Reapply "fix(ssd_streamer): remove obsolete buffer bounds check causing GLM-5.1 MoE crash" (#8)

This reverts commit 9b95713.

Author: Aegis AI Assistant

.github/workflows/downstream_integration.yml

@@ -32,11 +32,10 @@ jobs:
         working-directory: SwiftLM
 
       - name: Override mlx-swift dependency with PR checkout
-        # SwiftLM uses mlx-swift as a local path dependency (./mlx-swift submodule).
-        # SPM `package edit` cannot edit path dependencies, so we directly overwrite
-        # the submodule directory with our PR branch content.
+        # SPM `package edit` puts a local editable copy of the package in
+        # .build/checkouts; we then replace its contents with our PR branch.
         run: |
-          rsync -a --delete $GITHUB_WORKSPACE/mlx-swift-pr/ $GITHUB_WORKSPACE/SwiftLM/mlx-swift/
+          swift package edit mlx-swift --path $GITHUB_WORKSPACE/mlx-swift-pr
         working-directory: SwiftLM
 
       - name: Install MLX Metal library

Source/Cmlx/mlx/mlx/backend/metal/ssd_streamer.mm

@@ -39,9 +39,8 @@
 }
 
 void SSDStreamer::load_sync(off_t byte_offset, size_t length, void* dst_ptr) {
-    if (length > buffer_size_bytes_) {
-        throw std::invalid_argument("[SSDStreamer] Load length exceeds Pinned Buffer capacity.");
-    }
+    // The StreamBuffer pool was removed in favor of caller-provided unified memory,
+    // so we no longer bounds-check the length against an initial buffer size.
 
     // Synchronously read exactly byte_offset into the MLX allocator CPU mapped pointer.
     // By blocking the MLX graph evaluator thread, we implement absolute backpressure