Support RF-DETR Triton postprocess through 2XLarge

development/stream_interface/rfdetr_seg_trt_1080_benchmark.md

@@ -0,0 +1,70 @@
+# RF-DETR Seg TensorRT 1080p Variant Benchmark
+
+This note records the June 4, 2026 check for the largest RF-DETR segmentation
+variant that can run the `vehicles_1080p.mp4` stream workflow at 30 FPS on the
+Jetson Orin NX 8GB target used for PR 2405.
+
+## Context
+
+The public non-nano RF-DETR segmentation TensorRT packages are built for L4/T4,
+so they are not directly loadable on Jetson Orin. For this benchmark, local Orin
+FP16 TensorRT packages were compiled from the public ONNX packages and wired into
+the workflow as untracked local directories.
+
+The Triton sparse RLE postprocess path previously rejected non-nano mask sizes
+because it scanned the source mask with one Triton vector and capped source mask
+area below the `small` model's 96x96 mask. The current patch adds a tiled source
+mask bounds pass and raises the supported sparse path shape limit to RF-DETR Seg
+2XLarge's 192x192 mask with 300 queries and COCO class logits.
+
+## Benchmark Command
+
+Use the stream workflow with the optimization flags enabled:
+
+```bash
+env \
+  PYTHONPATH=/app/helloworld/inference/inference_models:/app/helloworld/inference \
+  USE_INFERENCE_MODELS=True \
+  ALLOW_INFERENCE_MODELS_UNTRUSTED_PACKAGES=True \
+  ALLOW_INFERENCE_MODELS_DIRECTLY_ACCESS_LOCAL_PACKAGES=True \
+  INFERENCE_MODELS_RFDETR_TRITON_POSTPROC_ENABLED=true \
+  INFERENCE_MODELS_RFDETR_TRITON_PREPROC_ENABLED=true \
+  RFDETR_PIPELINE_DEPTH=2 \
+  ENABLE_AUTO_CUDA_GRAPHS_FOR_TRT_BACKEND=true \
+  python development/stream_interface/rfdetr_nano_seg_trt_workflow.py \
+    --video_reference vehicles_1080p.mp4 \
+    --model_id rfdetr-seg-large/1 \
+    --backend trt
+```
+
+Change `--model_id` to the local package alias for each variant. A depth-3
+sanity run was also performed for `xlarge`.
+
+## Results
+
+| Variant | Input size | Pipeline depth | FPS |
+| --- | ---: | ---: | ---: |
+| `rfdetr-seg-small/1` | 384 | 2 | 63.85 |
+| `rfdetr-seg-large/1` | 504 | 2 | 35.49 |
+| `rfdetr-seg-xlarge/1` | 624 | 2 | 20.94 |
+| `rfdetr-seg-xlarge/1` | 624 | 3 | 20.91 |
+| `rfdetr-seg-2xlarge/1` | 768 | 2 | 12.90 |
+
+`large` is the largest tested non-nano RF-DETR Seg variant that clears 30 FPS on
+this 1080p workload with all optimization flags enabled. `xlarge` remains below
+30 FPS even when increasing pipeline depth from 2 to 3.
+
+## Verification
+
+The focused postprocess test suite passed after the 2XLarge shape-limit patch:
+
+```bash
+PYTHONPATH=/app/helloworld/inference/inference_models:/app/helloworld/inference \
+  python -m pytest tests/unit_tests/models/rfdetr/test_triton_postprocess.py
+```
+
+Result:
+
+```text
+24 passed, 23 warnings
+```