fxn 0.0.53__tar.gz → 0.0.55__tar.gz

{fxn-0.0.53 → fxn-0.0.55}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: fxn
-Version: 0.0.53
+Version: 0.0.55
 Summary: Run prediction functions locally in Python. Register at https://fxn.ai.
 Author-email: "NatML Inc." <hi@fxn.ai>
 License:                                  Apache License

{fxn-0.0.53 → fxn-0.0.55}/fxn/beta/__init__.py

@@ -6,8 +6,9 @@
 from .metadata import (
     CoreMLInferenceMetadata, LiteRTInferenceMetadata, LlamaCppInferenceMetadata,
     OnnxInferenceMetadata, OnnxRuntimeInferenceSessionMetadata, OpenVINOInferenceMetadata,
-    QnnInferenceMetadata, QnnInferenceBackend, QnnInferenceQuantization,
-    # Deprecated
-    ONNXInferenceMetadata, ONNXRuntimeInferenceSessionMetadata
+    QnnInferenceMetadata, QnnInferenceBackend, QnnInferenceQuantization, TensorRTInferenceMetadata    
 )
-from .services import RemoteAcceleration
+from .services import RemoteAcceleration
+
+# DEPRECATED
+from .metadata import ONNXInferenceMetadata, ONNXRuntimeInferenceSessionMetadata

{fxn-0.0.53 → fxn-0.0.55}/fxn/beta/cli/__init__.py

@@ -3,4 +3,5 @@
 #   Copyright © 2025 NatML Inc. All Rights Reserved.
 #
 
-from .llm import app as llm_app
+from .llm import app as llm_app
+from .mcp import app as mcp_app

fxn-0.0.55/fxn/beta/cli/mcp.py

@@ -0,0 +1,16 @@
+# 
+#   Function
+#   Copyright © 2025 NatML Inc. All Rights Reserved.
+#
+
+from pathlib import Path
+from typer import Argument, Option, Typer
+from typing_extensions import Annotated
+
+app = Typer(no_args_is_help=True)
+
+@app.command(name="serve", help="Start an MCP server.")
+def serve (
+    port: Annotated[int, Option(help="Port to start the server on.")] = 11436
+):
+    pass

{fxn-0.0.53 → fxn-0.0.55}/fxn/beta/metadata.py

@@ -3,28 +3,46 @@
 #   Copyright © 2025 NatML Inc. All Rights Reserved.
 #
 
-from os import PathLike
 from pathlib import Path
 from pydantic import BaseModel, BeforeValidator, ConfigDict, Field
 from typing import Annotated, Literal
 
 def _validate_torch_module (module: "torch.nn.Module") -> "torch.nn.Module": # type: ignore
     try:
-        from torch.nn import Module # type: ignore
+        from torch.nn import Module
         if not isinstance(module, Module):
-            raise ValueError(f"Expected torch.nn.Module, got {type(module)}")
+            raise ValueError(f"Expected `torch.nn.Module` model but got `{type(module).__qualname__}`")
         return module
     except ImportError:
-        raise ImportError("PyTorch is required to create this metadata but is not installed.")
+        raise ImportError("PyTorch is required to create this metadata but it is not installed.")
 
 def _validate_ort_inference_session (session: "onnxruntime.InferenceSession") -> "onnxruntime.InferenceSession": # type: ignore
     try:
-        from onnxruntime import InferenceSession # type: ignore
+        from onnxruntime import InferenceSession
         if not isinstance(session, InferenceSession):
-            raise ValueError(f"Expected onnxruntime.InferenceSession, got {type(session)}")
+            raise ValueError(f"Expected `onnxruntime.InferenceSession` model but got `{type(session).__qualname__}`")
         return session
     except ImportError:
-        raise ImportError("ONNXRuntime is required to create this metadata but is not installed.")
+        raise ImportError("ONNXRuntime is required to create this metadata but it is not installed.")
+
+def _validate_torch_tensor_args (args: list) -> list:
+    try:
+        from torch import Tensor
+        for idx, arg in enumerate(args):
+            if not isinstance(arg, Tensor):
+                raise ValueError(f"Expected `torch.Tensor` instance at `model_args[{idx}]` but got `{type(arg).__qualname__}`")
+        return args
+    except ImportError:
+        raise ImportError("PyTorch is required to create this metadata but it is not installed.")
+    
+def _validate_llama_cpp_model (model: "llama_cpp.llama.Llama") -> "llama_cpp.llama.Llama": # type: ignore
+    try:
+        from llama_cpp import Llama
+        if not isinstance(model, Llama):
+            raise ValueError(f"Expected `llama_cpp.llama.Llama` model but got `{type(model).__qualname__}`")
+        return model
+    except ImportError:
+        raise ImportError("Llama-cpp-python is required to create this metadata but it is not installed.")
 
