SAM3.1. size assert error during compile

Hi there,
trying to compile SAM3.1 as I used to with SAM3
```python
predictor = build_sam3_multiplex_video_predictor(
use_fa3=False,
compile=True,
)
```

However, upon calling
```python
for response in predictor.handle_stream_request(
request=dict(
type="propagate_in_video",
session_id=session_id,
start_frame_index=0,
max_frame_num_to_track=num_frames
)
):
pass
```

Some compilation steps are successful before I receive the following error:
```bash
File "/PROJ_HOME/test_sam3.py", line 543, in run_sam31
for response in predictor.handle_stream_request(
File "/ENV_HOME/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 38, in generator_context
response = gen.send(None)
File "/PROJ_HOME/subtrees/SAM3/sam3/model/sam3_base_predictor.py", line 90, in handle_stream_request
yield from self.propagate_in_video(
File "/PROJ_HOME/subtrees/SAM3/sam3/model/sam3_base_predictor.py", line 279, in propagate_in_video
for frame_idx, outputs in self.model.propagate_in_video(
File "/ENV_HOME/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 38, in generator_context
response = gen.send(None)
File "/PROJ_HOME/subtrees/SAM3/sam3/model/sam3_multiplex_tracking.py", line 2332, in propagate_in_video
yield from super().propagate_in_video(
File "/ENV_HOME/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 38, in generator_context
response = gen.send(None)
File "/PROJ_HOME/subtrees/SAM3/sam3/model/sam3_multiplex_tracking.py", line 366, in propagate_in_video
out = self._run_single_frame_inference(
File "/PROJ_HOME/subtrees/SAM3/sam3/model/sam3_multiplex_tracking.py", line 635, in _run_single_frame_inference
) = self._det_track_one_frame(
File "/PROJ_HOME/subtrees/SAM3/sam3/model/sam3_multiplex_base.py", line 447, in _det_track_one_frame
return self._det_track_one_frame_impl(
File "/PROJ_HOME/subtrees/SAM3/sam3/model/sam3_multiplex_base.py", line 527, in _det_track_one_frame_impl
self.run_tracker_propagation(
File "/PROJ_HOME/subtrees/SAM3/sam3/model/sam3_multiplex_base.py", line 810, in run_tracker_propagation
self._propogate_tracker_one_frame_local_gpu(
File "/PROJ_HOME/subtrees/SAM3/sam3/model/sam3_multiplex_base.py", line 1708, in _propogate_tracker_one_frame_local_gpu
for out in self.tracker.propagate_in_video(
File "/ENV_HOME/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 38, in generator_context
response = gen.send(None)
File "/PROJ_HOME/subtrees/SAM3/sam3/model/video_tracking_multiplex_demo.py", line 3447, in propagate_in_video
current_out, pred_masks = self._run_single_frame_inference(
File "/PROJ_HOME/subtrees/SAM3/sam3/model/video_tracking_multiplex_demo.py", line 2793, in _run_single_frame_inference
current_out = self.track_step(
File "/PROJ_HOME/subtrees/SAM3/sam3/model/video_tracking_multiplex.py", line 3382, in track_step
current_out, aux_out = self._track_step_aux(
File "/PROJ_HOME/subtrees/SAM3/sam3/model/video_tracking_multiplex.py", line 2057, in _track_step_aux
pix_feat_with_mem = self._prepare_memory_conditioned_features(
File "/PROJ_HOME/subtrees/SAM3/sam3/model/video_tracking_multiplex.py", line 1590, in _prepare_memory_conditioned_features
encoder_out = self.transformer.encoder(
File "/ENV_HOME/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "/ENV_HOME/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1786, in _call_impl
return forward_call(*args, **kwargs)
File "/PROJ_HOME/subtrees/SAM3/sam3/perflib/compile.py", line 113, in wrapper
return fn(*args, **kwargs)
File "/PROJ_HOME/subtrees/SAM3/sam3/perflib/compile.py", line 75, in compiled_fn_wrapper
result = compiled_fn(*args, **kwargs)
File "/ENV_HOME/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py", line 832, in compile_wrapper
return fn(*args, **kwargs)
File "/PROJ_HOME/subtrees/SAM3/sam3/model/decoder.py", line 1282, in forward
def forward(
File "/ENV_HOME/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py", line 1044, in _fn
return fn(*args, **kwargs)
File "/ENV_HOME/lib/python3.10/site-packages/torch/_functorch/aot_autograd.py", line 1130, in forward
return compiled_fn(full_args)
File "/ENV_HOME/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py", line 353, in runtime_wrapper
all_outs = call_func_at_runtime_with_args(
File "/ENV_HOME/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/utils.py", line 129, in call_func_at_runtime_with_args
out = normalize_as_list(f(args))
File "/ENV_HOME/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py", line 724, in inner_fn
outs = compiled_fn(args)
File "/ENV_HOME/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py", line 526, in wrapper
return compiled_fn(runtime_args)
File "/ENV_HOME/lib/python3.10/site-packages/torch/_inductor/output_code.py", line 613, in __call__
return self.current_callable(inputs)
File "/ENV_HOME/lib/python3.10/site-packages/torch/_inductor/utils.py", line 2962, in run
out = model(new_inputs)
File "/TORCHINDUCTOR_CACHE_DIR/wl/cwl2lkbm2n6yfn7ps7vfaw4wqz4n6m56rxcq7ga6qfv3d2dmalts.py", line 2884, in call
assert_size_stride(arg6_1, (s28, 1, 256), (256, 256, 1))
AssertionError: expected size 10368==5184, stride 256==256 at dim=0
This error most often comes from a incorrect fake (aka meta) kernel for a custom op.
Use torch.library.opcheck to test your custom op.
See https://pytorch.org/docs/stable/library.html#torch.library.opcheck
```

Any suggestions?

Reproduce the failure with build_sam3_multiplex_video_predictor(compile=True) and the propagate_in_video request. Start with sam3/perflib/compile.py and the decoder call in sam3/model/decoder.py, then trace the tensor through video_tracking_multiplex.py; the relevant assertion reports a size mismatch during the compiled encoder path. Done means SAM3.1 video propagation completes with compilation enabled without the assertion.

Written by the indexing model from the issue text.