Hackathon speedrun: max_new_tokens=32, seq_len=512 for 4-8x faster iterations

cloud_arena/llm_training.py

@@ -28,8 +28,8 @@ DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 
 # ── GPU Optimization Constants ────────────────────────────────────────────────
 GRAD_ACCUM_STEPS = 4       # accumulate gradients over N episodes before stepping
-MAX_SEQ_LEN = 1024         # longer context = more GPU work per forward pass
-MAX_GEN_TOKENS = 256       # more generation tokens per step
+MAX_SEQ_LEN = 512          # shorter context = O(N²) attention is 4× faster than 1024
+MAX_GEN_TOKENS = 32        # force brief output — just need the ACTION line, not essays
 
 
 def format_prompt(state_dict):