 class CoreMLInferenceMetadata (BaseModel):
     """
@@ -39,10 +57,15 @@ class CoreMLInferenceMetadata (BaseModel):
         description="PyTorch module to apply metadata to.",
         exclude=True
     )
-    model_args: list[object] = Field(
+    model_args: Annotated[list[object], BeforeValidator(_validate_torch_tensor_args)] = Field(
         description="Positional inputs to the model.",
         exclude=True
     )
+    output_keys: list[str] | None = Field(
+        default=None,
+        description="Model output dictionary keys. Use this if the model returns a dictionary.",
+        exclude=True
+    )
     model_config = ConfigDict(arbitrary_types_allowed=True, frozen=True)
 
 class OnnxInferenceMetadata (BaseModel):
@@ -58,10 +81,15 @@ class OnnxInferenceMetadata (BaseModel):
         description="PyTorch module to apply metadata to.",
         exclude=True
     )
-    model_args: list[object] = Field(
+    model_args: Annotated[list[object], BeforeValidator(_validate_torch_tensor_args)] = Field(
         description="Positional inputs to the model.",
         exclude=True
     )
+    output_keys: list[str] | None = Field(
+        default=None,
+        description="Model output dictionary keys. Use this if the model returns a dictionary.",
+        exclude=True
+    )
     model_config = ConfigDict(arbitrary_types_allowed=True, frozen=True)
 
 class OnnxRuntimeInferenceSessionMetadata (BaseModel):
@@ -96,10 +124,15 @@ class LiteRTInferenceMetadata (BaseModel):
         description="PyTorch module to apply metadata to.",
         exclude=True
     )
-    model_args: list[object] = Field(
+    model_args: Annotated[list[object], BeforeValidator(_validate_torch_tensor_args)] = Field(
         description="Positional inputs to the model.",
         exclude=True
     )
+    output_keys: list[str] | None = Field(
+        default=None,
+        description="Model output dictionary keys. Use this if the model returns a dictionary.",
+        exclude=True
+    )
     model_config = ConfigDict(arbitrary_types_allowed=True, frozen=True)
 
 class OpenVINOInferenceMetadata (BaseModel):
@@ -115,13 +148,18 @@ class OpenVINOInferenceMetadata (BaseModel):
         description="PyTorch module to apply metadata to.",
         exclude=True
     )
-    model_args: list[object] = Field(
+    model_args: Annotated[list[object], BeforeValidator(_validate_torch_tensor_args)] = Field(
         description="Positional inputs to the model.",
         exclude=True
     )
+    output_keys: list[str] | None = Field(
+        default=None,
+        description="Model output dictionary keys. Use this if the model returns a dictionary.",
+        exclude=True
+    )
     model_config = ConfigDict(arbitrary_types_allowed=True, frozen=True)
 
-QnnInferenceBackend = Literal["cpu", "gpu"] # `htp` coming soon
+QnnInferenceBackend = Literal["cpu", "gpu", "htp"]
 QnnInferenceQuantization = Literal["w8a8", "w8a16", "w4a8", "w4a16"]
 
 class QnnInferenceMetadata (BaseModel):
@@ -139,10 +177,15 @@ class QnnInferenceMetadata (BaseModel):
         description="PyTorch module to apply metadata to.",
         exclude=True
     )
-    model_args: list[object] = Field(
+    model_args: Annotated[list[object], BeforeValidator(_validate_torch_tensor_args)] = Field(
         description="Positional inputs to the model.",
         exclude=True
     )
+    output_keys: list[str] | None = Field(
+        default=None,
+        description="Model output dictionary keys. Use this if the model returns a dictionary.",
+        exclude=True
+    )
     backend: QnnInferenceBackend = Field(
         default="cpu",
         description="QNN backend to execute the model.",
@@ -155,13 +198,58 @@ class QnnInferenceMetadata (BaseModel):
     )
     model_config = ConfigDict(arbitrary_types_allowed=True, frozen=True)
 
