results

Recorded smoke tests

These files prove that a released RandKV revision ran against a public model checkpoint with real attention tensors. They are not performance benchmarks: there is no dense or scored-eviction baseline, warm-up protocol, or repeated trial distribution.

Qwen3-0.6B on Apple M4

• Result: qwen3-0.6b-mps-smoke.json
• Model revision: c1899de289a04d12100db370d81485cdf75e47ca
• Prompt tokens: 16
• Generated tokens: 64
• Persistent budget K: 32
• Buffer r: 8
• Eviction rounds: 5 per layer
• Final physical cache: 46 positions per layer

The final physical size is between eviction boundaries. The periodic policy returns to K + r = 40 positions after each round and may grow as high as K + 2r - 1 = 47 before the next compaction.

Dense vs RandKV Transformers microbenchmark

• Result: qwen3-0.6b-mps-microbenchmark.json
• Protocol: one warm-up per mode, three alternating-order trials, 128 generated tokens per trial, synchronized MPS timing
• Dense median: 38.53 tokens/second
• RandKV median: 35.48 tokens/second
• RandKV/dense throughput: 0.921x
• Final RandKV physical cache: 46 positions per layer after 143 total tokens

This single-request Apple M4 measurement is compatibility and adapter-overhead evidence only. The current Python gather path is slower than dense generation in this test. It is not a vLLM serving benchmark, a CUDA-kernel benchmark, or a model-quality evaluation, and it does not validate the paper's throughput claim.

Schema compatibility

Within schema v1, additive optional fields remain compatible. Removing or renaming fields, or changing their semantic meaning, requires a new schema version.

SmolLM2-135M on Windows CPU

• Result: smollm2-135m-smoke.json
• Model revision: 93efa2f097d58c2a74874c7e644dbc9b0cee75a2
• Environment: Windows 11, AMD64, Python 3.13.5, PyTorch 2.14.0+cpu, Transformers 5.16.1
• Hardware: 13th Gen Intel Core i7-13620H, 13.0 GB RAM
• Device: CPU
• Prompt tokens: 11
• Generated tokens: 128
• Persistent budget K: 64
• Buffer r: 8
• Eviction rounds: 9 per layer
• Final physical cache: 73 positions per layer

This smoke run verifies the Transformers adapter against the HuggingFaceTB/SmolLM2-135M full-attention decoder configuration. It records compatibility and eviction behavior only; it is not a throughput comparison, quality evaluation, or claim about the model family as a whole.