Can the larger model skip prefill?

Inference normally makes every model compute the prompt for itself. V0 tested a different path: prefill with Qwen3 0.6B, remove its source positional encoding, map the cached keys and values head by head, apply the target positions, and inject that state into Qwen3 1.7B before decode.

The useful result is not whether a tensor can be moved. It is whether the target model still behaves like the target model after the handoff.

One H100, disjoint fitting and evaluation data.

Source
Qwen3 0.6B
Target
Qwen3 1.7B
Hardware
1× NVIDIA H100 SXM5 80 GB
Precision
BF16
Mapper fit
50 training sequences
Selection
25 disjoint sequences
Selected map
Per-head linear map, k=1
Task checks
HellaSwag and PIQA

Before evaluating the transfer, the harness checked model and cache compatibility, cache serialization, BF16 cached decode, corruption detection, RoPE round trips, and deterministic mapper serialization. Those controls passed.

The learned map recovered most task performance.

Held-out behavior by cache variant
Variant Perplexity Next-token top-1 HellaSwag PIQA
Target full-prefill oracle 15.2157 100% 59.4% 74.0%
Direct source-cache injection 44.2169 2% 36.6% 60.0%
Learned mapped transfer, k=1 17.3496 61% 54.6% 70.2%
Stabilized mapped transfer, k=4 17.0744 62% 51.2% 68.4%

The selected k=1 map reduced perplexity from 44.2169 with direct injection to 17.3496 and raised next-token agreement from 2% to 61%, a 30× gain. It also recovered much of the oracle’s held-out HellaSwag and PIQA accuracy. A subsequent stabilized k=4 run reached 62% agreement, 17.0744 perplexity, and 76.4% normalized task retention, giving the next study two distinct mapper configurations to compare on fresh samples.

What “85.1% retained” means.

The headline is the mean task retention after normalizing each benchmark against its chance floor. A model does not get credit for performance that random guessing would produce. On this run, the mapped cache scored 54.6% on HellaSwag against the target oracle’s 59.4%, and 70.2% on PIQA against 74.0%.

For each task (mapped score − chance floor) ÷ (oracle score − chance floor)

This makes the number more comparable across multiple-choice tasks with different random baselines. Token-level agreement is reported separately so task retention and immediate fidelity remain visible together.

85.1% retained with a clear fidelity roadmap.

V0 reached the predeclared task-retention floor, stayed within the 1.25× perplexity limit, and improved every metric over direct injection. Next-token agreement reached 61%, establishing a 19-point path to the 80% milestone and a focused objective for multistep rollout work.

The protocol schedules formal latency acceptance after fidelity optimization. V0 establishes that cross-model transfer can preserve useful task performance and identifies the measurements the next model-pair study will optimize.

Expand fidelity, then quantify savings.

The next iteration expands calibration data, compares k=1 and k=4 on fresh disjoint samples, and tests richer mapping strategies. Larger compatible model pairs will make the latency study more representative of the inference savings this technique is intended to unlock.

The experiment sequence is task-level non-inferiority on representative traffic, stronger next-token agreement, stable rollouts, and then a synchronized latency comparison against ordinary target prefill.