, item: EvalInputItem) -> EvalOutputItem:  # noqa: D401
+        totals: list[int] = []
+        for step in (IntermediatePropertyAdaptor.from_intermediate_step(s) for s in item.trajectory):
+            if step.event_type == IntermediateStepType.LLM_END:
+                total_tokens = step.token_usage.total_tokens
+                # If framework doesn't set total, compute from prompt+completion
+                if total_tokens == 0:
+                    total_tokens = step.token_usage.prompt_tokens + step.token_usage.completion_tokens
+                totals.append(total_tokens)
+
+        avg_tokens = (sum(totals) / len(totals)) if totals else 0.0
+        reasoning = {
+            "num_llm_end": len(totals),
+            "totals": totals,
+        }
+        return EvalOutputItem(id=item.id, score=round(avg_tokens, 2), reasoning=reasoning)

nat/eval/runtime_evaluator/register.py

@@ -0,0 +1,100 @@
+# 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.
+
+from pydantic import Field
+
+from nat.builder.builder import EvalBuilder
+from nat.builder.evaluator import EvaluatorInfo
+from nat.cli.register_workflow import register_evaluator
+from nat.data_models.evaluator import EvaluatorBaseConfig
+from nat.eval.evaluator.evaluator_model import EvalInput
+from nat.eval.evaluator.evaluator_model import EvalOutput
+
+
+class AverageLLMLatencyConfig(EvaluatorBaseConfig, name="avg_llm_latency"):
+    """Mean difference between connected LLM_START and LLM_END events (same UUID)."""
+
+    max_concurrency: int = Field(default=8, description="Max concurrency for evaluation.")
+
+
+class AverageWorkflowRuntimeConfig(EvaluatorBaseConfig, name="avg_workflow_runtime"):
+    """Average workflow runtime per item (max timestamp - min timestamp)."""
+
+    max_concurrency: int = Field(default=8, description="Max concurrency for evaluation.")
+
+
+class AverageNumberOfLLMCallsConfig(EvaluatorBaseConfig, name="avg_num_llm_calls"):
+    """Average number of LLM calls per item (count of LLM_END)."""
+
+    max_concurrency: int = Field(default=8, description="Max concurrency for evaluation.")
+
+
+class AverageTokensPerLLMEndConfig(EvaluatorBaseConfig, name="avg_tokens_per_llm_end"):
+    """Average total tokens per LLM_END event (prompt + completion if available)."""
+
+    max_concurrency: int = Field(default=8, description="Max concurrency for evaluation.")
+
+
+@register_evaluator(config_type=AverageLLMLatencyConfig)
+async def register_avg_llm_latency_evaluator(config: AverageLLMLatencyConfig, builder: EvalBuilder):
+    from .evaluate import AverageLLMLatencyEvaluator
+
+    evaluator = AverageLLMLatencyEvaluator(max_concurrency=config.max_concurrency or builder.get_max_concurrency())
+
+    async def evaluate_fn(eval_input: EvalInput) -> EvalOutput:
+        return await evaluator.evaluate(eval_input)
+
+    yield EvaluatorInfo(config=config,
+                        evaluate_fn=evaluate_fn,
+                        description="Average LLM latency (s) from LLM_START to LLM_END")
+
+
+@register_evaluator(config_type=AverageWorkflowRuntimeConfig)
+async def register_avg_workflow_runtime_evaluator(config: AverageWorkflowRuntimeConfig, builder: EvalBuilder):
+    from .evaluate import AverageWorkflowRuntimeEvaluator
+
+    evaluator = AverageWorkflowRuntimeEvaluator(max_concurrency=config.max_concurrency or builder.get_max_concurrency())
+
+    async def evaluate_fn(eval_input: EvalInput) -> EvalOutput:
+        return await evaluator.evaluate(eval_input)
+
+    yield EvaluatorInfo(config=config, evaluate_fn=evaluate_fn, description="Average workflow runtime (s)")
+
+
+@register_evaluator(config_type=AverageNumberOfLLMCallsConfig)
+async def register_avg_num_llm_calls_evaluator(config: AverageNumberOfLLMCallsConfig, builder: EvalBuilder):
+    from .evaluate import AverageNumberOfLLMCallsEvaluator
+
+    evaluator = AverageNumberOfLLMCallsEvaluator(
+        max_concurrency=config.max_concurrency or builder.get_max_concurrency())
+
+    async def evaluate_fn(eval_input: EvalInput) -> EvalOutput:
+        return await evaluator.evaluate(eval_input)
+
+    yield EvaluatorInfo(config=config, evaluate_fn=evaluate_fn, description="Average number of LLM calls")
+
+
+@register_evaluator(config_type=AverageTokensPerLLMEndConfig)
+async def register_avg_tokens_per_llm_end_evaluator(config: AverageTokensPerLLMEndConfig, builder: EvalBuilder):
+    from .evaluate import AverageTokensPerLLMEndEvaluator
+
+    evaluator = AverageTokensPerLLMEndEvaluator(max_concurrency=config.max_concurrency or builder.get_max_concurrency())
+
+    async def evaluate_fn(eval_input: EvalInput) -> EvalOutput:
+        return await evaluator.evaluate(eval_input)
+
+    yield EvaluatorInfo(config=config,
+                        evaluate_fn=evaluate_fn,
+                        description="Average total tokens per LLM_END (prompt + completion)")

