Add pyrefly suppressions (3/n) (#164588)

Adds suppressions to pyrefly will typecheck clean: https://github.com/pytorch/pytorch/issues/163283

Test plan:
dmypy restart && python3 scripts/lintrunner.py -a
pyrefly check

step 1: uncomment lines in the pyrefly.toml file
step 2: run pyrefly check
step 3: add suppressions, clean up unused suppressions
before: https://gist.github.com/maggiemoss/bb31574ac8a59893c9cf52189e67bb2d

after:

 0 errors (1,970 ignored)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164588
Approved by: https://github.com/oulgen

torch/cuda/__init__.py

@@ -496,12 +496,14 @@ class cudaStatus:
 
 class CudaError(RuntimeError):
     def __init__(self, code: int) -> None:
+        # pyrefly: ignore  # missing-attribute
         msg = _cudart.cudaGetErrorString(_cudart.cudaError(code))
         super().__init__(f"{msg} ({code})")
 
 
 def check_error(res: int) -> None:
     r"""Raise an error if the result of a CUDA runtime API call is not success."""
+    # pyrefly: ignore  # missing-attribute
     if res != _cudart.cudaError.success:
         raise CudaError(res)
 
@@ -601,6 +603,7 @@ def get_device_capability(device: "Device" = None) -> tuple[int, int]:
     return prop.major, prop.minor
 
 
+# pyrefly: ignore  # not-a-type
 def get_device_properties(device: "Device" = None) -> _CudaDeviceProperties:
     r"""Get the properties of a device.
 
@@ -651,6 +654,7 @@ class StreamContext:
         self.idx = _get_device_index(None, True)
         if not torch.jit.is_scripting():
             if self.idx is None:
+                # pyrefly: ignore  # bad-assignment
                 self.idx = -1
 
         self.src_prev_stream = (
@@ -953,7 +957,9 @@ def _device_count_amdsmi() -> int:
             if raw_cnt <= 0:
                 return raw_cnt
             # Trim the list up to a maximum available device
+            # pyrefly: ignore  # bad-argument-type
             for idx, val in enumerate(visible_devices):
+                # pyrefly: ignore  # redundant-cast
                 if cast(int, val) >= raw_cnt:
                     return idx
     except OSError:
@@ -987,7 +993,9 @@ def _device_count_nvml() -> int:
             if raw_cnt <= 0:
                 return raw_cnt
             # Trim the list up to a maximum available device
+            # pyrefly: ignore  # bad-argument-type
             for idx, val in enumerate(visible_devices):
+                # pyrefly: ignore  # redundant-cast
                 if cast(int, val) >= raw_cnt:
                     return idx
     except OSError:
@@ -1203,7 +1211,9 @@ def _get_pynvml_handler(device: "Device" = None):
     if not _HAS_PYNVML:
         raise ModuleNotFoundError(
             "pynvml does not seem to be installed or it can't be imported."
+            # pyrefly: ignore  # invalid-inheritance
         ) from _PYNVML_ERR
+    # pyrefly: ignore  # import-error
     from pynvml import NVMLError_DriverNotLoaded
 
     try:
@@ -1220,6 +1230,7 @@ def _get_amdsmi_handler(device: "Device" = None):
     if not _HAS_PYNVML:
         raise ModuleNotFoundError(
             "amdsmi does not seem to be installed or it can't be imported."
+            # pyrefly: ignore  # invalid-inheritance
         ) from _PYNVML_ERR
     try:
         amdsmi.amdsmi_init()
@@ -1483,6 +1494,7 @@ def _get_rng_state_offset(device: Union[int, str, torch.device] = "cuda") -> int
     return default_generator.get_offset()
 
 
+# pyrefly: ignore  # deprecated
 from .memory import *  # noqa: F403
 from .random import *  # noqa: F403
 
@@ -1699,6 +1711,7 @@ def _register_triton_kernels():
     def kernel_impl(*args, **kwargs):
         from torch.sparse._triton_ops import bsr_dense_mm
 
+        # pyrefly: ignore  # not-callable
         return bsr_dense_mm(*args, skip_checks=True, **kwargs)
 
