Release built artifacts on partial failure + fix CUDA/SYCL bridge compile

.github/workflows/release.yml

@@ -160,16 +160,36 @@ jobs:
   release:
     name: Create Release
     needs: build
+    # Run even if some build targets failed, so the binaries that *did* build
+    # still get published. Including a status function overrides Actions'
+    # implicit success() gate, so `!cancelled()` runs even when a `needs` leg
+    # failed — while still skipping a manually cancelled run (unlike always(),
+    # which would cut a release from a cancelled build). Failed matrix legs
+    # upload no artifact, so they're simply absent from the release; the overall
+    # run stays red to keep the failures visible.
+    if: ${{ !cancelled() }}
     runs-on: ubuntu-latest
     steps:
       - uses: actions/download-artifact@v4
         with:
           merge-multiple: true
 
       - name: List artifacts
-        run: ls -la autocommit-*
+        id: artifacts
+        run: |
+          ls -la autocommit-* 2>/dev/null || true
+          count=$(ls autocommit-*.tar.gz 2>/dev/null | wc -l)
+          echo "count=${count}" >> "$GITHUB_OUTPUT"
+          echo "Found ${count} release artifact(s) to publish."
+
+      - name: Fail if nothing built
+        if: steps.artifacts.outputs.count == '0'
+        run: |
+          echo "::error::No build artifacts were produced — every target failed."
+          exit 1
 
       - uses: softprops/action-gh-release@v2
+        if: steps.artifacts.outputs.count != '0'
         with:
           generate_release_notes: true
           files: |

crates/llama-sys/src/autocommit_common_bridge.cpp

@@ -645,10 +645,12 @@ int autocommit_llama_params_fit(
         size_t * margins,
         uint32_t n_ctx_min,
         int log_level) {
-#ifdef LLAMA_CPP_PREBUILT
-    // llama_params_fit is not available in the b9837 binary release.
-    // Skip fitting and return success — the caller will load the model
-    // with whatever params were configured.
+    // llama_params_fit is not exposed by the pinned llama.cpp release (b9837)
+    // in either build mode: the binary release omits it, and in the from-source
+    // build the fitting logic lives in libcommon's internal common/fit.cpp with
+    // no matching public `llama_params_fit` symbol. Skip fitting and return
+    // success — the caller loads the model with whatever params were configured.
+    // This matches the behavior of the prebuilt backends that already ship.
     (void)path_model;
     (void)mparams;
     (void)cparams;
@@ -658,17 +660,6 @@ int autocommit_llama_params_fit(
     (void)n_ctx_min;
     (void)log_level;
     return 0;
-#else
-    return llama_params_fit(
-        path_model,
-        mparams,
-        cparams,
-        tensor_split,
-        tensor_buft_overrides,
-        margins,
-        n_ctx_min,
-        static_cast<enum ggml_log_level>(log_level));
-#endif
 }
 
 } // extern "C"