Receipts

A receipt is what the release gate measured for a recipe: speed at stated contexts, quality checks, and the image digest and model hash they were taken on. Each number below comes with its scope. The source files are in docs/receipts/engine-v0.1.0/.

Recipe	Receipt file
iamMrDuncan/Qwen3.8-27B-PN3-GGUF on p100	Qwen3.8-27B-PN3-GGUF-p100.md
iamMrDuncan/Qwen3.8-27B-UD-IQ3_XXS-Q3Floor-GGUF on p100	Qwen3.8-27B-UD-IQ3_XXS-Q3Floor-GGUF-p100.md

Two kinds of measurement

Kind	Who measured	On what	How
Research receipt	The decode hill climb’s judge	A source build of the engine commit, with the same CUDA configuration as the image, on one P100	Timed from outside the engine: each streamed token is stamped on the judge’s own clock. Greedy decoding.
Published image	tensorless bench	The released image digest, with exactly the recipe’s switches and flags, on one Tesla P100-PCIE-16GB	The median of 3 runs of 256 greedy tokens after a 429-token prompt. See Command line.

The two use different prompts, so a few percent of difference between them is expected.

Qwen3.8-27B-PN3

The engine source of v0.1.0 is the champion of an automated decode hill climb: research commit cfcab96e, exported unchanged except for six comment lines (see engine/PROVENANCE.md).

Speed, research receipt

Measure	Qwen3.8-27B-PN3	Q3Floor, same card, same cells	Scope
Decode, code prompt	30.06 tok/s	26.30	256 greedy tokens after a cold prompt of about 600 tokens; the median of 3 confirmation blocks
Decode, prose prompt	30.18 tok/s	26.31	the same, prose prompt
Decode, four days later on another card	29.95 / 29.96 tok/s	26.26 / 26.27	the same cells, the baseline of the next campaign
Prefill, code / prose	121.4 / 119.6 tok/s	—	a cold prompt of about 4,100 tokens, 8 micro-batches of 512
Decode at the start of the climb	7.74 / 7.73 tok/s	26.30 / 26.31	the same cells: 3.89× over the climb

Numerics held during the climb

Every step of the climb kept Qwen3.8-27B-PN3’s numerics where they started:

Check	Bar held
KL divergence to the starting point’s outputs, 128 forced prose tokens	within 2.593e-3
Same top token as the starting point	at 95% or more of positions
Decode-path perplexity	within 0.25%
Q3Floor’s greedy text	bit-identical
PN golden blocks, and one real tensor of each PN class	decoded bit-exactly on CPU and GPU

Quality, measured on the climb’s starting point

The climb held quality at its starting point. That starting point was gated against Q3Floor on one P100, each model with its own file:

Check	Qwen3.8-27B-PN3	Q3Floor	Bar
Perplexity, context 2,048 (6 chunks)	2.5494	2.5682	≤ 1.005 × Q3Floor: passes (0.993×)
Decode-path perplexity, code at 32K (512 forced tokens)	1.38147	1.37773	≤ 1.005 × Q3Floor: passes (1.003×)
Decode-path perplexity, prose at 32K	6.07793	5.65357	≤ 1.005 × Q3Floor: misses (1.075×), accepted
Needle recall at 1K–31K (32K context) and 1K–63K (65K context)	8 of 8	exact	exact
Task screen (26 cases)	23	23	at most 1 fewer than Q3Floor

The prose miss belongs to the quantization, not the kernels: the format’s own fp32 reference path scores 6.07572, so the fast kernels account for 0.5% of the gap. It was accepted.

The published image (2026-11-02)

ghcr.io/iammrduncan/tensorless-engine@sha256:35409ff97c2f518e468790c28bd19206ae777a945fcffe5775585e28fd9939a7, measured as 0.1.0-rc1, on one Tesla P100-PCIE-16GB that reported compute capability 6.0 and 16,269 MiB, with the model file checked against sha256 15ce2d44…:

Measure	Value
tensorless bench decode, median of 3 runs (256 tokens after a 429-token prompt)	30.03 tok/s (30.03 / 29.83 / 30.13)
Against the research receipt (30.06)	−0.1%
llama-server’s own decode timer, 3 runs of 256 tokens after a 45-token prompt	29.97 / 29.94 / 29.94 tok/s
Time to first token, 429-token prompt	4.46 s (median)
GPU memory, idle / generating	13,997 / 14,577 MiB
Load to listening, page cache warm	3.6 s

bench ran on another computer, through Wi-Fi, ssh and a port-forward to the server. The network adds latency to the time to first token, but not to the decode rate, which is timed between streamed tokens.

Not measured yet

• 32K and 65K context on this engine. At those contexts only the climb’s starting point was measured: 7.55 / 7.39 tok/s for Qwen3.8-27B-PN3, against 24.52 / 22.92 for Q3Floor. Attention costs a larger share of a 30 tok/s token than of a 7.5 tok/s one, so expect this engine’s long-context decode to be below 30.
• Speculative decoding for Qwen3.8-27B-PN3.

Qwen3.8-27B-UD-IQ3_XXS-Q3Floor

The published image (2026-11-02)

The same image digest, measured as 0.1.0-rc1 and released as 0.1.0-sm60, on one Tesla P100-PCIE-16GB, with exactly the recipe’s switches and flags. The model file was checked against sha256 4b4d6a39…, the same bytes as the Hugging Face upload at revision 11f76ec7.

Measure	Value
tensorless bench decode, median of 3 runs (256 tokens after a 429-token prompt)	26.18 tok/s (26.16 / 26.18 / 26.19)
Against the research receipt (26.26)	−0.3%
llama-server’s own decode timer, 3 runs of 256 tokens after a 45-token prompt	26.16 / 26.19 / 26.21 tok/s
Time to first token, 429-token prompt	3.85 s (median)
GPU memory, idle / generating	13,157 / 13,737 MiB
Load to listening, page cache warm	3.5 s

Research receipt

The hill climb’s judge decoded this model in every block, as the comparison model for Qwen3.8-27B-PN3.

Measure	Code	Prose	Scope
Decode, short context	26.26 tok/s	26.27 tok/s	256 greedy tokens after a cold prompt of about 600 tokens, on this engine commit
Decode at 32K	24.52 tok/s	24.54 tok/s	1,024 greedy tokens after a prompt filling the context, on the climb’s starting commit
Decode at 65K	22.92 tok/s	22.97 tok/s	the same

Its greedy text stayed bit-identical through the whole climb, and its speed was guarded within 3%.

Not measured yet

• 32K and 65K context on this engine commit.
• Speculative decoding (MTP) on this engine. The earlier kernel campaign ran it at about 28 tok/s on its own engine version; that mode is not in this recipe.