vaishaalli/stage2-gkd-mixed

Stage-2 GKD draft (TB:xLAM mix)

A draft model for speculative decoding against the frozen target Qwen/Qwen2.5-Coder-14B-Instruct (vLLM speculative decoding with method=draft_model), produced by Stage-2 on-policy GKD over a 1:1 TB:xLAM mix. Full project, training scripts, and the complete acceptance / wall-clock / exactness evaluation: Speculative_Decoding.

Qwen2.5-Coder-0.5B-Instruct after on-policy GKD (reverse-KL, draft samples / frozen target scores) over a 1:1 TB:xLAM context mix, initialized from the tb-kd checkpoint. Mixed Stage-2 draft from the Speculative_Decoding project.

Measured (k=5, greedy, 100 frozen prompts, instrumented HF loop)

eval	tau	alpha	exactness vs AR
xLAM-500	4.18	0.956	50/50 token-identical
TB-500 (held-out tools)	2.94	0.843	50/50 token-identical

Outputs are token-identical to plain greedy decoding of the target in the HF instrumented loop (exactness gates; the vLLM bf16 near-tie caveat is documented in the repo README). The vocabulary is zero-padded to 152,064 rows to satisfy vLLM s SpeculativeConfig same-vocab-size requirement; no real token id lives in the padding band, and the padding is gated on greedy parity (see src/serving/prepare_draft.py).

Benchmark context: vLLM 0.30.0, target bf16, k=5, batch 1. This model is a research artifact from a 1-day H100 study; no safety fine-tuning was performed (the base model s policies apply).