bench: RX 6700 XT 12 GB + Ryzen 5 5600X — Qwen3.6/LFM2.5/Tiel/Gemma suites at 32K+64K, draft A/B, practical appendix

bench/results/2026-10-22-community-rx6700xt-5600x-polystrata/PRACTICAL-EVALUATION.md

@@ -0,0 +1,70 @@
+# Practical evaluation, separately labeled (not the speed suite)
+
+This file is **not** part of the official benchmark suite and is not
+comparable to it. It records what this machine's campaign measured beyond
+speed, for readers deciding between models on 12 GB RDNA2 hardware.
+Methodology and caveats per section; no general accuracy claim is made.
+
+## A. Reliability: 6-hour day-desktop soak (2026-10-21)
+
+The ctx32768 Qwen server (the daily driver) served **483 health checks
+and traffic over 6 h under normal desktop use** (browser, editor) with
+**zero failures**, GPU telemetry live (`backend: hip`). Summary:
+`soak-summary.json` next to this file. The summary's `max_server_rss_kB`
+tracks the Python wrapper process, not the engine (engine RSS is tens of
+GB, mostly reclaimable page cache) — ignore that field. This is a single
+soak, not a statistical claim.
+
+## B. Blind agent screens (12-task easy set, 2026-10-20)
+
+An agent client (Nanobot, thinking on, temp 0.6, 20-tool-call cap)
+completed a 12-task office/research/synthesis suite twice, once per
+model, blind: rows and workspaces merged under symlinked names, graded
+against fixture-derived ground truths, grade reasons sealed by SHA before
+unblinding.
+
+| | Total | coding | research | synthesis | drafting |
+| --- | --- | --- | --- | --- | --- |
+| Qwen3.6 Q4_K_XL (poly) | **10/12** | 3/3 | 2/3 | **2/3** | 3/3 |
+| LFM2.5-24B Q4_K_XL (poly) | 8/12 | 3/3 | 2/3 | 0/3 | 3/3 |
+
+Both failed the same two research tasks (a procurement-revision conflict
+and a count-preservation check — the failure class: a fact wired to the
+wrong slot). LFM's synthesis failures included invented measurement
+values. Caveats: 1 repetition per task; same fixtures for both arms;
+graders saw neither model identity nor arm.
+
+## C. Hard-set screen (6 discriminating tasks, 2026-10-21/22)
+
+Six tasks built to discriminate (generator-derived ground truths: fixed
+totals, ID sets, a windowed timeline, a revision conflict, a sandboxed
+code task), screened blind the same way. The easy set's models all
+saturated it except Gemma:
+
+| Model | hard set | probe | gate |
+| --- | --- | --- | --- |
+| Qwen3.6 Q4_K_XL (driver) | **6/6** | 4/4 | 12/12 |
+| Tiel Coder Q3_K_XL | **6/6** | 4/4 | 11/12 |
+| Gemma-4 QAT (speed arm) | **3/6** | 4/4 | 10/12 |
+
+Gemma's failures: an arithmetic total (211 vs 216), an unwritten answer
+file, the sandbox code task, and a recovery sequence (twice). Caveats:
+1 repetition, 6 tasks, one failure was a degenerate generation; four
+independent signals agree on direction.
+
+## D. Same-day llama.cpp controls (methodology check)
+
+The matched harness (identical prompts across engines) ran same-day
+llama.cpp controls on this box: the current llama.cpp default
+(IQ3_XXS arm) measured 60.2/66.1 tok/s synth/prose against
+strata-poly's Qwen 75.4/73.5 on the same prompts; a fresh control the
+next day drifted ~2%, bounding day-to-day noise. These are custom
+matched probes, not suite numbers, and are reported only to show the
+speed ordering is not a harness artifact.
+
+## What was not tested
+
+Three repetitions (screens are 1x); other agent clients; other
+quantizations of the same models on this card; the draft/spec state
+separately (engine default); `--tail-skip 6`; hipBLASLt tuning for
+gfx1031; concurrency beyond the suite's 2-client stage.

