GPT-OSS-Lite Optimization Stack

GPT-OSS-Lite Optimization Stack An architecture diagram generated by Archify. Host data · mmap + boundary copies · Host budget · DataLoader workers Host data mmap + boundary copies DataLoader workers Training activations · checkpoint every 3rd block · Activation budget · CE chunks: 8192 tokens Training activations checkpoint every 3rd block CE chunks: 8192 tokens Decode KV · 6 rings + 6 global caches · Inference budget · logical ~1.998× at 128K Decode KV 6 rings + 6 global caches logical ~1.998× at 128K Compute dispatch · SDPA · compile requested · Compute budget · Triton MoE opt-in Compute dispatch SDPA · compile requested Triton MoE opt-in Precision + AdamW · BF16 CUDA autocast · Parameter / state budget · FP32 model / optimizer Precision + AdamW BF16 CUDA autocast FP32 model / optimizer Persistence + guard · atomic per-file writes · Reliability controls · periodic: no RNG/cursor Persistence + guard atomic per-file writes periodic: no RNG/cursor Host budget Activation budget Inference budget Compute budget Parameter / state budget Reliability controls Legend Backend Database Security

Conditional execution

  • • CUDA + training.compile request max-autotune; setup exceptions print and continue eager.
  • • SDPA backend selection depends on masks/hardware. Triton is separate, explicit opt-in.

Evidence limits

  • • 16–20 h on A100 80GB is an unverified target, not a measured run.
  • • Periodic checkpoints omit RNG and sampler state. No exact resume guarantee.

Source / configured baseline

  • • Revision 3c9675deaf16
  • • configs/pretrain_a100_502m.yaml
  • • ModelConfig defaults · no CLI overrides