swift-1.5-qwen3.8-27b-gsq-rco@iq2_s · server http://strata · run 2026-10-05 22:53:37 → 2026-10-05 22:55:47
OVERALL
Composite: 8.8 ✓ 8/8 tests
Speed (median): 27.2 tok/s
Thinking cost: 69,329 tokens · ≈2577s · starved: none
ENVIRONMENT & LOAD (verified)
27B at IQ2_S (MTP build) on a 12 GB 3080 Ti — FINAL VERIFIED STATE: single slot, 29,184 context, Q8_0 KV cache, all 63 layers forced to GPU (n_gpu_layers=63). Measured 10,725 / 12,288 MiB on the card → it FITS fully on the GPU, no CPU spill. (Earlier notes claiming 'spills to CPU' were WRONG — they assumed f16 KV and ignored the user's Q8_0 KV setting.) Every number below was read from the live machine (nvidia-smi / lms ps / llama.cpp log) or is an explicit observation.
| Model file actually loaded | Swift-1.5-Qwen3.8-27B-GSQ-RCO-IQ2_S-mtp.gguf — the MTP (multi-token-prediction) build, 8.95 GiB on disk. (NOT the plain iq2_s.) |
| Context | 29,184 tokens (kv_unified) |
| Parallel slots | 1 (fixed — was 2, which doubled the KV cache and caused two concurrent llama-server processes) |
| KV cache type | Q8_0 (user setting) — halves the f16 KV memory |
| GPU layer split | n_gpu_layers = 63 (ALL layers on GPU, hard-forced by user; llama.cpp log: 'failed to fit params to free device memory: n_gpu_layers already set by user to 63, abort') |
| GPU usage (measured, idle) | 10,725 / 12,288 MiB — FITS the 12 GB card with ~1.5 GiB headroom; 27 GiB system RAM free → no CPU offload at this config |
| GPU util / power (idle) | ~0% util, ~11 W (idle sample, not under load) |
| Concurrent load | llama-server PID 2040 on the Arc B70 is still running the benchmark's OWN reference model (Qwen3.8-27B Q4_K_M) on — same box. Does not touch the 3080 but shares CPU/RAM/PCIe. |
| Speed — earlier runs | 16K ≈ 31.5 tok/s (speed test) · 48K (Q8 KV) ≈ 4.6 tok/s · 64K ≈ 3.9 tok/s · games @16K ≈ 23 tok/s. NOTE: the 48K/64K slow runs predate the Q8-KV + single-slot config, so they are NOT comparable to the current setup. |
| Speed — current config | NOT yet measured at 29K + PAR 1 + Q8 KV. Needs a fresh speed test to anchor the record. |
Observations
· [user] GPU and CPU both finished loading at the same time → model spans both. (TRUE for the earlier PARALLEL-2 + f16-KV config (2 KV caches + f16 pushed past the card). RESOLVED after reload to PAR 1 + Q8 KV + forced 63 GPU layers: now fits on the GPU alone (10.7 GiB).)
· [user] 'you keep forgetting i have quantized the kv cache.' (ACKNOWLEDGED — Q8_0 KV halves the KV memory and is what makes 29K fit. Earlier spill math was wrong for ignoring it.)
· [user] 'it looks like you have loaded the model twice.' (CONFIRMED — was PARALLEL 2 (two llama-server processes, two KV caches). User ejected and reloaded to PARALLEL 1. Now one 3080 server (PID 6692) + the Arc reference (PID 2040).)
· [assistant] 0-CPU sample of the 3080 llama-server was taken while idle — not evidence of offload either way. (superseded by the log + nvidia-smi fit above)
Still to verify
· Fresh speed test at the FINAL config (29K + PAR 1 + Q8 KV) not yet taken — this is the clean number to anchor the record and compare against the earlier starved game runs.
· Whether the games (breakout/delta/fractal) succeed at this config — breakout failed (starved, 59 content tokens, no canvas) at the 16K/2-slot config; worth one clean re-test now that it's fully on the GPU.
· STATUS 2026-10-12: box strata retired (reloaded as strata Ubuntu) — these retests can no longer be run; record frozen as of retirement.
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 (152 bytes, full file beside this page)def fib(n):
if n < 0:
raise ValueError("n must be non-negative")
a, b = 0, 1
for _ in range(n):
a, b = b, a + b
return a reasoning.txt (78 bytes, full file beside this page)“All but 9 die” means 9 sheep do not die, so they remain.
Final number: **9**
constraints.txt (176 bytes, full file beside this page)The sky darkened as rain began to fall on the streets. Water streamed down the windows and pooled in quiet gutters. By evening, the world smelled of wet leaves and fresh stone.
json.txt (66 bytes, full file beside this page){"name":"Alice","age":30,"hobbies":["reading","hiking","cooking"]} creative.txt (582 bytes, full file beside this page)On the silent space station, master Voss found the last message in the archive, stamped with her own name: “If you are reading this, I am already gone.” She laughed, assuming a prank. The lights flickered. In the mirror of the airlock corridor, she saw herself aging rapidly, hair silvering, skin thinning. Alarms sang. Her console listed crew transfers to an unknown ship. “You will not remember,” it said. “You are being replaced by your future self.” Voss stared at her hands as they faded, and only then understood the message wasn’t a warning: it was her arrival from tomorrow.