bench/results/2026-10-22-community-rx6700xt-5600x-polystrata/README.md

@@ -0,0 +1,160 @@
+# Community benchmark on AMD Radeon RX 6700 XT 12 GB (gfx1031), Ryzen 5 5600X
+
+Measured 2026-10-20 to 2026-10-22 by mkw432. The official suite on four
+models (Qwen3.6-35B UD-Q4_K_XL, LFM2.5-24B UD-Q4_K_XL, Tiel Coder
+UD-Q3_K_XL, Gemma-4-26B-A4B QAT UD-Q4_K_XL) at 32K and 64K contexts, plus
+a same-machine re-run of the Qwen suites at current main (e97c8aa) against
+the original 0.1.45-c54f0eaf runs, a Gemma draft-pairing A/B, and a
+separately labeled practical evaluation appendix. This is the second
+report from an RX 6700 XT (after
+[2026-10-19-community-rx6700xt-polystrata](../2026-10-19-community-rx6700xt-polystrata/README.md),
+a Ryzen 5 3600 machine); the Qwen3.6 and LFM2.5 numbers reproduce that
+report's rows within a few percent on different RAM and CPU.
+
+Main limitation: single box, three days, one repetition of the optional
+correctness screens (not three); the quality appendix is not a benchmark
+suite and makes no claim beyond this hardware.
+
+## Hardware and software
+
+- GPU and VRAM: AMD Radeon RX 6700 XT 12 GB (gfx1031, PCI 1002:73DF,
+  display-attached); 12272 MiB VRAM reported by amdgpu sysfs; PCIe gen4 x16
+- CPU: AMD Ryzen 5 5600X (6 cores / 12 threads); installed RAM: 32 GB
+  (31.2 GiB reported); storage: NVMe SSD (models on a second internal
+  NVMe, ext4)
+- OS: Ubuntu, Linux 7.0.0-38-generic; GPU driver: amdgpu (in-kernel);
+  ROCm: AMD's Python wheels inside the project venv (no system rocm-smi)
+- PolyStrata: release engine 0.1.45 built locally for gfx1031 (HIP); the
+  Oct-20/21 runs used project commit c54f0eaf, the Oct-22 re-runs used
+  e97c8aa (the engine binary is unchanged; the newer commits touch CUDA
+  kernels and docs only)
+- Background workloads: the Oct-20 LFM and Qwen suites ran while a large
+  model download wrote to a second disk (labeled; see limitations); other
+  runs idle-ish desktop; no power limits set
+
+## Models and configurations
+
+All GGUFs SHA-256 verified the same day (hashes in
+[provenance.json](provenance.json)):
+
+| Model | File | Repository |
+| --- | --- | --- |
+| Qwen3.6-35B-A3B UD-Q4_K_XL | `Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf` | unsloth/Qwen3.6-35B-A3B-MTP-GGUF |
+| LFM2.5-24B-A2B UD-Q4_K_XL | `LFM2.5-24B-A2B-UD-Q4_K_XL.gguf` | unsloth/LFM2.5-24B-A2B-GGUF |
+| Tiel Coder 35B-A3B MTP UD-Q3_K_XL | `Tiel-Coder-35B-A3B-MTP-UD-Q3_K_XL.gguf` | peculiar-ragdoll/Tiel-Coder-35B-A3B-GGUF-MTP |
+| Gemma-4-26B-A4B it QAT UD-Q4_K_XL | `gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf` | unsloth/gemma-4-26B-A4B-it-GGUF |
+
+Common launch settings (the community-parity profile): `--kv-whole`
+`--conversation-cache-mib 0` `--vram-reserve-mib 1024`, `backend: hip` in
+the run config. LFM adds `--prompt-batch 2048` and its setup-written
+expert profile; Tiel adds `--threads 8` and its profile; Gemma adds
+`--threads 8` and a draft model (see the A/B below); Qwen adds
+`--tail-skip 0`. Contexts 32768 and 65536.
