ominix-cuda is the CUDA fork of OminiX-Ascend. The production stack is native C++ on ggml/ggml-cuda for NVIDIA GB10-class Blackwell systems.

No Python, PyTorch, diffusers, transformers, vLLM, or sglang is used in the production inference path. Dev-only conversion and inspection scripts may exist, but the runtime target is C++/CUDA direct dispatch.

Three authentic CUDA-native deliverables are in hand:

• QIE-Edit cat PNG — make the cat smile at 1024² / 20-step Euler on GB10 #2: 2:45 wall (165 s) via ominix-diffusion-cli --diffusion-fa (7.4× over the no-FA baseline). Cossim 1.0000 vs Ascend at n=1.
• TTS audio — text → 24 kHz WAV end-to-end via TalkerCudaEngine + CodePredictor + SpeechTokenizerDecoder vocoder. RTF 0.62 steady-state (1.74 cold). Sampling defeats greedy mode collapse; cross-prompt and cross-seed audio decorrelate.
• ASR transcript — Ellen audiobook 9.36 s WAV → 43-token plausible audiobook intro. RTF 0.37 wall. Audio encoder cossim 1.000000 across all 11 stages vs HF Python reference.

Headline numbers:

• TalkerCudaEngine: 53.85 TPS autoreg (Phase 2.3); 50.55 TPS with CUDA Graphs (28L Qwen3 body, F16, GEMM-bound)
• CodePredictor: 944 TPS hot loop (Phase 2.4); 10.3× wall reduction at Phase 2.9 via device F16 LM-head + CUDA Graphs
• Native QIE-Edit DiT: cossim 1.0000 vs Ascend at n=1, 1024² (byte-parity)
• Production CLI QIE-Edit: 7.53 s/step with --diffusion-fa (was 60.0 s/step naive F32)

The three demo scripts wrap the production smoke commands. They require SSH access to the GB10 hosts (zgx-3675, zgx-5b44) and the model assets in place — see SHIP_SUMMARY.md §4–§6 for build commands, model paths, and SSH config.

# Image edit (1024² / 20-step Euler, --diffusion-fa enabled, ~2:45 wall):
./scripts/demos/run_qie_edit.sh ./inputs/cat.jpg "make the cat smile"

# TTS (text → 24 kHz mono WAV, RTF 0.62 steady-state):
./scripts/demos/run_tts.sh "Hello world, this is the ominix CUDA TTS."

# ASR (WAV → transcript, RTF 0.37):
./scripts/demos/run_asr.sh ./samples/ellen_ref.wav

# Run all three back-to-back:
./scripts/demos/run_all.sh

• ggml/src/ggml-cuda/ — CUDA backend vendored from upstream ggml-org/llama.cpp tag b8532. Local CONV_1D and FLIP CUDA extensions preserved.
• tools/ominix_diffusion/ — stable-diffusion.cpp based image generation and image-edit CLI (ominix-diffusion-cli).
• tools/qwen_tts/ — TalkerCudaEngine (28L Qwen3), CodePredictor (5L Qwen3), SpeechTokenizerDecoder vocoder (RVQ + 2× upsample + 4 vocoder blocks + tanh).
• tools/qwen_asr/ — AsrCudaEngine + AudioEncoderCudaEngine; reuses TalkerCudaEngine verbatim for the 28L decoder.
• tools/qwen_image_edit/ — native ImageDiffusionCudaEngine (60-block DiT, F32-widened residual chain).
• docs/contracts/OMINIX_CUDA_CONTRACT.md — project contract and acceptance gates.
• SHIP_SUMMARY.md — full per-phase receipts, perf tables, build commands per host, run commands, model paths, known gaps.

Target hosts: GB10 #1 zgx-3675 (port 6222) for TTS/ASR, GB10 #2 zgx-5b44 (port 6022) for QIE-Edit. Ubuntu 24.04 aarch64 + CUDA 13.0.88 + sm_121a (Blackwell-A).

cmake -B build \
  -DGGML_CUDA=ON -DGGML_CANN=OFF \
  -DQWEN_TTS_CUDA=ON -DQWEN_IMAGE_EDIT_CUDA=ON \
  -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc)

CMake auto-promotes CMAKE_CUDA_ARCHITECTURES=121 to 121a. Expected linkage: libcudart.so.13, libcublas.so.13, libcublasLt.so.13, libcuda.so.1. No libascend*.

Phase 1 uses ominix-diffusion-cli as the CUDA sd.cpp baseline; production cat PNGs run through this CLI with --diffusion-fa enabled. Phases 2–4 added CUDA-native direct-dispatch engines:

• Phase 2 — TalkerCudaEngine + CodePredictor + SpeechTokenizerDecoderCudaEngine (text → 24 kHz speech)
• Phase 3 — ImageDiffusionCudaEngine (cossim 1.0000 vs Ascend at n=1; native E2E gated on cuDNN FMHA)
• Phase 4 — AsrCudaEngine + AudioEncoderCudaEngine (WAV → text, audio encoder byte-parity vs HF Python)

• Phase 2.6 (FP8/INT8 cuBLASLt) — deferred. Path to the 80–150 fps Talker contract bracket; CUDA Graph plumbing already in place.
• Phase 3.7 (text encoder + VAE FlashAttention) — next perf lever to clear the 140 s Phase 1 gate; ~50 s of the current 165 s wall is text encode + VAE encode/decode.
• Phase 4.6 (CPU mel parity) — currently 0.80 cossim vs HF Python (window/log-scale divergence); non-blocking, bypassed via OMINIX_ASR_USE_MEL_BIN when bit-parity matters.
• Native QIE-Edit E2E — naive F32 attention (~960 ms/block); gated on a future cuDNN FMHA pass for native-engine production runs.

See SHIP_SUMMARY.md for full per-phase receipts, perf numbers, build commands per host, and operational notes.