Repository navigation
bench: RX 6700 XT 12 GB + Ryzen 5 5600X — Qwen3.6/LFM2.5/Tiel/Gemma suites at 32K+64K, draft A/B, practical appendix - #71
Merged
Conversation
…uites at 32K+64K Nine official strata_bench.py runs (suite 1, --runs 3): Qwen3.6-35B UD-Q4_K_XL, LFM2.5-24B UD-Q4_K_XL, Tiel Coder UD-Q3_K_XL and Gemma-4 26B-A4B QAT UD-Q4_K_XL at 32768 and 65536 contexts, reserve 1024, plus a same-machine Qwen re-run at e97c8aa against the c54f0ea originals (no measurable change) and a Gemma draft-pairing A/B (QAT-paired draft +7.2% chat decode over the repo Q8_0 draft, prompt unchanged). Includes per-request results.json, provenance with model SHA-256s, and a separately labeled practical-evaluation appendix (6h/483-check soak, blind agent screens) that makes no claim as part of the speed suite. Results only; no engine or tool code changes.
Collaborator
|
Thanks, merged, and it has a line in COMMUNITY_BENCHMARKS.md right next to the first one for the RX 6700 XT. GPU tiles: setup wrote backend: hip for strata-poly only on Windows, fixed in v0.1.45-poly.3, running setup again will rewrite the config. Qwen3.6 Q6_K_XL illegal memory access on a 4.8K prompt, can you open an issue with the engine log? |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds official
tools/strata_bench.pyresults from an RX 6700 XT 12 GB (gfx1031), Ryzen 5 5600X and 32 GB RAM: Qwen3.6-35B UD-Q4_K_XL, LFM2.5-24B UD-Q4_K_XL, Tiel Coder UD-Q3_K_XL and Gemma-4 26B-A4B QAT UD-Q4_K_XL at 32K and 64K contexts (nine runs, suite 1,--runs 3), a same-machine Qwen re-run ate97c8aaagainst thec54f0eaforiginals, and a Gemma draft-pairing A/B. Every run has its per-requestresults.json;provenance.jsoncarries model SHA-256s (all verified same-day), settings, and the reproduce commands.Measured source: engine 0.1.45 built locally for gfx1031 (HIP, AMD's wheels); original suites at c54f0ea, Qwen re-runs at e97c8aa (the delta commits are CUDA-only, engine binary identical). Every profile reserves 1024 MiB VRAM; no memory fault in any of the nine runs including 64K contexts — consistent with the reserve-related fault discussed in #52.
Observations the report documents beyond the tables:
product_name, so the tool reports the generic "AMD Radeon" and the generated folder names sayunknown; folders were renamed torx6700xt-5600xhere (inner generated names untouched) and the card verified externally (PCI 1002:73DF, 12272 MiB). Same disclosure style as bench: RX 6700 XT 12 GB official PolyStrata results #52'scpu-named folders.PRACTICAL-EVALUATION.mdis a separately labeled appendix (6 h / 483-check soak, blind agent screens) kept outside the speed claims.Disclosures: the Oct-20 Qwen and LFM suites ran while a large download wrote to a second disk (the 4K/16K medians match the idle re-run, so no visible effect, but recorded); LFM's 32K agent stage did not complete (its tool-call format, same gap as the 10-19 report) and is listed as a skipped test; the LFM2.5 GGUF was deleted in a post-campaign disk cleanup, so its byte count is marked unavailable (SHA-256 recorded same-day).
Validation: files copied byte-for-byte from the campaign artifacts; report folder is 392 KB; no credentials or host names (the runner's own placeholder pass is intact). This PR contains only results and their metadata — no engine, setup, or benchmark-tool code changes, and no general accuracy claim.