Parallelize detection and map head execution using CUDA streams

[Copilot]

The detection and map heads in SparseDriveHead.forward() were executing sequentially. This change runs them in parallel using CUDA streams when both are enabled.

Changes

• Parallel execution: det_head and map_head now run on separate CUDA streams concurrently
• Device-aware: Checks feature_maps.is_cuda to only use streams when tensors are on GPU
• Fallback: Sequential execution preserved for CPU or when only one head is enabled

if with_det and with_map and is_cuda:
    det_stream = torch.cuda.Stream()
    map_stream = torch.cuda.Stream()
    
    with torch.cuda.stream(det_stream):
        det_output = self.det_head(feature_maps, metas)
    
    with torch.cuda.stream(map_stream):
        map_output = self.map_head(feature_maps, metas)
    
    torch.cuda.current_stream().wait_stream(det_stream)
    torch.cuda.current_stream().wait_stream(map_stream)

Original prompt

I want to execute the map and detection head in parallel. Currently it seems that they run sequential. Dont use multiprocessing for that but rather something like torch cuda stream

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.