Rohith Reddy 0bf88031f3 feat: support non-Triton platforms (aarch64/Jetson) — conditional torch.compile ... torch.compile requires Triton which is unavailable on aarch64 (NVIDIA Jetson, ARM servers). This makes compilation conditional: - nanochat/optim.py: Replace hardcoded @torch.compile decorators with _compile_decorator that becomes a no-op when Triton is not installed - scripts/base_train.py: Wrap torch.compile(model) in try/except - nanochat/common.py: Add Orin to peak FLOPS table (5.3 TFLOPS BF16) Tested end-to-end on NVIDIA Jetson Orin NX 8GB (Compute 8.7, CUDA 12.6, PyTorch 2.8.0) — pretrain and SFT both run successfully in eager mode. No changes to behavior on platforms where Triton is available.		2026-03-07 23:49:33 +05:30
..
base_eval.py	delete autocast, an unnecessary thorn in my side, manage dtypes directly	2026-03-04 23:55:30 +00:00
base_train.py	feat: support non-Triton platforms (aarch64/Jetson) — conditional torch.compile	2026-03-07 23:49:33 +05:30
chat_cli.py	delete autocast, an unnecessary thorn in my side, manage dtypes directly	2026-03-04 23:55:30 +00:00
chat_eval.py	delete autocast, an unnecessary thorn in my side, manage dtypes directly	2026-03-04 23:55:30 +00:00
chat_rl.py	delete autocast, an unnecessary thorn in my side, manage dtypes directly	2026-03-04 23:55:30 +00:00
chat_sft.py	delete autocast, an unnecessary thorn in my side, manage dtypes directly	2026-03-04 23:55:30 +00:00
chat_web.py	delete autocast, an unnecessary thorn in my side, manage dtypes directly	2026-03-04 23:55:30 +00:00
tok_eval.py	initial commit	2025-10-13 06:49:24 -07:00
tok_train.py	quick fix to not OOM main speedrun script	2026-01-26 22:31:42 +00:00