#5663 [Engine] ds4 won't start on R9700

closed · @AzurSaffaya · 3 コメント · GitHub で見る

Setup & installAMD / HIPNVIDIA / CUDADocumentationWindowsLinux

本文

### Problem
**Strata fails to launch DeepSeek-V4.1-Flash:**
> [X]  the model does not answer sensibly after the preparation: the engine stopped with an error (exit -11):

### Environment
**Strata 0.2.13** (**build 0.1.59**; python 3.12.15)
 [ok] **engine**: /home/xxx/.local/share/strata/install/engine-dsv41/**0.1.59-linux-x86_64-hip-rx9070**/strata-dsv41
Ubuntu 24.04
GPU 0: AMD Radeon RX 7900 series (gfx1100), 20 GB VRAM - on PCIe 4 x8
**GPU 1: AMD Radeon RX 9070 series / AI PRO R9700 (gfx1201), 32 GB VRAM - on nvme to PCIe 5 x4**
GPU 2: AMD Radeon RX 7900 series (gfx1100), 20 GB VRAM - on PCIe 4 x8
GPU 3: AMD Radeon RX 9070 series / AI PRO R9700 (gfx1201), 16 GB VRAM - on nvme to PCIe 5 x4
GPU 4: AMD Radeon (gfx1036), 0 GB VRAM - not supported

> [ok] GPU: AMD Radeon RX 9070 series / AI PRO R9700 (gfx1201), 31.9 GB VRAM, gfx1201 (AMD: docs/AMD_HIP.md)

[ok] RAM: 124 GB DDR 5 4800
[ok] CPU: AMD Ryzen 7 9800X3D 8-Core Processor (AVX-512)

### Reproduce
[TUI]

> strata launch

select model : DeepSeek-V4.1-Flash IST-DAS-2.25bpw
select card: GPU 2 - AI PRO R9700 (gfx1201) 32GB [Main]; All others [Off]
select text: 128K
Speed or Full quality?: Faster
Read Pictures: YES
How many requests: One at a time
Tune for the best speed: Quick
Advanced: No change

### Expected
Engine running

### Actual
Model: Preparing your model for Strata (one time)
`The model could not be installed. The model does not answer sensibly after the preparation: the engine stopped with an error (exit -11): ds4: prompt path: the chunked pass from 96 tokens, windows of 16 below | ds4: CUDA1 layers [0,39] + head: weights 5.86GB, experts 24.96 GB (2526 of 15360 in VRAM), caches 0.00 GB, 572 MiB free | ds4: 12834 routed experts on the CPU lane (124.9 GB, in place in the model's files), 8 threads.`

### Evidence
=== Step 6: preparing the model for Strata ===
`
  [ok] pack: /MODELS/DeepSeek4.1-Flash/packs/deepseek-v4.1-flash-ist-das-2.25bpw (checked: every file the engine needs is there)
  Testing the model with a short question (a few minutes: the engine reads the first experts from the disk) ...
  [X]  the model does not answer sensibly after the preparation: the engine stopped with an error (exit -11): ds4: prompt path: the chunked pass from 96 tokens, windows of 16 below | ds4: CUDA1 layers [0, 39] + head: weights 5.86 GB, experts 24.96 GB (2526 of 15360 in VRAM), caches 0.00 GB, 572 MiB free | ds4: 12834 routed experts on the CPU lane (124.9 GB, in place in the model's files), 8 threads
       delete /MODELS/DeepSeek4.1-Flash/packs/deepseek-v4.1-flash-ist-das-2.25bpw and run setup again so the pack is built fresh; if the test fails again, tell us (the log is in /home/xxx/.local/share/strata/install)
`

### Reproduction status
Reproduced locally with:
_Terminal 0.1.59
_AppImage 0.2.12

### **Notes**
Same problem has been happening for several versions of strata (last 4~5).
Same problem with: 9070 XT, R9700 + 9070 XT.

When running Qwen models: Strata terminal and App are confused by the 7900 XT + R9700 configuration and will download and install the engine for the 7900 XT even if the GPU selected for the model is R9700. And then will stop with error because of course that engine can't run on R9700. I have to manually drop the engine for R9700 in the components folder, there is no other way for Strata to run.
I do not know how to do the same for the ds4 engine, in case this is the problem. Where does it go? With what name?


関連リンク

インストール・モデル・リリースへのサイト内リンク。