inference-optimization/ctest-Qwen3-8B-speculator.dflash-fp8ablation-bf16

ctest: Qwen3-8B DFlash speculator — FP8 hidden-states ablation (bf16)

Test checkpoint — not a production release. Trained to validate the FP8-quantized hidden-states transfer backend added in vllm-project/speculators#1028 (draft, reviving #491).

• Verifier: Qwen/Qwen3-8B
• Algorithm: dflash
• Hidden-states precision during training: bf16 (existing bf16 file backend)
• Training data: 5K magpie + 5K ultrachat samples from inference-optimization/Dataset-Qwen3-235B-Instruct

This is one of a matched pair (...-fp8ablation-bf16 / ...-fp8ablation-fp8) trained identically except for the hidden-states transfer precision, to isolate the effect of FP8 quantization on speculator quality. See the sibling repo for the other precision.

Validation metrics (final epoch)

• loss_epoch: 0.670712
• full_acc_epoch: 0.190457
• position_1_acc_epoch: 0.671505
• position_2_acc_epoch: 0.470419
• position_3_acc_epoch: 0.345403
• position_4_acc_epoch: 0.259921
• position_5_acc_epoch: 0.199261
• position_6_acc_epoch: 0.156393
• position_7_acc_epoch: 0.125580
• position_8_acc_epoch: 0.105085
• position_9_acc_epoch: 0.090258
• position_10_acc_epoch: 0.081353
• position_11_acc_epoch: 0.075329
• position_12_acc_epoch: 0.070879
• position_13_acc_epoch: 0.068388
• position_14_acc_epoch: 0.065941
• position_15_acc_epoch: 0.063810
• eal_epoch: 1.165076

guidellm serving eval

See acceptance.csv in this repo for the full per-subset guidellm breakdown (9 subset rows).

Full ablation writeup, all 6 checkpoints' results, exact commands, and code: see the results package referenced from PR #1028.