[diffusion] docs: add verified DGX Spark recipe for Qwen-Image 2.1 (#40487)

Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>

docs/cookbook/diffusion/Qwen-Image/Qwen-Image-2.1.mdx

@@ -14,9 +14,10 @@ import { config } from '/src/snippets/configs/Qwen/qwen-image-2.1.jsx';
 
 Install the runtime dependencies with `uv pip install "sglang[diffusion]" --prerelease=allow`,
 then install this integration from its source checkout with
-`uv pip install -e "python[diffusion]"`. Use an authorized checkpoint directory in
-place of `/models/qwen-image-2.1`. The recipes below target NVIDIA CUDA on Linux;
-the hardware picker selects a tested single-GPU recipe for the full checkpoint.
+`uv pip install -e "python[diffusion]"`. The picker uses `Qwen/Qwen-Image-2.1`;
+you can also set a local checkpoint directory under **Variables**. The recipes
+target NVIDIA CUDA on Linux; the picker marks which single-GPU workloads have
+been verified with the full checkpoint.
 
 <Deployment config={config} />
 
@@ -27,7 +28,7 @@ steps, and output count. Set reference PNG paths under **Variables**; edits
 upload files from the machine running cURL, so they need not exist on the server.
 
 Hardware selection applies the recommended placement for that GPU. H200,
-B200, and RTX PRO 6000 96GB keep weights resident; RTX 5090 and RTX 4090
+B200, RTX PRO 6000 96GB, and DGX Spark keep weights resident; RTX 5090 and RTX 4090
 offload selected components to fit the full pipeline.
 Untested topologies and feature combinations remain selectable and are labeled
 **Unverified**. Invalid topology combinations disable Copy. This integration
@@ -58,6 +59,7 @@ execution, and full-image VAE decoding.
 | B200 192GB | Resident / FlashAttention | 2.46 s | 3.02 s | 38.5 GiB |
 | RTX PRO 6000 96GB | Resident / Torch SDPA | 8.03 s | 9.63 s | 38.4 GiB |
 | RTX 4090 24GB | DiT and VAE resident, encoder layerwise offload / FlashAttention | 18.68 s | 21.68 s | 22.7 GiB |
+| DGX Spark 128GB unified | Resident / Torch SDPA | 35.36 s | 42.23 s | — (unified) |
 
 Measured on 2026-09-20 at 1024×1024, 40 steps, CFG 1, and one RGBA PNG per
 request. Times are median HTTP latency after warmup, including PNG serialization
@@ -68,6 +70,25 @@ RTX 5090 uses DiT layerwise offload and Torch SDPA; its recipe has not been
 retested with the updated checkpoint. Both RTX 5090 and RTX PRO 6000 use SDPA
 when FlashAttention is selected in this runtime. CPU offload requires host RAM.
 
+### DGX Spark
+
+Select **DGX Spark** for one GB10 GPU on Linux ARM64 with CUDA 13. Use the
+source installation above. The recommended configuration keeps all components
+resident, uses native BF16/FP32 precision, and lets the runtime select Torch SDPA:
+
+```bash Command
+sglang serve \
+  --model-path Qwen/Qwen-Image-2.1 \
+  --performance-mode speed
+```
+
+The [128 GB unified memory](https://docs.nvidia.com/dgx/dgx-spark/hardware.html)
+is shared by the CPU and GPU. CPU offload is unnecessary for the verified
+single-image 1024×1024 workload. Keep full-image VAE decoding and eager execution.
+Generation, editing, transparent generation, and transparent editing were
+verified with PyTorch 2.13.0+cu130. Spark reports no separate VRAM usage in `nvidia-smi`.
+This recipe covers one Spark; multi-node deployment and batching remain unverified.
+
 ### Batching
 
 Keep **Request batching → Off** and **Outputs → 1** for interactive use.
@@ -203,7 +224,7 @@ See the [GGUF guide](/docs/sglang-diffusion/quantization#gguf).
 NVFP4 requires Blackwell and compatible ModelOpt exports. Select the component
 directories using `--component-paths.transformer` and/or
 `--component-paths.text_encoder`. Keep the FlashInfer backend at `auto` on
-RTX 5090 and RTX PRO 6000: TensorRT-LLM FP4 GEMM does not support SM120.
+RTX 5090, RTX PRO 6000, and DGX Spark: TensorRT-LLM FP4 GEMM does not support SM12.x.
 These GPUs remain unverified for this model's NVFP4 exports. See the
 [NVFP4 guide](/docs/sglang-diffusion/quantization#modelopt-nvfp4).