#6244 [GUI] Changing an option on the model screen silently drops drafting: dflash2 and resets STRATA_RESIDENT_PIN to 0

open · @ric20527 · 2 评论 · 去 GitHub 看

BenchmarksSetup & installAMD / HIPModels & quantsWindowsLinux

说明

### Area

[GUI]  (the desktop app)

### Desktop app, Terminal or source?

Desktop app (Windows, Linux or Mac)

### What happened

Changing an option on the model's settings screen rewrites the run config and silently drops two settings that are not on that screen: the drafting mode and a value in `env`.

I changed the context memory option (my app is in Spanish, where it is labelled "Memoria de contexto"). I set it back to the same value it had, so nothing should have changed. After that, both config files had been rewritten:

- `C:\Users\<user>\AppData\Local\Strata\install\strata-gyro-gyro-s.json`
- `C:\Users\<user>\.strata\data\native-tune\gyro-gyro-s.json`

Diff of each file, before → after (identical in both):

| key | before | after |
| --- | --- | --- |
| `drafting` | `"dflash2"` | *(key removed)* |
| `env.STRATA_RESIDENT_PIN` | `"1"` | `"0"` |
| `args` (drafting part) | `--draft dflash2 --draft-model …\qwen3.8-flash-next-dflash2.q4_0-selq4.pack --draft-fallback none --mtp-draft-vocab …\draft_vocab_latin.bin` | `--mtp D:\…\mtp\rt --spec-min-p 0.5` |

So the next load used the MTP draft (Auto) instead of DFlash2, and pinned residency was turned off. Nothing in the UI said so.

What was kept: `user_args` (same flags, reordered), the other `env` entries (`STRATA_HIP_WMMA`, `STRATA_PARTIAL_PIN`, `STRATA_PARTIAL_PIN_GIB`, `STRATA_UNBUFFERED_LOAD`, `STRATA_PREFILL_COPY_*`), `vram_elastic`, `expert_profile_save`.

Expected: changing one option leaves the settings that have nothing to do with it as they were. That matters most for `drafting`, which I set through the control's drafting apply (`strata drafting fetch … --mode dflash2` plus the control's apply). And `STRATA_RESIDENT_PIN` was only in `env`, which the app otherwise keeps as it is on every load.

Both settings were measured wins on this PC. DFlash2 gave +15–34 % on prompts against MTP, and resident pinning plus partial pinning took decode from 56 to 74 tok/s. So the silent reset costs real speed. I only noticed because I diff the config after each change.

Possibly related, but not the same path: #6203 (hand edits to `args` ignored by `engine_args()`) and #5351 (configuration spread across files).

### Smallest steps or command that reproduce it

1. Set a model to DFlash2 drafting (`"drafting": "dflash2"` in its run config) and add `"STRATA_RESIDENT_PIN": "1"` to `env`.
2. In the desktop app, open that model's settings and change the context memory option. Setting it back to its old value is enough.
3. Open `install\strata-<model>.json`: `drafting` is gone, `--draft dflash2 …` is replaced by `--mtp …`, and `STRATA_RESIDENT_PIN` is `"0"`.

I saw this once (2026-11-18) and restored the files from a backup. I kept a copy of the rewritten files, and I have not repeated the steps since.

### Reproduction status

Reproduced locally

### Strata version

Strata 0.2.16 · control 0.1.62 · engine 0.1.62

### Graphics card, memory and system

RX 9070 XT 16 GB (gfx1201) · 31.4 GB RAM · Windows 11 Pro 10.0.26300

### Model, size and settings

Qwen3.8-Flash-Next Gyro-S TQ1_0 · 131K context · pictures on (CPU) · DFlash2 draft, `--spec 8 --draft-conf 0.5` in `user_args`

本站相关内容

相关页面的快捷入口。