Merge branch 'develop' into feat/1066-executorch_export

Author: Borda

.github/workflows/ci-deps-resolution.yml

@@ -0,0 +1,52 @@
+name: Dependency resolution
+
+on:
+  push:
+    branches: ["main", "release/*", "develop"]
+    paths: ["pyproject.toml"]
+  pull_request:
+    paths: ["pyproject.toml"]
+
+permissions:
+  contents: read
+
+defaults:
+  run:
+    shell: bash
+
+concurrency:
+  group: deps-resolution-${{ github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
+jobs:
+  tflite-resolution:
+    name: Resolve [tflite] extra — Python ${{ matrix.python-version }}
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    strategy:
+      fail-fast: false
+      matrix:
+        # tflite deps carry python_version>='3.12' and <'3.13' markers; onnx2tf pins numpy==1.26.4
+        # exactly, which conflicts with ml-dtypes>=0.5.1 requiring numpy>=2.1.0 on 3.13.
+        # Add 3.13 and relax the <'3.13' upper bound when onnx2tf drops the exact numpy pin.
+        python-version: ["3.12"]
+    steps:
+      - name: 📥 Checkout the repository
+        uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6.0.1
+
+      - name: 🐍 Install uv and set Python version ${{ matrix.python-version }}
+        uses: astral-sh/setup-uv@61cb8a9741eeb8a550a1b8544337180c0fc8476b # v7.2.0
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: 🔍 Resolve [tflite] optional dependencies
+        run: |
+          uv pip compile pyproject.toml \
+            --extra tflite \
+            --python-version ${{ matrix.python-version }} \
+            --no-header \
+            --quiet
+
+      - name: Minimize uv cache
+        continue-on-error: true
+        run: uv cache prune --ci

CHANGELOG.md

@@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Fixed
+
+- `HungarianMatcher.forward()` now uses the configured `focal_alpha` in the focal classification matching cost. Previously the value was hardcoded to `0.25`, silently ignoring any non-default `focal_alpha` passed to the constructor or `build_matcher`. This misaligned the bipartite matching cost with the focal classification loss in `criterion.py`, which correctly used `self.focal_alpha`. ([#1147](https://github.com/roboflow/rf-detr/pull/1147))
+- `spatial_shapes` in `Transformer.forward()` is now built from symbolic `Shape` ops (`torch.stack` of per-level `torch._shape_as_tensor` slices) instead of `torch.empty` + in-place index assignment. The previous pattern emitted a `ScatterND` feeding a shape tensor (`level_start_index`), which TensorRT rejected with "IScatterLayer cannot be used to compute a shape tensor". This fix is required to export any RF-DETR model to a TensorRT engine. ([#1155](https://github.com/roboflow/rf-detr/pull/1155))
+
 ---
 
 ## [1.8.1] — 2026-06-19

docs/cookbooks/cards.yaml

@@ -1,4 +1,10 @@
 cards:
+  - href: inference-latency-benchmark/
+    name: "Inference Latency Benchmark"
+    labels: [INFERENCE, ONNX, GPU]
+    version: v1.8.1
+    author: Borda
+    description: "Benchmark detection, segmentation, and keypoint models across FP32, FP16+JIT, and ONNX Runtime — measures latency and FPS on GPU using CUDA events."
   - href: fine-tune_keypoints/
     name: "Fine-Tune RF-DETR Keypoint Detection"
     labels: [TRAINING, PYTORCH LIGHTNING, INFERENCE]