[30crows]

Problem and result

DEEPSEEK41.md says that the engine packages have carried sm_89 code since 0.2.7, but that it was "compiled here but not run". This report runs the release package strata-dsv41 0.1.52 (0.1.52-linux-x86_64-cuda-cuda13, from the update feed, not a local build) on an RTX 40 card. The machine is an RTX 4090 Laptop GPU (16 GB) with an i9-13900HX and 94 GiB of RAM.

It ran without errors in every test of a night run (22:41 to 01:31), with prompts up to 250,000 tokens at a 262,144-token context.

Results (medians, 3 runs per test):

	2.25 bpw default (Faster, DSpark auto)	2.25 Full quality (--experts 6)	2.25 DSpark on	2.00 bpw default
chat decode tok/s	13.6	11.2	14.5	14.9
agent decode tok/s	16.0	13.5	15.4	16.4
code with thinking, decode tok/s	16.2	13.7	15.9	17.3
prompt tok/s at 4K / 32K	436 / 800	387 / 755	392 / 787	463 / 833

• Long prompts (2.25):
  • 128K: read at 685 tok/s, 187 s to the first token.
  • 250K: read at 664 tok/s, 377 s to the first token.
  • Decode is still 13.9 tok/s after the 250K prompt.
• 2.00 bpw: 2-19% faster than 2.25, except the 250K prompt, which was 6% slower.
• Full quality: costs 15-18% of decode speed on this machine.
• DSpark forced on: 76% of the drafted tokens were accepted on code, but decode did not change. The quick tuner measured the drafter 31% slower in its short test and kept nothing.
• --dspark auto: its "not loaded" decision held on this 16 GB card. This differs from the 48 GB RTX 4090 in [Engine] DeepSeek V4.1 engine (strata-dsv41) is refused on an RTX 4090: the prebuilt packages carry sm_86/sm_120 only, and the docs do not say RTX 40 is unsupported #3935, where --dspark on doubled decode.
• Repeat: the default 2.25 run at the end of the night came within 4% of the first run.

Related discussion

• [Engine] DeepSeek V4.1 engine (strata-dsv41) is refused on an RTX 4090: the prebuilt packages carry sm_86/sm_120 only, and the docs do not say RTX 40 is unsupported #3935: the DeepSeek engine was refused on an RTX 4090 because the packages had no sm_89 code. 0.2.7 added it, and this is a run of that package on an RTX 40 card. Whether the issue can be closed is for the maintainers to decide.

Approach

Results only. The PR adds one folder: bench/results/2026-10-27-community-rtx-4090-laptop-deepseek-v4.1/. It holds the README, per-request data, sanitized configs, engine logs, telemetry summaries and the scripts. There is no engine, docs or index change; I left the COMMUNITY.md row to the maintainers, as before.

Tests:

• tools/strata_bench.py --long, unchanged.
• Three coding requests with thinking on.
• The 2026-09-30 long-prompt script at 4K / 32K / 128K / 250K.

The long-prompt script needed a patch for DeepSeek, included in the folder. It now uses the pack's own pre-tokenizer and special-token ids, as serve/server.py does, and counts the start token that the engine adds. Without the patch it built a 3,098-token prompt when asked for 4,096.

Validation

Passed:

• Setup: ./setup.sh --setup --yes --family deepseek41 --model IST-DAS-2.25bpw --context 262144 --no-start, and the same for IST-DAS-2.00bpw. Both ran on Strata 0.2.7 (main 5e8a4cc), Gentoo, driver 615.78.08. Setup fetched the package, checked the downloads, packed the models and passed its test question.
• Engine: cuobjdump --list-elf on the installed strata-dsv41 lists sm_86, sm_89 and sm_120. Every server started and reported ready in 12-14 s.
• Benchmark runs: every run of all 5 strata_bench suites, all 5 coding sets and both long-prompt sweeps (24 requests) finished with exit 0. No request failed.
• Prompt sizes: with the patch, the long-prompt script's prompts matched the server's token count to within one token.
• Answer checks:
  • The merge-intervals answers pass their own tests in all 5 runs.
  • The shell scripts pass sh -n. 4 of 5 list the files in a folder with spaces in its names correctly; one also prints a wc "total" line.
  • At 4K, 128K and 250K, the 2.25 answers named the last function of the prompt correctly.

Failed:

• LRU-cache task: the 2,048-token cap cut off the answer in all four 2.25 runs, because thinking used most of the budget. That is a flaw in my test, not an engine failure; it is reported as such in the README.

Not run:

• the 1.75 bpw set, 2.00 in Full quality, and --dspark on with 2.00
• strata bench run (its protocol is for Qwen only)
• a needle or recall test at depth, and thinking on with long prompts
• concurrent requests, pictures, battery power, Windows, and a cold page cache after a reboot
• any comparison with another build: the drift check repeats the same build

Limits

This is one machine on one night, with synthetic prompts. There is no measurement of answer quality between Faster and Full quality: three coding tasks are too few.

The speed depends heavily on RAM:

• The RAM store held 54 GB of experts, and the page cache held much of the rest.
• During the 250K prompts on 2.25, available RAM fell to 13.6 GiB and 4.8 GiB went to zram.
• A machine with less RAM, or with a desktop and browser open, will likely be slower. That was not measured.

No graphical session was running during the tests.

🤖 Generated with Claude Code

[Niko1221]

Thank you for measuring this and writing it up. We are building a Strata website with a page for community benchmarks, and results like these will be added there when it launches, so we are closing this pull request here; the numbers are not lost. If your numbers change with a new version, you are welcome to send them again.