+
+```text
+strata-poly --model Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf --max-context 32768 --tail-skip 0 --kv-whole --conversation-cache-mib 0 --vram-reserve-mib 1024
+```
+
+## Method
+
+`python tools/strata_bench.py --runs 3` through the HTTP API of a running
+server (the Oct-22 64K re-run started the server itself; its
+`engine-log-excerpt.txt` is included). Suite 1 as the guide describes:
+thinking off, temperature 0, unique prompt tags, warm-up discarded, 3 runs
+per test, medians and ranges. No other engine on the card during any run
+(one engine at a time; the daily-driver server was stopped for the re-run
+that started its own).
+
+## Results
+
+Full tables in each run folder's README; medians:
+
+| Profile | Stage | Date / commit | Short chat decode | 4K prompt | 16K prompt | 32K prompt |
+| --- | --- | --- | ---: | ---: | ---: | ---: |
+| Qwen3.6 Q4_K_XL, ctx32768 | official-standard | 10-20 / c54f0eaf | 79.7 | 1053.1 | 999.7 | — |
+| Qwen3.6 Q4_K_XL, ctx32768 | official-standard | 10-22 / e97c8aa | 82.2 | 1016.7 | 999.7 | — |
+| Qwen3.6 Q4_K_XL, ctx65536 | official-long | 10-20 / c54f0eaf | 81.4 | 1000.9 | 997.5 | 775.0 |
+| Qwen3.6 Q4_K_XL, ctx65536 | official-long | 10-22 / e97c8aa | 84.5 | 998.2 | 993.7 | 773.2 |
+| LFM2.5 Q4_K_XL, ctx32768 | official-standard | 10-20 / c54f0eaf | 88.5 | 1778.8 | 1414.0 | — |
+| LFM2.5 Q4_K_XL, ctx65536 | official-long | 10-20 / c54f0eaf | 86.4 | 1616.4 | 1407.0 | 1043.7 |
+| Tiel Coder Q3_K_XL, ctx32768 | official-standard | 10-21 / c54f0eaf | 73.1 | 1241.5 | 1031.6 | — |
+| Gemma-4 QAT Q4_K_XL, ctx32768, QAT-paired draft | official-standard | 10-21 / c54f0eaf | 87.7 | 1172.0 | 1009.3 | — |
+| Gemma-4 QAT Q4_K_XL, ctx32768, repo Q8_0 draft | official-control | 10-21 / c54f0eaf | 81.4 | 1186.5 | 1024.3 | — |
+
+All tok/s, engine-timed, medians of 3 runs (chat: median of 9 requests).
+Agent-session stages and per-turn tables are in each run README; LFM's
+32K agent stage did not complete (its Python-list tool-call format is
+rejected by the stage, same gap as the 10-19 report).
+
+Observations:
+
+1. **Reproduction of the 10-19 RX 6700 XT report** (different CPU and
+   RAM, same card): Qwen short-chat 79.7-84.5 vs their Q4_K_M control
+   rows in the 80s class; LFM 86-89 matches the reported LFM2.5 row
+   class; all four of our Qwen/LFM 32K and 64K rows land within a few
+   percent of that report.
+2. **Version delta on the same box, same prompts:** c54f0eaf → e97c8aa
+   changed nothing measurable on HIP (32K prompt 999.7 = 999.7; 32K-row
+   long 775.0 vs 773.2; chat deltas inside the run-to-run range).
+3. **Gemma draft pairing:** swapping the QAT-paired 252 MB draft for the
+   repo's Q8_0 draft (461 MB, the standard-model pairing) costs
+   −7.2% short-chat decode (87.7 → 81.4) with prompt throughput
+   unchanged (control ✓). Same model, same settings, only `--draft-model`
+   differs. QAT users should keep the QAT-paired draft.
