I traced the slowdown to Q/DQ placement and two TensorRT bugs, then wrote the placement myself on the FP16 graph. INT8 is now faster than FP16 at every size on an RTX 5070 (1.10–1.33×) and a Tesla T4 (1.11–1.21×), for 0.5–0.85 COCO AP points.
Chào, mình là Phú. Đây là ‘Byte of Me’ — một góc nhỏ trên internet, nơi mình chia sẻ hành trình và những gì mình đã làm.
You’ll never know how good you might have become unless you try.
— Mike Mentzer
Tôi luôn hứng thú với những dự án được đầu tư nghiêm túc và những mối hợp tác tốt đẹp.
I traced the slowdown to Q/DQ placement and two TensorRT bugs, then wrote the placement myself on the FP16 graph. INT8 is now faster than FP16 at every size on an RTX 5070 (1.10–1.33×) and a Tesla T4 (1.11–1.21×), for 0.5–0.85 COCO AP points.
33 pull request đã merge trên 3 dự án
roboflow/rf-detr
RF-DETR is a real-time object detection and segmentation model architecture developed by Roboflow, SOTA on COCO, designed for fine-tuning. [ICLR 2026]
perf(export): smooth the backbone norms and correct the MLP biases of TensorRT INT8 detection engines #1645
+2,125−205Xemperf(export): run the segmentation and keypoint INT8 patch embedding as a matrix multiply #1642
+344−207Xemroboflow/trackers
Trackers gives you clean, modular re-implementations of leading multi-object tracking algorithms released under the permissive Apache 2.0 license. You combine them with any detection model you already use.
sunsmarterjie/yolov12
[NeurIPS 2025] YOLOv12: Attention-Centric Real-Time Object Detectors

2020 — 2024
Khoa học Máy tính, lớp Cử nhân tài năng.
Hướng chính: Thị giác máy tính, và mô hình Thị giác–Ngôn ngữ.
Công bố bài báo: "Abnormality Detection In Medical Image Based On Visual-Language Model" | Code.


Đạt 28.65/30.00 điểm

2017 — 2020