-class LlamaCppInferenceMetadata (BaseModel): # INCOMPLETE
+CudaArchitecture = Literal[
+    "sm_80", "sm_86", "sm_87",  # Ampere (A100)
+    "sm_89",                    # Ada Lovelace (L40)
+    "sm_90",                    # Hopper (H100)
+    "sm_100",                   # Blackwell (B200)
+]
+
+TensorRTPrecision = Literal["fp32", "fp16", "int8", "int4"]
+
+class TensorRTInferenceMetadata (BaseModel):
+    """
+    Metadata required to lower a PyTorch model for inference on Nvidia GPUs with TensorRT.
+
+    Members:
+        model (torch.nn.Module): PyTorch module to apply metadata to.
+        model_args (tuple[Tensor,...]): Positional inputs to the model.
+        cuda_arch (CudaArchitecture): Target CUDA architecture for the TensorRT engine. Defaults to `sm_80` (Ampere).
+        precision (TensorRTPrecision): TensorRT engine inference precision. Defaults to `fp16`.
+    """
+    kind: Literal["meta.inference.tensorrt"] = "meta.inference.tensorrt"
+    model: Annotated[object, BeforeValidator(_validate_torch_module)] = Field(
+        description="PyTorch module to apply metadata to.",
+        exclude=True
+    )
+    model_args: Annotated[list[object], BeforeValidator(_validate_torch_tensor_args)] = Field(
+        description="Positional inputs to the model.",
+        exclude=True
+    )
+    output_keys: list[str] | None = Field(
+        default=None,
+        description="Model output dictionary keys. Use this if the model returns a dictionary.",
+        exclude=True
+    )
+    cuda_arch: CudaArchitecture = Field(
+        default="sm_80",
+        description="Target CUDA architecture for the TensorRT engine.",
+        exclude=True
+    )
+    precision: TensorRTPrecision = Field(
+        default="fp16",
+        description="TensorRT engine inference precision.",
+        exclude=True
+    )
+    model_config = ConfigDict(arbitrary_types_allowed=True, frozen=True)
+
+class LlamaCppInferenceMetadata (BaseModel):
     """
-    Metadata required to lower a GGUF model for LLM inference.
+    Metadata required to lower a Llama.cpp model for LLM inference.
     """
-    kind: Literal["meta.inference.gguf"] = "meta.inference.gguf"
-    model_path: Path = Field(
-        description="GGUF model path. The model must exist at this path in the compiler sandbox.",
+    kind: Literal["meta.inference.llama_cpp"] = "meta.inference.llama_cpp"
+    model: Annotated[object, BeforeValidator(_validate_llama_cpp_model)] = Field(
+        description="Llama model that metadata applies to.",
         exclude=True
     )
     model_config = ConfigDict(arbitrary_types_allowed=True, frozen=True)

fxn-0.0.55/fxn/cli/__init__.py

@@ -0,0 +1,99 @@
+# 
+#   Function
+#   Copyright © 2025 NatML Inc. All Rights Reserved.
+#
+
+import typer
+
+from ..logging import TracebackMarkupConsole
+from ..version import __version__
+
+from .auth import app as auth_app
+from .compile import compile_predictor, triage_predictor
+from .misc import cli_options
+from .predictions import create_prediction
+from .predictors import archive_predictor, delete_predictor, retrieve_predictor
+from .sources import retrieve_source
+from ..beta.cli import llm_app, mcp_app
+
+# Define CLI
+typer.main.console_stderr = TracebackMarkupConsole()
+app = typer.Typer(
+    name=f"Function CLI {__version__}",
+    no_args_is_help=True,
+    pretty_exceptions_show_locals=False,
+    pretty_exceptions_short=True,
+    add_completion=False
+)
+
+# Add top level options
+app.callback()(cli_options)
+
+# Predictions
+app.command(
+    name="predict",
+    help="Make a prediction.",
+    context_settings={ "allow_extra_args": True, "ignore_unknown_options": True },
+    rich_help_panel="Predictions"
+)(create_prediction)
+app.command(
+    name="source",
+    help="Retrieve the generated C++ code for a given prediction.",
+    rich_help_panel="Predictions"
+)(retrieve_source)
+
+# Predictors
+app.command(
+    name="compile",
+    help="Create a predictor by compiling a Python function.",
+    rich_help_panel="Predictors"
+)(compile_predictor)
+app.command(
+    name="retrieve",
+    help="Retrieve a predictor.",
+    rich_help_panel="Predictors"
+)(retrieve_predictor)
+app.command(
+    name="archive",
+    help="Archive a predictor." ,
+    rich_help_panel="Predictors"
+)(archive_predictor)
+app.command(
+    name="delete",
+    help="Delete a predictor.",
+    rich_help_panel="Predictors"
+)(delete_predictor)
+
+# Subcommands
+app.add_typer(
+    auth_app,
+    name="auth",
+    help="Login, logout, and check your authentication status.",
+    rich_help_panel="Auth"
+)
+app.add_typer(
+    llm_app,
+    name="llm",
+    hidden=True,
+    help="Work with large language models (LLMs).",
+    rich_help_panel="Beta"
+)
+app.add_typer(
+    mcp_app,
+    name="mcp",
+    hidden=True,
+    help="Provide prediction functions as tools for use by AI assistants.",
+    rich_help_panel="Beta"
+)
+
+# Insiders
+app.command(
+    name="triage",
+    help="Triage a compile error.",
+    rich_help_panel="Insiders",
+    hidden=True
+)(triage_predictor)
+
+# Run
+if __name__ == "__main__":
+    app()

