Refresh owned MLX dependency chain

CHANGELOG.md

@@ -20,6 +20,13 @@ The format is based on Keep a Changelog.
   fallback now consume the same source of truth, with runtime limit settings
   applied only when describing the active server.
 
+### Runtime
+
+- refreshed the owned MLX and mlx-swift forks onto current upstream cutoffs,
+  retained the scoped 1-bit, NVFP4, CUDA, stream-safety, and M4/H3 patches,
+  made the generated NAX optimizations source-reproducible, and regenerated
+  the provenance-locked Metal library from the new immutable revisions.
+
 ### Music
 
 - accelerated MiniMax Music 3 on Apple Silicon with a reachable 16,385-row

Package.swift

@@ -510,7 +510,7 @@ if !isLinuxPackage {
 var packageDependencies: [Package.Dependency] = (useLinuxPrebuiltMLX ? [] : [
   .package(
     url: "https://github.com/sawfwair/mlx-swift",
-    revision: "22c76cc40edbae2c2af72532678d8e5c06cbfa46"
+    revision: "3e6df6d8163a8f212061d15739eeeec12d5b89e3"
   )
 ]) + [
   .package(

Sources/MereRunCLI/Support/MLXBundleSupport.swift

@@ -27,8 +27,8 @@ enum MLXBundleSupport {
 
     static let expectedProvenance = MetallibProvenance(
       coreVersion: "0.32.1",
-      swiftRevision: "22c76cc40edbae2c2af72532678d8e5c06cbfa46",
-      kernelSourcesSHA256: "961bc42bb4161801fb8b0f7fd9d4570e7e86318f1644520bb39eae04eed9820c"
+      swiftRevision: "3e6df6d8163a8f212061d15739eeeec12d5b89e3",
+      kernelSourcesSHA256: "fb0c62d372d6aaa75edfbcb950d9dd797fce944a7df7bcde24dce2a672024be5"
     )
 
