NOTE: The Jupyter Notebook below is included in the Chimera SDK and can be run interactively by running the following CLI command:
$ quadric sdk notebook
From the Jupyter Notebook window in your browser, select the notebook named /quadric/sdk-cli/examples/models/yolo/yolor/yolor.ipynb.
YOLOR Object Detection on Chimera GPNPU
YOLOR (Wang, Yeh & Liao, 2021) is an anchor-based single-stage detector that adds an "implicit" knowledge branch alongside the standard explicit feature path — a shared backbone feeding multiple task heads. This notebook compiles the yolor_p6 variant end-to-end: pull the kadirnar/yolor-pip source, load the pretrained weights via huggingface_hub, export to ONNX via torch.onnx, split at the head boundary, quantize the backbone to INT8, compile the backbone with the Chimera Graph Compiler (CGC), and run ISS + ORT INT8 inferences side-by-side.
Why split the graph?
YOLOR's head fuses per-scale sigmoids and shape gymnastics that are cheaper to keep on the host CPU; the backbone is a dense stack of convolutions that compiles well to the Chimera GPNPU. Splitting the ONNX at the head boundary lets CGC focus on the convolutional backbone and lets ONNX Runtime handle the classical tail — each operator runs where it belongs.
Model: YOLOR-P6 (COCO, 640×640)
Note — split pipeline.
- On the Chimera GPNPU: the backbone
- On the host CPU: the detection head ONNX, NMS, and bounding-box visualization
The full pipeline can run end-to-end on the GPNPU — see
yolox_e2e_chipy.ipynborretina_net_e2e_chipy.ipynbfor examples that compose CCL custom ops with ChiPy to keep everything on-chip.
1. Setup
Imports and per-run configuration. YOLOR-pip source-tree download + patch, weights loading, ONNX export + split, calibration subset, and bounding-box visualization all live in yolor_helpers.py.
import gc
import os
## Suppress Ultralytics-style auto-pip-install if any transitive import trips it.
os.environ.setdefault("YOLO_AUTOINSTALL", "False")
import numpy as np
import onnx
from onnxruntime import InferenceSession
from PIL import Image
from sdk_cli.lib.inference import InferenceEngine, batch_inference
from sdk_cli.lib.quantize import QuantizedONNXModel, quantize_onnx_model
from tvm.contrib.epu.chimera_job.chimera_job import ChimeraJob
from yolor_helpers import (
DEFAULT_IMAGE_PATHS,
INPUT_SIZE,
MODEL_NAME,
build_calibration_subset,
display_detections,
export_onnx,
load_images,
run_postprocess,
split_backbone_and_head,
)
%matplotlib inline
2. Model Generation and ONNX Export
export_onnx() fetches the kadirnar/yolor-pip source tree (downloading and patching it if needed), loads the pretrained YOLOR-P6 weights via huggingface_hub, and writes yolor_p6.onnx via torch.onnx.export at the SDK's default opset.
onnx_file = export_onnx(MODEL_NAME)
print(f"Exported: {onnx_file.name}")
/usr/local/lib/python3.10/dist-packages/torch/functional.py:539: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at /pytorch/aten/src/ATen/native/TensorShape.cpp:3637.)
return _VF.meshgrid(tensors, **kwargs) # type: ignore[attr-defined]
/quadric/sdk-cli/examples/models/yolo/yolor/yolor/models/models.py:542: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
if not augment:
/quadric/sdk-cli/examples/models/yolo/yolor/yolor/utils/layers.py:124: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
if nx == na: # same shape
/quadric/sdk-cli/examples/models/yolo/yolor/yolor/models/models.py:396: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
if (self.nx, self.ny) != (nx, ny):
/quadric/sdk-cli/examples/models/yolo/yolor/yolor/models/models.py:360: TracerWarning: torch.tensor results are registered as constants in the trace. You can safely ignore this warning if you use this function to create tensors out of constant variables that would be the same every time you call this function. In any other case, this might cause the trace to be incorrect.