{fxn-0.0.53 → fxn-0.0.55}/fxn/cli/auth.py

@@ -12,7 +12,7 @@ from ..function import Function
 app = Typer(no_args_is_help=True)
 
 @app.command(name="login", help="Login to Function.")
-def login (
+def login(
     access_key: str=Argument(..., help="Function access key.", envvar="FXN_ACCESS_KEY")
 ):
     fxn = Function(access_key=access_key)
@@ -22,18 +22,18 @@ def login (
     print_json(data=user)
 
 @app.command(name="status", help="Get current authentication status.")
-def auth_status ():
+def auth_status():
     fxn = Function(get_access_key())
     user = fxn.users.retrieve()
     user = user.model_dump() if user else None
     print_json(data=user)
 
 @app.command(name="logout", help="Logout from Function.")
-def logout ():
+def logout():
     _set_access_key(None)
     print("Successfully logged out of Function")
 
-def get_access_key () -> str:
+def get_access_key() -> str:
     """
     Get the CLI access key.
 
@@ -46,7 +46,7 @@ def get_access_key () -> str:
     with open(credentials_path) as f:
         return f.read()
 
-def _set_access_key (key: str):
+def _set_access_key(key: str):
     """
     Set the CLI access key.
 

{fxn-0.0.53 → fxn-0.0.55}/fxn/cli/compile.py

@@ -9,9 +9,11 @@ from inspect import getmembers, getmodulename, isfunction
 from pathlib import Path
 from pydantic import BaseModel
 from rich import print as print_rich
+from rich.panel import Panel
 import sys
 from typer import Argument, Option
 from typing import Callable, Literal
+from typing_extensions import Annotated
 from urllib.parse import urlparse, urlunparse
 
 from ..client import FunctionAPIError
@@ -21,16 +23,39 @@ from ..sandbox import EntrypointCommand
 from ..logging import CustomProgress, CustomProgressTask
 from .auth import get_access_key
 
-class CompileError (Exception):
-    pass
-
-def compile_predictor (
+def compile_predictor(
     path: str=Argument(..., help="Predictor path."),
     overwrite: bool=Option(False, "--overwrite", help="Whether to delete any existing predictor with the same tag before compiling."),
 ):
     run_async(_compile_predictor_async(path, overwrite=overwrite))
 
-async def _compile_predictor_async (
+def triage_predictor(
+    reference_code: Annotated[str, Argument(help="Predictor compilation reference code.")]
+):
+    fxn = Function(get_access_key())
+    error = fxn.client.request(
+        method="GET",
+        path=f"/predictors/triage?referenceCode={reference_code}",
+        response_type=_TriagedCompileError
+    )
+    user_panel = Panel(
+        error.user,
+        title="User Error",
+        title_align="left",
+        highlight=True,
+        border_style="bright_red"
+    )
+    internal_panel = Panel(
+        error.internal,
+        title="Internal Error",
+        title_align="left",
+        highlight=True,
+        border_style="gold1"
+    )
+    print_rich(user_panel)
+    print_rich(internal_panel)
+
+async def _compile_predictor_async(
     path: str,
     *,
     overwrite: bool
@@ -78,11 +103,11 @@ async def _compile_predictor_async (
                         task_queue.push_log(event)
                     elif isinstance(event, _ErrorEvent):
                         task_queue.push_error(event)
-                        raise CompileError(event.data.error)
+                        raise _CompileError(event.data.error)
     predictor_url = _compute_predictor_url(fxn.client.api_url, spec.tag)
     print_rich(f"\n[bold spring_green3]🎉 Predictor is now being compiled.[/bold spring_green3] Check it out at [link={predictor_url}]{predictor_url}[/link]")
 
