[Bug] SAM3 fails on NVIDIA GB10 (sm_121) with "no kernel image is available for execution on the device"

**Description**
On a desktop system with an NVIDIA GB10 GPU (compute capability 12.1, sm_121) and Ubuntu 24.04.3 LTS on arm64, running SAM3 causes a CUDA runtime crash during model initialization. The logs say that the current PyTorch build only supports CUDA capabilities sm_80–sm_90, and then SAM3 fails with CUDA error: no kernel image is available for execution on the device when creating tensors on cuda.

**Environment**
SAM3 version / commit:

**OS**: Ubuntu 24.04.3 LTS (arm64)

**Kernel**: Linux 6.14.0-1013-nvidia

**GPU:** NVIDIA GB10 (compute capability 12.1 / sm_121)

**Hardware**: Dell Pro Max with GB10

**Python**: 3.10 (Conda env sam3_clean)

**SAM3 install method**: `pip / git clone + pip install -e .`

Steps to Reproduce
Create and activate a clean environment:

bash
`conda create -n sam3_clean python=3.10`
`conda activate sam3_clean`
# Install PyTorch and SAM3 dependencies as in the SAM3 documentation
Run the SAM3‑based script:

bash
python backend/src/run_upload.py( this is a script to upload vectors to weavite db )
The script starts, logs that the feature extractor and SAM3 are initialized on CUDA, and then crashes during model construction.

**Expected Behavior**
SAM3 should either:

Run successfully on NVIDIA GB10 when a compatible GPU build of PyTorch is installed, or

Detect that the GPU architecture (sm_121) is unsupported by the current PyTorch build and fail fast with a clear error or allow explicit CPU fallback.

**Actual Behavior**
At startup, warnings indicate that:

The GPU is NVIDIA GB10 with CUDA capability 12.1.

The installed PyTorch build supports CUDA capabilities only from 8.0 to 9.0.

The GB10 GPU is therefore not compatible with this PyTorch installation.

Later, during SAM3 model construction, the program crashes with:

torch.AcceleratorError: CUDA error: no kernel image is available for execution on the device

The error is raised in PositionEmbeddingSine when calling something like:

**python**
torch.zeros((1, 1) + size, device=dev)
where dev is set to "cuda".

**Additional Context**
Other components (e.g., ResNet feature extractor) log successful initialization on CUDA before the crash, so the GPU is visible but cannot execute kernels with this PyTorch build.

It is unclear whether SAM3 currently supports NVIDIA GB10 / sm_121 on arm64, and if a specific PyTorch + CUDA combination is recommended for this hardware.

Any guidance on:

Supported architectures for SAM3, and

Recommended versions / builds for running on GB10 (arm64),
would be very helpful.

Reproduce the failure with backend/src/run_upload.py in the stated Ubuntu arm64 and NVIDIA GB10 environment, then inspect the SAM3 model initialization around PositionEmbeddingSine and the device selection. Compare the installed PyTorch build's supported architectures with sm_121. Done means establishing a supported configuration or providing a clear unsupported-GPU failure or explicit CPU fallback.

Written by the indexing model from the issue text.