Fix device_id access and retain CudaContext on frame

- Add _device_id field to VideoFrame, set it in from_dlpack and HW decode path
- Store CudaContext on frame (_cuda_ctx) to prevent premature GC
- Fix plane.py to use frame._device_id

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: WyattBlue

av/codec/hwaccel.py

@@ -130,7 +130,7 @@ def __init__(
         self.allow_software_fallback = allow_software_fallback
 
         self.options = {} if not options else dict(options)
-        if self._device_type == HWDeviceType.cuda and self.is_hw_owned == True:
+        if self._device_type == HWDeviceType.cuda and self.is_hw_owned:
             self.options.setdefault("primary_ctx", "1")
         self.flags = 0 if not flags else flags
         self.ptr = cython.NULL
@@ -174,7 +174,6 @@ def create(self, codec: Codec) -> HWAccel:
             device=self._device,
             allow_software_fallback=self.allow_software_fallback,
             options=self.options,
-            output_format=self.output_format,
         )
         ret._initialize_hw_context(codec)
         return ret

av/video/codeccontext.py

@@ -128,6 +128,7 @@ def _transfer_hwframe(self, frame: Frame):
             return frame
 
         if self.hwaccel_ctx.is_hw_owned:
+            cython.cast(VideoFrame, frame)._device_id = self.hwaccel_ctx.device_id
             return frame
 
         frame_sw: Frame = self._alloc_next_frame()

av/video/frame.pxd

@@ -24,9 +24,11 @@ cdef class VideoFrame(Frame):
     # We don't ever actually access it directly.
     cdef uint8_t *_buffer
     cdef object _np_buffer
+    cdef object _cuda_ctx
 
     cdef VideoReformatter reformatter
     cdef readonly VideoFormat format
+    cdef readonly int _device_id
 
     cdef _init(self, lib.AVPixelFormat format, unsigned int width, unsigned int height)
     cdef _init_user_attributes(self)

av/video/frame.py

@@ -628,10 +628,7 @@ def to_ndarray(self, channel_last=False, **kwargs):
 
         frame: VideoFrame = self.reformat(**kwargs2)
         if frame.ptr.hw_frames_ctx:
-            raise ValueError(
-                "Cannot convert a hardware frame to numpy directly. "
-                "Specify a software format (e.g. format='rgb24') or decode with HWAccel(output_format='sw')."
-            )
+            raise ValueError("Cannot convert a hardware frame to numpy directly.")
 
         import numpy as np
 
@@ -1542,6 +1539,8 @@ def from_dlpack(
                 frames_ref = ctx.get_frames_ctx(sw_fmt, width, height)
                 frame.ptr.format = get_pix_fmt(b"cuda")
                 frame.ptr.hw_frames_ctx = frames_ref
+                frame._device_id = device_id
+                frame._cuda_ctx = ctx
             else:
                 frame.ptr.format = sw_fmt
 

av/video/plane.py

@@ -83,7 +83,7 @@ def __dlpack_device__(self):
                 raise NotImplementedError(
                     "DLPack export is only implemented for CUDA hw frames"
                 )
-            return (kCuda, self.frame.device_id)
+            return (kCuda, self.frame._device_id)
         return (kCPU, 0)
 
     def __dlpack__(self, stream: int | None = None):
@@ -109,7 +109,7 @@ def __dlpack__(self, stream: int | None = None):
             )
             sw_fmt = frames_ctx.sw_format
             device_type = kCuda
-            device_id = self.frame.hwaccel_ctx.device_id
+            device_id = self.frame._device_id
         else:
             sw_fmt = cython.cast(lib.AVPixelFormat, self.frame.ptr.format)
             device_type = kCPU