Vulkan crash on Intel iGPU with PocketTTS for longer inputs (ADL GT2, Mesa)

[AdityaAWaghmare]

Vulkan crash on Intel iGPU with PocketTTS for longer inputs

Description

PocketTTS works correctly on my Intel integrated GPU for short inputs, but crashes with a Vulkan workgroup-limit assertion when processing the following longer input.

The same input works correctly when using my NVIDIA RTX 3050 Laptop GPU.

Environment

• OS: Ubuntu Linux

• Backend: Vulkan

• Model: PocketTTS (GGUF)

• Reference voice: demo_1_man

• GPU 0: Intel(R) Graphics (ADL GT2)

  • Mesa Intel open-source Vulkan driver

  • fp16: 1

  • bf16: 0

  • warp size: 32

  • int dot: 0

  • matrix cores: none

• GPU 1: NVIDIA GeForce RTX 3050 Laptop GPU

Reproduction input

Timothy Donald Cook (born November 1, 1960) is an American business executive who served as the chief executive officer (CEO) of Apple Inc. from 2011 to 2026, and has been the company's executive chairman since September 2026. Cook joined the technology company in 1998 as senior vice president for worldwide operations, and later served as vice president for worldwide sales and operations and then as chief operating officer under co-founder Steve Jobs. He was appointed chief executive in 2011, when Jobs resigned due to failing health. Under Cook, Apple grew into one of the world's most valuable companies; in 2025 it was the second largest technology company by revenue in the world, at US$416 billion.[1] Cook led a succession plan that saw John Ternus succeed him as CEO.

The reference voice used was:

demo_1_man

Observed behavior

A short sentence works successfully on the Intel iGPU, and is actually faster than CPU execution.

With the longer input above:

Backend | Result -- | -- CPU | Works Intel iGPU / Vulkan | Crashes NVIDIA RTX 3050 / Vulkan | Works

The Intel Vulkan backend aborts with:

/home/runner/work/audio.cpp/audio.cpp/external/ggml/src/ggml-vulkan/ggml-vulkan.cpp:6819:
GGML_ASSERT(wg0 <= ctx->device->properties.limits.maxComputeWorkGroupCount[0] &&
            wg1 <= ctx->device->properties.limits.maxComputeWorkGroupCount[1] &&
            wg2 <= ctx->device->properties.limits.maxComputeWorkGroupCount[2]) failed
Aborted (core dumped)

Potentially relevant commit

I noticed commit 43833e8749f24466b680266e9d805b7d58386748, "Apply upstream Vulkan fill dispatch fix (#508)".

That commit changes the Vulkan GGML_OP_FILL dispatch to distribute work over two dimensions specifically to avoid exceeding maxComputeWorkGroupCount.

The error I'm seeing is also a maxComputeWorkGroupCount assertion, and the problem only appears once the workload becomes sufficiently large, so this may be related. I have not yet determined whether the operation triggering my crash is actually GGML_OP_FILL.

Version information

I also built audio.cpp from the latest repository a few days before reporting this, so my local build may predate the above fix.

The crash was also observed using the audio-v0.7.3-bin-ubuntu-x64-vulkan binary release.

Questions

Could this be another instance of the Vulkan dispatch/workgroup-limit issue addressed by #508?

In particular, it would be useful to determine:

1. Which GGML/Vulkan operation is triggering the assertion.

2. Whether GGML_OP_FILL is involved.

3. Whether the Intel Vulkan implementation requires additional handling for large dispatches.

4. Whether this is already fixed on current audio.cpp main.

I can provide additional logs or test the latest main if useful.

## Vulkan crash on Intel iGPU with PocketTTS for longer inputs

Description

PocketTTS works correctly on my Intel integrated GPU for short inputs, but crashes with a Vulkan workgroup-limit assertion when processing the following longer input.

The same input works correctly when using my NVIDIA RTX 3050 Laptop GPU.

