Problemi con CUDA RTX5070

[dozeen]

Segmentation fault on RTX 5070 (SM120) under WSL2 before model loading

Environment

• OS: Ubuntu 24.04 running under WSL2
• Host OS: Windows 11 Pro
• CPU: AMD Ryzen 7 5700X
• GPU: NVIDIA GeForce RTX 5070 (Compute Capability 12.0 / SM120)
• NVIDIA Driver: 595.79
• CUDA Toolkit: 13.2
• nvcc: 13.2.51

Repository:

80ebbc396aee40eedc1d829222f3362d10fa4c6c

Problem

Both ds4 and ds4-server immediately crash with a segmentation fault when using the CUDA backend.

Example:

./ds4 --cuda -p "Hello"

Output:

ds4: Linux cuda backend set oom_score_adj=1000
Segmentation fault (core dumped)

The crash happens before the model is loaded.

The same happens with:

./ds4-server --cuda --ctx 4096

What I already verified

• Fresh clone of the repository
• Built from source successfully
• Rebuilt after replacing

targets/sbsa-linux/lib

with

targets/x86_64-linux/lib

inside the Makefile.

• Compiled using

make cuda CUDA_ARCH=sm_120

• Model path is correct.
• CPU backend works correctly.
• CUDA Toolkit installation appears correct.
• cudaGetDeviceCount() and cudaSetDevice() both work in a standalone CUDA test program.

Example:

cudaGetDeviceCount(...)
cudaSetDevice(0)

returns success.

GDB backtrace

The crash occurs during CUDA initialization.

cudaSetDevice()

↓

libcuda.so

↓

libnvidia-ptxjitcompiler.so

↓

SIGSEGV

Backtrace:

#0 libnvidia-ptxjitcompiler.so
#1 __cuda_CallJitEntryPoint
#2 libcuda.so
#3 cudaSetDevice
#4 ds4_gpu_init()
#5 ds4_engine_open()
#6 main()

strace

The last operations before the crash are loading

/lib/x86_64-linux-gnu/libnvidia-ptxjitcompiler.so.580.159.03

Immediately afterwards:

SIGSEGV si_addr=0x30

Additional information

The crash happens on two independent clones of the repository.

The crash also occurs before loading the GGUF model, so it does not appear to be model-related.

Since a standalone CUDA program works correctly, I suspect this may be:

• a CUDA JIT issue affecting SM120 / RTX 5070,
• or an interaction between DS4 CUDA initialization and the CUDA 13.2 runtime under WSL2.

If there is anything else I can test or any additional logging you would like, I'd be happy to help.

[dan64]

Hi antirez,

To democratise the use of AI it would be very helpful if you could complete the porting of DS4 to Linux and WLS2/Windows, since the number of users running these OS it's enormously higher than the users running Apple Silicon or DGX Spark.

[dan64]

Hi antirez,

I was able to use your GGUF in Windows 11 (112GB DDR5 RAM) and RTX5070 TI (16GB VRAM).
I used the following fork of your project: Fringe210/llama.cpp-deepseek-v4-flash-cuda.
Here a small sample session

PS C:\AiProjects> .\llama.cpp-deepseek-v4-flash-cuda\build\bin\Release\llama-cli.exe -ngl auto --cpu-moe -cnv -m .\gguf\DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix.gguf
ggml_cuda_init: found 1 CUDA devices (Total VRAM: 16302 MiB):
  Device 0: NVIDIA GeForce RTX 5070 Ti, compute capability 12.0, VMM: yes, VRAM: 16302 MiB

Loading model...


▄▄ ▄▄
██ ██
██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
                                    ██    ██
                                    ▀▀    ▀▀

build      : b8936-44c7b01de
model      : DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix.gguf
modalities : text

available commands:
  /exit or Ctrl+C     stop or exit
  /regen              regenerate the last response
  /clear              clear the chat history
  /read <file>        add a text file
  /glob <pattern>     add text files using globbing pattern


> che versione sei ?

[Start thinking]
Sto ricevendo una domanda in italiano: "che versione sei ?" che significa "che versione sei?" (quale versione sei?). L'utente sta chiedendo la mia versione. Devo rispondere in modo chiaro e diretto. Sono DeepSeek, l'ultima versione è DeepSeek-V2 o forse DeepSeek-V3? In realtà, DeepSeek ha rilasciato DeepSeek-V2 e poi DeepSeek-V3, ma come assistente devo specificare la mia versione. La risposta standard è: "Sono DeepSeek, l'ultima versione disponibile. Se hai bisogno di dettagli specifici, posso fornirti informazioni aggiornate." Ma devo essere preciso: sono DeepSeek-V3? Oppure DeepSeek-R1? L'utente chiede "che versione sei?" e io devo rispondere con la mia identità. Come assistente AI, posso dire "Sono DeepSeek, un assistente AI creato da DeepSeek (深度求索). La mia versione corrente è DeepSeek-V3. Se hai bisogno di dettagli specifici, posso fornirti informazioni aggiornate." Tuttavia, per evitare confusione, rispondo in modo semplice.
[End thinking]

Sono DeepSeek, l'ultima versione disponibile al momento del mio addestramento (DeepSeek-V3). Se hai bisogno di dettagli specifici sulla versione o sul modello, posso fornirti informazioni aggiornate. Come posso aiutarti?

[ Prompt: 1.2 t/s | Generation: 8.8 t/s ]

The generation speed is not too bad, taking into account that I don't have unified memory on my PC. The total RAM used by the client is 93 GB (o/w 9-12GB are related to Windows environment). This is expected given that the GGUF size is about 84GB. The VRAM (16 GB) is fully used while GPU is used at 30% and CPU at 60%.

Some questions:

1. are you planning to include the Fringe210/llama.cpp-deepseek-v4-flash-cuda fork in your ds4 project ?
2. do you have any suggestion to improve the performance on my setup ?

Grazie,
Dan