self.ng = torch.tensor(ng, dtype=torch.float)
/quadric/sdk-cli/examples/models/yolo/yolor/yolor/models/models.py:360: TracerWarning: Converting a tensor to a Python float might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
self.ng = torch.tensor(ng, dtype=torch.float)
Exported: yolor_p6.onnx
3. Split ONNX
YOLOR splits into a backbone (convolutional, GPNPU-friendly) and a head (anchor decode + per-scale detection ops, better on the host CPU). We cut the graph at the four module_list head-boundary edges and keep the backbone for CGC compilation; the head piece is run as an ONNX-Runtime reference.
backbone_onnx_file, head_onnx_file = split_backbone_and_head(onnx_file, MODEL_NAME)
print(f"Backbone: {backbone_onnx_file.name}")
print(f"Head: {head_onnx_file.name}")
Backbone: yolor_p6-backbone.onnx
Head: yolor_p6-head.onnx
4. Quantization
The backbone is quantized to INT8 with asymmetric activations using the COCO-like subset of QuadricCalibration for tensor-range statistics. The QuantizedONNXModel return value carries both the quantized ONNX and its companion .tranges file.
calibration_subset = build_calibration_subset(INPUT_SIZE)
quantized_onnx_model: QuantizedONNXModel = quantize_onnx_model(
str(backbone_onnx_file),
calibration_subset,
asymmetric_activation=True,
)
print(f"Quantized ONNX: {quantized_onnx_model.model_path.name}")
print(f"Tensor ranges: {quantized_onnx_model.tensor_ranges_path.name}")
gc.collect()
2026-06-19 03:55 - INFO - sdk - quantize - ONNX model shapes inferred.
2026-06-19 03:55 - DEBUG - sdk - quantize - Forcing node types: []
2026-06-19 03:55 - DEBUG - sdk - quantize - ONNX Node types excluded from quantization: ['Softmax', 'Sigmoid', 'QuadricCustomOp']
2026-06-19 03:55 - DEBUG - sdk - quantize - ONNX Node names excluded from quantization: ['', '/module_list.195/activation/Sigmoid', '/module_list.103/activation/Mul', '/module_list.82/activation/Mul', '/module_list.160/activation/Mul', '/module_list.40/activation/Mul', '/module_list.218/activation/Sigmoid', '/module_list.198/activation/Mul', '/module_list.190/activation/Mul', '/module_list.44/activation/Sigmoid', '/module_list.17/activation/Mul', '/module_list.97/activation/Mul', '/module_list.47/activation/Sigmoid', '/module_list.21/activation/Mul', '/module_list.179/activation/Mul', '/module_list.125/activation/Mul', '/module_list.166/activation/Mul', '/module_list.51/activation/Sigmoid', '/module_list.224/activation/Sigmoid', '/module_list.60/activation/Sigmoid', '/module_list.101/activation/Mul', '/module_list.126/activation/Sigmoid', '/module_list.140/activation/Sigmoid', '/module_list.79/activation/Sigmoid', '/module_list.45/activation/Sigmoid', '/module_list.159/activation/Sigmoid', '/module_list.132/activation/Sigmoid', '/module_list.169/activation/Mul', '/module_list.182/activation/Sigmoid', '/module_list.163/activation/Sigmoid', '/module_list.179/activation/Sigmoid', '/module_list.115/activation/Sigmoid', '/module_list.86/activation/Mul', '/module_list.161/activation/Sigmoid', '/module_list.186/activation/Mul', '/module_list.22/activation/Sigmoid', '/module_list.66/activation/Sigmoid', '/module_list.6/activation/Mul', '/module_list.163/activation/Mul', '/module_list.55/activation/Sigmoid', '/module_list.144/activation/Mul', '/module_list.43/activation/Sigmoid', '/module_list.39/activation/Mul', '/module_list.137/activation/Mul', '/module_list.172/activation/Sigmoid', '/module_list.9/activation/Mul', '/module_list.37/activation/Sigmoid', '/module_list.39/activation/Sigmoid', '/module_list.138/activation/Sigmoid', '/module_list.16/activation/Mul', '/module_list.145/activation/Mul', '/module_list.132/activation/Mul', '/module_list.187/activation/Mul', '/module_list.78/activation/Mul', '/module_list.44/activation/Mul', '/module_list.34/activation/Sigmoid', '/module_list.93/activation/Mul', '/module_list.230/activation/Mul', '/module_list.55/activation/Mul', '/module_list.72/activation/Mul', '/module_list.71/activation/Sigmoid', '/module_list.167/activation/Sigmoid', '/module_list.192/activation/Sigmoid', '/module