-def _load_predictor_func (path: str) -> Callable[...,object]:
+def _load_predictor_func(path: str) -> Callable[...,object]:
     if "" not in sys.path:
         sys.path.insert(0, "")
     path: Path = Path(path).resolve()
@@ -97,7 +122,7 @@ def _load_predictor_func (path: str) -> Callable[...,object]:
     main_func = next(func for _, func in getmembers(module, isfunction) if hasattr(func, "__predictor_spec"))
     return main_func
 
-def _compute_predictor_url (api_url: str, tag: str) -> str:
+def _compute_predictor_url(api_url: str, tag: str) -> str:
     parsed_url = urlparse(api_url)
     hostname_parts = parsed_url.hostname.split(".")
     if hostname_parts[0] == "api":
@@ -107,32 +132,39 @@ def _compute_predictor_url (api_url: str, tag: str) -> str:
     predictor_url = urlunparse(parsed_url._replace(netloc=netloc, path=f"{tag}"))
     return predictor_url
 
-class _Predictor (BaseModel):
+class _Predictor(BaseModel):
     tag: str
 
-class _LogData (BaseModel):
+class _LogData(BaseModel):
     message: str
     level: int = 0
     status: Literal["success", "error"] = "success"
     update: bool = False
 
-class _LogEvent (BaseModel):
+class _LogEvent(BaseModel):
     event: Literal["log"]
     data: _LogData
 
-class _ErrorData (BaseModel):
+class _ErrorData(BaseModel):
     error: str
 
-class _ErrorEvent (BaseModel):
+class _ErrorEvent(BaseModel):
     event: Literal["error"]
     data: _ErrorData
 
+class _CompileError(Exception):
+    pass
+
+class _TriagedCompileError(BaseModel):
+    user: str
+    internal: str
+
 class ProgressLogQueue:
 
-    def __init__ (self):
+    def __init__(self):
         self.queue: list[tuple[int, CustomProgressTask]] = []
 
-    def push_log (self, event: _LogEvent):
+    def push_log(self, event: _LogEvent):
         # Check for update
         if event.data.update and self.queue:
             current_level, current_task = self.queue[-1]
@@ -149,15 +181,15 @@ class ProgressLogQueue:
         task.__enter__()
         self.queue.append((event.data.level, task))
 
-    def push_error (self, error: _ErrorEvent):
+    def push_error(self, error: _ErrorEvent):
         while self.queue:
             _, current_task = self.queue.pop()
             current_task.__exit__(RuntimeError, None, None)
 
-    def __enter__ (self):
+    def __enter__(self):
         return self
     
-    def __exit__ (self, exc_type, exc_value, traceback):
+    def __exit__(self, exc_type, exc_value, traceback):
         while self.queue:
             _, current_task = self.queue.pop()
             current_task.__exit__(None, None, None)

{fxn-0.0.53 → fxn-0.0.55}/fxn/cli/misc.py

@@ -9,22 +9,22 @@ from webbrowser import open as open_browser
 
 from ..version import __version__
 
-def _explore (value: bool):
+def _explore(value: bool):
     if value:
         open_browser("https://fxn.ai/explore")
         raise Exit()
 
-def _learn (value: bool):
+def _learn(value: bool):
     if value:
         open_browser("https://docs.fxn.ai")
         raise Exit()
 
-def _version (value: bool):
+def _version(value: bool):
     if value:
         print(__version__)
         raise Exit()
 
-def cli_options (
+def cli_options(
     explore: bool = Option(None, "--explore", callback=_explore, help="Explore predictors on Function."),
     learn: bool = Option(None, "--learn", callback=_learn, help="Learn about Function."),
     version: bool = Option(None, "--version", callback=_version, help="Get the Function CLI version.")

