Updates

Snapshots, not law. The home status card is the current recommendation. Subscribe:RSS.

Dated notes

Oct 12, 2026 · labeled current that week

Consumer stacks — Qwen3.6 daily driver, Flash agent, calculator

North star is now the two stacks from our test rig. llama.cpp Vulkan daily driver; Strata for Flash; /calculator and /stacks are first-class pages.

The site is a consumer AMD ROCm notebook, not a withdrawn-small-model tracker.

Daily driver on our test rig: Qwen3.6-35B-A3B Unsloth UD-Q5_K_XL, llama.cpp Vulkan, one Radeon AI PRO R9700, 262144 context, f16 KV, 0% system RAM offload.

Fast agent: Qwen 3.8 Flash Unsloth UD-Q6_K_XL on Strata v0.1.42 at 262K. Measured on that rig ~32.6 tok/s versus llama.cpp ~16 tok/s. llama.cpp Q6 is the fallback.

New pages: /stacks and the flagship /calculator. Exact serve flags and measured VRAM stay TODO until the rig owner fills them.

Oct 12, 2026 · labeled current that week

ROCm 10.1.0 notes, CLI v0.2.0, error paste, version matrix

10.1.0 lists R9700 as gfx1201 and Radeon Ubuntu 26.04.1 / 24.04.5 / 22.04.5. Published ROCm CLI is v0.2.0, not a 1.0.0 tag. New pages for search, errors, backends, quants, RSS.

Cited from the 10.1.0 release notes and the 10.1 blog.

Install walkthrough: Install ROCm. Matrix cells that we have not measured say no data: /updates/matrix. RSS: /updates/rss.xml.

Sep 22, 2026 · archived snapshot

vLLM ROCm Docker try-tag v0.30.0

Dated snapshot. AMD-validated vLLM 0.27 / ROCm 10.0.0 unchanged. Community image moved to vllm/vllm-openai-rocm:v0.30.0. Named Qwen defaults from that week were later withdrawn.

Dated digest (2026-09-22, America/Chicago). ROCm 10.0.0 is still the R9700 (gfx1201) inference pin — release notes and the live compatibility matrix. AMD’s 10.0 ecosystem still lists vLLM 0.27.

Community cutting-edge Docker tag moved: vllm/vllm-openai-rocm:v0.30.0 (Docker Hub, pushed ~2026-09-22Z). Keep v0.29.0 as the previous try-tag (wheels.vllm.ai/rocm/0.29.0/rocm723 still documents that week). Do not mix 0.27 and 0.30 in one venv. Default PyPI vllm is still CUDA.

Named Qwen defaults from this week were later withdrawn. Qwen3.5 / 3.6 27B and 35B-A3B stay multi-GPU territory — a public llama.cpp multi-GPU writeup is not a single-card recipe.

Honesty unchanged: TORCH_BLAS_PREFER_HIPBLASLT=0 on training resets; SGLang AITER may fail; prefer llama.cpp HIP/Vulkan over Ollama on gfx1201.

Sep 21, 2026 · archived snapshot

Weekly refresh — vLLM 0.27 reliability (dated)

Dated snapshot. ROCm 10.0.0 unchanged for gfx1201. Reliability serve is AMD-aligned vLLM 0.27. vLLM 0.29 ROCm Docker remains the community cutting-edge pin (no newer stable that week). Named Qwen defaults from that week were later withdrawn. R9700 training KIs documented.

This is a dated weekly card, not a new Core SDK line. ROCm 10.0.0 is still the R9700 inference pin (release notes). AMD’s 10.0 matrix still lists vLLM 0.27. Upstream vLLM 0.29.0 ROCm extras (vllm/vllm-openai-rocm:v0.29.0, wheels.vllm.ai/rocm/0.29.0/rocm723) stay as the community serve try — no newer stable ROCm vLLM found this morning.

Named Qwen defaults from this week were later withdrawn — do not treat them as current. Training/fine-tune on this SKU can GPU-reset (TORCH_BLAS_PREFER_HIPBLASLT=0); SGLang AITER may fail. Community training writeups often use ROCm 7.2.1 — keep inference on 10.0. TheRock gfx1201 perf logs are collection, not a bench table.

Sep 9, 2026 · archived snapshot

vLLM 0.29.0 ROCm extras

Upstream vLLM 0.29 ships ROCm wheels and Docker. Extra index like wheels.vllm.ai/rocm/0.29.0/rocm723; image vllm/vllm-openai-rocm:v0.29.0. Plain pip install vllm is CUDA. AMD 10.0 docs still documented 0.27.

This is the upstream snapshot people were running the week of 8–9 Sep 2026 — not AMD’s 10.0 “supported” 0.27 pair. Install with the ROCm extra index or the -rocm Docker tag. 0.30rc existed as bleeding. Re-check vLLM’s GPU install before you copy a wheel URL from memory.

Sep 2, 2026 · archived snapshot

ROCm 7.14.1 patch line

2 Sep 2026 patch on the 7.14 TheRock line. Still a valid pin if your wheel/image was built against it. Not a reason to install Ubuntu archive 7.1.

7.14 is when TheRock went production. 7.14.1 is the 2 Sep 2026 patch train. If a container tag says rocm7.14.1, stay there on purpose. Do not “upgrade” by sprinkling 10.0 libraries into that image.

Aug 26, 2026 · archived snapshot

ROCm Core SDK 10.0.0

Production TheRock line. R9700 / gfx1201 first-class. amdrocm-* packages, one gfx family, scoped /opt/rocm prefix. AMD’s notes pair PyTorch 2.13 and vLLM 0.27 — that is the supported combo, not a forever pin.

AMD’s 10.0.0 notes (26 Aug 2026) put Radeon AI PRO R9700 on the device table as gfx1201. Framework table in that release: vLLM 0.27, SGLang 0.5.15, PyTorch 2.13. Use the versioned release notes and transition guide if you are reconstructing this week.

Version matrix