RT-DETRv2-S on a Pixel 8a — both transformer graphs on CompiledModel GPU
RT-DETRv2 (Baidu, 2024 — PekingU/rtdetr_v2_r18vd) object
detection, converted to LiteRT and running 100% on the CompiledModel GPU (ML Drift) on a phone,
with no CPU/ONNX fallback.
RT-DETRv2 is a transformer detector (ResNet18-vd backbone + a hybrid AIFI/CCFM encoder + a plain
deformable-attention DETR decoder). Off-the-shelf it is GPU-incompatible (deformable grid_sample →
GATHER_ND, two-stage query selection → TOPK/GATHER). Here it is converted with litert-torch and
split into two GPU graphs with a host step between them, so both transformer graphs run on the GPU.
image[1,3,640,640]
→[GPU Graph A]→ enc_class, memory_raw
→[host: top-300 by max class score; per-token tail on the 300 selected (fp32):
target = enc_output(valid·memory_raw) (Linear + LayerNorm)
ref = enc_bbox_head(target) + anchors (3-layer MLP)]
→[GPU Graph B (memory_raw, target, ref)]→ boxes[1,300,4], logits[1,300,80]
→[host: sigmoid + threshold + cxcywh→xyxy + light NMS]→ detections
The two-stage query selection (TOPK/GATHER) has no GPU op, but the proposal grid is
image-independent, so the model splits there. The per-token tail (enc_output + enc_bbox_head) runs on
the host over the 300 selected tokens (exact, since per-token ops commute with the gather).
Why the per-token tail is on the host — a Mali 3D-sequence fan-out bug
Both graphs convert GPU-clean, but a naïve Graph A (emitting enc_class/enc_coord/output_memory/memory_raw
together) silently produced wrong boxes on device — large objects vanished while small ones stayed
perfect. A 3-D token tensor [1,N,256] (from conv.flatten(2).transpose(1,2)) that is both a graph
output and consumed by another node — or that fans out to several consumers — gets clobbered on the
longer branch (4-D conv-map outputs are fine). output_memory fed both heads; the 3-layer box head lost,
so its reference-box deltas collapsed to ~0. Fix: Graph A emits only the two fp16-clean leaves
(enc_class + memory_raw×2) and the per-token tail moves to the host.
Minimal usage
Android (Kotlin, CompiledModel GPU)
kotlin
1val ga = CompiledModel.create(context.assets,"rtdetr_graphA_fp16.tflite",2 CompiledModel.Options(Accelerator.GPU),null)3val gb = CompiledModel.create(context.assets,"rtdetr_graphB_fp16.tflite",4 CompiledModel.Options(Accelerator.GPU),null)5val aIn = ga.createInputBuffers();val aOut = ga.createOutputBuffers()6val bIn = gb.createInputBuffers();val bOut = gb.createOutputBuffers()7aIn[0].writeFloat(chw)// [1,3,640,640] RGB in [0,1], NCHW8ga.run(aIn, aOut)// -> enc_class[1,8400,80], memory_raw*2[1,8400,256]9// host step: /2 -> top-300 -> per-token tail (host_params.bin) -> target[1,300,256], ref[1,300,4]10// (resolve buffer slots by float size; full math in the Python below / litert-samples object_detection)11bIn[0].writeFloat(memory); bIn[1].writeFloat(target); bIn[2].writeFloat(ref)12gb.run(bIn, bOut)13val boxes = bOut[0].readFloat()// [1,300,4] cxcywh in [0,1]14val logits = bOut[1].readFloat()// [1,300,80] -> sigmoid + threshold + light NMS
Python (desktop verification)
python
1import numpy as np
2from PIL import Image
