Tensorflow integration

README.md

@@ -32,6 +32,13 @@ Call `client.instrument()` to activate auto-tracking for a supported library. Mo
 
 See the `examples/` folder for complete working examples.
 
+### Integration initialization
+
+Initialize integrations at process startup, before model loading begins. Instrumentation patches are applied per process and should be installed before imports and constructor calls on instrumented libraries.
+
+For high-priority paths, keep explicit registration with `client.load(...)` or `client.register_model(...)` as a fallback when model creation does not go through a patched API.
+For an explicit fallback pattern, see `examples/gguf_gemma_manual_example.py`.
+
 ### PyTorch (custom models)
 
 PyTorch models are user-defined subclasses, so there is no single constructor to patch. Use `client.load()` to time construction and track load/unload automatically; inference is tracked via forward hooks once the model is registered.
@@ -67,6 +74,19 @@ outputs = session.run(None, {"input": image})  # tracked automatically
 
 See `examples/onnx_example.py` for a complete example.
 
+### TensorFlow
+
+```python
+import tensorflow as tf
+
+client.instrument("tensorflow")
+
+model = tf.keras.models.load_model("model.keras")  # tracked automatically
+output = model(batch, training=False)  # tracked automatically
+```
+
+See `examples/tensorflow_example.py` for a complete example.
+
 ### timm
 
 ```python

examples/tensorflow_example.py

@@ -0,0 +1,51 @@
+# /// script
+# requires-python = ">=3.10"
+# dependencies = ["wildedge-sdk", "tensorflow", "numpy"]
+#
+# [tool.uv.sources]
+# wildedge-sdk = { path = "..", editable = true }
+# ///
+"""TensorFlow integration example. Run with: uv run tensorflow_example.py."""
+
+from __future__ import annotations
+
+from pathlib import Path
+from tempfile import TemporaryDirectory
+
+import numpy as np
+import tensorflow as tf
+
+import wildedge
+
+client = wildedge.WildEdge(
+    app_version="1.0.0",  # set WILDEDGE_DSN env var
+)
+client.instrument("tensorflow")
+
+
+def build_and_save_model(save_path: Path) -> None:
+    model = tf.keras.Sequential(
+        [
+            tf.keras.layers.Input(shape=(16,)),
+            tf.keras.layers.Dense(32, activation="relu"),
+            tf.keras.layers.Dense(8),
+        ]
+    )
+    # Trigger variable creation before save.
+    _ = model(np.zeros((1, 16), dtype=np.float32))
+    model.save(save_path)
+
+
+with TemporaryDirectory() as temp_dir:
+    model_path = Path(temp_dir) / "demo_model.keras"
+    build_and_save_model(model_path)
+
+    # load_model is auto-instrumented by client.instrument("tensorflow")
+    loaded = tf.keras.models.load_model(model_path)
+
+    batch = np.random.randn(4, 16).astype(np.float32)
+    output = loaded(batch, training=False)
+    print("output shape:", tuple(output.shape))
+
+
+client.close()

tests/test_integrations.py

@@ -1,4 +1,4 @@
-"""Tests for ONNX, GGUF, PyTorch, and Keras integration extractors."""
+"""Tests for ONNX, GGUF, and Keras integration extractors."""
 
 from __future__ import annotations
 
@@ -24,15 +24,6 @@
 from wildedge.integrations.onnx import (
     _detect_quantization as onnx_detect_quantization,
 )
-from wildedge.integrations.pytorch import (
-    PytorchExtractor,
-)
-from wildedge.integrations.pytorch import (
-    _detect_accelerator as torch_detect_accelerator,
-)
-from wildedge.integrations.pytorch import (
-    _detect_quantization as torch_detect_quantization,
-)
 from wildedge.model import ModelHandle, ModelInfo
 
 # ---------------------------------------------------------------------------
@@ -97,41 +88,6 @@ def __call__(self, *args, **kwargs):
         raise RuntimeError("cuda oom")
 
 
-# PyTorch
-
-
-class _TorchBase:
-    """Looks like torch.nn.Module to the MRO check."""
-
-
-_TorchBase.__name__ = "Module"
-_TorchBase.__module__ = "torch.nn.modules.module"
-
-
-class _FakeParam:
-    class _Device:
-        type = "cpu"
-
-    device = _Device()
-    dtype = "torch.float32"
-
-
-class _FakeTorchModel(_TorchBase):
-    def parameters(self):
-        yield _FakeParam()
-
-    def modules(self):
-        return iter([self])
-
-    def register_forward_pre_hook(self, hook):
-        self._pre_hook = hook
-        return MagicMock()
-
-    def register_forward_hook(self, hook):
-        self._post_hook = hook
-        return MagicMock()
-
-
 # Keras
 
 
@@ -309,78 +265,6 @@ def test_install_hooks_tracks_error_on_exception(self, publish_spy):