     /// Relationship between a bundle's metallib version stamp

THIRD_PARTY_NOTICES.md

@@ -716,10 +716,10 @@ SOFTWARE.
 - source project: [`sawfwair/mlx-swift`](https://github.com/sawfwair/mlx-swift),
   based on upstream [`ml-explore/mlx-swift`](https://github.com/ml-explore/mlx-swift)
   0.32.1
-- pinned package revision: `22c76cc40edbae2c2af72532678d8e5c06cbfa46`
-- embedded MLX revision: `2a977448c20451eb4f100691bd2f6513e610f150`
+- pinned package revision: `3e6df6d8163a8f212061d15739eeeec12d5b89e3`
+- embedded MLX revision: `b57bd7640f3f7c743b76a58478faaf1e8ee084f2`
 - generated-kernel source SHA-256:
-  `961bc42bb4161801fb8b0f7fd9d4570e7e86318f1644520bb39eae04eed9820c`
+  `fb0c62d372d6aaa75edfbcb950d9dd797fce944a7df7bcde24dce2a672024be5`
 - license: MIT
 
 ```

Tests/MereRunCLITests/LinuxNativeBridgeTests.swift

@@ -110,6 +110,9 @@ final class LinuxNativeBridgeTests: XCTestCase {
         XCTAssertTrue(packageScript.contains("cuda_runtime_root"))
         XCTAssertTrue(packageScript.contains("export CUDA_HOME=\"$cuda_runtime_root\""))
         XCTAssertTrue(packageScript.contains("export CUDA_PATH=\"$CUDA_HOME\""))
+        XCTAssertTrue(packageScript.contains("cuda_runtime_library_candidates"))
+        XCTAssertTrue(packageScript.contains("$cuda_root/targets/$cuda_target/lib"))
+        XCTAssertTrue(packageScript.contains("LD_LIBRARY_PATH=\"$LD_LIBRARY_PATH:$cuda_runtime_library\""))
         XCTAssertTrue(packageScript.contains("MERERUN_LLAMA_CLI"))
         XCTAssertTrue(packageScript.contains("$llama_prefix/bin/llama-cli"))
         XCTAssertTrue(packageScript.contains("cuda_root/include/cccl"))

docs/mlx-swift-fork.md

@@ -1,23 +1,62 @@
 # mlx-swift fork policy and compiled-call overhead
 
-mere-run pins the public `sawfwair/mlx-swift` fork based on upstream 0.32.1.
-That fork carries the Linux package bridge and Prism low-bit compatibility
-needed by this repository. Its embedded `sawfwair/mlx` revision also provides
-native affine 1-bit CUDA quantize, dequantize, and QMV execution, plus the
-Metal affine 1-bit path and the generation-17 NAX correctness gate. The fork
-retains upstream's NVFP4 split-K fix and drops the old duplicate `qmv_wide`
-implementation now supplied upstream. Changes stay scoped to their bit width,
-group size, quantization mode, and backend so stock 2-bit and wider models keep
-their existing paths.
+mere-run pins the public `sawfwair/mlx-swift` fork at
+`3e6df6d8163a8f212061d15739eeeec12d5b89e3`. It is rebased onto upstream
+`mlx-swift` `da318704cc0e972b61dcca43c62cd15e545362ae`, including the upstream
+`MLXArray` finalizer fix and generated-source-list maintenance. The embedded
+`sawfwair/mlx` revision is
+`b57bd7640f3f7c743b76a58478faaf1e8ee084f2`, based on upstream MLX
+`bd5c3a2b170bb95340482e35b2a49fb08aea4de3` and retaining the 0.32.1 ABI.
+
+The owned patch stack carries the Linux/CUDA package bridge, executor-safe
+Swift streams, native affine 1-bit CUDA quantize/dequantize/QMV execution, the
+Metal affine 1-bit path, custom quantized-kernel headers, NVFP4 staging, the
+generation-17 NAX correctness gate, and M4/H3 tuning. The obsolete unaligned
+1-bit fast-kernel tail was not replayed because upstream now requires aligned
+fast dispatch; the fork instead tests matching host/kernel alignment directly.
+Changes stay scoped to their bit width, group size, quantization mode, and
+backend so stock 2-bit and wider models keep their existing paths.
 
 The current pin also lets an MLXFast custom Metal kernel explicitly request
 the core quantized helper headers. The source marker
 `MLX_INCLUDE_FP_QUANTIZED_HEADERS` exposes the NVFP4 helpers, while
 `MLX_INCLUDE_AFFINE_QUANTIZED_HEADERS` exposes the affine helpers. Kernels
-without either marker compile from the unchanged default header set. The
-embedded MLX core revision is `9bee51623d5a34806821e0f414fae293f90dda19`;
-the mlx-swift revision pinned by mere.run is
-`a9485c38b82b8e0cee76e8f1f5a3e5bf6f36d543`.
+without either marker compile from the unchanged default header set. The NAX
+attention and gather-tile optimizations now live in MLX core source, so
