#6174 [Docs] Mac: source setup gives Q2_0 the DFlash2 drafter, but MAC.md says only the two Unsloth sets get it

open · @m-wells · 1 评论 · 去 GitHub 看

BenchmarksSetup & installModels & quantsDocumentationWindows

说明

### Area

[Docs]  (documentation)

### Desktop app, Terminal or source?

Source (START-HERE.bat, setup.sh or a build)

### What happened

On an M2 Max, source setup gave Q2_0 the DFlash2 drafter, while docs/MAC.md says only the two Unsloth sets get it.

`./setup.sh --yes --family qwen --model Q2_0 --vision no --context 262144 --no-start --no-browser` downloaded the DFlash2 drafter ("Downloading the DFlash2 drafter (1.2 GB, once)"), and the server started with:

```
[strata] drafting: DFlash2 alone (--draft dflash2, --draft-conf 0.5)
```

IQ3_XXS on the same Mac contradicts itself within one run. Setup printed:

```
[ok] the experts will stream from the SSD: no MTP draft layer (one token per step is faster there); ./setup.sh --setup on a Mac with more memory adds it
```

and its server then started with the same `drafting: DFlash2 alone` line.

docs/MAC.md, "Install" step 6, says: "For UD-Q4_K_XL and UD-IQ4_XS-v31 it also downloads the DFlash2 drafter ... IQ3_S and the other sets keep the MTP." It also says a streaming Mac below 96 GB runs "one token per step". The docstring of `fetch_drafter` in metal/setup_mac.py says the same as step 6.

From reading the code (unconfirmed static analysis): `fetch_drafter` asks `drafting.choice_of("auto", ..., backend="metal")`, which returns DFlash2 by the general rule ("Auto is DFlash2 wherever the engine has it") for every set without a `SET_DEFAULTS` entry. On Metal only `iq3_s` is pinned to the MTP, so Q2_0 and IQ3_XXS get DFlash2 too.

So either the docs are out of date, or Q2_0 and IQ3_XXS on Metal were meant to keep the MTP (or no draft when streaming) like IQ3_S. I did not compare DFlash2 against the MTP on this Mac. For reference, Q2_0 in memory decoded 43.3-44.1 tok/s and read 391.6-391.7 tok/s on an 11K-token prompt (greedy, thinking off, 512 tokens), and streamed IQ3_XXS 21.6-22.0 and 243.8-252.4.

### Smallest steps or command that reproduce it

On a 64 GB Apple Silicon Mac: `./setup.sh --yes --family qwen --model Q2_0 --vision no --no-start --no-browser`, start `run-q2_0.sh`, read the drafting line.

### Reproduction status

Reproduced locally

### Strata version

v0.2.15 (commit c775728), Metal engine 0.1.61, built from source

### Graphics card, memory and system

Apple M2 Max, 64 GB unified memory (iogpu.wired_limit_mb 59392), macOS 26.7.3

### Model, size and settings

ISTA-DASLab Qwen3.8-Flash-Next GSQ-RCO Q2_0 (resident) and IQ3_XXS (streamed), context 262144, pictures off.

本站相关内容

相关页面的快捷入口。