Skip to content

armgit / dspy

Public Imported from GitHub
0

DSPy: The framework for programming—not prompting—language models

Python

ci(llm): remove AMX backend variant to fix segfault on Sapphire Rapids runners (#10079)

Isaac Miller6f68dcdb3ef4
diff --git a/.github/workflows/run_tests.yml b/.github/workflows/run_tests.yml
index 49396ffb2..9018e369e 100644
--- a/.github/workflows/run_tests.yml
+++ b/.github/workflows/run_tests.yml
@@ -150,6 +150,12 @@ jobs:
-p 11434:11434 \
-v "$GITHUB_WORKSPACE/ollama-data:/root/.ollama" \
ollama/ollama:0.31.2
+ # Some GitHub runners are Sapphire-Rapids-class machines with AMX. llama.cpp's
+ # AMX-repacked weight path segfaults there during model load (llama-server dies
+ # with SIGSEGV in warm-up), while all non-AMX runners pass. Removing the AMX
+ # backend variant makes ggml's autodetection fall back to the icelake kernels
+ # that every other runner already uses.
+ docker exec ollama rm /usr/lib/ollama/libggml-cpu-sapphirerapids.so
timeout 60 bash -c 'until curl -f http://localhost:11434/api/version; do sleep 2; done'
- name: Pull LLM
if: steps.cache-ollama.outputs.cache-hit != 'true'