         assert publish_spy.events[0]["event_type"] == "error"
 
 
-# ---------------------------------------------------------------------------
-# PyTorch
-# ---------------------------------------------------------------------------
-
-
-class TestPytorchExtractor:
-    extractor = PytorchExtractor()
-
-    def test_can_handle_torch_module(self):
-        assert self.extractor.can_handle(_FakeTorchModel()) is True
-
-    def test_can_handle_rejects_plain_object(self):
-        assert self.extractor.can_handle(object()) is False
-
-    def test_detect_accelerator_reads_parameter_device(self):
-        model = _FakeTorchModel()
-        assert torch_detect_accelerator(model) == "cpu"
-
-    def test_detect_accelerator_cuda(self):
-        model = _FakeTorchModel()
-        model.parameters = lambda: iter([MagicMock(device=MagicMock(type="cuda"))])
-        assert torch_detect_accelerator(model) == "cuda"
-
-    def test_detect_accelerator_no_parameters_falls_back(self):
-        model = _FakeTorchModel()
-        model.parameters = lambda: iter([])
-        assert isinstance(torch_detect_accelerator(model), str)
-
-    def test_detect_quantization_by_module_name(self):
-        model = _FakeTorchModel()
-
-        class QuantizedLinear:
-            pass
-
-        QuantizedLinear.__module__ = "torch.nn.quantized"
-        model.modules = lambda: iter([QuantizedLinear()])
-        model.parameters = lambda: iter([])
-        assert torch_detect_quantization(model) == "int8"
-
-    def test_detect_quantization_by_param_dtype(self):
-        model = _FakeTorchModel()
-        model.modules = lambda: iter([])
-        model.parameters = lambda: iter([MagicMock(dtype="torch.float16")])
-        assert torch_detect_quantization(model) == "f16"
-
-    def test_extract_info_uses_class_name_as_model_id(self):
-        model = _FakeTorchModel()
-        model_id, info = self.extractor.extract_info(model, {})
-        assert model_id == "_FakeTorchModel"
-        assert info.model_format == "pytorch"
-
-    def test_extract_info_override_model_id(self):
-        model = _FakeTorchModel()
-        model_id, _ = self.extractor.extract_info(model, {"id": "my-resnet"})
-        assert model_id == "my-resnet"
-
-    def test_install_hooks_publishes_inference(self, publish_spy):
-        model = _FakeTorchModel()
-        handle = make_handle(publish_spy)
-        self.extractor.install_hooks(model, handle)
-        model._pre_hook(model, (None,))
-        model._post_hook(model, (None,), None)
-        assert len(publish_spy.events) == 1
-        assert publish_spy.events[0]["event_type"] == "inference"
-
-    def test_install_hooks_sets_detected_accelerator(self, publish_spy):
-        model = _FakeTorchModel()
-        handle = make_handle(publish_spy)
-        self.extractor.install_hooks(model, handle)
-        assert handle.detected_accelerator == "cpu"
-
-
 # ---------------------------------------------------------------------------
 # Keras
 # ---------------------------------------------------------------------------

tests/test_integrations_pytorch.py

@@ -0,0 +1,165 @@
+"""Tests for PyTorch integration extractor."""
+
+from __future__ import annotations
+
+from unittest.mock import MagicMock
+
+from wildedge.integrations.pytorch import (
+    PytorchExtractor,
+)
+from wildedge.integrations.pytorch import (
+    _detect_accelerator as torch_detect_accelerator,
+)
+from wildedge.integrations.pytorch import (
+    _detect_quantization as torch_detect_quantization,
+)
+from wildedge.model import ModelHandle, ModelInfo
+
+
+def make_handle(publish_spy) -> ModelHandle:
+    info = ModelInfo(
+        model_name="test",
+        model_version="1.0",
+        model_source="local",
+        model_format="test",
+    )
+    return ModelHandle(model_id="m", info=info, publish=publish_spy)
+
+
+class _TorchBase:
+    """Looks like torch.nn.Module to the MRO check."""
+
+
+_TorchBase.__name__ = "Module"
+_TorchBase.__module__ = "torch.nn.modules.module"
+
+
+class _FakeParam:
+    class _Device:
+        type = "cpu"
+
+    device = _Device()
+    dtype = "torch.float32"
+
+
+class _FakeTorchModel(_TorchBase):
+    def parameters(self):
+        yield _FakeParam()
+
+    def modules(self):
+        return iter([self])
+
+    def register_forward_pre_hook(self, hook):
+        self._pre_hook = hook
+        return MagicMock()
+
+    def register_forward_hook(self, hook):