+4. **Concurrent clients:** 2 clients total 65.0 tok/s on Qwen at ctx65536
+   (below one client; `--batch` was not enabled on these servers).
+
+## Correctness and limitations
+
+See [PRACTICAL-EVALUATION.md](PRACTICAL-EVALUATION.md) for a separately
+labeled, non-speed evaluation (blind agent screens and a day-desktop
+soak) — it is not part of the official suite and no general accuracy
+claim is made from it.
+
+- The suite is a speed check; correctness is not measured by it.
+- Oct-20 LFM and Qwen 32K suites ran while a large model download wrote
+  to a second disk; prompt rates at 4K/16K match the idle re-run
+  (1053.1/999.7 vs 1016.7/999.7), so the interference is not visible in
+  the reported medians, but the condition is disclosed.
+- Single machine; three days; day-to-day drift on this box measured at
+  ~1-2% on a same-configuration llama.cpp control.
+- GPU naming: this kernel's amdgpu sysfs exposes no `product_name`, so
+  the server and `tools/strata_bench.py` (which now read sysfs) report
+  the generic "AMD Radeon", and the report folder names say `unknown`
+  (renamed to rx6700xt here; original generated names kept inside).
+  The card is verified externally as an RX 6700 XT: PCI ID 1002:73DF,
+  12272 MiB VRAM, gfx1031 engine build.
+- No power limits, no clock capture; the card is display-attached.
+- Failures: none in the official suites; one skipped test (the LFM 32K
+  agent stage, above).
+
+## Notes for RX 6700 XT / RDNA2 / 32 GB users
+
+Collected during the same campaign; all on this hardware only:
+
+- `--vram-reserve-mib 1024` matters: with the default reserve the expert
+  cache grows into the last free MiB and a later request can fault in
+  ROCm's buffer fill (the failure reported in #52's review thread). All
+  rows here used an explicit reserve; zero memory faults across nine
+  suite runs including 64K contexts.
+- Qwen3.6 **Q6_K_XL is not usable on a 12/32 GB box**: native Strata
+  cannot place it (32.6 GB file, placement timed out at >900 s even with
+  all background load stripped — nominal fit is not placement slack),
+  and strata-poly loads it (19 s, 2,344/10,496 experts resident) but
+  faults with `illegal memory access` during a ~4.8K-token prefill.
+  Short prompts generate fine (~25-37 tok/s). Q4_K_XL remains the pick.
+- Qwen3.6 MTP lives **inside** the GGUF (~768 MB of the file); files
+  from the plain `Qwen3.6-35B-A3B-GGUF` repo lack
+  `nextn_predict_layers` and the engine refuses them — use the
+  `Qwen3.6-35B-A3B-MTP-GGUF` repo.
+- The web UI's GPU tiles need `"backend": "hip"` in the run config;
+  setup-written configs lacked it and the tiles fell back to NVML
+  ("not readable").
+- No hipBLASLt tuning table ships for gfx1031; low-RAM-mode streaming
+  models (IQ2/IQ1-class Flash-Next files) show prompt throughput 4-10x
+  below the Q4-class rows here. `tools/hip/tune_hipblaslt` on this card
+  is untested territory.
+- Native Strata pins ~19-29 GB host RAM at load (32 GB boxes: bench it
+  solo or watch the kernel OOM-killer); strata-poly's experts are
+  page-cache based and tolerate background load — that difference is
+  why every tabled number here is a poly number.

