[ROCm] gfx1201 (Navi 48) support - compilation guide and known issues

[UFO0506]

System Info

bitsandbytes ROCm 編譯

踩坑 1：pip 安裝的版本不支援 ROCm

直接 pip install bitsandbytes 安裝的版本沒有 ROCm backend，必須從源碼編譯。

踩坑 2：multi-backend-refactor 分支有 bug

multi-backend-refactor 分支編譯出的 .so 缺少 kOptimizer32bit1StateI12hip_bfloat16Li2E 符號，導致 undefined symbol 錯誤。必須用 main 分支。

踩坑 3：gfx1153 無效 target ID

ROCm 6.3 的 CMake 模組內建了一個無效的 target gfx1153，必須在 cmake 前設定環境變數覆蓋：

export AMDGPU_TARGETS="gfx1201"

正確的編譯步驟

cd ~
git clone https://github.com/bitsandbytes-foundation/bitsandbytes.git
cd bitsandbytes
 
export PATH=/opt/rocm/bin:/usr/local/bin:$PATH
export AMDGPU_TARGETS="gfx1201"
 
cmake -DCOMPUTE_BACKEND=hip \
      -DCMAKE_BUILD_TYPE=Release \
      -DCMAKE_HIP_ARCHITECTURES="gfx1201" \
      -S . -B build
 
cmake --build build -j$(nproc)
pip3.11 install -e . --no-cache-dir
 
# 驗證
python3.11 -c "import bitsandbytes as bnb; print(bnb.__version__)"

踩坑 4：editable install 路徑問題

pip install -e . 之後，bitsandbytes.__path__ 可能指向上層目錄而非模組目錄，導致 import 失敗。

解決方法：手動加入 PYTHONPATH：

echo 'export PYTHONPATH=/root/bitsandbytes:$PYTHONPATH' >> ~/.bashrc

踩坑 5：BNB_BACKEND 環境變數

推論時需要明確告知 bitsandbytes 使用 ROCm backend：

import os
os.environ['BNB_BACKEND'] = 'rocm'

---

Reproduction

Environment

• GPU: AMD Radeon AI PRO R9700 (Navi 48 / gfx1201 / 32GB VRAM)
• Host: Proxmox VE 9.0.3 (LXC Container, Ubuntu 22.04)
• ROCm: 6.3.0
• Python: 3.11.15
• PyTorch: 2.9.1+rocm6.3
• bitsandbytes: 0.50.0.dev0 (compiled from source)

Problem

pip-installed bitsandbytes has no ROCm backend for gfx1201 (Navi 48).
Must compile from source. Several issues encountered during compilation.

Issue 1: gfx1153 invalid target ID

ROCm 6.3 CMake module includes gfx1153 which is an invalid target ID,
causing cmake to fail:
clang++: error: invalid target ID 'gfx1153'

Fix: Set environment variable before cmake:

export AMDGPU_TARGETS="gfx1201"

Issue 2: multi-backend-refactor branch missing symbol

Compiling from multi-backend-refactor branch produces a .so missing
kOptimizer32bit1StateI12hip_bfloat16Li2E symbol:
OSError: undefined symbol: _Z36__device_stub__kOptimizer32bit1StateI12hip_bfloat16Li2E...

Fix: Use main branch instead.

Working Compilation Steps for gfx1201

git clone https://github.com/bitsandbytes-foundation/bitsandbytes.git
cd bitsandbytes

export AMDGPU_TARGETS="gfx1201"
export PATH=/opt/rocm/bin:$PATH

cmake -DCOMPUTE_BACKEND=hip \
      -DCMAKE_BUILD_TYPE=Release \
      -DCMAKE_HIP_ARCHITECTURES="gfx1201" \
      -S . -B build

cmake --build build -j$(nproc)
pip install -e .

Issue 3: editable install path problem

After pip install -e ., bitsandbytes.__path__ points to parent directory
instead of module directory, causing import to fail silently.