_list.230/activation/Sigmoid', '/module_list.22/activation/Mul', '/module_list.67/activation/Sigmoid', '/module_list.27/activation/Sigmoid', '/module_list.13/activation/Sigmoid', '/module_list.58/activation/Sigmoid', '/module_list.89/activation/Mul', '/module_list.129/activation/Sigmoid', '/module_list.124/activation/Mul', '/module_list.115/activation/Mul', '/module_list.154/activation/Mul', '/module_list.12/activation/Sigmoid', '/module_list.49/activation/Mul', '/module_list.76/activation/Mul', '/module_list.124/activation/Sigmoid', '/module_list.187/activation/Sigmoid', '/module_list.195/activation/Mul', '/module_list.91/activation/Sigmoid', '/module_list.192/activation/Mul', '/module_list.10/activation/Sigmoid', '/module_list.141/activation/Sigmoid', '/module_list.52/activation/Mul', '/module_list.16/activation/Sigmoid', '/module_list.94/activation/Sigmoid', '/module_list.170/activation/Sigmoid', '/module_list.144/activation/Sigmoid', '/module_list.1/activation/Mul', '/module_list.24/activation/Mul', '/module_list.137/activation/Sigmoid', '/module_list.72/activation/Sigmoid', '/module_list.224/activation/Mul', '/module_list.189/activation/Mul', '/module_list.81/activation/Mul', '/module_list.21/activation/Sigmoid', '/module_list.85/activation/Sigmoid', '/module_list.135/activation/Mul', '/module_list.153/activation/Mul', '/module_list.113/activation/Sigmoid', '/module_list.183/activation/Sigmoid', '/module_list.54/activation/Mul', '/module_list.66/activation/Mul', '/module_list.200/activation/Sigmoid', '/module_list.180/activation/Sigmoid', '/module_list.196/activation/Mul', '/module_list.157/activation/Mul', '/module_list.145/activation/Sigmoid', '/module_list.197/activation/Mul', '/module_list.5/activation/Sigmoid', '/module_list.30/activation/Sigmoid', '/module_list.18/activation/Mul', '/module_list.140/activation/Mul', '/module_list.170/activation/Mul', '/module_list.135/activation/Sigmoid', '/module_list.128/activation/Sigmoid', '/module_list.51/activation/Mul', '/module_list.1/activation/Sigmoid', '/module_list.82/activation/Sigmoid', '/module_list.173/activation/Sigmoid', '/module_list.90/activation/Sigmoid', '/module_list.20/activation/Mul', '/module_list.193/activation/Mul', '/module_list.202/activation/Sigmoid', '/module_list.7/activation/Mul', '/module_list.212/activation/Sigmoid', '/module_list.57/activation/Sigmoid', '/module_list.28/activation/Mul', '/module_list.74/activation/Mul', '/module_list.199/activation/Mul', '/module_list.96/activation/Mul', '/module_list.3/activation/Mul', '/module_list.58/activation/Mul', '/module_list.176/activation/Mul', '/module_list.57/activation/Mul', '/module_list.7/activation/Sigmoid', '/module_list.93/activation/Sigmoid', '/module_list.85/activation/Mul', '/module_list.174/activation/Mul', '/module_list.193/activation/Sigmoid', '/module_list.112/activation/Sigmoid', '/module_list.174/activation/Sigmoid', '/module_list.148/activation/Sigmoid', '/module_list.154/activation/Sigmoid', '/module_list.196/activation/Sigmoid', '/module_list.200/activation/Mul', '/module_list.97/activation/Sigmoid', '/module_list.143/activation/Sigmoid', '/module_list.171/activation/Mul', '/module_list.126/activation/Mul', '/module_list.48/activation/Sigmoid', '/module_list.30/activation/Mul', '/module_list.161/activation/Mul', '/module_list.94/activation/Mul', '/module_list.103/activation/Sigmoid', '/module_list.105/activation/Sigmoid', '/module_list.34/activation/Mul', '/module_list.143/activation/Mul', '/module_list.24/activation/Sigmoid', '/module_list.156/activation/Sigmoid', '/module_list.87/activation/Sigmoid', '/module_list.127/activation/Sigmoid', '/module_list.10/activation/Mul', '/module_list.61/activation/Sigmoid', '/module_list.160/activation/Sigmoid', '/module_list.60/activation/Mul', '/module_list.151/activation/Mul', '/module_list.104/activation/Mul', '/module_list.125/activation/Sigmoid', '/module_list.70/activation/Mul', '/module_list.78/activation/Sigmoid', '/module_list.176/activation/Sigmoid', '/module_list.212/activation/Mul', '/module