fix: use DEBUG=1 and MAKEFLAGS=-j1 in CI image to prevent EXLA OOM

The EXLA Makefile uses -O3 by default which causes massive memory usage
during C++ compilation. DEBUG=1 replaces -O3 with -g, drastically
reducing peak RAM. Combined with MAKEFLAGS=-j1 to serialize jobs, this
should fit within CI runner memory limits.

.forgejo/workflows/build.yaml

@@ -155,7 +155,13 @@ jobs:
               sh -euc '\
               mkdir -p /app && cd /app && tar xf - && \
               mix deps.get && \
-              cp -a /opt/exla-cache/cache /app/deps/exla/cache && \
+              if [ -d /opt/exla-cache/cache/0.13.0 ]; then \
+                rm -rf /app/deps/exla/cache && \
+                cp -a /opt/exla-cache/cache /app/deps/exla/cache; \
+                echo "Restored pre-built EXLA cache"; \
+              else \
+                echo "No pre-built EXLA cache found, will compile from source"; \
+              fi && \
               mix ecto.create --quiet && \
               mix ecto.migrate --quiet && \
               mix test'

Dockerfile.ci

@@ -34,9 +34,11 @@ COPY vendor /tmp/exla_warm/vendor
 RUN mkdir -p /tmp/exla_warm/config && \
     printf 'import Config\nconfig :nx, :default_backend, EXLA.Backend\n' > /tmp/exla_warm/config/config.exs
 
-# Compile all deps. EXLA's cached_make downloads the XLA archive, compiles
-# the C++ NIF, and caches libexla.so at ~/.cache/xla/exla/{key}/.
-RUN mix deps.get && mix deps.compile
+# Compile all deps. DEBUG=1 tells the EXLA Makefile to use -g instead
+# of -O3 — this avoids the heavy optimization passes that OOM-kill g++
+# on low-memory CI runners. MAKEFLAGS=-j1 serializes make jobs. The
+# resulting unoptimized NIF is fine for tests.
+RUN MAKEFLAGS="-j1" DEBUG=1 mix deps.get && MAKEFLAGS="-j1" DEBUG=1 mix deps.compile
 
 # Save the full build cache (libexla.so + .o files + xla_extension lib)
 # so Make finds all targets already built when restored in the test step.