Run ced.cpp on any ggml GPU backend

[localai-org-maint-bot]

GPU builds of ced.cpp compiled but never used the GPU: the runner always ran on a static CPU backend and created and freed a graph allocator for every graph. This PR makes the model run on CUDA, Vulkan and Metal, and keeps the CPU path bit-identical.

What changed

• One backend per loaded model (ced::Backend). It picks the first GPU or integrated GPU, or the device named in CED_DEVICE (cpu, CUDA0, Vulkan0, MTL0), and falls back to the CPU. One ggml_gallocr lives for the model's lifetime. A graph goes through ggml_backend_sched with a CPU fallback only when the device has no kernel for one of its ops.
• Weights on the device. On a GPU the weights are uploaded to one device buffer at load. The mel window, filterbank and init_bn stats keep a host copy, and init_bn is folded once at load.
• One graph per chunk in classify (embed, blocks and head), instead of two graphs with a host round trip. The parity entry points (embed_from_input_values, forward_from_tokens) are unchanged.
• Faster mel frontend. The filterbank product skips the zero bins of each mel filter. Output is bit-identical. With the encoder on a GPU the host mel was most of the latency.
• ced-cli info and ced-cli bench print the device. README documents CED_DEVICE, and docs/BENCHMARKS.md has a GPU section.
• Test tolerance. GPU matmuls move intermediate activations by up to 1.15e-2. The per-stage parity checks use a 2e-2 tolerance when the device is not the CPU. The CPU tolerances and every end-to-end probability check are unchanged.

Results

ctest is 7/7 on every backend.

Backend	Machine	vs CPU	ced-base f32, 36 s clip
CPU	x86 devbox, M4	all 527 probs bit-identical to main; speed unchanged (±0.5%)
Metal	Apple M4	same top-5 tags; f32 within 1.4e-4	99 ms (CPU: 1152 ms)
Vulkan (RADV)	Radeon 8060S	same top-5 tags; f32 within 1.7e-4	71 ms (CPU: 446 ms)
CUDA 13	NVIDIA GB10	same top-5 tags; f32 within 1.3e-4	62 ms

Top-5 tags were compared for tiny and base at f32, f16 and q8_0 on a 6 s and a 36 s clip. Quantized models move more on GPU (up to about 1e-2 for tiny q8_0), because the CPU also quantizes the activations of q8_0 matmuls. For small models the host mel frontend is still most of the GPU latency. A GPU mel frontend is a possible follow-up.

How to verify

cmake -B build -DCED_BUILD_TESTS=ON -DCED_GGML_VULKAN=ON   # or CUDA / METAL
cmake --build build -j
ctest --test-dir build --output-on-failure                 # 7/7, logs "device: ..."
build/examples/cli/ced-cli bench models/ced-base-f32.gguf clip.wav --iters 30
CED_DEVICE=cpu build/examples/cli/ced-cli bench models/ced-base-f32.gguf clip.wav --iters 5

🤖 Generated with Claude Code