_list.116/activation/Sigmoid', '/module_list.37/activation/Mul', '/module_list.45/activation/Mul', '/module_list.25/activation/Mul', '/module_list.36/activation/Mul', '/module_list.171/activation/Sigmoid', '/module_list.148/activation/Mul', '/module_list.33/activation/Sigmoid', '/module_list.141/activation/Mul', '/module_list.147/activation/Sigmoid', '/module_list.70/activation/Sigmoid', '/module_list.52/activation/Sigmoid', '/module_list.122/activation/Sigmoid', '/module_list.189/activation/Sigmoid', '/module_list.182/activation/Mul', '/module_list.63/activation/Mul', '/module_list.96/activation/Sigmoid', '/module_list.173/activation/Mul', '/module_list.100/activation/Mul', '/module_list.158/activation/Sigmoid', '/module_list.183/activation/Mul', '/module_list.79/activation/Mul', '/module_list.63/activation/Sigmoid', '/module_list.112/activation/Mul', '/module_list.121/activation/Sigmoid', '/module_list.36/activation/Sigmoid', '/module_list.185/activation/Sigmoid', '/module_list.87/activation/Mul', '/module_list.90/activation/Mul', '/module_list.159/activation/Mul', '/module_list.13/activation/Mul', '/module_list.89/activation/Sigmoid', '/module_list.20/activation/Sigmoid', '/module_list.166/activation/Sigmoid', '/module_list.61/activation/Mul', '/module_list.147/activation/Mul', '/module_list.184/activation/Sigmoid', '/module_list.127/activation/Mul', '/module_list.28/activation/Sigmoid', '/module_list.12/activation/Mul', '/module_list.76/activation/Sigmoid', '/module_list.100/activation/Sigmoid', '/module_list.198/activation/Sigmoid', '/module_list.113/activation/Mul', '/module_list.48/activation/Mul', '/module_list.74/activation/Sigmoid', '/module_list.105/activation/Mul', '/module_list.164/activation/Mul', '/module_list.180/activation/Mul', '/module_list.202/activation/Mul', '/module_list.119/activation/Mul', '/module_list.0/Reorg/Concat', '/module_list.31/activation/Mul', '/module_list.129/activation/Mul', '/module_list.49/activation/Sigmoid', '/module_list.131/activation/Sigmoid', '/module_list.218/activation/Mul', '/module_list.131/activation/Mul', '/module_list.75/activation/Sigmoid', '/module_list.116/activation/Mul', '/module_list.17/activation/Sigmoid', '/module_list.81/activation/Sigmoid', '/module_list.164/activation/Sigmoid', '/module_list.186/activation/Sigmoid', '/module_list.64/activation/Sigmoid', '/module_list.6/activation/Sigmoid', '/module_list.27/activation/Mul', '/module_list.151/activation/Sigmoid', '/module_list.2/activation/Sigmoid', '/module_list.33/activation/Mul', '/module_list.121/activation/Mul', '/module_list.199/activation/Sigmoid', '/module_list.25/activation/Sigmoid', '/module_list.40/activation/Sigmoid', '/module_list.18/activation/Sigmoid', '/module_list.142/activation/Sigmoid', '/module_list.86/activation/Sigmoid', '/module_list.158/activation/Mul', '/module_list.184/activation/Mul', '/module_list.64/activation/Mul', '/module_list.119/activation/Sigmoid', '/module_list.153/activation/Sigmoid', '/module_list.190/activation/Sigmoid', '/module_list.5/activation/Mul', '/module_list.75/activation/Mul', '/module_list.91/activation/Mul', '/module_list.142/activation/Mul', '/module_list.156/activation/Mul', '/module_list.47/activation/Mul', '/module_list.128/activation/Mul', '/module_list.177/activation/Sigmoid', '/module_list.169/activation/Sigmoid', '/module_list.122/activation/Mul', '/module_list.167/activation/Mul', '/module_list.3/activation/Sigmoid', '/module_list.101/activation/Sigmoid', '/module_list.138/activation/Mul', '/module_list.177/activation/Mul', '/module_list.185/activation/Mul', '/module_list.2/activation/Mul', '/module_list.9/activation/Sigmoid', '/module_list.157/activation/Sigmoid', '/module_list.67/activation/Mul', '/module_list.104/activation/Sigmoid', '/module_list.71/activation/Mul', '/module_list.172/activation/Mul', '/module_list.43/activation/Mul', '/module_list.197/activation/Sigmoid', '/module_list.54/activation/Sigmoid', '/module_list.31/activation/Sigmoid']
2026-06-19 03:55 - INFO - sdk - quantize - Starting quantization...
