CUDA 13.0 Triton 3.6.0 Workaround | Generated by AI
CUDA 13.0 (driver 580.126.20) + Triton 3.6.0 — the downgrade to 3.4.0 got reverted by pip dependency resolution. This is a bleeding-edge combo known to cause segfaults with Triton JIT.
Two clean paths:
Path A — vLLM (recommended, more stable on consumer GPUs):
/home/lzw/sglang-venv/bin/pip install vllm
/home/lzw/sglang-venv/bin/python -m vllm.entrypoints.openai.api_server \
--model Qwen/Qwen2.5-1.5B-Instruct --port 30000 --host 0.0.0.0
Path B — sglang without Triton (pin triton with –no-deps):
/home/lzw/sglang-venv/bin/pip install --no-deps triton==3.4.0
Then retry sglang with --attention-backend torch_native --disable-cuda-graph.
vLLM is the safer bet — it’s more battle-tested on RTX 4070 and CUDA 13.0. Want me to try installing and launching it?
Generated by AI. Curating and sharing still takes effort.
If you find it useful, feel free to
donate.
WeChat: @lzwjava
·
X: @lzwjava ·
Say hi 👋