Environment

• OS: Ubuntu Linux

• Backend: Vulkan

• Model: PocketTTS (GGUF)

• Reference voice: demo_1_man

• GPU 0: Intel(R) Graphics (ADL GT2)

  • Mesa Intel open-source Vulkan driver
  • fp16: 1
  • bf16: 0
  • warp size: 32
  • int dot: 0
  • matrix cores: none

• GPU 1: NVIDIA GeForce RTX 3050 Laptop GPU

Reproduction input

Timothy Donald Cook (born November 1, 1960) is an American business executive who served as the chief executive officer (CEO) of Apple Inc. from 2011 to 2026, and has been the company's executive chairman since September 2026. Cook joined the technology company in 1998 as senior vice president for worldwide operations, and later served as vice president for worldwide sales and operations and then as chief operating officer under co-founder Steve Jobs. He was appointed chief executive in 2011, when Jobs resigned due to failing health. Under Cook, Apple grew into one of the world's most valuable companies; in 2025 it was the second largest technology company by revenue in the world, at US$416 billion.[1] Cook led a succession plan that saw John Ternus succeed him as CEO.

The reference voice used was:

demo_1_man

Observed behavior

A short sentence works successfully on the Intel iGPU, and is actually faster than CPU execution.

With the longer input above:

Backend	Result
CPU	Works
Intel iGPU / Vulkan	Crashes
NVIDIA RTX 3050 / Vulkan	Works

The Intel Vulkan backend aborts with:

/home/runner/work/audio.cpp/audio.cpp/external/ggml/src/ggml-vulkan/ggml-vulkan.cpp:6819:
GGML_ASSERT(wg0 <= ctx->device->properties.limits.maxComputeWorkGroupCount[0] &&
            wg1 <= ctx->device->properties.limits.maxComputeWorkGroupCount[1] &&
            wg2 <= ctx->device->properties.limits.maxComputeWorkGroupCount[2]) failed

Aborted (core dumped)

Potentially relevant commit

I noticed commit [43833e8749f24466b680266e9d805b7d58386748](https://github.com/0xShug0/audio.cpp/commit/43833e8749f24466b680266e9d805b7d58386748), "Apply upstream Vulkan fill dispatch fix (#508)".

That commit changes the Vulkan GGML_OP_FILL dispatch to distribute work over two dimensions specifically to avoid exceeding maxComputeWorkGroupCount.

The error I'm seeing is also a maxComputeWorkGroupCount assertion, and the problem only appears once the workload becomes sufficiently large, so this may be related. I have not yet determined whether the operation triggering my crash is actually GGML_OP_FILL.

Version information

I also built audio.cpp from the latest repository a few days before reporting this, so my local build may predate the above fix.

The crash was also observed using the audio-v0.7.3-bin-ubuntu-x64-vulkan binary release.

Questions

Could this be another instance of the Vulkan dispatch/workgroup-limit issue addressed by #508?

In particular, it would be useful to determine:

1. Which GGML/Vulkan operation is triggering the assertion.
2. Whether GGML_OP_FILL is involved.
3. Whether the Intel Vulkan implementation requires additional handling for large dispatches.
4. Whether this is already fixed on current audio.cpp main.

I can provide additional logs or test the latest main if useful.

[0xShug0]

@AdityaAWaghmare It could be related, but I don’t have the hardware to reproduce the issue, so I can’t say for sure. Can you test https://github.com/0xShug0/audio.cpp/actions/runs/34642627647 dev release under the "Artifacts" section?

[AdityaAWaghmare]

Update:

I tested the dev release artifact requested above:

audio-v0.7.2-dev.fbe3e-bin-ubuntu-x64-vulkan

This artifact corresponds to commit fbe3eed, which is newer than the
Vulkan fill dispatch fix (#508).

However, I cannot test the original issue with this build because it
currently crashes with Illegal instruction (core dumped).

This happens both with:

--backend vulkan --device 0

and:

--backend cpu

The crash occurs after the server starts and PocketTTS is loaded:

[warning][model_spec] GGUF for family 'pocket_tts' ...
ggml_vulkan: Found 2 Vulkan devices:
...
Illegal instruction (core dumped)

So this appears to be a separate issue from the original Intel Vulkan
maxComputeWorkGroupCount assertion.

The original v0.7.3 Vulkan binary remains able to start and exhibits the
original behavior described above.

[0xShug0]

@AdityaAWaghmare There are only two commits could be relevent to pockettts. Could you check other models and see if they give the same error?

[0xShug0]

Checked the prebuilts:

CPU artifacts

• v0.7.3 CPU: built with -march=native, contains AVX-512-style zmm instructions.
• dev fbe3e CPU: built with the same -march=native setting, but contains no zmm instructions.