WARNING:root:Please use QuantFormat.QDQ for activation type QInt8 and weight type QInt8. Or it will lead to bad performance on x64.
2026-06-19 03:55 - INFO - sdk - quantize - Quantization completed! Quantized model saved to /quadric/sdk-cli/examples/models/yolo/yolor/yolor_p6-backbone_OpSet16_optimized_asym_int8_q.onnx
2026-06-19 03:55 - INFO - sdk - quantize - ONNX full precision model size: 140.39MB
2026-06-19 03:55 - INFO - sdk - quantize - ONNX quantized model size: 35.45MB
2026-06-19 03:55 - INFO - sdk - quantize - ONNX model shapes inferred.
2026-06-19 03:55 - INFO - sdk - quantize - ONNX Model with well-defined shapes has been saved at `/quadric/sdk-cli/examples/models/yolo/yolor/yolor_p6-backbone_OpSet16_optimized_asym_int8_q_shaped.onnx`.
2026-06-19 03:55 - DEBUG - sdk - quantize - Checking for FLOAT/FLOAT16 types...
2026-06-19 03:55 - INFO - sdk - quantize - Checking for remaining FLOAT/FLOAT16 types.
2026-06-19 03:55 - INFO - sdk - quantize - Model still has FLOAT/FLOAT16 types after quantization. Creating ranges for floating point tensors using calibration data...
2026-06-19 03:55 - INFO - sdk - quantize - Saved computed tensor ranges to /quadric/sdk-cli/examples/models/yolo/yolor/yolor_p6-backbone_OpSet16_optimized_asym_int8_q_shaped.tranges.
2026-06-19 03:55 - INFO - sdk - quantize -
╒═════════════════════════════════════════════════════════════════════════════════════════════════════════╤════════════════════════════════════════════════════════════════════════════════════════════════════════════╕
│ Quantized ONNX Model │ Tensor Ranges File │
╞═════════════════════════════════════════════════════════════════════════════════════════════════════════╪════════════════════════════════════════════════════════════════════════════════════════════════════════════╡
│ /quadric/sdk-cli/examples/models/yolo/yolor/yolor_p6-backbone_OpSet16_optimized_asym_int8_q_shaped.onnx │ /quadric/sdk-cli/examples/models/yolo/yolor/yolor_p6-backbone_OpSet16_optimized_asym_int8_q_shaped.tranges │
╘═════════════════════════════════════════════════════════════════════════════════════════════════════════╧════════════════════════════════════════════════════════════════════════════════════════════════════════════╛
Quantized ONNX: yolor_p6-backbone_OpSet16_optimized_asym_int8_q_shaped.onnx
Tensor ranges: yolor_p6-backbone_OpSet16_optimized_asym_int8_q_shaped.tranges
19197
5. Compilation
ChimeraJob wraps the full compile pipeline: graph analyze, kernel selection, memory planning, and code generation. Printing the job after ISS has run renders a summary table that includes the target configuration and cycle counts.
cgc_job = ChimeraJob(model_p=str(quantized_onnx_model.model_path))
cgc_job.compile(quiet=True)
6. Inference
batch_inference runs the compiled backbone against multiple images in parallel through two engines: CHIMERA_ORT_INT8 (ONNX Runtime reference) and CHIMERA_ISS_INT8 (the Chimera GPNPU cycle-accurate simulator). Both drive the same head ONNX on the host for the classical tail.