bench/results/2026-10-22-community-rx6700xt-5600x-polystrata/index.json

@@ -0,0 +1,73 @@
+{
+  "source_commits": {
+    "original_runs": "c54f0eaf6200a9207f13f900a5fc772d437d6f7a",
+    "rerun_runs": "e97c8aa47ca2c01025accdd32d07a01536d794ce"
+  },
+  "runs": [
+    {
+      "profile": "gemma4-qat-ud-q4_k_xl-qatdraft-reserve1024-ctx32768",
+      "stage": "official-standard",
+      "status": "completed",
+      "reports": 1,
+      "path": "runs/gemma4-qat-ud-q4_k_xl-qatdraft-reserve1024-ctx32768/official-standard"
+    },
+    {
+      "profile": "gemma4-qat-ud-q4_k_xl-repodraft-reserve1024-ctx32768",
+      "stage": "official-control",
+      "status": "completed",
+      "reports": 1,
+      "path": "runs/gemma4-qat-ud-q4_k_xl-repodraft-reserve1024-ctx32768/official-control"
+    },
+    {
+      "profile": "lfm25-ud-q4_k_xl-reserve1024-ctx32768",
+      "stage": "official-standard",
+      "status": "completed",
+      "reports": 1,
+      "path": "runs/lfm25-ud-q4_k_xl-reserve1024-ctx32768/official-standard"
+    },
+    {
+      "profile": "lfm25-ud-q4_k_xl-reserve1024-ctx65536",
+      "stage": "official-long",
+      "status": "completed",
+      "reports": 1,
+      "path": "runs/lfm25-ud-q4_k_xl-reserve1024-ctx65536/official-long"
+    },
+    {
+      "profile": "qwen36-ud-q4_k_xl-reserve1024-ctx32768",
+      "stage": "official-standard",
+      "status": "completed",
+      "reports": 1,
+      "path": "runs/qwen36-ud-q4_k_xl-reserve1024-ctx32768/official-standard"
+    },
+    {
+      "profile": "qwen36-ud-q4_k_xl-reserve1024-ctx32768-rerun-e97c8aa",
+      "stage": "official-standard",
+      "status": "completed",
+      "reports": 1,
+      "path": "runs/qwen36-ud-q4_k_xl-reserve1024-ctx32768-rerun-e97c8aa/official-standard"
+    },
+    {
+      "profile": "qwen36-ud-q4_k_xl-reserve1024-ctx65536",
+      "stage": "official-long",
+      "status": "completed",
+      "reports": 1,
+      "path": "runs/qwen36-ud-q4_k_xl-reserve1024-ctx65536/official-long"
+    },
+    {
+      "profile": "qwen36-ud-q4_k_xl-reserve1024-ctx65536-rerun-e97c8aa",
+      "stage": "official-long",
+      "status": "completed",
+      "reports": 1,
+      "path": "runs/qwen36-ud-q4_k_xl-reserve1024-ctx65536-rerun-e97c8aa/official-long"
+    },
+    {
+      "profile": "tiel-ud-q3_k_xl-reserve1024-ctx32768",
+      "stage": "official-standard",
+      "status": "completed",
+      "reports": 1,
+      "path": "runs/tiel-ud-q3_k_xl-reserve1024-ctx32768/official-standard"
+    }
+  ],
+  "tool_sha256": "bd65126bed1f8e5ff3648f5d65c230ea7566219a443272115c1d622735d02be5",
+  "tool_sha_note": "sha256 of tools/strata_bench.py at e97c8aa (rerun runs); original runs used the same file at c54f0ea (runner-side AMD naming fix present in both)"
+}

bench/results/2026-10-22-community-rx6700xt-5600x-polystrata/provenance.json

@@ -0,0 +1,82 @@
+{
+  "source_commit_note": "original suites (Oct-20/21) at c54f0eaf6200a9207f13f900a5fc772d437d6f7a; Qwen re-runs (Oct-22) at e97c8aa47ca2c01025accdd32d07a01536d794ce; the engine binary (0.1.45, gfx1031, built from c54f0eaf) is identical across both \u2014 the delta commits are CUDA-only",