Fix:

export PYTHONPATH=/root/bitsandbytes:$PYTHONPATH

Issue 4: BNB_BACKEND not auto-detected

Even with ROCm available, must explicitly set:

import os
os.environ['BNB_BACKEND'] = 'rocm'

Verification

python3 -c "
import sys
sys.path.insert(0, '/root/bitsandbytes')
import bitsandbytes as bnb
print(bnb.__version__)
print(bnb.supported_torch_devices)
"
# Output: 0.50.0.dev0

Expected behavior

bitsandbytes should automatically detect ROCm backend and support gfx1201 (Navi 48)
without requiring manual compilation from source or manual PYTHONPATH configuration.

Ideally:

1. pip install bitsandbytes should include ROCm support for gfx1201
2. AMDGPU_TARGETS should be auto-detected during cmake
3. BNB_BACKEND should be auto-detected when ROCm is available

[Kaihui-AMD]

Hi @UFO0506,
Thanks for the detailed writeup. A few clarifications based on testing on the same hardware (R9700 / gfx1201, ROCm 7.2.3):
Issue 1 (gfx1153): This affects ROCm 6.3.x only. On ROCm 7.2.x the compiler accepts gfx1153 without error and gfx1201 is already in bitsandbytes' default architecture list, so no AMDGPU_TARGETS override is needed.

Issue 3 (editable install path): Not reproducible on current main with ROCm 7.2.3 — bitsandbytes.path points correctly to the module directory after pip install -e ..

Issue 4 (BNB_BACKEND): os.environ['BNB_BACKEND'] = 'rocm' has no effect — bitsandbytes doesn't read this env var. The backend auto-detects via torch.version.hip, no manual setting needed.

[UFO0506]

Hi @Kaihui-AMD,

Thank you for testing on the same hardware and providing those clarifications!

I've updated the guide based on your feedback:

• Marked AMDGPU_TARGETS, PYTHONPATH, and BNB_BACKEND workarounds as ROCm 6.3-only
• Added a ROCm 7.2.3 simplified path where most workarounds are no longer needed
• Added a version comparison table so users can quickly see what applies to them

Hoping this helps other R9700 users skip the painful parts. Thanks again! 🙏

[UFO0506]

AMD Radeon AI PRO R9700 (gfx1201) ROCm 訓練環境完整踩坑紀錄

作者：UFO Homelab
環境：Proxmox VE 9.0.3 + LXC Container + ROCm 6.3 → 7.2.3 + LLaMA-Factory
模型：Qwen2.5-Coder-32B-Instruct QLoRA 4bit
目的：讓後來的 R9700 用戶不用重踩相同的坑

感謝 AMD 工程師 @chejh-amd 和 @Kaihui-AMD 在 GitHub issue 上提供的寶貴回饋與修正。

---

⚡ 速查：ROCm 版本對照

項目	ROCm 6.3	ROCm 7.2.3+
gfx1201 官方支援	部分	✅ 完整
AMDGPU_TARGETS 需要手動設定	✅ 需要	❌ 不需要
bitsandbytes PYTHONPATH 需要手動設定	✅ 需要	❌ 不需要
BNB_BACKEND 環境變數有效	✅ 需要	❌ 無效（自動偵測）
flash-attn composable_kernel	❌ 編譯失敗	❌ 仍失敗
SDPA（AOTriton）加速	❌ 不支援	✅ 支援
訓練速度（32B QLoRA）	~147 t/s	~323 t/s

---

十二、致謝

感謝以下 AMD 工程師在 GitHub issue 上提供技術指引與修正：

• @chejh-amd：指出 flash-attn 的正確替代方案（SDPA + AOTriton），並建議升級至 ROCm 7.2。
• @Kaihui-AMD：在相同硬體（R9700 / gfx1201 / ROCm 7.2.3）上驗證，並指出 ROCm 7.2.3 已修復多個 workaround。

---

整理自 2026 年 5 月的實際踩坑經驗
AMD Radeon AI PRO R9700 / gfx1201 / ROCm 6.3 → 7.2.3 / LLaMA-Factory 0.9.5