#1964 V0.1.43 Decode performance degradation on 4070ti + 14700 + 64 RAM, Linux IQ3_XXS

说明

Something changed in 43 that made decode on this setup to be about a third of what it was in 0.1.42.

Bench.sh output for v42:
```
Results (median, range over the runs):
  chat decode tok/s                     96.50   (80.10 - 108.00, n=9)
  chat first token (s)                   0.34   (0.30 - 0.39, n=9)
  4K prompt tok/s                     2181.30   (2143.90 - 2188.80, n=3)
  4K first token (s)                     1.86   (1.85 - 1.89, n=3)
  16K prompt tok/s                    3136.60   (3090.30 - 3139.70, n=3)
  16K first token (s)                    5.14   (5.14 - 5.22, n=3)
  agent cache hit %                     68.77   (68.70 - 68.80, n=3)
  agent re-read tokens/session           4.00   (3.00 - 74.00, n=3)
  agent turn first token (s)             1.94   (1.34 - 2.15, n=15)
  agent decode tok/s                   104.90   (73.20 - 131.40, n=15)
```
v43:
```
Results (median, range over the runs):
  chat decode tok/s                     37.90   (33.70 - 59.30, n=9)
  chat first token (s)                   0.37   (0.32 - 0.52, n=9)
  4K prompt tok/s                     2118.70   (2117.20 - 2173.80, n=3)
  4K first token (s)                     1.92   (1.87 - 1.92, n=3)
  16K prompt tok/s                    3140.70   (3095.00 - 3141.20, n=3)
  16K first token (s)                    5.14   (5.13 - 5.21, n=3)
  agent cache hit %                     68.73   (68.60 - 68.81, n=3)
  agent re-read tokens/session          74.00   (4.00 - 74.00, n=3)
  agent turn first token (s)             1.93   (1.33 - 2.01, n=15)
  agent decode tok/s                    45.90   (39.80 - 62.80, n=15)
```

"agent re-read tokens/session" jumps out. Though the range seems to be about the same.

Startup Json same for both:
```
{
 "exe": "/opt/Servarr/llm/Strata/engine/strata",
 "args": [
  "--pack",
  "/opt/Servarr/llm/Strata-data/packs/iq3_xxs",
  "--native",
  "/opt/Servarr/llm/Strata-data/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf",
  "--ple-gguf",
  "/opt/Servarr/llm/Strata-data/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00002-of-00002.gguf",
  "--expert-profile",
  "/opt/Servarr/llm/Strata/data/expert-profile.bin",
  "--expert-cache",
  "auto",
  "--prefill",
  "auto:32768",
  "--spec",
  "4",
  "--mtp",
  "/opt/Servarr/llm/Strata-data/mtp/rt",
  "--max-context",
  "262144",
  "--kv",
  "int8",
  "--kv-resident",
  "32768",
  "--vision",
  "--vram-reserve-mib",
  "600",
  "--pool-affinity",
  "auto",
  "--pool-workers",
  "15",
  "--coupled-draft",
  "--pcie-frac",
  "0.24",
  "--spec-min-p",
  "0.70"
 ],
 "cwd": "/opt/Servarr/llm/Strata",
 "tokenizer": "/opt/Servarr/llm/Strata-data/packs/iq3_xxs/tokenizer",
 "model_name": "qwen3.8-flash-next-iq3_xxs",
 "log": "/opt/Servarr/llm/Strata/strata-iq3_xxs.log",
 "lib_dirs": [
  "/usr/local/cuda-13.0/bin",
  "/usr/local/cuda-13.0/lib64"
 ],
 "port": 8080,
 "gpu": 0,
 "gpus_asked": true,
 "draft_vocab": "cyrillic",
 "fit_max_tokens": true,
 "expert_profile_save": "/opt/Servarr/llm/Strata-data/iq3_xxs-profile.bin",
 "vision": {
  "exe": "/opt/Servarr/llm/Strata/engine/strata-vision",
  "mmproj": "/opt/Servarr/llm/Strata-data/models/mmproj-Qwen3.8-Flash-Next-BF16.gguf",
  "model": "/opt/Servarr/llm/Strata-data/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf",
  "gpu": true,
  "max_tokens": 3072
 },
 "sampling": {
  "temperature": 1.0,
  "top_p": 0.95,
  "top_k": 20,
  "min_p": 0.0
 }
}
```
Startup sh also the same for both:
```
#!/bin/sh
cd "/opt/Servarr/llm/Strata"
export STRATA_EXCHANGE_ROTATE=1
export STRATA_KV_PREFETCH=1
export STRATA_PREFILL_CPU_SHARE=auto
export STRATA_SPEC_COUPLED=1
export STRATA_SPEC_GUMBEL=1
exec "/opt/Servarr/llm/Strata/.venv/bin/python" "/opt/Servarr/llm/Strata/serve/server.py" "--engine" "strata" "--config" "/opt/Servarr/llm/Strata/strata-iq3_xxs.json" "--host" "0.0.0.0" "--port" "8080" "--api-key" "xxx" "--open"
```
Startup logs attached:
[v43.log](https://github.com/user-attachments/files/33306188/v43.log)
[v42.log](https://github.com/user-attachments/files/33306189/v42.log)

Suspected VRAM overflow with new changed, did not pan out, changing   --vram-reserve-mib to 1000 did not produce better speeds.