+regenerating mlx-swift's AOT sources reproduces them instead of depending on
+generated-file-only commits.
+
+## Refresh procedure
+
+Refresh the dependency chain from the bottom up and publish it in the same
+order:
+
+1. Record immutable upstream cutoffs for both repositories and classify every
+   fork-only commit as replay, replace with upstream, or drop with a regression
+   test. Rebase the MLX core fork first.
+2. Build an installable MLX wheel from the rebased core, run the full core gate,
+   and exercise every owned quantization mode on its actual backend. Do not
+   treat source compilation as runtime proof.
+3. Update mlx-swift's MLX submodule to the reviewed core revision, regenerate
+   AOT sources, regenerate a second time to prove the tree is idempotent, and
+   run the full Xcode test plan so the Metal shader bundle is present.
+4. Publish the reviewed core revision, then pin that immutable revision from
+   mlx-swift and publish the reviewed Swift revision. Never make mere.run depend
+   on an unpushed or floating dependency ref.
+5. Pin the immutable mlx-swift revision in `Package.swift` and
+   `Package.resolved`, rebuild the vendored Metal library, and update all three
+   provenance fields: Swift revision, core version, and generated-source hash.
+6. Run `./scripts/check.sh`, followed by both supported runtime gates:
+   `MERERUN_RUN_E2E=core ./scripts/check.sh` and
+   `MERERUN_RUN_E2E=installed ./scripts/check.sh`. Include a real generation
+   smoke for any model family implicated by the changed stream, quantization,
+   or kernel paths.
+7. Open dependency-ordered draft pull requests (MLX, mlx-swift, mere.run), link
+   them explicitly, and merge only after the downstream pin and platform CI are
+   green against the exact advertised SHAs.
+
+Perform this audit at least once per upstream minor release, and sooner for a
+security fix or a correctness/performance change in a path mere.run owns.
 
 A separate measured one-line compiled-call optimization exists on staging
 branches but is **deliberately not included in the pin**. The rest of this
@@ -75,8 +114,8 @@ architecture exercises it" is past the threshold to hold it back.
    *independent* compiled functions (and new shapes of shared ones) from
    multiple threads. Repeated calls to a single closure — including the
    existing `MLXCompiledFunctionOverheadTests` micro-bench — cannot catch
-   this class. (Prerequisite: fix the `default.metallib` packaging issue that
-   currently blocks Metal-backed concurrency tests in the test bundle.)
+   this class. Run it through Xcode, which builds the Metal shader bundle;
+   command-line SwiftPM alone cannot build that bundle.
 
 Do **not** upstream or re-pin the lock removal alone.
 

docs/runtime/geo.md

@@ -1,7 +1,7 @@
 # Geospatial Runtime
 
 `mere.run geo` is a local-first inference boundary for Earth-observation and
-humanitarian workflows. Any 
+humanitarian workflows. Any
 workflow can prepare the documented tensors, run the native model locally, and
 carry the resulting candidates or embeddings into its own evidence process.
 

scripts/e2e_smoke.sh

@@ -180,7 +180,8 @@ core_suite() {
   if require_model "speech-asr-qwen3" && [[ -f "$speech_tts_wav" ]]; then
     run_step "speech_transcribe_qwen3" 300 \
       "$MERERUN_BIN" speech transcribe "$speech_tts_wav" --backend qwen --model speech-asr-qwen3 --no-timestamps --quiet
-    assert_contains "speech_transcribe_qwen3_output" "$OUT_DIR/speech_transcribe_qwen3.stdout" "local inference smoke test"
+    assert_contains "speech_transcribe_qwen3_output" "$OUT_DIR/speech_transcribe_qwen3.stdout" \
+      "local inference[^[:alnum:]]+smoke[^[:alnum:]]+test"
   fi
 
   if require_model "text-chat-q36-nano"; then

scripts/package-linux.sh

@@ -246,6 +246,7 @@ mlx_swift_cuda_link_flags() {
   local local_openblas_root="$native_root/deps/apt-root"
   local cudnn_library_path="${CUDNN_LIBRARY_PATH:-}"
   local cuda_library_path="${CUDA_LIBRARY_PATH:-}"
+  local cuda_library_paths=()
   local cuda_stub_library_path="${CUDA_STUB_LIBRARY_PATH:-}"
   local flags=()
 
@@ -292,25 +293,32 @@ mlx_swift_cuda_link_flags() {
       fi
