ATOM (atom-ai): code snapshot (100fae0 + batched-train), README with measured numbers, curated checkpoints
9f1076a verified | == host: 169f95f090e6 2026-09-30T20:59:23+00:00 | |
| == GPU | |
| NVIDIA GeForce RTX 5090, 32607 MiB, 580.82.09 | |
| | NVIDIA-SMI 580.82.09 Driver Version: 580.82.09 CUDA Version: 13.0 | | |
| == disk (repo fs) | |
| overlay 300G 250M 300G 1% / | |
| == CPU/RAM | |
| 24 | |
| total used free shared buff/cache available | |
| Mem: 125 10 81 0 34 113 | |
| python 3.10.13 | |
| torch 2.11.0+cu128 cuda build 12.8 cuda available True | |
| gpu0: NVIDIA GeForce RTX 5090 31.4 GiB sm_120 | |
| matmul ok True | |
| == batched equivalence tests (CPU + CUDA) | |
| ........... [100%] | |
| =============================== warnings summary =============================== | |
| test/test_batched_train.py::BatchIndependenceTests::test_b4_equals_four_b1 | |
| /root/atom-ai/test/test_batched_train.py:235: UserWarning: Converting a tensor with requires_grad=True to a scalar may lead to unexpected behavior. | |
| Consider using tensor.detach() first. (Triggered internally at /pytorch/torch/csrc/autograd/generated/python_variable_methods.cpp:836.) | |
| self.assertLessEqual(float((os_["byte_logits"][0] - o["byte_logits"][b]).abs().max()), TOL) | |
| -- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html | |
| 11 passed, 1 warning in 74.21s (0:01:14) | |
| == short GPU benchmark (bytes/s vs B) | |
| batched B= 64 88.3 ticks/s 5653 bytes/s wall (median 5583) 5654 bytes/cpu-s peakRSS=1900MB cuda_peak=918MB | |
| batched B= 256 87.6 ticks/s 22438 bytes/s wall (median 22351) 22445 bytes/cpu-s peakRSS=1902MB cuda_peak=3604MB | |
| batched B= 1024 83.9 ticks/s 85881 bytes/s wall (median 84573) 85908 bytes/cpu-s peakRSS=1912MB cuda_peak=14383MB | |
| == setup OK | |