Tech spike: four smoke-test runs and a divergence-stop test

Pipeline works on MPS (~105 s per 300-step run), results are reproducible, and the divergence stop works. Not experimental evidence. Claude-Session: https://claude.ai/code/session_01S8KN7NuUCm8nXLmy6eDuVv
2026-10-28 18:14:26 +00:00
1 parent 55ba128dcc
commit 84b04d12e9
@@ -0,0 +1,86 @@
+{
+  "run_id": "20261028-180840_spike-qkon-lr1e-3",
+  "name": "spike-qkon-lr1e-3",
+  "note": "tech spike smoke test (not an experiment)",
+  "status": "completed",
+  "started": "2026-10-28T18:08:40+00:00",
+  "command": "/Users/nik/ISE/9.CS6513/thesis/06-experiments/.venv/bin/python /Users/nik/ISE/9.CS6513/thesis/06-experiments/nanogpt/train.py configs/spike.py --out_dir=/Users/nik/ISE/9.CS6513/thesis/06-experiments/runs/20261028-180840_spike-qkon-lr1e-3 --data_root=/Users/nik/ISE/9.CS6513/thesis/06-experiments/data --qk_norm=True --learning_rate=1e-3",
+  "config_file": "configs/spike.py",
+  "overrides": [
+    "qk_norm=True",
+    "learning_rate=1e-3"
+  ],
+  "config_requested": {
+    "dataset": "tinystories_spike",
+    "device": "mps",
+    "dtype": "float32",
+    "compile": false,
+    "n_layer": 4,
+    "n_head": 4,
+    "n_embd": 256,
+    "block_size": 256,
+    "bias": false,
+    "dropout": 0.0,
+    "qk_norm": true,
+    "batch_size": 32,
+    "gradient_accumulation_steps": 1,
+    "learning_rate": 0.001,
+    "max_iters": 300,
+    "lr_decay_iters": 300,
+    "min_lr": 0.0001,
+    "warmup_iters": 30,
+    "weight_decay": 0.1,
+    "grad_clip": 1.0,
+    "eval_interval": 100,
+    "eval_iters": 20,
+    "log_interval": 10,
+    "always_save_checkpoint": false,
+    "seed": 1337
+  },
+  "git": {
+    "commit": "55ba128dcca8310f85e99e172fe920a8c53a3a79",
+    "branch": "main",
+    "dirty": false
+  },
+  "nanogpt_upstream_commit": "3adf61e154c3fe3fca428ad6bc3818b27a3b8291",
+  "data": {
+    "dataset": "TinyStories V2 (GPT-4)",
+    "source": "https://huggingface.co/datasets/roneneldan/TinyStories/resolve/main/",
+    "tokenizer": "tiktoken gpt2",
+    "files": {
+      "TinyStoriesV2-GPT4-valid.txt": "6874bae9a4c1a4e7edcf0e53b86c17817e9cf881fc75ff2368da457b80c0585d"
+    },
+    "split": "valid file split 90/10 by story order: 24867 train / 2763 val stories",
+    "tokens": {
+      "train": 4920360,
+      "val": 548664
+    },
+    "train_bin_bytes": 9840720,
+    "val_bin_bytes": 1097328
+  },
+  "environment": {
+    "host": "Cormacs-MacBook-Pro.local",
+    "platform": "macOS-15.7.7-arm64-arm-64bit",
+    "machine": "arm64",
+    "cpu": "Apple M4 Pro",
+    "python": "3.12.14",
+    "torch": "2.15.0",
+    "mps_available": true,
+    "cuda_available": false
+  },
+  "finished": "2026-10-28T18:10:25+00:00",
+  "wall_seconds": 104.8,
+  "summary": {
+    "params_millions_nanogpt_count": 16.03,
+    "first_train_loss": 10.9141,
+    "last_train_loss": 3.3557,
+    "last_eval": {
+      "iter": 300,
+      "train_loss": 3.3683,
+      "val_loss": 3.4304
+    },
+    "best_val_loss": 3.4304,
+    "diverged_at_iter": null,
+    "returncode": 0
+  }
+}