Site-wide newcomer FAQ: /faq. This page is ROCm-specific — what people get wrong, then short answers with cites.
Common misconceptions
“ROCm is just CUDA for AMD.”
ROCm is AMD’s own open platform.[1]HIP is deliberately CUDA-like and can port a lot of code, but AMD says it is not a drop-in CUDA replacement.[2]Wheels, gfx targets, and tools are a different world. Budget time for matching versions, not for a one-line extra on a CUDA tutorial.
“ROCm doesn’t work on consumer cards.”
Out of date as a blanket claim. ROCm 10.1.0’s hardware table lists Radeon AI PRO R9700 and the RX 9070 series as gfx1201 (RDNA 4), plus a long RDNA 3 / RDNA 2 PRO list.[3]What is still fair: it is not CUDA, packaging is the pain, and “official” is a table. Single-GPU local inference on a listed Ubuntu is the path this site writes.
“If you use Vulkan, ROCm failed.”
llama.cpp has a Vulkan backend and a HIP backend.[9]Our 35B daily driver is Vulkan on purpose. HIP/ROCm is what you need for PyTorch, vLLM, and the AMD-documented AI ecosystem. When each wins.
“Set HSA_OVERRIDE_GFX_VERSION and move on.”
That variable is a community workaround (classic case: gfx1031 pretending to be gfx1030).[7]It is unofficial, process-wide, and can fault. Current gfx1201 cards on ROCm 10.x should not need it. gfx page.
“Two R9700s will double tok/s.”
Not automatically. Multi-GPU is RCCL, topology, and a backend that knows how to split. Our daily driver is one card. House rule: one LLM per GPU unless you deliberately combine. Use the calculator split toggle before you buy a second card for capacity.
“apt install rocm from Ubuntu is the install.”
Archive packages lag the 10.x / TheRock line. Use AMD’s10.1.0 install docsor the ROCm CLI, not a mix of archive + leftover runfile + CLI.[6]Walkthrough.
“The new CLI is ROCm 1.0 / CLI v1.0.”
The 10.1 blog names ROCm CLI v0.2.0.[5]Check releases for tags. Do not write v1.0.0 from memory.
“I have an RX 9700.”
You don’t, in AMD’s current lineup. The 32 GB workstation card is theRadeon AI PRO R9700.[8]The gaming sibling is the RX 9070 series (also gfx1201, less VRAM).Directory.
FAQ
Can I stay on Windows?
AMD documents ROCm as Linux and Windows.[1]10.1 adds a WSL2 technical preview (standard Linux packages in the guest; profiling / KFD-dependent tooling not supported at that launch, per the notes).[3][5]For a first weekend of tokens, llama.cpp Vulkan is the honest Windows path. A full HIP + PyTorch home lab is still the listed-Ubuntu story we write inInstall ROCm.
Which PyTorch / vLLM numbers are “the” numbers?
Dated snapshots, not eternal law. 10.1.0 notes: PyTorch 2.14.0, vLLM 0.29.0 (notes say this replaces 2.11.0 / 0.27.0).[3]10.0.0 notes: PyTorch 2.13.0, vLLM 0.27.0.[4]Community try tags move faster. The home card is this week. Do not mix pins in one venv. Default PyPI vLLM is CUDA.
Do I need ROCm to run Qwen locally?
No, if llama.cpp Vulkan sees the discrete GPU and VRAM moves. Yes, if you want PyTorch, vLLM, or HIP llama.cpp against /opt/rocm. Chooser.
Can I fine-tune like a 4090 blog post?
PyTorch ROCm exists. Fine-tune cookbooks are thinner than CUDA. This site has a dated note about training/fine-tune GPU resets on R9700 in the 10.0 era and aTORCH_BLAS_PREFER_HIPBLASLT=0 suggestion — treat that as a snapshot, not a current KI we have re-verified today.
Prefer inference on the pin you already trust. Ask theerror paste if a train run dies. Thetracker row stays TODO until that KI URL is copied.
rocm-smi or amd-smi?
The 10.1 blog says the rocm-smi deprecation period is complete andamd-smi is the successor.[5]New scripts should call amd-smi. Old blog posts will still sayrocm-smi.
Sources
Every version and support claim on this page points here. If a sentence has no mark, treat it as lab opinion or a TODO — not as AMD policy.