all_image_paths, all_images = load_images(INPUT_SIZE, DEFAULT_IMAGE_PATHS)
num_images = len(all_images)
engines = {
InferenceEngine.CHIMERA_ORT_INT8: cgc_job,
InferenceEngine.CHIMERA_ISS_INT8: cgc_job,
}
head_session = InferenceSession(onnx.load(str(head_onnx_file)).SerializeToString())
outputs_per_engine = {}
threads = min(num_images, 6)
for engine, job in engines.items():
outputs_per_engine[engine] = batch_inference(
engine, job, all_images, head_session=head_session, threads=threads
)
2026-06-19 04:04 - WARNING - epu - chimera_job - ORT is not threadsafe -- forcing single threaded batch execution
100%|████████████████████████████████████████████| 3/3 [00:02<00:00, 1.20it/s]
Processing: 100%|████████████████████████████████| 3/3 [02:43<00:00, 54.47s/it]
7. Run Statistics
plot_run_statistics() renders the per-kernel cycle breakdown for the compiled backbone.
## Re-enable inline display in case the yolor-pip import switched the
## matplotlib backend to Agg.
%matplotlib inline
print(cgc_job)
cgc_job.plot_run_statistics()
╒═════════════════════╤═════════════════════════════════════════════════════════════════════════════════════════════════════════╕
│ Module Name │ yolor_p6_backbone_OpSet16_optimized_asym_int8_q_shaped_QC_U_1d7_16MB_4kB_128GBps_128GBps_16_OFF_x1_x1 │
├─────────────────────┼─────────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ ONNX File │ /quadric/sdk-cli/examples/models/yolo/yolor/yolor_p6-backbone_OpSet16_optimized_asym_int8_q_shaped.onnx │
├─────────────────────┼─────────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ Product Target │ QC-U │
├─────────────────────┼─────────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ Number of Cores │ 1 │
├─────────────────────┼─────────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ ISS Clock Frequency │ 1.700 │
├─────────────────────┼─────────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ L2M Size │ 16MB │
├─────────────────────┼─────────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ LRM Size │ 4kB │
├─────────────────────┼─────────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ External Read BW │ 128GBps │
├─────────────────────┼─────────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ External Write BW │ 128GBps │
├─────────────────────┼─────────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ MACS per PE │ 16 │
├─────────────────────┼─────────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ Max L2M │ 15.801MB │
├─────────────────────┼─────────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ Max LRM │ 2.875kB │
├─────────────────────┼─────────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ Max Temp Ext Bytes │ 0.000MB │
├─────────────────────┼─────────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ Network GMACs │ 39.997 │
╘═════════════════════╧═════════════════════════════════════════════════════════════════════════════════════════════════════════╛
╒════╤════════╤══════════════════════════════════════════╤══════════════════╤══════════════════════════╤═══════╕
│ │ Type │ Name │ shape │ type │ mse │
╞════╪════════╪══════════════════════════════════════════╪══════════════════╪══════════════════════════╪═══════╡
│ 0 │ Input │ inputs0 │ [1, 3, 640, 640] │ tensor[FixedPoint32<27>] │ n/a │
├────┼────────┼──────────────────────────────────────────┼──────────────────┼──────────────────────────┼───────┤
│ 1 │ Output │ /module_list.212/activation/Mul_output_0 │ [1, 256, 80, 80] │ tensor[FixedPoint32<25>] │ 0.020 │
├────┼────────┼──────────────────────────────────────────┼──────────────────┼──────────────────────────┼───────┤
│ 2 │ Output │ /module_list.218/activation/Mul_output_0 │ [1, 384, 40, 40] │ tensor[FixedPoint32<26>] │ 0.011 │
├────┼────────┼──────────────────────────────────────────┼──────────────────┼──────────────────────────┼───────┤
