Merge remote-tracking branch 'upstream/master' into maintenance/merge…

…-master-20240212

.github/workflows/tests.yml

@@ -87,7 +87,9 @@ jobs:
 
       - name: Run mypy
         run: python -m mypy thinc --no-implicit-reexport
-        if: matrix.python_version != '3.6'
+        if: |
+          matrix.python_version != '3.6' &&
+          matrix.python_version != '3.7'
 
       - name: Delete source directory
         run: rm -rf thinc

requirements.txt

@@ -25,7 +25,7 @@ pytest-cov>=2.7.0,<5.0.0
 coverage>=5.0.0,<8.0.0
 mock>=2.0.0,<3.0.0
 flake8>=3.5.0,<3.6.0
-mypy>=1.0.0,<1.1.0; python_version >= "3.7"
+mypy>=1.5.0,<1.6.0; platform_machine != "aarch64" and python_version >= "3.8"
 types-mock>=0.1.1
 types-contextvars>=0.1.2; python_version < "3.7"
 types-dataclasses>=0.1.3; python_version < "3.7"

thinc/backends/cupy_ops.py

@@ -1,7 +1,7 @@
 import numpy
 
 from .. import registry
-from ..compat import cupy, cupyx
+from ..compat import cublas, cupy, cupyx
 from ..types import DeviceTypes
 from ..util import (
     is_cupy_array,
@@ -257,7 +257,7 @@ def clip_gradient(self, gradient, threshold):
         # implementation.
         def frobenius_norm(X):
             X_vec = X.reshape(-1)
-            return cupy.cublas.nrm2(X_vec)
+            return cublas.nrm2(X_vec)
 
         grad_norm = cupy.maximum(frobenius_norm(gradient), 1e-12)
         gradient *= cupy.minimum(threshold, grad_norm) / grad_norm

thinc/compat.py

@@ -4,9 +4,11 @@
 
 try:  # pragma: no cover
     import cupy
+    import cupy.cublas
     import cupyx
 
     has_cupy = True
+    cublas = cupy.cublas
     cupy_version = Version(cupy.__version__)
     try:
         cupy.cuda.runtime.getDeviceCount()
@@ -20,6 +22,7 @@
     else:
         cupy_from_dlpack = cupy.fromDlpack
 except (ImportError, AttributeError):
+    cublas = None
     cupy = None
     cupyx = None
     cupy_version = Version("0.0.0")

thinc/shims/torchscript.py

@@ -30,7 +30,7 @@ class TorchScriptShim(PyTorchShim):
 
     def __init__(
         self,
-        model: Optional["torch.ScriptModule"],
+        model: Optional["torch.jit.ScriptModule"],
         config=None,
         optimizer: Any = None,
         mixed_precision: bool = False,

thinc/tests/layers/test_basic_tagger.py

@@ -60,7 +60,6 @@ def get_shuffled_batches(Xs, Ys, batch_size):
         yield list(batch_X), list(batch_Y)
 
 
-@pytest.mark.slow
 @pytest.mark.parametrize(
     ("depth", "width", "vector_width", "nb_epoch"), [(2, 32, 16, 5)]
 )