+        self._post_hook = hook
+        return MagicMock()
+
+
+class TestPytorchExtractor:
+    extractor = PytorchExtractor()
+
+    def test_can_handle_torch_module(self):
+        assert self.extractor.can_handle(_FakeTorchModel()) is True
+
+    def test_can_handle_rejects_plain_object(self):
+        assert self.extractor.can_handle(object()) is False
+
+    def test_detect_accelerator_reads_parameter_device(self):
+        model = _FakeTorchModel()
+        assert torch_detect_accelerator(model) == "cpu"
+
+    def test_detect_accelerator_cuda(self):
+        model = _FakeTorchModel()
+        model.parameters = lambda: iter([MagicMock(device=MagicMock(type="cuda"))])
+        assert torch_detect_accelerator(model) == "cuda"
+
+    def test_detect_accelerator_no_parameters_falls_back(self):
+        model = _FakeTorchModel()
+        model.parameters = lambda: iter([])
+        assert isinstance(torch_detect_accelerator(model), str)
+
+    def test_detect_quantization_by_module_name(self):
+        model = _FakeTorchModel()
+
+        class QuantizedLinear:
+            pass
+
+        QuantizedLinear.__module__ = "torch.nn.quantized"
+        model.modules = lambda: iter([QuantizedLinear()])
+        model.parameters = lambda: iter([])
+        assert torch_detect_quantization(model) == "int8"
+
+    def test_detect_quantization_by_param_dtype(self):
+        model = _FakeTorchModel()
+        model.modules = lambda: iter([])
+        model.parameters = lambda: iter([MagicMock(dtype="torch.float16")])
+        assert torch_detect_quantization(model) == "f16"
+
+    def test_detect_quantization_by_param_dtype_bf16(self):
+        model = _FakeTorchModel()
+        model.modules = lambda: iter([])
+        model.parameters = lambda: iter([MagicMock(dtype="torch.bfloat16")])
+        assert torch_detect_quantization(model) == "bf16"
+
+    def test_detect_quantization_by_param_dtype_qint(self):
+        model = _FakeTorchModel()
+        model.modules = lambda: iter([])
+        model.parameters = lambda: iter([MagicMock(dtype="torch.qint8")])
+        assert torch_detect_quantization(model) == "int8"
+
+    def test_detect_quantization_by_param_dtype_quint(self):
+        model = _FakeTorchModel()
+        model.modules = lambda: iter([])
+        model.parameters = lambda: iter([MagicMock(dtype="torch.quint8")])
+        assert torch_detect_quantization(model) == "int8"
+
+    def test_detect_quantization_by_param_dtype_int8(self):
+        model = _FakeTorchModel()
+        model.modules = lambda: iter([])
+        model.parameters = lambda: iter([MagicMock(dtype="torch.int8")])
+        assert torch_detect_quantization(model) == "int8"
+
+    def test_detect_quantization_returns_none_when_unknown(self):
+        model = _FakeTorchModel()
+        model.modules = lambda: iter([])
+        model.parameters = lambda: iter([MagicMock(dtype="torch.float32")])
+        assert torch_detect_quantization(model) is None
+
+    def test_detect_quantization_returns_none_on_exception(self):
+        model = _FakeTorchModel()
+        model.modules = lambda: iter([])
+
+        def broken_parameters():
+            raise RuntimeError("broken params")
+
+        model.parameters = broken_parameters
+        assert torch_detect_quantization(model) is None
+
+    def test_extract_info_uses_class_name_as_model_id(self):
+        model = _FakeTorchModel()
+        model_id, info = self.extractor.extract_info(model, {})
+        assert model_id == "_FakeTorchModel"
+        assert info.model_format == "pytorch"
+
+    def test_extract_info_override_model_id(self):
+        model = _FakeTorchModel()
+        model_id, _ = self.extractor.extract_info(model, {"id": "my-resnet"})
+        assert model_id == "my-resnet"
+
+    def test_install_hooks_publishes_inference(self, publish_spy):
+        model = _FakeTorchModel()
+        handle = make_handle(publish_spy)
+        self.extractor.install_hooks(model, handle)
+        model._pre_hook(model, (None,))
+        model._post_hook(model, (None,), None)
+        assert len(publish_spy.events) == 1
+        assert publish_spy.events[0]["event_type"] == "inference"
+
+    def test_install_hooks_sets_detected_accelerator(self, publish_spy):
+        model = _FakeTorchModel()
+        handle = make_handle(publish_spy)
+        self.extractor.install_hooks(model, handle)
+        assert handle.detected_accelerator == "cpu"