│ 3 │ Output │ /module_list.224/activation/Mul_output_0 │ [1, 512, 20, 20] │ tensor[FixedPoint32<26>] │ 0.004 │
├────┼────────┼──────────────────────────────────────────┼──────────────────┼──────────────────────────┼───────┤
│ 4 │ Output │ /module_list.230/activation/Mul_output_0 │ [1, 640, 10, 10] │ tensor[FixedPoint32<27>] │ 0.001 │
╘════╧════════╧══════════════════════════════════════════╧══════════════════╧══════════════════════════╧═══════╛
Post-ISS Report 1.7 GHz ***
Fully placed-and-routed gate simulation:
╒══════════════════════════════════╤═════════╕
│ Latency (ms) │ 4.97 │
├──────────────────────────────────┼─────────┤
│ FPS │ 201.34 │
├──────────────────────────────────┼─────────┤
│ Average Power @ 3nm SSGNP (mW) │ 2134.54 │
├──────────────────────────────────┼─────────┤
│ FPS per Watt @ 3nm SSGNP (FPS/W) │ 94.32 │
├──────────────────────────────────┼─────────┤
│ Ext Rd Bytes (MB) │ 40.22 │
├──────────────────────────────────┼─────────┤
│ Ext Wr Bytes (MB) │ 9.62 │
├──────────────────────────────────┼─────────┤
│ Avg Ext Rd BW (GBps) │ 7.91 │
├──────────────────────────────────┼─────────┤
│ Avg Ext Wr BW (GBps) │ 1.89 │
├──────────────────────────────────┼─────────┤
│ MAC Utilization │ 28.91% │
╘══════════════════════════════════╧═════════╛
*** Data generated using 7nm SSGNP gatesim and scaled to 3nm
[SDK-CLI] : TotalCycles: 8,443,465
[SDK-CLI] : Executions/second: 201.34
compute : ▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 1.565M
data_array : ▇▇▇▇▇▇▇▇▇▇ 1.056M
mac : ▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 4.881M
data_external: ▇ 115.762K
data_ocm : ▇▇▇▇▇▇▇▇ 822.59K
for more information check run directory: /quadric/sdk-cli/examples/models/yolo/yolor/ccl_build/yolor_p6_backbone_OpSet16_optimized_asym_int8_q_shaped_QC_U_1d7_16MB_4kB_128GBps_128GBps_16_OFF_x1_x1/run/20260619_040448_2a1a3d
2026-06-19 04:07 - INFO - epu - chimera_job - Combined plots generated and saved to:
/quadric/sdk-cli/examples/models/yolo/yolor/ccl_build/yolor_p6_backbone_OpSet16_optimized_asym_int8_q_shaped_QC_U_1d7_16MB_4kB_128GBps_128GBps_16_OFF_x1_x1/run/20260619_040448_2a1a3d/data/yolor_p6_backbone_OpSet16_optimized_asym_int8_q_shaped_QC_U_1d7_16MB_4kB_128GBps_128GBps_16_OFF_x1_x1.combined.png
'/quadric/sdk-cli/examples/models/yolo/yolor/ccl_build/yolor_p6_backbone_OpSet16_optimized_asym_int8_q_shaped_QC_U_1d7_16MB_4kB_128GBps_128GBps_16_OFF_x1_x1/run/20260619_040448_2a1a3d/data'
Decode the network outputs and overlay the detections on the input image.
detections_per_engine = run_postprocess(outputs_per_engine, all_image_paths, INPUT_SIZE)
display_detections(all_image_paths, detections_per_engine)
Summary
| Model | YOLOR-P6 (COCO, 640×640) |
| Pipeline | Source download → weights via huggingface_hub → ONNX export → split → quantize (INT8 asymmetric) → CGC compile → ISS + ORT inference → YOLOR NMS → bbox overlay |
| Compiled on GPNPU | Convolutional backbone |
| On host CPU | YOLOR head ONNX, classical NMS, bounding-box visualization |
| Calibration | QuadricCalibration COCO-like subset |
Key takeaways
- Splitting the ONNX at the
module_listhead boundary keeps CGC focused on GPNPU-friendly convolutional ops while the anchor-decode + NMS tail runs on the host. - The SDK exposes the same
ChimeraJobhandle to bothCHIMERA_ORT_INT8(ONNX Runtime reference) andCHIMERA_ISS_INT8(ISS), making side-by-side validation a one-dict change. - The YOLOR "implicit" branch folds cleanly into the backbone graph after
onnxsim, so quantization and compilation treat it as ordinary convolutional structure.
Citation
@article{wang2021yolor,
title = {You Only Learn One Representation: Unified Network for Multiple Tasks},
author = {Wang, Chien-Yao and Yeh, I-Hau and Liao, Hong-Yuan Mark},
journal = {arXiv preprint arXiv:2105.04206},
year = {2021}
}