     @_WrappedTritonKernel

torch/cuda/_utils.py

@@ -279,6 +279,7 @@ class _CudaModule:
             return self._kernels[name]
 
         # Import the CUDA library inside the method
+        # pyrefly: ignore  # missing-module-attribute
         from torch.cuda._utils import _get_gpu_runtime_library
 
         libcuda = _get_gpu_runtime_library()

torch/cuda/amp/__init__.py

@@ -1,3 +1,4 @@
+# pyrefly: ignore  # deprecated
 from .autocast_mode import autocast, custom_bwd, custom_fwd
 from .common import amp_definitely_not_available
 from .grad_scaler import GradScaler

torch/cuda/graphs.py

@@ -259,6 +259,7 @@ class graph:
         self.cuda_graph.capture_begin(
             # type: ignore[misc]
             *self.pool,
+            # pyrefly: ignore  # bad-keyword-argument
             capture_error_mode=self.capture_error_mode,
         )
 
@@ -524,6 +525,7 @@ def make_graphed_callables(
     ) -> Callable[..., object]:
         class Graphed(torch.autograd.Function):
             @staticmethod
+            # pyrefly: ignore  # bad-override
             def forward(ctx: object, *inputs: Tensor) -> tuple[Tensor, ...]:
                 # At this stage, only the user args may (potentially) be new tensors.
                 for i in range(len_user_args):
@@ -535,6 +537,7 @@ def make_graphed_callables(
 
             @staticmethod
             @torch.autograd.function.once_differentiable
+            # pyrefly: ignore  # bad-override
             def backward(ctx: object, *grads: Tensor) -> tuple[Tensor, ...]:
                 assert len(grads) == len(static_grad_outputs)
                 for g, grad in zip(static_grad_outputs, grads):
@@ -548,7 +551,9 @@ def make_graphed_callables(
                 # Input args that didn't require grad expect a None gradient.
                 assert isinstance(static_grad_inputs, tuple)
                 return tuple(
-                    b.detach() if b is not None else b for b in static_grad_inputs
+                    # pyrefly: ignore  # bad-argument-type
+                    b.detach() if b is not None else b
+                    for b in static_grad_inputs
                 )
 
         def functionalized(*user_args: object) -> object:

torch/cuda/memory.py

@@ -770,6 +770,7 @@ def list_gpu_processes(device: "Device" = None) -> str:
             import pynvml  # type: ignore[import]
         except ModuleNotFoundError:
             return "pynvml module not found, please install pynvml"
+        # pyrefly: ignore  # import-error
         from pynvml import NVMLError_DriverNotLoaded
 
         try:
@@ -852,6 +853,7 @@ def _record_memory_history_legacy(
     _C._cuda_record_memory_history_legacy(  # type: ignore[call-arg]
         enabled,
         record_context,
+        # pyrefly: ignore  # bad-argument-type
         trace_alloc_max_entries,
         trace_alloc_record_context,
         record_context_cpp,

torch/cuda/nvtx.py

@@ -53,6 +53,7 @@ def range_start(msg) -> int:
     Args:
         msg (str): ASCII message to associate with the range.
     """
+    # pyrefly: ignore  # missing-attribute
     return _nvtx.rangeStartA(msg)
 
 
@@ -63,6 +64,7 @@ def range_end(range_id) -> None:
     Args:
         range_id (int): an unique handle for the start range.
     """
+    # pyrefly: ignore  # missing-attribute
     _nvtx.rangeEnd(range_id)
 
 
@@ -83,6 +85,7 @@ def _device_range_start(msg: str, stream: int = 0) -> object:
         msg (str): ASCII message to associate with the range.
         stream (int): CUDA stream id.
     """
+    # pyrefly: ignore  # missing-attribute
     return _nvtx.deviceRangeStart(msg, stream)
 
 
@@ -95,6 +98,7 @@ def _device_range_end(range_handle: object, stream: int = 0) -> None:
         range_handle: an unique handle for the start range.
         stream (int): CUDA stream id.
     """
+    # pyrefly: ignore  # missing-attribute
     _nvtx.deviceRangeEnd(range_handle, stream)