Optimize inference on Strix Halo

[fedeizzo]

Originated from #170

What does this PR do?

• .devops/nix/package.nix, flake.nix, CMakeLists.txt: add an opt-in rocm-gfx1151 build and ENGINE_HIP_STRIX_HALO_OPTIMIZATIONS CMake option. package.nix automatically deduces strixHaloOptimizations when rocmGpuTargets == [ "gfx1151" ]. It enables the Strix Halo path only for HIP with CUDA-family fast paths and rocBLAS, keeping rocBLAS because hipBLASLt regressed latency on this hardware.
• include/engine/framework/core/module.h, src/framework/modules/{conv_modules.cpp,optimizations/fast_projection_modules.cpp}: update uses_ggml_cuda_family_backend to allow HIP to use the verified GGML CUDA-family ConvTranspose1d col2im and packed-projection fast paths used by Qwen3-TTS, Higgs-Audio-TTS, Confucius4-TTS, and Irodori-TTS. Other backends keep their existing behavior.
• include/engine/models/qwen3_tts/tokenizer_speech_decoder.h, src/models/qwen3_tts/tokenizer_speech_decoder.cpp: retain common speech decoder graphs on Strix Halo HIP builds, add upload_static_inputs() to restore static inputs (positions_ and mask_) prior to replaying retained graphs, and remove redundant backend synchronizations after graph execution.
• tests/unittests/test_conv_transpose_fast_path.cpp, tests/unittests/test_qwen_decoder_packed_projections.cpp: add HIP fast-path classification, parity execution, and backend rejection unit tests.

All Strix-specific behavior is disabled by default. Generic CPU, CUDA, Vulkan, Metal, and generic HIP builds retain their current defaults.

Motivation

Qwen3-TTS repeatedly rebuilt stable decoder graphs on Strix Halo, adding graph construction overhead and variable compute latency. Reusing those graphs removes graph construction overhead and stabilizes speech decoder execution, but reuse is only correct when static graph inputs (positions and attention masks) are restored before every graph replay.

Additionally, expanding uses_ggml_cuda_family_backend allows HIP devices to share the verified CUDA col2im transposed convolution and packed projection fast paths across multiple model architectures (Qwen3-TTS, Higgs-Audio-TTS, Confucius4-TTS, Irodori-TTS).

Gains

Empirical performance and numerical parity results on AMD Radeon 8060S Graphics (Strix Halo gfx1151, 128 GB VRAM):

Workload / Benchmark	Baseline / Reference	ROCm (gfx1151 Fast Path)	Gain / Parity Result
ConvTranspose1d col2im block	2.15 ms	1.77 ms	1.21x speedup (max diff 4.65e-9)
FastPackedProjection4Module	Rejected (CPU)	Executed (HIP)	Verified PASSED (ok)
Qwen3-TTS Speech Decoder	Rebuilt graph per chunk	Retained graph + input restore	840.89 ms compute time
Higgs-Audio-TTS (4B Model)	Standard pipeline	HIP Fast Paths	1.81x Real-Time Speed (RTF 0.55)

All unit tests and end-to-end multi-model inference runs (Qwen3-TTS and Higgs-Audio-TTS) produced clean audio output with floating-point parity down to float epsilon (~4.65e-9).

How to use it

Use the specialized Nix development shell:

nix develop .#rocm-gfx1151

Or configure manually:

cmake -S . -B build/rocm-gfx1151 -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DENGINE_ENABLE_HIP=ON \
  -DGPU_TARGETS=gfx1151 \
  -DENGINE_HIP_STRIX_HALO_OPTIMIZATIONS=ON
cmake --build build/rocm-gfx1151 --parallel

Run audiocpp_cli normally with the HIP backend. The decoder cache activates automatically for the specialized build when the runtime backend is HIP.

For generic ROCm behavior, use nix develop .#rocm or leave ENGINE_HIP_STRIX_HALO_OPTIMIZATIONS=OFF.

AI usage disclosure

AI was used for the following purposes:

• understand project structure and validate hypothesis about code organization
• help defining benchmarking scripts (not included in the PR)
• clean-up code after manual definition

AI was NOT used for the following purposes:

• optimization ideas
• basic implementation
• initial exploration of similar projects like llama-cpp and CrispASR

[0xShug0]

@fedeizzo Thanks for the update. The overall direction looks good, but I found a few things that should be fixed before merge.

1. CMakeLists.txt adds tests/unittests/test_qwen3_tts_options_config.cpp, but that file is missing. I validated
   ENGINE_BUILD_TESTS=ON + AUDIOCPP_MODEL_SET=custom + AUDIOCPP_MODELS=qwen3_tts fails at CMake generate with Cannot find source file.

2. It regresses Metal ConvTranspose1d fast-path eligibility. Current main allows Metal for ConvTranspose1d col2im fast path based on the tests in https://github.com/0xShug0/audio.cpp/blob/main/docs/reports/metal_convtranspose_performance.md . PR changes it to uses_ggml_cuda_family_backend(), which returns only CUDA/HIP. Note that it's not safe to allow Metal to use FastPackedProjection4Module.

3. Maybe rename uses_ggml_cuda_family_backend to something more explicit, such as uses_ggml_cuda_or_hip_backend?

I'll test Qwen3-related changes later.

Update: The Qwen3-related changes in this PR look safe based on my tests.

[fedeizzo]

@0xShug0 done 👍

[0xShug0]

@fedeizzo Thanks! I'm just waiting for the Nix CPU check. It keeps failing...

Update: Merged!