PaddleRec -- AMD ROCm / AMD Instinct Enablement

This is the AMD ROCm enablement port of PaddleRec, delivered on the port/PaddleRec branch of AMD-AIOSS/hipSHIFT.

PaddleRec is a large-scale recommendation framework built on PaddlePaddle. This port validates it on AMD Instinct GPUs (MI300X / gfx942, MI350X-MI355X / gfx950) using ROCm 10.0.0 and a ROCm-enabled PaddlePaddle build.

Runs on AMD Instinct

GPU	Architecture	Status
AMD Instinct MI300X	gfx942	Validated -- 5/5 passed
AMD Instinct MI350X / MI355X	gfx950	Validated -- 4/4 passed
AMD Instinct MI450 (FFM simulator)	gfx1250	Validated (functional sim) -- 4/4 passed

ROCm Install Path

PaddleRec rides PaddlePaddle, which has no public AMD-Instinct wheel on PyPI. The install is source-based:

1. Prerequisites

• AMD Instinct GPU (MI300X, MI350X/MI355X, or MI450)
• Ubuntu 24.04, Python 3.11
• AMD GPU driver (amdgpu-dkms); add yourself to render/video groups

sudo apt install libatomic1 libquadmath0 python3.11 python3.11-venv
sudo usermod -a -G render,video $LOGNAME
# log out and back in (or reboot) to apply group membership

2. Install ROCm 10.0.0 (public pip wheels)

python3.11 -m venv .venv && source .venv/bin/activate
pip install --index-url https://stable.repo.amd.com/rocm/whl-next/ "rocm[libraries,devel,device-gfx942]"
# for MI350X/MI355X (gfx950): replace device-gfx942 with device-gfx950

3. Build PaddlePaddle-on-ROCm from source

There is no public AMD-Instinct PaddlePaddle wheel (the paddlepaddle-rocm PyPI channel targets Hygon DCU, not AMD Instinct). Build from the AMD ROCm fork:

git clone -b port/Paddle https://github.com/AMD-AIOSS/hipSHIFT.git paddle-rocm && cd paddle-rocm
# follow the build instructions in paddle-rocm/hipshift/release_docs/installation.md

4. Install PaddleRec from this branch

# (after building and installing PaddlePaddle-on-ROCm above)
git clone -b port/PaddleRec https://github.com/AMD-AIOSS/hipSHIFT.git paddlerec && cd paddlerec
pip install -r requirements-rocm.txt
pip install .

5. Verify

rocminfo | grep -i 'Marketing Name'   # expect: AMD Instinct MI300X (or MI355X)

import paddle
print('Compiled with ROCm:', paddle.device.is_compiled_with_rocm())  # expect: True
paddle.set_device('gpu:0')
a = paddle.randn([512, 512])
print('GPU matmul result:', float((a @ a).abs().mean()))  # expect: a positive float

Train a DNN ranking model on GPU:

python -u tools/trainer.py -m models/rank/dnn/config.yaml \
    -o runner.use_gpu=True runner.epochs=3
# Expected: epoch 2 AUC approaching 1.0, RC=0

AMD-specific files in this branch

File	Purpose
requirements-rocm.txt	ROCm-compatible requirements (replaces CUDA deps)
requirements-rocm-optional.txt	Optional ROCm extras (faiss-cpu fallback for ANN retrieval)
amd-constraints.txt	Version pins for ROCm 10.0.0 + AMD Instinct
Dockerfile.rocm	ROCm base Dockerfile for AMD Instinct
VALIDATION_NOTES.md	Per-arch validation results and gotchas

Documentation

Full AMD AIOSS release documentation is under hipshift/release_docs/:

• Installation
• Compatibility Matrix
• Build & Validation Report
• Benchmark Report
• Release Notes

License

Apache-2.0. See upstream LICENSE. AMD modifications: Copyright (C) 2026 Advanced Micro Devices, Inc. All rights reserved.