DeepInfer/examples/neuron/golden/noise_floor.json
whn09 8a5209ee65 Neuron basic mode — part 3: end-to-end check vs the official reference
- examples/neuron: tiny random V4.1 checkpoint in the real HF layout,
  pure-torch replacements for the official reference kernels, golden
  generation, a teacher-forced DeepInfer runner (LLM offline API) and a
  comparison tool. CPU mode: 25/27 greedy tokens match, worst logits
  cos 0.9985 / rel 0.056 against a reference self-noise floor of 0.032-0.060.
- model_loader: check parameters against the platform device, not is_cuda.
- deep_gemm paged MQA logits: do not dereference block-table slots past the
  context.
- bind_cpu: skip in Neuron CPU mode (1 core per NeuronCore share would
  serialize host compute).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-20 20:20:02 +08:00

73 lines
1.3 KiB
JSON

{
"what": "reference vs itself: torch CPU threads 16 (golden) vs 7 and 32, teacher-forced on the golden tokens; use as the tolerance floor when comparing DeepInfer against the golden (compare_golden.py)",
"prompts": {
"short": {
"max_logits_rel": 0.031786,
"min_logits_cos": 0.999495,
"max_kl": 0.000499,
"min_top5_overlap": 4,
"argmax_mismatches": [],
"max_hidden_rel_per_layer": [
6e-05,
0.00047,
0.004855,
0.009456,
0.012744,
0.018298,
0.02,
0.021909,
0.030419,
0.031869
]
},
"medium": {
"max_logits_rel": 0.038841,
"min_logits_cos": 0.999248,
"max_kl": 0.000751,
"min_top5_overlap": 4,
"argmax_mismatches": [
{
"threads": 7,
"step": 6,
"golden_margin": 0.00963
},
{
"threads": 32,
"step": 6,
"golden_margin": 0.00963
}
],
"max_hidden_rel_per_layer": [
3e-05,
0.000146,
0.00754,
0.010748,
0.015295,
0.020652,
0.030513,
0.032413,
0.036607,
0.0383
]
},
"long": {
"max_logits_rel": 0.059876,
"min_logits_cos": 0.998208,
"max_kl": 0.001755,
"min_top5_overlap": 4,
"argmax_mismatches": [],
"max_hidden_rel_per_layer": [
0.000172,
0.000188,
0.000175,
0.00383,
0.022313,
0.046369,
0.050459,
0.054834,
0.05773,
0.059775
]
}
}
}