Is “R9700 Pro” a real SKU?
The product is the AMD Radeon AI PRO R9700. There is no consumer RX 9700 in AMD’s current lineup. The gaming sibling is the RX 9070 series. AMD GPUs.
Will ROCm work for local LLMs on Ubuntu in 2026?
Yes, for single-GPU inference, if you pin a current AMD Core SDK (10.1 line as of the weekly card) and a matching runtime. It is still not CUDA. Packaging is the hard part. Longer answers: What is ROCm and the ROCm FAQ.
Can I apt install rocm from Ubuntu?
You can. You should not, if you want Core SDK 10.0/10.1, gfx1201, or a current vLLM ROCm image. Canonical’s archive has lagged (e.g. 7.1 on 26.04). Add AMD’s 10.1 repo or use ROCm CLI v0.2.0 — never both plus archive rocm. Install. ROCm.
Windows?
Vulkan (llama.cpp, some UIs) is the honest path. Full ROCm userspace is a Linux product. Limited PyTorch-on-Windows exists; do not expect rocm-smi culture.
Which Qwen should I run first?
Qwen3.6-35B-A3B Unsloth UD-Q5_K_XL on one R9700 via llama.cpp Vulkan, 262K context, f16 KV, no RAM offload. Measured on our rig, 2026-10: 100.8 tok/s short, ~88 at 32K, ~50 at 256K. Fast agents: Qwen 3.8 Flash UD-Q6_K_XL on Strata v0.1.42. Cards: stacks. Fit check: calculator. Abliterated is educational only.
vLLM 0.27, 0.30, or 0.31?
Consumers should not start here. 0.27 + PyTorch 2.13 is still AMD’s 10.0 supported combo. Community / ROCm 0.31 is this week’s serve try (0.30 is the previous try-tag). Do not mix pins in one venv. Default PyPI vLLM is CUDA.
Can I fine-tune like a 4090 blog post?
PyTorch ROCm exists. ROCm 10.0 has a known issue: training/fine-tune GPU resets on R9700 — try TORCH_BLAS_PREFER_HIPBLASLT=0. SGLang AITER may fail on this SKU. Community training writeups often use ROCm 7.2.1; keep inference on 10.0. The cookbook is thinner than CUDA.
What is Radeon Stack Lab?
An Ubuntu companion app (Origin repo / local install — no public domain here) with a living catalog. Profile qwen-reliable is the daily driver (Qwen3.6-35B-A3B Q5). qwen-fast-agent is Qwen 3.8 Flash on Strata. Stack Lab.
Do I need a harness?
No. You need tokens on the GPU. Hermes or OpenClaw come after a loopback /v1 works. Local Hermes can draft the weekday YAML scrape; it does not publish.
Is OpenClaw the same as Hermes?
No. Hermes is an agent product from Nous. OpenClaw is a multi-channel self-hosted gateway. Harnesses.
Will two R9700s double speed?
Not automatically. Expect RCCL and topology work. Measure. Our daily driver is one card. The house rule is one LLM per GPU unless you deliberately combine. Use the calculator split toggle before you buy a second card for capacity.
Do buy links pay you?
As an Amazon Associate I earn from qualifying purchases. Some Newegg links may use Rakuten Advertising and some B&H links may use B&H’s program when those IDs are set. Micro Center and AMD Store are not affiliate links. We never show a live or stale price. Disclosure.