{fxn-0.0.53 → fxn-0.0.55}/fxn/cli/predictions.py

@@ -24,7 +24,7 @@ def create_prediction (
 ):
     run_async(_predict_async(tag, quiet=quiet, context=context))
 
-async def _predict_async (tag: str, quiet: bool, context: Context):
+async def _predict_async(tag: str, quiet: bool, context: Context):
     # Preload
     with CustomProgress(transient=True, disable=quiet):
         fxn = Function(get_access_key())
@@ -42,16 +42,7 @@ async def _predict_async (tag: str, quiet: bool, context: Context):
             prediction = fxn.predictions.create(tag, inputs=inputs)
     _log_prediction(prediction)
 
-def _parse_value (value: str):
-    """
-    Parse a value from a CLI argument.
-
-    Parameters:
-        value (str): CLI input argument.
-
-    Returns:
-        bool | int | float | str | Path: Parsed value.
-    """
+def _parse_value (value: str) -> float | int | bool | str | Image.Image | BytesIO:
     # Boolean
     if value == "true":
         return True
@@ -81,14 +72,14 @@ def _parse_value (value: str):
     # String
     return value
 
-def _log_prediction (prediction: Prediction):
+def _log_prediction(prediction: Prediction):
     images = [value for value in prediction.results or [] if isinstance(value, Image.Image)]
     prediction.results = [_serialize_value(value) for value in prediction.results] if prediction.results is not None else None
     print_json(data=prediction.model_dump())
     for image in images:
         image.show()
 
-def _serialize_value (value):
+def _serialize_value(value) -> str:
     if isinstance(value, ndarray):
         return array_repr(value)
     if isinstance(value, Image.Image):

{fxn-0.0.53 → fxn-0.0.55}/fxn/cli/predictors.py

@@ -10,7 +10,7 @@ from ..function import Function
 from ..logging import CustomProgress, CustomProgressTask
 from .auth import get_access_key
 
-def retrieve_predictor (
+def retrieve_predictor(
     tag: str=Argument(..., help="Predictor tag.")
 ):
     with CustomProgress(transient=True):
@@ -20,7 +20,7 @@ def retrieve_predictor (
             predictor = predictor.model_dump() if predictor else None
             print_json(data=predictor)
 
-def archive_predictor (
+def archive_predictor(
     tag: str=Argument(..., help="Predictor tag.")
 ):
     with CustomProgress():
@@ -34,7 +34,7 @@ def archive_predictor (
                 path=f"/predictors/{tag}/archive"
             )
 
-def delete_predictor (
+def delete_predictor(
     tag: str=Argument(..., help="Predictor tag.")
 ):
     with CustomProgress():

{fxn-0.0.53 → fxn-0.0.55}/fxn/cli/sources.py

@@ -14,9 +14,9 @@ from ..function import Function
 from ..logging import CustomProgress, CustomProgressTask
 from .auth import get_access_key
 
-def retrieve_source (
-    predictor: Annotated[str, Option(help="Predictor tag.")] = None,
+def retrieve_source(
     prediction: Annotated[str, Option(help="Prediction identifier. If specified, this MUST be from a prediction returned by the Function API.")] = None,
+    predictor: Annotated[str, Option(help="Predictor tag. If specified, a prediction will be made with this predictor before retrieving the source.")] = None,
     output: Annotated[Path, Option(help="Path to output source file.")] = Path("predictor.cpp")
 ):
     if not ((predictor is not None) ^ (prediction is not None)):
@@ -37,7 +37,7 @@ def retrieve_source (
             source.code = str(output.resolve())
             print_json(data=source.model_dump(mode="json", by_alias=True))
 
-class _PredictionSource (BaseModel):
+class _PredictionSource(BaseModel):
     tag: str
     target: str
     code: str

{fxn-0.0.53 → fxn-0.0.55}/fxn/compile.py

@@ -14,10 +14,10 @@ from typing import Any, Callable, Literal, ParamSpec, TypeVar, cast
 from .beta import (
     CoreMLInferenceMetadata, LiteRTInferenceMetadata, LlamaCppInferenceMetadata,
     OnnxInferenceMetadata, OnnxRuntimeInferenceSessionMetadata, OpenVINOInferenceMetadata,
-    QnnInferenceMetadata
+    QnnInferenceMetadata, TensorRTInferenceMetadata
 )
 from .sandbox import Sandbox
-from .types import AccessMode
+from .types import PredictorAccess
 
 CompileTarget = Literal[
     "android",
@@ -36,7 +36,8 @@ CompileMetadata = (
     OnnxInferenceMetadata               |
     OnnxRuntimeInferenceSessionMetadata |
     OpenVINOInferenceMetadata           |
-    QnnInferenceMetadata
+    QnnInferenceMetadata                |
+    TensorRTInferenceMetadata
 )
 
