Segmentation fault under WSL with AMD

[jdluzen]

I just found this, thanks so much for this amazing project!
I'm running under WSL on an AMD 7900 XTX.
I started out by trying to compile and run with Vulkan, but unfortunately, I could never get it off the CPU driver (llvmpipe). I didn't realize I could use DGGML_HIP. It compiled much easier, but unfortunately when running

./ace-qwen3 --request /models/acestepcpp/mytest.json --model /models/acestepcpp/acestep-5Hz-lm-4B-BF16.gguf

I get this:

[Request] parsed /models/acestepcpp/mytest.json (4 fields)
[Request] seed=-1
  caption:    <my caption>
  lyrics:     0 bytes
  bpm=0 dur=360 key= ts= lang=unknown
  lm: temp=0.85 cfg=2.0 top_p=0.90 top_k=0
  dit: steps=8 guidance=1.0 shift=3.0
  audio_codes: (none)
[BPE] Loaded from GGUF: 151643 vocab, 151387 merges
ggml_cuda_init: found 1 ROCm devices:
  Device 0: AMD Radeon RX 7900 XTX, gfx1100 (0x1100), VMM: no, Wave Size: 32
[Load] LM backend: ROCm0 (CPU threads: 16)
[GGUF] /models/acestepcpp/acestep-5Hz-lm-4B-BF16.gguf: 398 tensors, data at offset 5346240
[LM-Config] 36L, H=2560, V=217204, Nh=32, Nkv=8, D=128, tied=1
[Qwen3] Attn: Q+K+V fused
[Qwen3] MLP: gate+up fused
[WeightCtx] Loaded 398 tensors, 7991.3 MB into backend
[LM-KV] Allocated 2 sets x 36 layers (4D batched), 2304.0 MB
[FSM] Prefix trees: bpm=301, dur=601, key=654, lang=18, tsig=5 nodes
[Fill] lyrics=generate metas=fill gaps | 57 tokens, CFG: 1.00, N=1
Segmentation fault (core dumped)

Any ideas? In the meantime, I'll try to build a debug version and see if that has any more details.

[jdluzen]

Something new after the debug build:
ggml_backend_sched_alloc_splits: failed to allocate graph, reserving (backend_ids_changed = 1)

[ServeurpersoCom]

Interesting, Apart from all my Linux machines I've never tested it under WSL! I use MSVC. I'll take a look!

D:\A\acestep.cpp>toolchain.cmd
**********************************************************************
** Visual Studio 2022 Developer Command Prompt v17.13.6
** Copyright (c) 2022 Microsoft Corporation
**********************************************************************
[vcvarsall.bat] Environment initialized for: 'x64'

D:\A\acestep.cpp>type toolchain.cmd
@echo off

"C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Auxiliary\Build\vcvars64.bat"

D:\A\acestep.cpp>type build.cmd
@echo off

cmake -B build-msvc -DGGML_CUDA=ON -DGGML_VULKAN=ON
cmake --build build-msvc --config Release -j 24

D:\A\acestep.cpp>

[ServeurpersoCom]

git pull, and retry if there are no more segfaults, we will see that it is a memory allocation problem.

[jdluzen]

Thanks for the quick commits, unfortunately same thing. I attempt to build it in a Dockerfile and then run in a container.
I was able to compile in Win11 using Vulkan which runs amazing!

[ServeurpersoCom]

It works perfectly with WSL2 and Vulkan. I haven't been able to reproduce the segfault, ROCm (HIP) also works very well under Linux, never tested HIP under WSL2...