[Examples] add Python runner for BuddyDeepSeekR1 E2E inference

[GuoningHuang]

Summary

Add a Python AOT end-to-end runner for examples/BuddyDeepSeekR1 to execute exported prefill/decode subgraphs.

Why

• Simpler than C++ for runtime orchestration (prefill/decode loop, stop conditions, sampling).
• Faster iteration: change runtime behavior without rebuilding C++.
• More flexible runtime features (chat template, sampling params, eos/length controls).
• Updated examples/BuddyDeepSeekR1/README.md with usage instructions.

performance

The performance is comparable to the C++ version:

[zhanghb97]

(torch2.8) zhb@s306-gpu-p1-102:~/buddy-mlir$ python3 examples/BuddyDeepSeekR1/run-subgraphs-python.py \
  --prompt "Hello, who are you?" \
  --export-subgraphs
[Python] runtime: aot
Traceback (most recent call last):
  File "/home/zhb/buddy-mlir/examples/BuddyDeepSeekR1/run-subgraphs-python.py", line 391, in <module>
    main()
  File "/home/zhb/buddy-mlir/examples/BuddyDeepSeekR1/run-subgraphs-python.py", line 377, in main
    _run_aot(
  File "/home/zhb/buddy-mlir/examples/BuddyDeepSeekR1/run-subgraphs-python.py", line 163, in _run_aot
    runtime_lib = ctypes.CDLL(str(runtime_so), mode=rtld_global)
  File "/home/zhb/miniconda3/envs/torch2.8/lib/python3.10/ctypes/__init__.py", line 374, in __init__
    self._handle = _dlopen(self._name, mode)
OSError: /home/zhb/buddy-mlir/build/examples/BuddyDeepSeekR1/libdeepseek_forward_runtime.so: undefined symbol: _mlir_ciface_rtclock

Does the process use rtclock? It looks like after generating the C interface, it wasn't linked.

[GuoningHuang]

(torch2.8) zhb@s306-gpu-p1-102:~/buddy-mlir$ python3 examples/BuddyDeepSeekR1/run-subgraphs-python.py \
  --prompt "Hello, who are you?" \
  --export-subgraphs
[Python] runtime: aot
Traceback (most recent call last):
  File "/home/zhb/buddy-mlir/examples/BuddyDeepSeekR1/run-subgraphs-python.py", line 391, in <module>
    main()
  File "/home/zhb/buddy-mlir/examples/BuddyDeepSeekR1/run-subgraphs-python.py", line 377, in main
    _run_aot(
  File "/home/zhb/buddy-mlir/examples/BuddyDeepSeekR1/run-subgraphs-python.py", line 163, in _run_aot
    runtime_lib = ctypes.CDLL(str(runtime_so), mode=rtld_global)
  File "/home/zhb/miniconda3/envs/torch2.8/lib/python3.10/ctypes/__init__.py", line 374, in __init__
    self._handle = _dlopen(self._name, mode)
OSError: /home/zhb/buddy-mlir/build/examples/BuddyDeepSeekR1/libdeepseek_forward_runtime.so: undefined symbol: _mlir_ciface_rtclock

Does the process use rtclock? It looks like after generating the C interface, it wasn't linked.

I have updated the corresponding README. You can try running:

python3 run-subgraphs-python.py --prompt "hello" --artifact-dir ../../build/examples/BuddyDeepSeekR1 --llvm-build-dir ../../llvm/build --omp-num-threads 48 --omp-proc-bind close

in the buddy-mlir/examples/BuddyDeepSeekR directory.