 P = ParamSpec("P")
@@ -51,7 +52,7 @@ class PredictorSpec (BaseModel):
     sandbox: Sandbox = Field(description="Sandbox to compile the function.")
     targets: list[str] | None = Field(description="Targets to compile this predictor for. Pass `None` to compile for our default targets.")
     metadata: list[object] = Field(default=[], description="Metadata to use while compiling the function.")
-    access: AccessMode = Field(description="Predictor access.")
+    access: PredictorAccess = Field(description="Predictor access.")
     card: str | None = Field(default=None, description="Predictor card (markdown).")
     media: str | None = Field(default=None, description="Predictor media URL.")
     license: str | None = Field(default=None, description="Predictor license URL. This is required for public predictors.")
@@ -65,7 +66,7 @@ def compile (
     trace_modules: list[ModuleType]=[],
     targets: list[CompileTarget]=None,
     metadata: list[CompileMetadata]=[],
-    access: AccessMode=AccessMode.Private,
+    access: PredictorAccess="private",
     card: str | Path=None,
     media: Path=None,
     license: str=None,
@@ -81,7 +82,7 @@ def compile (
         trace_modules (list): Modules to trace and compile.
         targets (list): Targets to compile this predictor for. Pass `None` to compile for our default targets.
         metadata (list): Metadata to use while compiling the function.
-        access (AccessMode): Predictor access.
+        access (PredictorAccess): Predictor access.
         card (str | Path): Predictor card markdown string or path to card.
         media (Path): Predictor thumbnail image (jpeg or png) path.
         license (str): Predictor license URL. This is required for public predictors.

{fxn-0.0.53 → fxn-0.0.55}/fxn/sandbox.py

@@ -57,6 +57,8 @@ class EntrypointCommand (UploadableCommand):
 class PipInstallCommand (BaseModel):
     kind: Literal["pip_install"] = "pip_install"
     packages: list[str]
+    index_url: str | None
+    flags: str
 
 class AptInstallCommand (BaseModel):
     kind: Literal["apt_install"] = "apt_install"
@@ -133,14 +135,25 @@ class Sandbox (BaseModel):
         )
         return Sandbox(commands=self.commands + [command])
 
-    def pip_install (self, *packages: str) -> Sandbox:
+    def pip_install (
+        self,
+        *packages: str,
+        index_url: str=None,
+        flags: str=""
+    ) -> Sandbox:
         """
         Install Python packages in the sandbox.
 
         Parameters:
             packages (list): Packages to install.
+            index_url (str | None): Index URL to search for package.
+            flags (str): Additional flags to pass to `pip`.
         """
-        command = PipInstallCommand(packages=packages)
+        command = PipInstallCommand(
+            packages=packages,
+            index_url=index_url,
+            flags=flags
+        )
         return Sandbox(commands=self.commands + [command])
 
     def apt_install (self, *packages: str) -> Sandbox:

{fxn-0.0.53 → fxn-0.0.55}/fxn/services/prediction.py

@@ -117,6 +117,21 @@ class PredictionService:
                 with prediction:
                     yield self.__to_prediction(tag, prediction)
 
+    def delete (self, tag: str) -> bool:
+        """
+        Delete a predictor that is loaded in memory.
+
+        Parameters:
+            tag (str): Predictor tag.
+
+        Returns:
+            bool: Whether the predictor was successfully deleted from memory.
+        """
+        if tag not in self.__cache:
+            return False
+        with self.__cache.pop(tag):
+            return True
+
     def __create_raw_prediction (
         self,
         tag: str,
@@ -245,7 +260,7 @@ class PredictionService:
                     tmp_file.write(chunk)
                     completed += len(chunk)
                     task.update(total=size, completed=completed)
-            Path(tmp_file.name).replace(path)
+        Path(tmp_file.name).replace(path)
         return path
 
     def __get_resource_path (self, resource: PredictionResource) -> Path:

{fxn-0.0.53 → fxn-0.0.55}/fxn/types/__init__.py

@@ -5,5 +5,5 @@
 
 from .dtype import Dtype
 from .prediction import Acceleration, Prediction, PredictionResource
-from .predictor import AccessMode, EnumerationMember, Parameter, Predictor, PredictorStatus, Signature
+from .predictor import EnumerationMember, Parameter, Predictor, PredictorAccess, PredictorStatus, Signature
 from .user import User

{fxn-0.0.53 → fxn-0.0.55}/fxn/types/predictor.py

@@ -3,28 +3,15 @@
 #   Copyright © 2025 NatML Inc. All Rights Reserved.
 #
 
-from enum import Enum
 from pydantic import AliasChoices, BaseModel, ConfigDict, Field
-from typing import Any
+from typing import Any, Literal
 
 from .dtype import Dtype
 from .user import User
 
-class AccessMode (str, Enum):
-    """
-    Predictor access mode.
-    """
-    Public = "PUBLIC"
-    Private = "PRIVATE"
+PredictorAccess = Literal["public", "private", "unlisted"]
 