3from ai_edge_litert.interpreter import Interpreter
45NP_, NQ, NC, H =8400,300,80,2566img = Image.open("photo.jpg").convert("RGB").resize((640,640))7x =(np.asarray(img, np.float32)/255.0).transpose(2,0,1)[None]# [1,3,640,640], [0,1] only89# host_params.bin (fp32 LE): enc_output W[256,256],b,gamma,beta · bbox-MLP W0,b0,W1,b1,W2[4,256],b2 · valid[8400] · anchors[8400,4]10p = np.fromfile("host_params.bin", np.float32); o =011deftake(*s):12global o; n =int(np.prod(s)); v = p[o:o+n].reshape(s); o += n;return v
13eoW, eoB, eoG, eoBe = take(H, H), take(H), take(H), take(H)14W0, b0, W1, b1, W2, b2 = take(H, H), take(H), take(H, H), take(H), take(4, H), take(4)15valid, anchors = take(NP_), take(NP_,4)1617defrun(path, feeds):# feed/fetch tensors by shape (converter slot order is arbitrary)18 it = Interpreter(model_path=path); it.allocate_tensors()19for d in it.get_input_details(): it.set_tensor(d["index"], feeds[tuple(d["shape"][1:])])20 it.invoke();return{tuple(d["shape"][1:]): it.get_tensor(d["index"])for d in it.get_output_details()}2122a = run("rtdetr_graphA_fp16.tflite",{(3,640,640): x})23enc_cls, mem = a[(NP_, NC)][0], a[(NP_, H)][0]/2.0# Graph A emits memory_raw*2 — undo2425top = np.argsort(-enc_cls.max(-1))[:NQ]# top-300 by max class logit26t =(valid[top,None]* mem[top]) @ eoW.T + eoB # per-token tail: enc_output Linear...27t =(t - t.mean(-1, keepdims=True))/ np.sqrt(t.var(-1, keepdims=True)+1e-5)* eoG + eoBe # ...+ LayerNorm28h = np.maximum(t @ W0.T + b0,0); h = np.maximum(h @ W1.T + b1,0)29ref = h @ W2.T + b2 + anchors[top]# enc_bbox_head MLP + anchors3031b = run("rtdetr_graphB_fp16.tflite",32{(NP_, H): mem[None],(NQ, H): t[None].astype(np.float32),(NQ,4): ref[None].astype(np.float32)})33boxes, logits = b[(NQ,4)][0], b[(NQ, NC)][0]# cxcywh in [0,1] / 80-way logits34labels =open("coco_labels.txt").read().splitlines()35score =1/(1+ np.exp(-logits.max(-1))); cls = logits.argmax(-1)36for q in np.where(score >0.4)[0]:# + light NMS (IoU 0.7) in a real app37 cx, cy, w, hh = boxes[q]38print(f"{labels[cls[q]]:12s}{score[q]:.2f} xyxy=({cx-w/2:.3f},{cy-hh/2:.3f},{cx+w/2:.3f},{cy+hh/2:.3f})")
On-device (Pixel 8a, Tensor G3 — verified)
Both graphs run 100% GPU-resident (LITERT_CL): Graph A fully delegated, Graph B 704/704. The device
chain reproduces the PyTorch detections exactly — COCO val giraffe image 7/7, cats image
(000000039769) 6/6, every box at IoU 0.98–1.00 with matching class and score.
End-to-end ~615 ms/frame on a Pixel 8a: Graph B's deformable decoder over RT-DETR's 8400 tokens /
80×80 levels is ~350 ms of GPU compute (the GATHER-free tent-matmul grid_sample turns an O(points)
gather into an O(H·W) matmul). So this model is accurate and fully-GPU but not real-time on this
device; it suits still-image / snapshot detection. (A real-time camera demo of the same family is
RF-DETR Nano, whose single small
deformable level runs at ~9 fps.)
Preprocessing / outputs
Input: square resize to 640×640, RGB, [0,1] rescale only (no ImageNet normalization), NCHW.
Output: Graph B boxes are cxcywh normalized to [0,1]; logits are 80-way (contiguous COCO id 0–79). Host applies sigmoid + score threshold + cxcywh→xyxy + light NMS.