     done
   fi
-  if [[ -z "$cuda_library_path" ]]; then
+  if [[ -n "$cuda_library_path" ]]; then
+    cuda_library_paths+=("$cuda_library_path")
+  else
     local cuda_library_candidates=()
     local cuda_root
     while IFS= read -r cuda_root; do
-      cuda_library_candidates+=("$cuda_root/lib64")
+      cuda_library_candidates+=("$cuda_root/lib64" "$cuda_root/lib")
       local cuda_target
       for cuda_target in "${cuda_target_names[@]}"; do
         cuda_library_candidates+=("$cuda_root/targets/$cuda_target/lib")
       done
     done < <(cuda_toolkit_root_candidates)
     for candidate in "${cuda_library_candidates[@]}"; do
       if [[ -f "$candidate/libcublasLt.so" || -f "$candidate/libnvrtc.so" || -f "$candidate/libcudart.so" ]]; then
-        cuda_library_path="$candidate"
-        break
+        cuda_library_paths+=("$candidate")
       fi
     done
   fi
-  if [[ -z "$cuda_stub_library_path" &&
-        ( -z "$cuda_library_path" || ! -f "$cuda_library_path/libcuda.so" ) ]]; then
+  local cuda_driver_link_available=0
+  for candidate in "${cuda_library_paths[@]}"; do
+    if [[ -f "$candidate/libcuda.so" ]]; then
+      cuda_driver_link_available=1
+      break
+    fi
+  done
+  if [[ -z "$cuda_stub_library_path" && "$cuda_driver_link_available" != "1" ]]; then
     local cuda_stub_candidates=()
     local cuda_root
     while IFS= read -r cuda_root; do
@@ -327,13 +335,13 @@ mlx_swift_cuda_link_flags() {
       fi
     done
   fi
-  if [[ -n "$cuda_library_path" ]]; then
-    flags+=("-L" "$cuda_library_path")
-    if [[ "$(basename "$cuda_library_path")" != "stubs" ]]; then
-      flags+=("-Xlinker" "-rpath" "-Xlinker" "$cuda_library_path")
+  for candidate in "${cuda_library_paths[@]}"; do
+    flags+=("-L" "$candidate")
+    if [[ "$(basename "$candidate")" != "stubs" ]]; then
+      flags+=("-Xlinker" "-rpath" "-Xlinker" "$candidate")
     fi
-  fi
-  if [[ -n "$cuda_stub_library_path" && "$cuda_stub_library_path" != "$cuda_library_path" ]]; then
+  done
+  if [[ -n "$cuda_stub_library_path" ]]; then
     flags+=("-L" "$cuda_stub_library_path")
   fi
   if [[ -d /usr/lib/$deb_multiarch ]]; then
@@ -493,6 +501,29 @@ for cuda_cccl_include in "${cuda_cccl_candidates[@]}"; do
   fi
 done
 export LD_LIBRARY_PATH="$payload_dir/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
+if [[ -f "$payload_dir/.mererun-linux-cuda" ]]; then
+  cuda_runtime_library_candidates=()
+  for cuda_root in \
+    "${CUDA_HOME:-}" \
+    "${CUDA_PATH:-}" \
+    /usr/local/cuda \
+    /usr/local/cuda-* \
+    /usr
+  do
+    [[ -n "$cuda_root" && -d "$cuda_root" ]] || continue
+    cuda_runtime_library_candidates+=("$cuda_root/lib64" "$cuda_root/lib")
+    for cuda_target in sbsa-linux aarch64-linux x86_64-linux; do
+      cuda_runtime_library_candidates+=("$cuda_root/targets/$cuda_target/lib")
+    done
+  done
+  for cuda_runtime_library in "${cuda_runtime_library_candidates[@]}"; do
+    [[ -d "$cuda_runtime_library" ]] || continue
+    case ":$LD_LIBRARY_PATH:" in
+      *":$cuda_runtime_library:"*) ;;
+      *) export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:$cuda_runtime_library" ;;
+    esac
+  done
+fi
 exec "$payload_dir/mere.run-bin" "$@"
 WRAPPER
 chmod +x "$payload_dir/mere.run"

scripts/prepare-linux-native.sh

@@ -885,6 +885,7 @@ mlx_swift_cuda_link_flags() {
   local local_openblas_root="$native_root/deps/apt-root"
   local cudnn_library_path="${CUDNN_LIBRARY_PATH:-}"
   local cuda_library_path="${CUDA_LIBRARY_PATH:-}"
+  local cuda_library_paths=()
   local cuda_stub_library_path="${CUDA_STUB_LIBRARY_PATH:-}"
   local flags=()
 
@@ -931,25 +932,32 @@ mlx_swift_cuda_link_flags() {
       fi
     done
   fi
-  if [[ -z "$cuda_library_path" ]]; then
+  if [[ -n "$cuda_library_path" ]]; then
+    cuda_library_paths+=("$cuda_library_path")
+  else
     local cuda_library_candidates=()
     local cuda_root
     while IFS= read -r cuda_root; do
-      cuda_library_candidates+=("$cuda_root/lib64")
+      cuda_library_candidates+=("$cuda_root/lib64" "$cuda_root/lib")
       local cuda_target
       for cuda_target in "${cuda_target_names[@]}"; do
