#2096 "--prefill", "auto" with value 1,024 resulted in significant t/sec drop

Descrição

With "--prefill",  "auto" if engine decide to use a value like 1024 token generation will drop from 35 to 12 t/sec (from 26 to 4 t/sec pre Strata v0.1.43) (5,309 tokens prompt or most of pi sessions resume) .  Using "--prefill", "auto:512" will do.
Maybe its worth a note or calibrate entry.

# Community benchmark on NVIDIA GeForce RTX 3060 Ti

Measured on 2026-10-11 with `tools/strata_bench.py` (suite 1, 3 runs per test, medians with ranges). Speeds are the engine's own timings; first-token times are measured by the client over HTTP. Speed only: no correctness check was run.

## Hardware and software

- GPU and VRAM: NVIDIA GeForce RTX 3060 Ti, 8 GiB VRAM
- CPU: AMD Ryzen 7 2700X Eight-Core Processor (8 cores, 16 threads); installed RAM: 63.9 GiB; storage: not measured
- OS: Windows 11 (10.0.26200); GPU driver: 591.86; CUDA: 13.1
- Strata: project 0.1.43, commit 7695859a; engine 0.1.43
- Background workloads and power limits: not recorded

## Model and configuration

- Model and quantization: qwen3.8-flash-next-iq3_xxs
- Context: 131072 tokens; parallel slots: 1
- Settings: the install's own configuration, unchanged; thinking off for the suite; temperature 0

```text
strata.exe --pack iq3_xxs --native Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf --ple-gguf Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00002-of-00002.gguf --expert-profile expert-profile.bin --expert-cache auto --prefill auto --spec 4 --mtp rt --max-context 131072 --kv int8 --kv-resident 32768 --pcie-frac 0.36 --spec-min-p 0.70
```

## Method

Run by `python tools/strata_bench.py` against the HTTP API of a Strata server (started by the benchmark from the install configuration); model loading is not timed. One warm-up chat and a 1K prompt are discarded first. Every prompt starts with a unique tag, so the conversation cache cannot help the fresh-prompt tests.

- short chat: the three questions of `tools/ab_engine.py`, up to 100 new tokens, 3 runs
- 4K / 16K prompts: generated word lists sized by a calibration request, 24 new tokens, 3 fresh prompts each
- agent session: system prompt and four tools, then 5 turns that each add about 1500 words of a project file and continue from the model's own earlier answers (the way agent clients resend a conversation), 64 new tokens per turn, 3 sessions; the cache numbers come from the answer's usage (`cached_tokens`) and timings (`prompt_n`)
- concurrency (when run): N requests sent at the same moment, total tokens over the wall time of the burst, latency over all requests

## Results

| Configuration | Actual prompt tokens | Reused tokens | Generated tokens | Runs | Prompt tok/s median and range | Decode tok/s median and range | TTFT seconds median and range |
| --- | ---: | ---: | ---: | ---: | --- | --- | --- |
| short chat | 46 (43-51) | 0 | 100 (100-100) | 9 | not measured | 37.9 (32.5-44.9) | 1.08 (0.97-1.16) |
| 4K prompt, fresh | 4019 (4019-4019) | 0 (0-0) | 24 cap | 3 | 374.9 (372.4-375.5) | not measured | 10.76 (10.74-10.84) |
| 16K prompt, fresh | 16058 (16058-16058) | 0 (0-0) | 24 cap | 3 | 370.5 (369.6-371.0) | not measured | 43.39 (43.34-43.50) |
| agent session, follow-up turns | 14462 (14452-14515) at the last turn | see below | up to 64 | 3 | see below | 35.2 (21.2-38.4) | 16.50 (7.96-18.88) |

Ranges are the lowest and highest of the runs; `Runs` counts requests that returned a token.

### Agent session (conversation cache)

Cache hit over the follow-up turns: 68.8 % (68.7-68.9); earlier-turn tokens read again per session: 3 (2-74) (earlier turns that the cache did not hold; the 13633 (13633-13635) new tokens of the files are read once); first turn (nothing cached) first token: 3.89 s (3.87-3.90).

| Turn (run 1) | Prompt tokens | Reused | Read | Prompt tok/s | First token (s) | Decode tok/s |
| ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| 0 | 609 | 0 | 609 | 159.0 | 3.868 | 34.2 |
| 1 | 3730 | 647 | 3083 | 185.2 | 16.694 | 37.8 |
| 2 | 5263 | 3794 | 1469 | 185.0 | 7.984 | 32.4 |
| 3 | 8707 | 5327 | 3380 | 185.4 | 18.291 | 35.6 |
| 4 | 11611 | 8751 | 2860 | 178.0 | 16.119 | 21.2 |
| 5 | 14462 | 11619 | 2843 | 172.9 | 16.499 | 29.8 |

Per-request data: `results.json`. Engine log lines from the start: `engine-log-excerpt.txt` (when this run started the server).

## Correctness and limitations

Not measured: this is a speed suite. The conversation uses the model's own replies, so token counts differ a little between runs. Machine details come from the server, `nvidia-smi`/`rocm-smi` and the OS; background load and power limits are not recorded. Total time of the suite: 472 s.