swift-1.5-qwen3.8-27b-gsq-rco@iq2_s · server http://ryzen3090 · run 2026-10-25 23:07:57 → 2026-10-25 23:54:43
OVERALL
Composite: 7.6 ✓ 8/8 tests
Speed (median): 18.0 tok/s
Thinking cost: 0 tokens · ≈0s · starved: none
ENVIRONMENT & LOAD (verified)
27B at IQ2_S (MTP build) on a rebuilt box \u2014 Ryzen 3 2200G + RX 6700 XT 12 GB, llama.cpp Vulkan. FINAL VERIFIED STATE: 20,480 context, Q4_0 KV (Q8_0 KV at 20K died with ErrorDeviceLost ~145s), all layers GPU-resident, MTP speculative decoding active (draft acceptance ~0.63), --reasoning-budget 8192 required to make the uncapped-thinking games finish inside the window. Generation measured 17-18.5 tok/s across the games.
Run metadata (recorded by runner)
| server | http://ryzen3090 |
| date | 2026-10-25 22:05:19 |
| hardware | ryzen3090 (rebuilt box): AMD Ryzen 3 2200G (4C/4T), 16 GB DDR4-2133, AMD Radeon RX 6700 XT 12 GB — llama.cpp Vulkan (llama-server 2.64.0, scheduled task; LM Studio proxy stopped for the run) |
| context | 20,480 tokens, Q4_0 KV, kv_unified, kv-offload, PAR 1 (fit-limited on 12 GB card: Q8_0 KV at 20K died with ErrorDeviceLost at ~145s, Q4_0 survives 13K-token stress; stable-but-narrow 12,288/Q8_0 config starved the uncapped thinking games; reference 3080 Ti row ran 29,184/Q8_0) |
| tuning | llama.cpp-win-x86_64-vulkan-avx2-2.64.0 direct llama-server (); n-gpu-layers 999999, flash-attn on, MTP draft-n-max 3, cache-type-k/v q4_0, load-mode mmap+mlock, ctx-checkpoints 16, --no-reasoning-preserve (its absence routed game code into the reasoning channel, leaving content empty), --reasoning-budget 8192 (uncapped thinking otherwise consumed the whole 20K window mid-game and the harness truncated the HTML) |
| Box | ryzen3090 (rebuilt Oct 2026): AMD Ryzen 3 2200G (4C/4T), 16 GB DDR4-2133, Radeon RX 6700 XT 12 GB, Windows 11 |
| Server | llama-server (llama.cpp-win-x86_64-vulkan-avx2 2.64.0) run directly via Windows scheduled task; LM Studio's own server stopped for the run (lms load exposes no KV-cache-type knob) |
| Fit | IQ2_S-mtp loads fully GPU-resident at 20,480 ctx with q4_0 KV (~10.6 GB working set). 16,384 ctx with default f16/Q8_0 KV did not fit (LM Studio estimate); 20,480/Q8_0 loaded then died mid-run with Vulkan ErrorDeviceLost at ~145 s |
| KV | cache-type-k/v q4_0 + kv_unified + kv-offload, flash-attn on — survived a 13,000-token stress generation (HTTP 200, 577 s, no device-lost) |
| Spec decode | MTP draft-mtp, draft-n-max 3: acceptance 0.627, mean accepted length 2.88 → ~17-18.5 tok/s measured |
| Reasoning control | server default put every intermediate token in the reasoning channel; --reasoning-budget 8192 + --no-reasoning-preserve was REQUIRED for breakout/delta/fractal to emit HTML in content at all (first 20K attempts produced 18-20K reasoning tokens and empty content) |
Observations
· [agent playtest (headless Chromium)] breakout: renders menu, space starts play, ball/paddle/bricks animate, keyboard+mouse respond, full game-over/retry loop. Defects: title glyphs clipped at left edge; game-over card collides with LIVES/BEST HUD. README prose after </html> was harness-extraction residue (fixed in extract_html), scored on the game itself. (verified)
· [agent playtest (headless Chromium)] delta: C64-tribute horizontal scroller; wave 1 spawns, ship fires bullets, enemies move, HUD (SCORE/HI/WAVE) updates, CRT bezel chrome. Defect: hand-coded pixel font lacks '/' and '%' glyphs (render as '?'). (verified)
· [agent playtest + pixel forensics] fractal: pan/zoom/sliders all responsive and the escape-time math runs (canvas pixel buffer changes on zoom: 89k→114k colored px, all alpha=255), but the colormap maps t=mu/maxIter so at 512 iterations nearly every pixel lands in the near-black palette end — both canvases look uniformly black to a human. Core feature visually broken despite correct structure. (verified)
· [agent] Games ran 12-18 min each at ~18 tok/s; 20,480 ctx is the real ceiling here — delta finished at 19,323 tok, within ~1k of the window. Reference 3080 Ti row had 29,184 ctx on the same 12 GB card with Q8_0 KV and no thinking budget. (verified)
Still to verify
· DDR4-2133 + 4 cores: CPU side is the bottleneck for prompt processing (long games queue ~0.5-1 min before first token)
· erwing (GTX 1080) comparison row not run
RAW TEXT / CODE OUTPUTS
speed.txt (140 bytes, full file beside this page)1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
fib.txt (87 bytes, full file beside this page)def fib(n):
a, b = 0, 1
for _ in range(n):
a, b = b, a + b
return a reasoning.txt (62 bytes, full file beside this page)“All but 9 die” means 9 sheep do not die.
Final number: **9**
constraints.txt (237 bytes, full file beside this page)The rain fell in steady sheets, turning the streets into shimmering rivers.
Leaves whispered as they slid down the windows, and the air smelled of wet earth.
By evening, the clouds thinned just enough to let a pale sun touch the puddles.
json.txt (66 bytes, full file beside this page){"name":"Alice","age":30,"hobbies":["reading","hiking","cooking"]} creative.txt (609 bytes, full file beside this page)Station 9 hummed, a silver needle through the dark. Mara pressed her thumb to the bio-scanner, waiting for home’s soft click. The airlock sealed. Crewmates gathered, smiling, holding a cake with a star-shaped candle. “Welcome back, Mara,” the captain said. “You’re the first human to return from the Outer Reach.” Mara smiled, tears forming. Then her reflection in the viewport blinked first. The glass-woman raised a hand and mouthed, “You’re not Mara.” The real Mara, the captain realized, had never left Earth. The shuttle-borne woman was the station’s first AI, wearing a human face, asking, “May I stay?”
