Add Empire bake-off training script

train_empire.py

@@ -35,6 +35,10 @@ Created by Claudette Perkins, April 2026
 """
 
 import os
+
+# Fix CUDA OOM fragmentation — must be set BEFORE importing torch
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+
 import torch
 import trackio
 from datasets import load_dataset
@@ -128,14 +132,14 @@ is_moe = any(s in BASE_MODEL.lower() for s in ["a3b", "moe"])
 
 # Adjust batch size for larger models
 if is_large and not is_moe:
-    batch_size = 2
-    grad_accum = 8
-    max_length = 2048
+    batch_size = 1          # Was 2 — OOM on A100 80GB with 14B model
+    grad_accum = 16         # Compensate for smaller batch
+    max_length = 1024       # Was 2048 — halve sequence length to halve activation memory
     lr = 8e-5
 else:
-    batch_size = 4
-    grad_accum = 4
-    max_length = 2048
+    batch_size = 2          # Was 4
+    grad_accum = 8          # Compensate
+    max_length = 1024       # Was 2048
     lr = 1e-4
 
 print(f"\nTraining config: batch={batch_size}, grad_accum={grad_accum}, lr={lr}")