         cuda_library_candidates+=("$cuda_root/targets/$cuda_target/lib")
       done
     done < <(cuda_toolkit_root_candidates)
     for candidate in "${cuda_library_candidates[@]}"; do
       if [[ -f "$candidate/libcublasLt.so" || -f "$candidate/libnvrtc.so" || -f "$candidate/libcudart.so" ]]; then
-        cuda_library_path="$candidate"
-        break
+        cuda_library_paths+=("$candidate")
       fi
     done
   fi
-  if [[ -z "$cuda_stub_library_path" &&
-        ( -z "$cuda_library_path" || ! -f "$cuda_library_path/libcuda.so" ) ]]; then
+  local cuda_driver_link_available=0
+  for candidate in "${cuda_library_paths[@]}"; do
+    if [[ -f "$candidate/libcuda.so" ]]; then
+      cuda_driver_link_available=1
+      break
+    fi
+  done
+  if [[ -z "$cuda_stub_library_path" && "$cuda_driver_link_available" != "1" ]]; then
     local cuda_stub_candidates=()
     local cuda_root
     while IFS= read -r cuda_root; do
@@ -966,13 +974,13 @@ mlx_swift_cuda_link_flags() {
       fi
     done
   fi
-  if [[ -n "$cuda_library_path" ]]; then
-    flags+=("-L" "$cuda_library_path")
-    if [[ "$(basename "$cuda_library_path")" != "stubs" ]]; then
-      flags+=("-Xlinker" "-rpath" "-Xlinker" "$cuda_library_path")
+  for candidate in "${cuda_library_paths[@]}"; do
+    flags+=("-L" "$candidate")
+    if [[ "$(basename "$candidate")" != "stubs" ]]; then
+      flags+=("-Xlinker" "-rpath" "-Xlinker" "$candidate")
     fi
-  fi
-  if [[ -n "$cuda_stub_library_path" && "$cuda_stub_library_path" != "$cuda_library_path" ]]; then
+  done
+  if [[ -n "$cuda_stub_library_path" ]]; then
     flags+=("-L" "$cuda_stub_library_path")
   fi
   if [[ -d /usr/lib/$deb_multiarch ]]; then

scripts/test-package-linux.sh

@@ -50,6 +50,7 @@ if [[ "${1:-}" == "__print-env" ]]; then
   printf 'MERERUN_CUDA_CCCL_INCLUDE_PATH=%s\n' "${MERERUN_CUDA_CCCL_INCLUDE_PATH:-}"
   printf 'MERERUN_MLX_CUDA_JIT_INCLUDE_PATH=%s\n' "${MERERUN_MLX_CUDA_JIT_INCLUDE_PATH:-}"
   printf 'CPATH=%s\n' "${CPATH:-}"
+  printf 'LD_LIBRARY_PATH=%s\n' "${LD_LIBRARY_PATH:-}"
   exit 0
 fi
 echo "mere.run package fixture"
@@ -129,7 +130,9 @@ fi
 
 cuda_home_fixture="$fixture_root/cuda-13.0"
 cuda_cccl_fixture="$cuda_home_fixture/targets/sbsa-linux/include/cccl"
-mkdir -p "$cuda_cccl_fixture/cuda/std"
+cuda_lib64_fixture="$cuda_home_fixture/lib64"
+cuda_sbsa_lib_fixture="$cuda_home_fixture/targets/sbsa-linux/lib"
+mkdir -p "$cuda_cccl_fixture/cuda/std" "$cuda_lib64_fixture" "$cuda_sbsa_lib_fixture"
 wrapper_env_output="$(CUDA_HOME="$cuda_home_fixture" "$payload_dir/mere.run" __print-env)"
 if ! grep -q "^MERERUN_CUDA_CCCL_INCLUDE_PATH=$cuda_cccl_fixture$" <<<"$wrapper_env_output"; then
   echo "[test-package-linux] launcher did not export the target-specific CUDA CCCL include root:" >&2
@@ -183,6 +186,13 @@ if ! grep -q "$cuda_payload_dir/include" <<<"$cuda_wrapper_env_output" ||
   printf '%s\n' "$cuda_wrapper_env_output" >&2
   exit 1
 fi
+if ! grep -q "^LD_LIBRARY_PATH=$cuda_payload_dir/lib" <<<"$cuda_wrapper_env_output" ||
+   ! grep -q "$cuda_lib64_fixture" <<<"$cuda_wrapper_env_output" ||
+   ! grep -q "$cuda_sbsa_lib_fixture" <<<"$cuda_wrapper_env_output"; then
+  echo "[test-package-linux] CUDA launcher did not add both generic and SBSA toolkit libraries to LD_LIBRARY_PATH:" >&2
+  printf '%s\n' "$cuda_wrapper_env_output" >&2
+  exit 1
+fi
 if ! dpkg-deb --field "$cuda_deb" Depends | grep -q 'libcufft-13-0'; then
   echo "[test-package-linux] expected CUDA .deb dependencies to include libcufft-13-0:" >&2
   dpkg-deb --field "$cuda_deb" Depends >&2

vendor/mlx-swift_Cmlx.bundle/Contents/Resources/default.metallib.version

@@ -1,6 +1,6 @@
 mlx-core-version: 0.32.1
 mlx-swift-version: unknown
-mlx-swift-revision: 22c76cc40edbae2c2af72532678d8e5c06cbfa46
-kernel-sources-sha256: 961bc42bb4161801fb8b0f7fd9d4570e7e86318f1644520bb39eae04eed9820c
-built-at: 2026-08-08T20:47:45Z
+mlx-swift-revision: 3e6df6d8163a8f212061d15739eeeec12d5b89e3
+kernel-sources-sha256: fb0c62d372d6aaa75edfbcb950d9dd797fce944a7df7bcde24dce2a672024be5
+built-at: 2026-08-14T11:27:04Z
 metal-compiler: Apple metal version 32023.883 (metalfe-32023.883)