Hi Antonio, Thank you for the response !

Hi @AntonioDeDomenico, thanks for the earlier clarifications. A few follow-ups specifically on Phase 3 execution, to help us package submissions correctly:

1. Is the Phase 3 GPU NVIDIA (CUDA) or Huawei Ascend (CANN)? This directly changes what we can ship — CUDA-compiled artifacts (AWQ kernels, Flash-Attention, custom Triton) won't run on Ascend, and vice versa. Even just confirming the ecosystem (without disclosing the exact chip) would be enormously helpful.
2. Will the you serve the base model yourself (e.g., via vLLM, SGLang, MindIE) and we just submit LoRA weights + main.py, or do we submit a full runnable environment (Dockerfile, requirements.txt, our own serving code)?
3. If we submit code + weights only: what Python, PyTorch, and CUDA/CANN versions will be installed in the execution environment? This lets us pin requirements.txt to match.
4. Is the Agent Tool Server in Phase 3 the same HF-Spaces-CPU deployment as Phase 2, or hosted on dedicated infrastructure during grading? This affects our tool-call latency budget planning.

I understand you may not want to disclose exact GPU model or VRAM. Even partial answers on the ecosystem (CUDA vs CANN) and the submission contract (weights-only vs full-container) are what we need most. Thanks!

1. CUDA 2. Please submit the full runnable environment with one-click button flow and make sure the environment can run successfully. The GPU deployment code and the Agent execution code need to be separated, because one is deployed on the GPU server and the other on the CPU server. 3. Please refer to question 2 4. Please do not worry about tool-call latency. The latency will be short enough not to affect the overall runtime.