summaryrefslogtreecommitdiff
path: root/machines/kusanagi
diff options
context:
space:
mode:
authorHenry J. Webster <hwebs@hwebs.info>2026-08-03 21:38:08 -0500
committerHenry J. Webster <hwebs@hwebs.info>2026-08-03 21:39:51 -0500
commit71e36370770acc7498748a890cd0365b3d351027 (patch)
tree60c4180135368f928bdc651f7b5d769886c56441 /machines/kusanagi
parente9b9671d238bb3de779c0dbf4a518b41280be778 (diff)
kusanagi/ollama: fix ROCm GPU fallback to CPU for Gemma
The RDNA3 card reports as gfx1101/gfx1102/gfx11-generic, which the ROCm build bundled with ollama-rocm has no kernels for, so ollama found no usable device and ran 100% on CPU after the nixpkgs bump. Force it to present as gfx1100 via HSA_OVERRIDE_GFX_VERSION. Also set OLLAMA_FLASH_ATTENTION=1 so the existing q8_0 KV cache actually engages instead of silently falling back to f16 and inflating VRAM use. Assisted-by: Claude-code:Opus-4.8
Diffstat (limited to 'machines/kusanagi')
-rw-r--r--machines/kusanagi/default.nix8
1 files changed, 8 insertions, 0 deletions
diff --git a/machines/kusanagi/default.nix b/machines/kusanagi/default.nix
index 5e6dc1e..a22843d 100644
--- a/machines/kusanagi/default.nix
+++ b/machines/kusanagi/default.nix
@@ -212,6 +212,14 @@
package = pkgs.ollama-rocm;
environmentVariables = {
HIP_VISIBLE_DEVICES = "0";
+ # RDNA3 card reports as gfx1101/gfx1102/gfx11-generic, which the ROCm
+ # build bundled with ollama-rocm doesn't have kernels for -> it finds no
+ # usable device and silently runs 100% on CPU. Force it to present as
+ # gfx1100 (7900 XT/XTX), which is fully supported and binary-compatible.
+ HSA_OVERRIDE_GFX_VERSION = "11.0.0";
+ # KV-cache quantization (below) is a no-op without flash attention; ollama
+ # falls back to f16 KV cache, inflating VRAM use and forcing CPU offload.
+ OLLAMA_FLASH_ATTENTION = "1";
OLLAMA_CONTEXT_LENGTH = "32768";
OLLAMA_KV_CACHE_TYPE = "q8_0";
};