DroidSpeak · independent mechanism baseline · October 2026

Reuse most sender KV.
Recompute the receiver layers that matter.

Our DroidSpeak baseline selected four receiver layers out of 32 and passed its held-out quality rule on a paper-named Mistral adapter pair. The receiver reused sender KV before and after that interval. It did not translate every KV tensor and did not train a new checkpoint.

This result validates the partial-recomputation mechanism in one process. It does not reproduce the paper's remote transfer pipeline or its end-to-end speedups.

Current decisionMechanism quality: pass. Production system and distributed speedup: not yet tested.
4 OF 32 LAYERS RECOMPUTED50 PROFILE + 50 HELD-OUT RECORDSFULL-KV PARITY: EXACTLOCAL TIMING ONLY

What DroidSpeak changes

The key-value (KV) cache stores the attention keys and values for prior tokens. Two models can have the same architecture and token vocabulary but different weights, such as two adapters on one base model. Their caches are not normally interchangeable. DroidSpeak profiles where the receiver must regain its own representation, then recomputes only one contiguous layer interval.

1 · Sender prefillCache the whole prompt except its final input token. Record KV and possible transition embeddings.
2 · Offline profileTest all contiguous two-layer-boundary intervals and measure answer F1 and local time to first token.
3 · SelectChoose the fewest receiver layers that stay within 5% of full-receiver profile F1.
4 · Receiver handoffAssemble reused and recomputed KV, then process the final input token to obtain receiver logits.

The cache covers the whole prompt, not only the passage. Holding back the last input token is part of the generation contract. The receiver processes that token after the hybrid cache is ready and uses its output logits for the first generated token.

The profiler chose four receiver layers

The profile used 50 HotpotQA multi-hop question-answering records. It tested 137 intervals on two-layer boundaries, including full reuse and full receiver prefill. Token F1 measures normalized word overlap with the reference answer on a scale from 0 to 1. The selection rule allowed at most a 5% relative F1 drop from full receiver prefill.

Full receiver
0.2067

Mean profile token F1 and the quality reference.

Quality floor
0.1963

95% of full-receiver profile F1.

Selected [14,18)
0.2030

Four recomputed layers. This meets the floor.

Direct full reuse
0.0767

No receiver recomputation. This fails badly.

This is an offline profile decision, not a trained model. The selected interval is configuration for this model pair. No learned DroidSpeak weights or checkpoint were produced.

The selected interval passed on disjoint records

A separate set of 50 HotpotQA records tested the selected interval. These records were not used to choose it. The full receiver reached 0.2887 F1, which set a held-out 5%-drop floor of 0.2742.

Held-out mean token F1

Selected DroidSpeak exceeded the receiver-relative floor

selected0.3149
full receiver0.2887
direct reuse0.2263
full sender0.0320
QUALITY DECISION: PASS

The selected interval reached 109.1% of the full-receiver F1 point estimate and exceeded the required floor. Direct reuse reached only 78.4% of full receiver and failed the floor.

The point estimate above full receiver is not evidence that DroidSpeak improves model quality. This is a small, single-split evaluation, and the selection used a related profile set from the same dataset.

Full receiver time to first token
0.2218 s

Native receiver prefill in the same process.

Selected time to first token
0.0551 s

Hybrid assembly and four-layer recomputation.

Local ratio
4.03×

Full-receiver TTFT divided by selected TTFT.

Direct reuse TTFT
0.0273 s

Fastest local control, but below the quality floor.

Do not read 4.03× as a DroidSpeak system speedup. The timing is from one in-process Transformers run. It excludes sender cache production, network transfer, remote cache lookup, serialization, scheduling, and communication overlap. It is useful as a compute-path signal only.

Paper versus this work

We implemented the paper's central partial-recomputation method. The evaluation is a bounded mechanism study, not a paper-scale system reproduction.

ItemDroidSpeak paperThis workStatus
Core methodReuse sender KV outside one contiguous receiver interval. Transfer a transition embedding and recompute the interval.The same cache ownership and interval-recomputation contract.Matched
Model pairIncludes Mistral-7B with public conllpp and glue_sst2 adapters. Exact repository revisions are not reported.The same named base and adapters, using resolved public revisions.Names match
ProfilerSearches contiguous intervals and uses a receiver-relative quality budget.All 137 intervals on two-layer boundaries; 50 profile records; 5% relative F1 budget.Method matched
EvaluationLong-input and cross-dataset experiments at paper scale.50 disjoint held-out HotpotQA records, 2,048-token passage limit, 32-token answer limit.Smaller and narrower
Selected intervalThe paper shows pair-dependent intervals and an example near 11 recomputed layers.Layers [14,18): four of 32 layers. The exact paper checkpoints and sample IDs were unavailable.Not directly comparable
SystemDistributed sender and receiver with transfer and compute overlap.One-process Transformers baseline. No remote data path and no overlap.Not reproduced
WHAT THE RESULT SUPPORTS

For one same-base adapter pair, direct cache reuse was poor, while a profiled four-layer receiver interval restored held-out quality under the stated rule. The full-receiver interval produced bit-exact KV tensors relative to ordinary receiver prefill.

WHAT IT DOES NOT SUPPORT

It does not establish arbitrary cross-family sharing, paper-scale generalization, remote-transfer speedup, production readiness, or an exact reproduction of the authors' interval.

LMCache can carry the cache, but it does not supply DroidSpeak

A separate control used one LMCache server and two same-model serving processes. The second process retrieved 7,168 cached tokens, and both processes produced the same completion. This verifies shared same-model transport.

VERIFIED TRANSPORT

LMCache can move and retrieve a compatible KV cache across serving processes. The cache server reported the retrieval in 0.003 seconds in this bounded control.

MISSING CROSS-MODEL BRIDGE

Stock LMCache does not select a DroidSpeak interval, transfer the arbitrary transition embedding, run only that receiver interval, or assemble sender and receiver layers into one cache.

System role: LMCache is a useful transport substrate. A production integration still needs model-pair namespaces, compatibility checks, transition-embedding transport, partial receiver execution, hybrid-cache assembly, and transfer/recomputation overlap.

Source and claim boundary

The public code contains the in-process mechanism, exhaustive interval profiler, exact full-recompute audit, scoring, deterministic data split, and LMCache transport control. Raw run records are not part of the public code tree.

Run shape: 16-bit brain floating point (BF16), scaled dot-product attention, one 48 GB accelerator, 2,048-token passage limit, and at most 32 generated tokens. No DroidSpeak training checkpoint exists because the method profiles an interval rather than fitting new weights.

Practical statusResearch mechanism implemented and quality-qualified for one adapter pair; distributed serving integration remains open.