Jun-LoRA-12B-MTP-GGUF

A trained MTP / speculative-decoding draft model for efficiencyx/Jun-LoRA-12B-GGUF. Not a standalone chat model — it only exists to propose tokens that Jun then verifies.

• jun-drafter-qat600-q4_k_m.gguf — 327 MB, Q4_K_M
• Initialized from google/gemma-4-12B-it-qat-q4_0-unquantized-assistant, matching Jun's own lineage (unsloth/gemma-4-12B-it-qat-q4_0-unquantized)
• 4 layers, hidden 1024, backbone hidden 3840
• 600 steps on a 1500-sample in-character roleplay corpus, A100 80GB

Speed

RTX 3060, ollama, Jun 12B Q4_K_M as the target, medians over 6 runs:

setup	tok/s
Jun alone, no drafter	36.28
this drafter, draft_num_predict=1	45.48 (+25.4%)
this drafter, n=2	41.85
this drafter, n=3	40.79
this drafter, n=4	36.73

Use n=1. The usual "draft 2-3 tokens" advice loses on this hardware: each extra token in the verify batch costs ~10ms on a 3060, so deeper drafts pay more to verify than they save.

Acceptance

Accepted tokens per target forward, higher is better. On an A100 at bf16 the training gain is clear:

	bf16 (A100)
stock drafter	2.10
trained, 300 steps	2.81
trained, 600 steps	2.89

Acceptance improved monotonically with training. On a 3060 the end-to-end tok/s above is bounded by verify cost rather than by acceptance, so the headroom this buys shows up on hardware where the target forward is the cheaper half.

Lineage matters An earlier run initialized from the

non-qat assistant — same backbone_hidden_size, so every shape check passed — and lost to stock outright at Q4_K_M. A drafter reads the target's last-layer activations directly; "same size" is not "same body".

Quantization

Drafter quantization barely moves acceptance (bf16 2.18, Q8_0 2.16, Q4_K_M 2.16 accepted tokens per target forward) but strongly moves draft time (8817 / 4935 / 4352 ms for the same work). Q4_K_M is the right format for a drafter.