--gpu ignores integrated GPUs (ggml DEVICE_TYPE_IGPU) and silently falls back to CPU #14

Title: --gpu ignores integrated GPUs (ggml DEVICE_TYPE_IGPU) and silently falls back to CPU

Summary

On a laptop whose only Vulkan device is an integrated GPU, zonos2-cli --gpu prints
zonos2: backend = CPU and runs on the CPU, with no warning. The Vulkan device is detected
(ggml_vulkan: Found 1 Vulkan devices), but it is never selected.

Cause

ggml-vulkan reports integrated GPUs as GGML_BACKEND_DEVICE_TYPE_IGPU
(ggml/src/ggml-vulkan/ggml-vulkan.cpp, at the pinned ggml 3af5f576:
return ctx->is_integrated_gpu ? GGML_BACKEND_DEVICE_TYPE_IGPU : GGML_BACKEND_DEVICE_TYPE_GPU;).
zonos2.cpp only looks up GGML_BACKEND_DEVICE_TYPE_GPU:

• src/zonos2.cpp:77 and src/dac.cpp:53:
  if (use_gpu) dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
• src/zonos2.cpp:88 and src/dac.cpp:57:
  m.is_gpu = (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU);

so an iGPU is skipped and the CPU device is used.

Environment

Reproduce

zonos2-cli zonos2-q8_0.gguf --tts "<a 14-word English sentence>" out.wav --dac dac.gguf --gpu --seed 1 --max 800

Release binary:

ggml_vulkan: 0 = Intel(R) Graphics (Intel Corporation) | uma: 1 | ...
zonos2: backend = CPU | Intel(R) Core(TM) Ultra 5 125U
generate(kv): 480 decode steps in 21.033s ... RTF=3.77
dac: H=481 eos=471 -> 241152 samples = 5.47s, rms=0.0241 peak=0.2220

The CPU output is correct (faster-whisper large-v3-turbo transcribes the sentence verbatim) and
identical to the same run on another x86 CPU (eos=471, rms=0.0241).

Experiment with a two-line patch

A local build of adce4ffb (Vulkan SDK 1.4.313.2, MSVC 14.44, -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=OFF) with this change in both zonos2.cpp and dac.cpp:

     if (use_gpu) dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
+    if (use_gpu && !dev) dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_IGPU);
     if (!dev)    dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
...
-    m.is_gpu  = (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU);
+    m.is_gpu  = (ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_CPU);

does select the iGPU, but the output is wrong:

zonos2: backend = Vulkan0 | Intel(R) Graphics
generate(kv): 799 decode steps in 43.257s = 18.5 frames/s (real-time = 86.1), RTF=4.66
dac: loaded 128 tensors on GPU
dac: H=800 eos=-1 -> 409600 samples = 9.29s, rms=0.0846 peak=0.5774

No EOS (runs to the frame cap), and faster-whisper recognises only the first two words (and
one of them wrongly), followed by unintelligible audio. This looks like the same Vulkan divergence as #11 (RADV) and #10
(NVIDIA on Windows), now also on an Intel iGPU, so it is probably not specific to one driver.

Suggestions

1. Accept GGML_BACKEND_DEVICE_TYPE_IGPU for --gpu (or at least print a warning when --gpu
   was requested and no GPU device was chosen), so the fallback is not silent.
2. Until the Vulkan divergence (Vulkan backend: Japanese generation often never emits EOS and runs to the frame cap (CUDA/CPU do not reproduce) #10, **Vulkan backend on RADV/RDNA3: wrong pitch, no EOS, and non-deterministic output — CPU backend is bit-identical** #11) is fixed, it may help to note in the README that the
   Vulkan output can differ from CPU, and that CPU is the reference path.