llama 3.3 70b lora - warmup schedule affects which layer collapses first

1. trained llama 3.3 70b lora (rank 32, code task, 18k examples) with three different warmup schedules: no warmup, 50 steps, 200 steps 2. tracked per-layer gradient norms and found that no warmup causes layer 22 to collapse first (step 440), 50 step warmup moves collapse to layer 19 (step 620), 200 step warmup trains clean past 1000 steps with no collapse the gradient norms in layers 18-24 look completely different depending on warmup length. without warmup those layers hit peak gradient norm around step 200 and then explode. with 200 step warmup they stay stable the whole time. is this a known thing or did we stumble into something weird?