File size: 1,159 Bytes
80692f2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 | """Shape sweep for Kahan-corrected softmax.
The point of this problem is numerical accuracy on long reductions. Shapes
mix typical LLM vocab sizes with deliberately adversarial regimes:
- small vocab (sanity check; naive fp32 should pass)
- Llama3 vocab 128K (real-world, where fp16 accumulation starts to drift)
- 256K (DeepSeek-V3 / Gemma-3 class vocab; naive fp16 sum DOES drift past
the 1e-5 tolerance — this row is what proves Kahan was needed)
- extreme-logit edge case (large positive logits stress max-subtract +
summation; if the implementation accidentally exps before subtracting
max, this row overflows)
The 'extreme' flag is read by check.py to switch input generation to a
distribution that produces a few very large logits per row.
"""
SHAPES = [
{"batch": 32, "vocab": 4096, "extreme": False}, # sanity
{"batch": 16, "vocab": 32768, "extreme": False}, # GPT-2 class
{"batch": 8, "vocab": 131072, "extreme": False}, # Llama3 vocab
{"batch": 4, "vocab": 262144, "extreme": False}, # 256K — Kahan needed
{"batch": 8, "vocab": 131072, "extreme": True}, # extreme logits edge
]
|