Conversion notes
Converted with litert-torch (NCHW preserved — onnx2tf destroys ViT attention). Re-authoring
(per-graph tflite-vs-torch correlation 1.0): deformable grid_sample → a GATHER/CAST-free tent-matmul,
MSDeformAttn ≤4D, baked AIFI sine pos-embed, ResNet18-vd stem zero-pad maxpool (the -inf-pad maxpool
lowers to a Mali-rejected PADV2), a down-scaled fp16-safe LayerNorm, and the 3D-fan-out fix above
(emit clean leaves + host-side per-token tail).
A runnable Android sample (CompiledModel GPU) and the conversion scripts are in the official
ai-edge-litert/litert-samplesobject_detection
example.
Performance
Measured on a Pixel 8a (Tensor G3, Android 16) with the standard TFLite benchmark_model tool — 10 warm-up runs then 50 timed runs, reported as the tool's mean.
Any on-device figure recorded when this model shipped came from a different runtime. It was taken through LiteRT's own CompiledModel accelerator (logcat reports it as LITERT_CL), which is the path the Kotlin sample app and the LiteRT API use, and it appears elsewhere on this card. The rows above are the classic TFLite OpenCL delegate, measured with a tool anyone can download and re-run. The two are not comparable, so read the rows above as a reproducible floor rather than as this model's speed on LiteRT.
On this delegate the CPU is the faster choice for rtdetr_graphA_fp16.tflite (548.0 ms on CPU against 925.2 ms on GPU) — worth knowing before you reach for the GPU on a mid-range phone.
Note that the GPU does not take the whole graph here (184 / 365 in rtdetr_graphA_fp16.tflite, 95 / 704 in rtdetr_graphB_fp16.tflite); the remainder runs on the CPU and the split costs a per-partition round trip.
Snapdragon NPU (Hexagon)
rtdetr_graphA_fp16.tflite — the NPU is 1.09x faster than the GPU (21.13 ms against 23.07 ms) and loads 18.15x faster (122 ms against 2218 ms).
rtdetr_graphB_fp16.tflite — the NPU is 1.26x faster than the GPU (132.6 ms against 166.6 ms) and loads 8.22x faster (220 ms against 1808 ms).
file
backend
compiled
inference (median / min)
load
rtdetr_graphA_fp16.tflite
NPU (Hexagon v81)
on-device JIT
21.13 ms / 19.24 ms
122 ms
rtdetr_graphA_fp16.tflite
GPU (Adreno)
—
23.07 ms / 19.95 ms
2218 ms
rtdetr_graphB_fp16.tflite
NPU (Hexagon v81)
on-device JIT
132.6 ms / 124.4 ms
220 ms
rtdetr_graphB_fp16.tflite
GPU (Adreno)
—
166.6 ms / 165.7 ms
1808 ms
Measured on a Samsung Galaxy S26 (Snapdragon 8 Elite Gen 5 / SM8850, Hexagon v81, Android 16) with LiteRT CompiledModel 2.2.0, one accelerator per process, 5 warm-up runs then N=50 timed runs, median reported. Every run held thermal status NONE throughout. Headroom 0.76–0.81, where 1.0 is the throttling threshold.
The NPU rows ran the published file unchanged. LiteRT compiled it for the Hexagon on the device at first load. Those first compiles took 5.8 s to 113 s here. The load column above is the cached load every later run pays. Recipe and the runtime libraries it needs: NPU guide.
Measured on a Raspberry Pi 5 Model B Rev 1.1 (8 GB, Raspberry Pi OS 64-bit) with the LiteRT benchmark_model tool from litert-cli-nightly 0.2.0.dev20260805: CPU inference (XNNPACK, 4 threads), 3 invocations per file of 10 warm-up plus 50 timed runs (the tool caps a phase at 150 s, so very slow graphs run fewer — the Runs column is the actual timed total). The latency is the median across invocations; the spread is the min–max over all timed runs. No thermal throttling occurred during these runs (vcgencmd get_throttled stayed 0x0).
