This section preserves the 2026-09-22 audit. For the latest per-cell disposition and 2026-09-23 follow-up evidence, use the [current 60-cell E2E matrix](../../benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md). The HX370 AMD NPU now has a placement- and quality-verified **Spark 1.7B output-head component** through Omni, while complete Spark generation on that route remains unverified; see the [component record](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_amd_npu_output_head/README.md). Native Windows CPU and Radeon 890M completed [scoped whole-model Spark text requests](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/README.md) with the same Q4_K_M GGUF in standalone llama.cpp, then [complete requests through a bounded Omni stage](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_omni_llamacpp/README.md). The Omni route verifies pinned artifacts, actual placement, one-slot memory admission, terminal events and cancellation drain. Incremental streaming, restart after cancellation, memory peaks and broader quality remain open. These results advance the M3 hardware/backend evidence without claiming a joint multi-accelerator plan.
The [Qwen3-TTS CPU+Radeon whole-session path](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/README.md) also completed two named text-to-audio inputs and 20 serial Omni requests with exact PCM parity to its standalone record. It verifies pinned executable and weights, logged hybrid placement, one-slot shared-RAM admission, terminal audio and cancellation drain. Playable streaming, post-cancel restart, loading/iGPU peaks, sustained use and broader speech quality remain open; RTF still exceeds 1. No joint CPU+iGPU+NPU claim follows from this result.
The [native Windows CPU Qwen3-TTS Omni path](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_cpu/README.md) completed two named text-to-WAV inputs and 20 serial Omni requests with exact PCM parity to the standalone PyTorch record. It verifies pinned weights, CPU-only placement, a 10 GiB host reservation, a 4 GiB pre-load refusal, terminal audio and cancellation drain. Request-wall p50/p95 were 15.32/15.60 s for 4.56 s of audio. Playable streaming, post-cancel restart, loading peak, broad voice quality and sustained use remain open. This CPU route does not qualify AMD or mobile execution.
| Spark-X2.5-1.7B INT8 | CPU execution; accelerators unused | WSL / vLLM CPU | Fresh 12-prompt acceptance passed on this branch with the existing vLLM 0.28 CPU wheel; historical token parity is separate. |
| Spark-X2.5-1.7B INT8 | CPU execution; no NVIDIA dependency | Native Windows | Model path not yet qualified; CPU Add graph success is not text-model support. |
| Qwen3-TTS-12Hz-0.6B CustomVoice | CPU + RTX 5090 Laptop GPU; both stages on GPU | WSL / vLLM CUDA | Current branch: real-weight audio streaming completes; playback stalls and remaining M2 gates stay open. |
| Qwen3-TTS-12Hz-0.6B CustomVoice | CPU + RTX 5090 Laptop GPU; both stages on GPU | Native Windows / vLLM CUDA | Current branch: real-weight audio streaming completes; playback stalls and remaining M2 gates stay open. |
| Qwen3-TTS-12Hz-0.6B CustomVoice, Q8_0 GGUF talker + F16 codec | CPU + Radeon 890M iGPU; code predictor on CPU FP32, talker/codec on Radeon | Native Windows / standalone CrispASR Vulkan | Two named text-to-WAV requests and a 20-request serial resident profile passed; [placement and quality evidence](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_radeon890m_joint/README.md). Still outside Omni, slower than real time and unqualified for streaming or sustained use. |
| Qwen3.8-27B vision tower, FP32 ONNX | CPU + Radeon 890M iGPU | Windows / ORT DirectML | Historical component validation only; not the complete 27B model or a newly qualified graph-stage vision-to-language pipeline. |
| Qwen3.8-27B vision tower, tested A16W8 exports | CPU + AMD NPU | Windows / ORT VitisAI | Tested whole-tower artifacts were rejected by NPU partitioning. This does not rule out other exports or smaller components. |
| Qwen3.8-27B complete multimodal pipeline | CPU/iGPU/NPU, with optional discrete GPU | Target-specific | Fresh public Omni startup fails pipeline resolution on Windows/WSL; earlier bare-vLLM short-text loading is separate and not multimodal validation. |
The abstraction supports these single-machine compositions. This table separates
topology/admission support from successful execution of a particular model; it
does not promise that every listed accelerator can run every model.
| Composition | Resource accounting | Current execution evidence |
|---|---|---|
| CPU only | Host RAM; WSL quota when applicable | ORT CPU graphs through public Omni on Windows and WSL. Native Windows CPU text remains unqualified. |
| CPU + integrated GPU | Shared host RAM | Radeon 890M DirectML graph through Omni. This extraction's probe is not a complete vision/language pipeline. |
| CPU + integrated NPU | Shared host RAM | AMD NPU VitisAI graph through Omni; permitted CPU partitions are explicit in the placement report. |
| CPU + integrated GPU + integrated NPU | Shared RAM plus optional observed package/power/bandwidth relationships | Topology and admission contracts tested; individual laptop routes execute. Simultaneous full-model overlap is not qualified. |
| CPU + integrated GPU + integrated NPU + discrete GPU | Shared RAM and separate discrete VRAM | Current Ryzen AI 9 HX370 / Radeon 890M / AMD NPU / RTX 5090 Laptop machine. Spark CUDA text and native two-stage TTS run on Windows and WSL; using all accelerators together is not qualified. |
| CPU + discrete GPU, with integrated devices unused or absent | Host RAM plus discrete VRAM | Native CUDA text/TTS execution uses this subset. NVIDIA is optional in the abstraction. |
Mobile/embedded SoCs fit the CPU + integrated GPU + NPU composition, but their
local controller and model paths still require device-specific qualification.
Galaxy S25 AI Hub inference/profiling passed for one Spark attention component;
this is component evidence, not full local mobile generation or streaming.
AI Hub is a test facility, never a deployment dependency. Intel, Apple and other
vendor combinations are not certified by the AMD/NVIDIA/Qualcomm results.
`DeviceDescriptor` separates physical device identity, execution domain and
integration. A CPU, integrated GPU and NPU can all reference `machine:ram`.
An additional discrete GPU references its own `machine:vram:<physical-id>`.
A WSL allocation can consume both `machine:ram` and `wsl:quota` constraints;
those are two limits on the same allocation, not two amounts of physical RAM.
Use the same physical GPU ID when Windows and WSL expose the same adapter.
Package IDs, bandwidth groups and power domains are optional observed relationships.
Missing values mean unknown. Fitting memory does not authorize overlap on a shared
package: selection and overlap still require model quality and complete workload
measurements. Synthetic topology tests cover CPU-only and all four integrated /
discrete combinations; they are not hardware certification.
## Deploy a graph
Register a `PipelineConfig` with `StageExecutionType.GRAPH`, `model_stage="graph"`
and an appropriate final output type. Set `async_chunk=False`. A graph following
another stage must have a model-owned `custom_process_input_func` that returns
`{"tensors": {name: numpy_array}}`. Names, dtypes and shapes must match the
manifest's placement examples. Graph stages do not tokenize prompts or apply
model-specific transformations.
Graph-containing v1 pipelines must be a linear chain with exactly one terminal
output at the last stage. Fan-out, joins and multiple final outputs are refused
before loading; those require reference-counted consumer leases.
Each stage in a graph-containing plan declares the same capacities and its own