+  "build": {
+    "source": "local-hip",
+    "backend": "hip",
+    "version": "0.1.45",
+    "archs": [
+      "gfx1031"
+    ],
+    "vision": "none",
+    "rocm": "AMD Python wheels in the project venv (rocm SDK gfx103X libs)"
+  },
+  "models": [
+    {
+      "file": "Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf",
+      "repo": "unsloth/Qwen3.6-35B-A3B-MTP-GGUF",
+      "bytes": 22853663008,
+      "sha256": "55983c5a75a1ab969824077b3bb3de4146e82a9234072b48ad4e8f92ad3fe9f1"
+    },
+    {
+      "file": "LFM2.5-24B-A2B-UD-Q4_K_XL.gguf",
+      "repo": "unsloth/LFM2.5-24B-A2B-GGUF",
+      "sha256": "a235d23638cc5549cb1ff61da7522c8a7e97f056f27093462a1e56f35aba91d8",
+      "bytes": "not recorded (file deleted in post-campaign cleanup; sha256 verified same day)",
+      "sha_note": "sha256 recorded same day (2026-10-20) by the campaign harness before the file was deleted in a post-campaign disk cleanup"
+    },
+    {
+      "file": "Tiel-Coder-35B-A3B-MTP-UD-Q3_K_XL.gguf",
+      "repo": "peculiar-ragdoll/Tiel-Coder-35B-A3B-GGUF-MTP",
+      "bytes": 17234935648,
+      "sha256": "46b8fbeba6e888c9a101e49cc717baa9bb8dfa0d203ec7b3185288159543ec14"
+    },
+    {
+      "file": "gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf",
+      "repo": "unsloth/gemma-4-26B-A4B-it-GGUF",
+      "sha256": "a7c5bc715f5ff8e99a3e8901ce7d2b42b402c669bf24f7c5250747633d0f5891",
+      "bytes": 14249047104
+    },
+    {
+      "file": "mtp-gemma-4-26B-A4B-it.gguf (QAT-paired draft, 252 MB)",
+      "repo": "unsloth/gemma-4-26B-A4B-it-GGUF",
+      "sha256": "7272d97595f0d4c74bd7b623492b7dbdaafd8b7c72f329a8270ba4eca68f768a",
+      "bytes": 251939328
+    },
+    {
+      "file": "MTP/mtp-gemma-4-26B-A4B-it-Q8_0.gguf (repo draft, 461 MB)",
+      "repo": "unsloth/gemma-4-26B-A4B-it-GGUF",
+      "sha256": "6326fb9f5e487aa8dcdd313a091e3c67724cb2a666ec3b7d2895b5b26d93ed1b",
+      "bytes": 461766816
+    }
+  ],
+  "settings_common": [
+    "--kv-whole",
+    "--conversation-cache-mib 0",
+    "--vram-reserve-mib 1024",
+    "backend hip"
+  ],
+  "settings_extra": {
+    "qwen36": [
+      "--tail-skip 0"
+    ],
+    "lfm25": [
+      "--prompt-batch 2048",
+      "setup-written expert-profile.txt"
+    ],
+    "tiel": [
+      "--threads 8",
+      "setup-written expert-profile.txt"
+    ],
+    "gemma": [
+      "--threads 8",
+      "--draft-model <see models>"
+    ]
+  },
+  "gpu_identity": {
+    "reported_by_tool": "AMD Radeon (generic; no amdgpu product_name sysfs in this kernel)",
+    "verified_externally": "AMD Radeon RX 6700 XT, PCI 1002:73DF, 12272 MiB VRAM, gfx1031",
+    "folder_rename_note": "report folders renamed *-unknown-* -> rx6700xt-5600x; the inner generated names and READMEs are unmodified"
+  },
+  "reproduce": "python tools/strata_bench.py --url http://127.0.0.1:8091 --config <config> --runs 3 [--long --clients 2 for the 64K stages]",
+  "campaign_reference": "Full campaign record (matched probes, llama.cpp controls, quality screens, incidents) kept in the contributor's local campaign notes; the distilled version is in this folder's README.md and PRACTICAL-EVALUATION.md"
+}