binarynoise

Svelte has https://www.npmjs.com/package/@svelte-plugins/tooltips but that doesn't seem to get used yet.

The math-mode rule seems broken, it requires semicolons or line breaks before the `it`s now

``` systemd[1]: Started Ollama Service. sudo[179945]: pam_unix(sudo:session): session closed for user root ollama[179970]: time=2025-11-08T20:52:13.208+01:00 level=INFO source=routes.go:1525 msg="server config" env="map[CUDA_VISIBLE_DEVICES: GGML_VK_VISIBLE_DEVICES: GPU_DEVICE_ORDINAL: HIP_VISIBLE_DEVICES: HSA_OVERRIDE_GFX_VERSION: HTTPS_PROXY: HTTP_PROXY: NO_PROXY: OLLAMA_CONTEXT_LENGTH:4096 OLLAMA_DEBUG:INFO OLLAMA_FLASH_ATTENTION:true OLLAMA_GPU_OVERHEAD:3000000000...

As I told, I set `OLLAMA_GPU_OVERHEAD:3000000000` to get rid of the crashes. Yes this is a custom build but almost back to upstream (I increased the timeouts to talk to...

`Linux 6.12.57-1-lts #1 SMP PREEMPT_DYNAMIC Mon, 03 Nov 2025 14:27:55 +0000 x86_64 GNU/Linux`

It seems that for some models, even 3GB of overhead is not enough and they still run out of VRAM. As I thought, this workaround does not really help solve...

I temporarily downgraded my kernel to `Linux 6.11.0-arch1-1 #1 SMP PREEMPT_DYNAMIC Sun, 15 Sep 2024 18:38:36 +0000 x86_64 GNU/Linux`, the amdgpu warnings went away, the crashes stayed. ``` systemd[1]: Started...

Yes, now the reported memory matches the sysfs node. However, it still allocates more memory than calculated: With the 3GB overhead, it allocates 14.8 GiB instead of the 11.1 GiB...

I think I should rename this to "more memory allocated than available/calculated". My computer gets crashed unless I manually add a RAM overhead so it can use 6GB more than...

It doesn't really crash, the PC just hangs up. The last thing I see in `journalctl` before it freezes is that the model gets loaded (this may take a while)...