Vulkan artifacts

• v0.7.3 Vulkan: built with -march=native, contains no zmm instructions.
• dev fbe3e Vulkan: built with the same -march=native setting, contains many zmm instructions.

-march=native depends on the specific GitHub runner CPU for that job. CPU and Vulkan jobs are separate jobs and may run on different CPU SKUs. So the same CMake/build command can produce different CPU ISA in different artifacts.

[0xShug0]

@AdityaAWaghmare Please try both the v0.7.4 Vulkan and Vulkan Portable builds and let me know if either of them works.

[AdityaAWaghmare]

Update: reproduced on v0.7.4 portable build and with Qwen3-TTS

I tested the dev/release builds requested above.

The non-portable Ubuntu Vulkan builds (v0.7.2-dev and v0.7.4) exit with:

Illegal instruction (core dumped)

on my system, so I tested the portable v0.7.4 Vulkan build instead.

Using:

audio-v0.7.4-bin-ubuntu-x64-vulkan-portable/audiocpp_server \
    --ui --ui-management --backend vulkan --device 0

the server starts successfully and detects:

0 = Intel(R) Graphics (ADL GT2) (Intel open-source Mesa driver)
1 = NVIDIA GeForce RTX 3050 Laptop GPU (NVIDIA)

With PocketTTS and the same long input from the original report, I can
reproduce the original assertion:

/home/runner/work/audio.cpp/audio.cpp/external/ggml/src/ggml-vulkan/ggml-vulkan.cpp:6819:
GGML_ASSERT(wg0 <= ctx->device->properties.limits.maxComputeWorkGroupCount[0] &&
            wg1 <= ctx->device->properties.limits.maxComputeWorkGroupCount[1] &&
            wg2 <= ctx->device->properties.limits.maxComputeWorkGroupCount[2]) failed

I also tested Qwen3-TTS with the same v0.7.4 portable Vulkan build and
Intel device (--device 0). It produces the exact same assertion.

This suggests that the problem is not specific to PocketTTS. Two
different TTS models trigger the same Vulkan workgroup-count assertion
on the Intel iGPU.

For reference, the original behavior is:

Intel Vulkan + PocketTTS (short input)  -> works
Intel Vulkan + PocketTTS (long input)   -> maxComputeWorkGroupCount assertion
Intel Vulkan + Qwen3-TTS                -> maxComputeWorkGroupCount assertion

The NVIDIA RTX 3050 was able to run the corresponding PocketTTS workload
successfully.

I also noticed commit 43833e8
("Apply upstream Vulkan fill dispatch fix (#508)"), which changes
GGML_OP_FILL dispatch to avoid exceeding maxComputeWorkGroupCount.

Since the v0.7.4 portable build still reproduces the assertion, this may
be a different Vulkan dispatch path hitting the same underlying limit.

It may be useful to identify which operation is generating the
out-of-range wg0/wg1/wg2 values. I can test further or provide
additional logs if needed.

[0xShug0]

@AdityaAWaghmare seems this issue has been reported elsewhere (everywhere) localai-org/vibevoice.cpp#9. More likely the issue is deep in GGML or the Intel iGPU backend/Mesa driver. Could you look into similar reports and see if there are any known workarounds? You could also try applying any relevant upstream patches and building locally to see whether they fix the issue. I don’t have access to an Intel iGPU device to reproduce and debug it myself, so iteration on my side will be slow.

[AdityaAWaghmare]

@0xShug0 Sure, I am currently looking into it and have a prospective fix. Everything is working now, I am just testing it a bit just in case.

[AdityaAWaghmare]

Hey @0xShug0 , quick heads-up: I tracked down the Vulkan crash here and opened a fix in #542.

The issue was the COL2IM_1D Vulkan implementation flattening the [T_out, OC] output into a single X dispatch dimension, which can exceed maxComputeWorkGroupCount[0] on some Vulkan devices. I changed it to a 2D T_out × OC dispatch.

I tested the fix on Intel ADL GT2 with both PocketTTS and Qwen3-TTS using the long input from this issue. Both now complete without the assertion. Also verified PocketTTS on my RTX 3050.

PR: #542