-
Fix: total_mem -> total_memory"
-
Add quick_test.py - standalone timing test, prints immediately, no imports from other files"
-
Fix: force unbuffered stdout for Kaggle notebook output + use print() with flush"
-
Add --log_batch_every: prints loss every N batches within an epoch"
-
Add --num_gpus: splits tasks across GPUs in parallel processes"
-
Add --log_every arg to control epoch print frequency (default 100)
-
Fix: add build_trm_nano_onnx function (alias used by train_task.py)
-
Add solved_tasks.json - 51 tasks solved by analytical solvers
-
Replace train_task.py with full batch runner: loops all 400 tasks, skips solved, logs everything, uses manual ONNX builder
-
Add manual ONNX builder with weight reuse - fits 1.44MB with full h=96 T=3 n=4"
-
Add TRM-Nano README with architecture details and usage"
-
Add TRM-Nano per-task training script with augmentation and ONNX export
-
Add TRM-Nano model implementation (h=96, 4 heads, 2 layers, unrolled recursion)"
-
Update classifier to support Kilo local server
-
Update TODO with TRM findings, fix outdated info
-
Add TRM research findings and architecture notes
-
Add composition solvers for chained transforms
-
Add DeepSeek task classifier for LLM-guided solver routing
-
reorg kilo skills folder
Roger MT
-
Add TRM solver skill for agent guidance
-
Upload trm_solver/__init__.py
-
Add TRM solver README
-
Add Kilo Bridge for LLM-driven ARC task analysis and ONNX export
-
Add TRM solver NN executor with 14 transform types and ONNX export
