feat(bpu): D-Robotics RDK BPU backend with on-board compilation

[lvyufeng]

Summary

Merges D-Robotics RDK BPU support into torch-fl as ACCELERATOR=bpu — one torch-fl package, no separate torch_bpu repo, no VM dependency.

Key changes

• Build system: CMake + setup.py recognize ACCELERATOR=bpu, skip CUDA/kernel requirements
• Runtime: UCP allocator (hbUCPMallocCached) for real device memory, device/stream/event stubs
• Eager mode: All ops fall back to CPU via cpu_fallback; convolution_overrideable has explicit wrappers (the boxed fallback cannot serve it — it redispatches the same op and lands back on the raising stub)
• Compile backend: torch.compile(backend="bpu") auto-registered, partitions aten graph → ONNX → hbdk4 → .hbm → hbm_runtime
• Quantization: Q/DQ insertion is mandatory (hbdk4 lowers float conv to CPU); default scale 0.05, overridable via calibration
• Weight freezing: Parameters baked into artifact to avoid crossing the boundary on every call
• Zero-copy: flagos tensors wrap UCP storage in numpy arrays in place (D→H copy eliminated; quantization and output alloc still copy)
• On-board hbdk4, on the stock kernel: the x86_64 compiler runs under box64. No VM, no cross-compile host, and no kernel rebuild — the earlier conclusion that a 4 KB-page kernel was required turned out to be wrong. The real blocker was box64's version: the packaged 0.2.6 hard-codes a 4 KB page size, while current box64 (0.4+) reads it at runtime and maps 4 KB-aligned x86 segments onto 64 KB pages itself. scripts/setup_bpu_hbdk4.sh automates the setup.

Testing

• 65 BPU unit tests pass (partition, qdq, decompose, freeze, splice, eager device, zero-copy, x86 env)
• Eager: elementwise, matmul, reduction, conv2d forward+backward all match CPU
• Compile: correct graph partitioning, graceful CPU fallback when hbdk4 unavailable
• Clean process exit (caching allocator deliberately skips block release at exit to avoid a double-free race with libhbucp teardown; _empty_cache() before exit frees the same blocks cleanly, which confirms only the ordering is at fault)
• ruff check / ruff format --check clean
• 4 tests in test_profiler_privateuse1.py fail on this board — they need libcupti and real GPU kernel events. That file is untouched by this branch; the failures come from feat(profiler): torch-cuda parity for PrivateUse1 — flows, device time, kernel metadata #55 and are not reproducible without a GPU.

torch-2.13 compat fix (affects all platforms)

PyTorch 2.13 removed named tensors (at::DimnameList no longer exists), but checked-in generated files still declared 16 named-tensor factory variants. This blocked every platform on 2.13. Fixed by adding them to codegen_skip_ops.txt and surgically deleting from the four generated files (pure deletions, 0 additions, verified with git diff --numstat). Did not re-run full codegen: CI pins torch 2.10, and a full regen would bake in 2.13-only signature drift + silently empty flaggems_python_kernels.cc when flag_gems is not installed.

⚠️ This board has no CUDA, so I could not verify CUDA compilation after the generated-file edits. Recommend running CI before merge.

Docs

• docs/bpu.md: architecture rationale, on-board hbdk4 setup, quantization, calibration, env vars, known limits
• README.md: BPU build instructions, new env vars

Benchmark

6-layer conv stack @ 224×224, measured on-board with artifacts compiled on-board: 3.75 ms (BPU) vs 72.06 ms (eager CPU) = 19.2×, int8 quantized with ~3% relative error. Toy nets are a wash — submission overhead dominates.

Dependencies

None beyond the board image. libhbucp/libbpu ship at /usr/hobot/lib with headers at /usr/include/hobot, so there is no SDK root to configure. hbdk4 is optional: without it the backend warns and runs every partition on the CPU, so the install stays usable.