Update for pyo3 0.21. (#1985)

* Update for pyo3 0.21.

* Also adapt the RL example.

* Fix for the pyo3-onnx bindings...

* Print details on failures.

* Revert pyi.

candle-examples/Cargo.toml

@@ -25,7 +25,7 @@ hf-hub = { workspace = true, features = ["tokio"] }
 image = { workspace = true }
 intel-mkl-src = { workspace = true, optional = true }
 num-traits = { workspace = true }
-pyo3 = { version = "0.20.0", features = ["auto-initialize"], optional = true }
+pyo3 = { version = "0.21.0", features = ["auto-initialize"], optional = true }
 rayon = { workspace = true }
 rubato = { version = "0.15.0", optional = true }
 safetensors = { workspace = true }

candle-examples/examples/reinforcement-learning/gym_env.rs

@@ -42,7 +42,7 @@ impl GymEnv {
     /// Creates a new session of the specified OpenAI Gym environment.
     pub fn new(name: &str) -> Result<GymEnv> {
         Python::with_gil(|py| {
-            let gym = py.import("gymnasium")?;
+            let gym = py.import_bound("gymnasium")?;
             let make = gym.getattr("make")?;
             let env = make.call1((name,))?;
             let action_space = env.getattr("action_space")?;
@@ -66,10 +66,10 @@ impl GymEnv {
     /// Resets the environment, returning the observation tensor.
     pub fn reset(&self, seed: u64) -> Result<Tensor> {
         let state: Vec<f32> = Python::with_gil(|py| {
-            let kwargs = PyDict::new(py);
+            let kwargs = PyDict::new_bound(py);
             kwargs.set_item("seed", seed)?;
-            let state = self.env.call_method(py, "reset", (), Some(kwargs))?;
-            state.as_ref(py).get_item(0)?.extract()
+            let state = self.env.call_method_bound(py, "reset", (), Some(&kwargs))?;
+            state.bind(py).get_item(0)?.extract()
         })
         .map_err(w)?;
         Tensor::new(state, &Device::Cpu)
@@ -81,8 +81,10 @@ impl GymEnv {
         action: A,
     ) -> Result<Step<A>> {
         let (state, reward, terminated, truncated) = Python::with_gil(|py| {
-            let step = self.env.call_method(py, "step", (action.clone(),), None)?;
-            let step = step.as_ref(py);
+            let step = self
+                .env
+                .call_method_bound(py, "step", (action.clone(),), None)?;
+            let step = step.bind(py);
             let state: Vec<f32> = step.get_item(0)?.extract()?;
             let reward: f64 = step.get_item(1)?.extract()?;
             let terminated: bool = step.get_item(2)?.extract()?;

candle-examples/examples/reinforcement-learning/vec_gym_env.rs

@@ -24,13 +24,13 @@ fn w(res: PyErr) -> candle::Error {
 impl VecGymEnv {
     pub fn new(name: &str, img_dir: Option<&str>, nprocesses: usize) -> Result<VecGymEnv> {
         Python::with_gil(|py| {
-            let sys = py.import("sys")?;
+            let sys = py.import_bound("sys")?;
             let path = sys.getattr("path")?;
             let _ = path.call_method1(
                 "append",
                 ("candle-examples/examples/reinforcement-learning",),
             )?;
-            let gym = py.import("atari_wrappers")?;
+            let gym = py.import_bound("atari_wrappers")?;
             let make = gym.getattr("make")?;
             let env = make.call1((name, img_dir, nprocesses))?;
             let action_space = env.getattr("action_space")?;
@@ -60,10 +60,10 @@ impl VecGymEnv {
 
     pub fn step(&self, action: Vec<usize>) -> Result<Step> {
         let (obs, reward, is_done) = Python::with_gil(|py| {
-            let step = self.env.call_method(py, "step", (action,), None)?;
-            let step = step.as_ref(py);
+            let step = self.env.call_method_bound(py, "step", (action,), None)?;
+            let step = step.bind(py);
             let obs = step.get_item(0)?.call_method("flatten", (), None)?;
-            let obs_buffer = pyo3::buffer::PyBuffer::get(obs)?;
+            let obs_buffer = pyo3::buffer::PyBuffer::get_bound(&obs)?;
             let obs: Vec<u8> = obs_buffer.to_vec(py)?;
             let reward: Vec<f32> = step.get_item(1)?.extract()?;
             let is_done: Vec<f32> = step.get_item(2)?.extract()?;

candle-pyo3/Cargo.toml

@@ -20,10 +20,10 @@ candle-nn = { workspace = true }
 candle-onnx = { workspace = true, optional = true }
 half = { workspace = true }
 intel-mkl-src = { workspace = true, optional = true }
-pyo3 = { version = "0.20.0", features = ["extension-module", "abi3-py38"] }
+pyo3 = { version = "0.21.0", features = ["extension-module", "abi3-py38"] }
 
 [build-dependencies]
-pyo3-build-config = "0.20"
+pyo3-build-config = "0.21"
 
 [features]
 default = []

candle-pyo3/py_src/candle/nn/__init__.pyi

@@ -0,0 +1,19 @@
+# Generated content DO NOT EDIT
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union, Sequence
+from os import PathLike
+from candle.typing import _ArrayLike, Device, Scalar, Index, Shape
+from candle import Tensor, DType, QTensor
+
+@staticmethod
+def silu(tensor: Tensor) -> Tensor:
+    """
+    Applies the Sigmoid Linear Unit (SiLU) function to a given tensor.
+    """
+    pass
+
+@staticmethod
+def softmax(tensor: Tensor, dim: int) -> Tensor:
+    """
+    Applies the Softmax function to a given tensor.#
+    """
+    pass