Issues / #629
#629 Re-running setup with a different --context silently discards mcp_servers, mcp and sampling from the run config
closed · @jmaher-fas · 1 Kommentare · Auf GitHub
Setup & installServer & APIMulti-GPUNVIDIA / CUDAModels & quantsSecurityWindows
Beschreibung
**Title:** Re-running setup with a different `--context` silently discards `mcp_servers`, `mcp` and `sampling` from the run config
### What happened
I had a working install (`qwen` / `IQ3_S`, two GPUs) and had hand-edited
`strata-iq3_s.json` to add three blocks that the installer does not write itself:
```json
"sampling": { "temperature": 1.0, "top_p": 0.95, "top_k": 20, "min_p": 0.0 },
"mcp_servers": { "files": {...}, "browser": {...}, "search": {...} },
"mcp": { "timeout_s": 60, "max_result_chars": 16000, "max_rounds": 8 }
```
I then re-ran setup to change only the context length:
```
START-HERE.bat --yes --family qwen --model IQ3_S --gpus 0,1 --context 131072 --no-start
```
Setup correctly reused the downloaded model and rewrote the start script. It also **rewrote
`strata-iq3_s.json` from scratch**, so all three blocks were gone. The file came back with only
the generated keys (`exe`, `args`, `cwd`, `tokenizer`, `model_name`, `log`, `lib_dirs`, `port`,
`gpu`, `gpus_asked`, `layer_split`).
Nothing warned that a config containing user-authored blocks was about to be replaced, and the
loss is quiet: the model loads fine, it just decodes greedy (no `sampling`) and has no MCP tools
in the chat page (no `mcp_servers`) until you notice.
I only recovered because I had copied the file to `strata-iq3_s.json.preserve` beforehand, on the
suspicion that this might happen.
### Why it is easy to hit
The documentation encourages putting these blocks in this file — `DETAILS.md` documents
`"sampling"` and the `"mcp_servers"` / `"mcp"` blocks as belonging in the run config — and
changing context later is a natural thing to do (`MODELS.md` discusses context choice as a
tradeoff worth revisiting). So the two instructions combine into data loss.
### Reproduce
1. Install any model.
2. Add a `"sampling"` block (or `"mcp_servers"`) to `strata-<model>.json`.
3. Re-run setup with a different `--context`.
4. The block is gone.
### Suggested fix, roughly in order of effort
- **Merge rather than replace.** When regenerating `strata-<model>.json`, carry over any key the
generator does not own — `sampling`, `mcp_servers`, `mcp`, and anything else unrecognised.
Setup already knows exactly which keys it writes, so the rest can be preserved by construction.
- **Or back it up and say so**: write `strata-<model>.json.bak` and print one line —
"your previous run config was saved to ...; sampling / mcp_servers were not carried over".
- **Or warn before overwriting** when the existing file contains keys the generator will not
re-create, and ask (or require `--force`).
The first seems closest to what a user expects: the flags passed to setup describe how the model
is loaded, not how it samples or what tools it may use.
### Environment
Windows 11, NVIDIA 2x RTX 5070 Ti (driver 616.92), engine 0.1.38 CUDA 13.0,
`qwen` / `IQ3_S`, two-GPU layer split. Strata cloned from `main` on 2026-10-03.
---
*Attribution: this was found and diagnosed by Claude (Anthropic's Claude Code) while it was
setting up an MCP-driven worker loop against Strata on my machine — it hit the wipe, recovered
from its own backup, and worked out the cause and the suggested fixes above. I'm filing it under
my account and can relay or dig into any follow-up questions.*
Mehr auf der Site
Links zu Install, Modellen, Releases.