nat/experimental/test_time_compute/functions/plan_select_execute_function.py

@@ -97,7 +97,11 @@ async def plan_select_execute_function(config: PlanSelectExecuteFunctionConfig,
                          f"function without a description.")
 
     # Get the function dependencies of the augmented function
-    function_used_tools = builder.get_function_dependencies(config.augmented_fn).functions
+    function_dependencies = builder.get_function_dependencies(config.augmented_fn)
+    function_used_tools = set(function_dependencies.functions)
+    for function_group in function_dependencies.function_groups:
+        function_used_tools.update(builder.get_function_group_dependencies(function_group).functions)
+
     tool_list = "Tool: Description\n"
 
     for tool in function_used_tools:

nat/front_ends/fastapi/dask_client_mixin.py

@@ -0,0 +1,43 @@
+# 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 typing
+from abc import ABC
+from collections.abc import AsyncGenerator
+from contextlib import asynccontextmanager
+
+if typing.TYPE_CHECKING:
+    from dask.distributed import Client
+
+
+class DaskClientMixin(ABC):
+
+    @asynccontextmanager
+    async def client(self, address: str) -> AsyncGenerator["Client"]:
+        """
+        Async context manager for obtaining a Dask client connection.
+
+        Yields
+        ------
+        Client
+            An active Dask client connected to the scheduler. The client is automatically closed when exiting the
+            context manager.
+        """
+        from dask.distributed import Client
+        client = await Client(address=address, asynchronous=True)
+
+        yield client
+
+        await client.close()

nat/front_ends/fastapi/fastapi_front_end_config.py

@@ -197,9 +197,20 @@ class FastApiFrontEndConfig(FrontEndBaseConfig, name="fastapi"):
     port: int = Field(default=8000, description="Port to bind the server to", ge=0, le=65535)
     reload: bool = Field(default=False, description="Enable auto-reload for development")
     workers: int = Field(default=1, description="Number of workers to run", ge=1)
-    max_running_async_jobs: int = Field(default=10,
-                                        description="Maximum number of async jobs to run concurrently",
-                                        ge=1)
+    scheduler_address: str | None = Field(
+        default=None,
+        description=("Address of the Dask scheduler to use for async jobs. If None, a Dask local cluster is created. "
+                     "Note: This requires the optional dask dependency to be installed."))
+    db_url: str | None = Field(
+        default=None,
+        description=
+        "SQLAlchemy database URL for storing async job metadata, if unset a temporary SQLite database is used.")
+    max_running_async_jobs: int = Field(
+        default=10,
+        description=(
+            "Maximum number of async jobs to run concurrently, this controls the number of dask workers created. "
+            "This parameter is only used when scheduler_address is `None` and a Dask local cluster is created."),
+        ge=1)
     step_adaptor: StepAdaptorConfig = StepAdaptorConfig()
 
     workflow: typing.Annotated[EndpointBase, Field(description="Endpoint for the default workflow.")] = EndpointBase(