"""Bounded Modal L4 CUDA preflight. Not a behavioral experiment. Run only after `modal token info` and `modal billing rates` succeed. No volume or endpoint is created. Image build may incur CPU/network charges. """ import json import modal app = modal.App('modelclinic-gpu-preflight') image = modal.Image.debian_slim(python_version='3.11').pip_install('torch==2.8.0') @app.function(image=image, gpu=['L4','A10'], timeout=120, scaledown_window=2) def gpu_check(): import torch assert torch.cuda.is_available(), 'CUDA unavailable' assert torch.cuda.is_bf16_supported(), 'BF16 unavailable' a = torch.ones((256, 256), device='cuda', dtype=torch.bfloat16) result = a @ a torch.cuda.synchronize() assert float(result[0, 0]) == 256.0 free, total = torch.cuda.mem_get_info() return {'gpu':torch.cuda.get_device_name(), 'vram_total_bytes':total, 'vram_free_bytes':free, 'allocated_bytes':torch.cuda.memory_allocated(), 'torch_version':torch.__version__, 'cuda_version':torch.version.cuda, 'bf16_supported':torch.cuda.is_bf16_supported(), 'tensor_result':float(result[0,0])} @app.local_entrypoint() def main(): from pathlib import Path result = gpu_check.remote() Path('results/checkpoint_analysis/modal_gpu_smoke.json').write_text(json.dumps(result, indent=2)) print(json.dumps(result, indent=2))