-class PredictorStatus (str, Enum):
-    """
-    Predictor status.
-    """
-    Compiling = "COMPILING"
-    Active = "ACTIVE"
-    Invalid = "INVALID"
-    Archived = "ARCHIVED"
+PredictorStatus = Literal["compiling", "active", "archived"]
 
 class EnumerationMember (BaseModel):
     """
@@ -79,7 +66,7 @@ class Predictor (BaseModel):
         owner (User): Predictor owner.
         name (str): Predictor name.
         status (PredictorStatus): Predictor status.
-        access (AccessMode): Predictor access.
+        access (PredictorAccess): Predictor access.
         signature (Signature): Predictor signature.
         created (str): Date created.
         description (str): Predictor description.
@@ -91,7 +78,7 @@ class Predictor (BaseModel):
     owner: User = Field(description="Predictor owner.")
     name: str = Field(description="Predictor name.")
     status: PredictorStatus = Field(description="Predictor status.")
-    access: AccessMode = Field(description="Predictor access.")
+    access: PredictorAccess = Field(description="Predictor access.")
     signature: Signature = Field(description="Predictor signature.")
     created: str = Field(description="Date created.")
     description: str | None = Field(default=None, description="Predictor description.")

{fxn-0.0.53 → fxn-0.0.55}/fxn/version.py

@@ -3,4 +3,4 @@
 #   Copyright © 2025 NatML Inc. All Rights Reserved.
 #
 
-__version__ = "0.0.53"
+__version__ = "0.0.55"

{fxn-0.0.53 → fxn-0.0.55}/fxn.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: fxn
-Version: 0.0.53
+Version: 0.0.55
 Summary: Run prediction functions locally in Python. Register at https://fxn.ai.
 Author-email: "NatML Inc." <hi@fxn.ai>
 License:                                  Apache License

{fxn-0.0.53 → fxn-0.0.55}/fxn.egg-info/SOURCES.txt

@@ -19,6 +19,7 @@ fxn/beta/client.py
 fxn/beta/metadata.py
 fxn/beta/cli/__init__.py
 fxn/beta/cli/llm.py
+fxn/beta/cli/mcp.py
 fxn/beta/llm/__init__.py
 fxn/beta/llm/server.py
 fxn/beta/services/__init__.py

fxn-0.0.53/fxn/cli/__init__.py

@@ -1,53 +0,0 @@
-# 
-#   Function
-#   Copyright © 2025 NatML Inc. All Rights Reserved.
-#
-
-import typer
-
-from ..logging import TracebackMarkupConsole
-from ..version import __version__
-
-from .auth import app as auth_app
-from .compile import compile_predictor
-from .misc import cli_options
-from .predictions import create_prediction
-from .predictors import archive_predictor, delete_predictor, retrieve_predictor
-from .sources import retrieve_source
-from ..beta.cli import llm_app
-
-# Define CLI
-typer.main.console_stderr = TracebackMarkupConsole()
-app = typer.Typer(
-    name=f"Function CLI {__version__}",
-    no_args_is_help=True,
-    pretty_exceptions_show_locals=False,
-    pretty_exceptions_short=True,
-    add_completion=False
-)
-
-# Add top level options
-app.callback()(cli_options)
-
-# Add subcommands
-app.add_typer(auth_app, name="auth", help="Login, logout, and check your authentication status.")
-app.add_typer(llm_app, name="llm", hidden=True, help="Work with large language models (LLMs).")
-
-# Add top-level commands
-app.command(
-    name="predict",
-    help="Make a prediction.",
-    context_settings={ "allow_extra_args": True, "ignore_unknown_options": True }
-)(create_prediction)
-app.command(
-    name="compile",
-    help="Create a predictor by compiling a Python function."
-)(compile_predictor)
-app.command(name="retrieve", help="Retrieve a predictor.")(retrieve_predictor)
-app.command(name="archive", help="Archive a predictor.")(archive_predictor)
-app.command(name="delete", help="Delete a predictor.")(delete_predictor)
-app.command(name="source", help="Retrieve the generated native code for a given predictor.")(retrieve_source)
-
-# Run
-if __name__ == "__main__":
-    app()