Support configurable mask grids in RF-DETR Triton postprocessing

docs/instance_segmentation_mask_resolution.md

@@ -111,8 +111,20 @@ the new opt-in flag.
 Values from 0 to 1 interpolate model-grid and image-grid dimensions. A lower
 factor does not always mean a smaller grid or faster execution. `accurate`
 selects factor 1.0, `fast` selects 0.0, and only `tradeoff` reads `tradeoff_factor`.
-Damian's Triton optimization is separate; factors other than 1.0 currently use
-the fallback post-processor.
+The RF-DETR Triton RLE post-processor supports native and upsampled mask grids
+at every factor. Boxes and masks share the encoded grid in `mask_size`, including
+for empty and deferred results. It preserves fractional float32 boxes on grids
+that differ from the image and rounds to integers only on the image grid.
+It uses the same target-size rounding as the reference path. Antialiased downsampling, unsupported preprocessing
+transforms, and other existing compatibility limits use the reference post-processor.
+The existing Triton enablement flag still controls this path; no execution plan is
+required. Sparse-record capacity limits are unchanged, including the existing
+deferred-mode overflow error.
+
+The first request for a new mask grid may compile Triton kernels and initialize
+interpolation tables. Warm each expected image shape and mask-resolution factor
+before serving latency-sensitive requests. Warm inference measurements exclude
+this first-use time and memory cost.
 
 Workflow versions v1–v4 retain their compatibility guards. Workflow v5,
 reduced-mask downstream handling, and full tensor support are deferred to a

inference_models/docs/changelog.md

@@ -72,10 +72,15 @@
   COCO `size`. Manually constructed reduced dense detections must supply
   `image_size` for image-space conversion because it cannot be inferred from the
   reduced tensor.
-- The RF-DETR fused Triton post-processor supports factor `1.0` only. Other
-  factors use the fallback post-processor so the requested resolution is honored.
-  Reduced-resolution Triton support and its benchmarking are deferred to a
-  separate PR.
+- The RF-DETR fused Triton RLE post-processor now honors `masks_resolution_factor`
+  for native and upsampled mask grids, including multiclass and deferred streaming
+  results. Boxes and masks share the selected grid, with float32 fractional boxes
+  when that grid differs from the image and final integer rounding otherwise.
+  Empty and deferred results preserve the same grid and box dtype. Antialiased
+  downsampling and previously unsupported transforms use the reference path.
+  Multiclass results respect RF-DETR's query-count cap even when `max_detections`
+  is larger. Multiclass and deferred selection use global top-k before remapping,
+  without a four-class-per-query cutoff, and encode each selected query mask once.
 - The `inference_models` adapter now rejects unknown `mask_decode_mode` values
   with `InvalidMaskDecodeArgument`; it previously ignored the mode. The legacy
   backend already rejected invalid modes. Validation also runs without opt-in.