Vela 2.0 decoder export parity (GPU): stage2/ckpt-003000

Export vela2_9B_nb_20261003; reference = the research decoder scorer (dec_scorer_d3_20261001.D3Scorer: render_row(train=False), tree forward, DEV2.0 head, span head v2, windowed c), device cuda. Rows: choice:domain_dev 15, choice:factcheck_dev 10, choice:modality_dev 10, choice:feedback_dev 10, choice:oasst1 8, choice:boolq 8, choice:snli 8, choice:arc 6, choice:massive 6, score:helpsteer2 10, score:jsts 8, score:klue_sts 6, set:shield 40, span:pii_dev 30, span:pii_longdev_4096 4, span:pii_longdev_8192 3, span:halu 20, span:toxic 10, span:fewnerd 6, span:multiconer 6, span:tydiqa_goldp 6 = 230 real dev rows.

Probabilities are the calibrated ones the module returns (choice / score softmax over the options at T_type, set sigmoid at T_set, span sigmoid per word x label at T_span). Decisions: argmax (choice / score), set options above the calibrated threshold, decoded span sets (label, start, end) at the calibrated threshold (PII: length rule + sparse gate).

Arms: ref_bf16 = research scorer, bf16 autocast (the evaluation setting); ref_fp32_r = research scorer in FP32 with its backbone Linear / conv1d weights rounded to bf16 (= the stored release weights); ref_fp32 = research scorer in FP32, unrounded weights; exp_* = this export (vela2_inference + modeling_vela2).

Rendering

rows	sequences (incl. windows)	mismatches	parts ids	blocks	word positions / offsets	window count
230	260	0	0	0	0	0

Outputs

comparison	type	n	max abs prob diff	max abs logit diff	argmax / decision diffs
ref_bf16_vs_exp_bf16	choice	169	0	0	0
ref_bf16_vs_exp_bf16	score	101	0	0	0
ref_bf16_vs_exp_bf16	set	26	0	0	0
ref_bf16_vs_exp_bf16	span (36360 word units)	85	0	0	0 span-set diffs (408 spans), 0 unit-offset mismatches, PII gate fired on 36 rows
ref_fp32_r_vs_exp_fp32	choice	169	0	0	0
ref_fp32_r_vs_exp_fp32	score	101	0	0	0
ref_fp32_r_vs_exp_fp32	set	26	0	0	0
ref_fp32_r_vs_exp_fp32	span (36360 word units)	85	0	0	0 span-set diffs (405 spans), 0 unit-offset mismatches, PII gate fired on 36 rows
ref_fp32_vs_exp_fp32	choice	169	0.0065	0.0495	0
ref_fp32_vs_exp_fp32	score	101	0.00149	0.0191	0
ref_fp32_vs_exp_fp32	set	26	0.00174	0.0295	0
ref_fp32_vs_exp_fp32	span (36360 word units)	85	0.0189	0.664	3 span-set diffs (405 spans), 0 unit-offset mismatches, PII gate fired on 36 rows

Decoder cross-check (export decoder vs span_decode_r4 on the same reference probabilities): 0 row(s) differ.

Bundled vs single (export, FP32)

60 multi-question rows (choice:oasst1 8, choice:massive 6, score:helpsteer2 10, set:shield 26, constructed:domain+4signals+pii 10): every question run alone (its own sequence) vs bundled. Choice / score / set: 251 questions, 0 bit-identical, max |logit diff| 0.00593. Span: 10 rows, 0 bit-identical, max |logit diff| 0.0138. Noise-floor control (the same bundled rows, one row per batch instead of batched with others): max |logit diff| 0.00418 (span 0.0117); decision flips bundled vs single: 1. Differences of this size come from floating-point batch shapes (padding, GEMM sizes), not from one question block seeing another.

SystemOne vs predict()

6 SystemOne requests (choice / noul / score; span + choice over a 3-field JSON state; set): max |diff| between the SystemOne answers and the predict() probabilities they map from 0, 0 decision mismatches. HTTP route function (system_one_request, what vela2_serve.py calls): ok=True, other model name rejected=True.

Arm run times

ref_bf16 47.5 s, ref_fp32_r 69.7 s, ref_fp32 61.1 s, exp_bf16 18.9 s, exp_fp32 59.9 s

Notes

Bar: rendering identical; with ref_bf16 vs exp_bf16 (same kernels, same weights as autocast sees them, same batches) the outputs should be identical up to kernel nondeterminism; ref_fp32_r vs exp_fp32 checks the FP32 path on identical weights.
ref_fp32 vs exp_fp32 is what storing the backbone matmul weights in bf16 changes when running in FP32 (no effect under bf16 autocast).
Choice / score probabilities are over the real options (the convention the temperatures were fitted on); [ABS] is reported separately (abstain_probability). The research scorer slices [ABS] off, so its calibration is unfitted.
Generic span threshold span:* = 0.5 and the GLiNER2-style default question texts for open-label classify / extract are export defaults, not dev-fitted.