diff --git a/mlx/backend/metal/eval.cpp b/mlx/backend/metal/eval.cpp
index 6f55976efe..603f1c83c2 100644
--- a/mlx/backend/metal/eval.cpp
+++ b/mlx/backend/metal/eval.cpp
@@ -18,18 +18,37 @@ void new_stream(Stream s) {
   encoders.try_emplace(s.index, d, s.index, d.residency_set());
 }
 
-inline void check_error(MTL::CommandBuffer* cbuf) {
+// Capture (don't throw) a Metal command-buffer error from an async
+// completion handler. The handler runs on a Metal-managed dispatch
+// thread; throwing through Objective-C frames hits _objc_terminate ->
+// abort() and crashes the whole process. Storing the message and
+// re-throwing on the next user-thread eval()/finalize() call keeps
+// the error catchable by Python / C++ callers.
+inline void capture_async_error(const Stream& s, MTL::CommandBuffer* cbuf) {
   if (cbuf->status() == MTL::CommandBufferStatusError) {
     std::ostringstream msg;
     msg << "[METAL] Command buffer execution failed: "
         << cbuf->error()->localizedDescription()->utf8String();
-    throw std::runtime_error(msg.str());
+    scheduler::capture_error(s, msg.str());
+  }
+}
+
+// Re-throw any previously-captured async Metal error on the calling
+// thread. Called at the entry of every user-facing eval()/finalize()
+// so failures from the prior step surface before more work is queued.
+inline void throw_if_captured(const Stream& s) {
+  auto msg = scheduler::take_error(s);
+  if (!msg.empty()) {
+    throw std::runtime_error(msg);
   }
 }
 
 void eval(array& arr) {
   auto pool = metal::new_scoped_memory_pool();
   auto s = arr.primitive().stream();
+  // Surface any async Metal error captured by a prior step's completion
+  // handler before queuing more work on this stream.
+  throw_if_captured(s);
   auto& encoder = metal::get_command_encoder(s);
   auto* command_buffer = encoder.get_command_buffer();
 
@@ -63,32 +82,45 @@ void eval(array& arr) {
     command_buffer->addCompletedHandler(
         [s, buffers = std::move(buffers)](MTL::CommandBuffer* cbuf) {
           scheduler::notify_task_completion(s);
-          check_error(cbuf);
+          capture_async_error(s, cbuf);
         });
     encoder.commit();
   } else {
     command_buffer->addCompletedHandler(
-        [buffers = std::move(buffers)](MTL::CommandBuffer* cbuf) {
-          check_error(cbuf);
+        [s, buffers = std::move(buffers)](MTL::CommandBuffer* cbuf) {
+          capture_async_error(s, cbuf);
         });
   }
 }
 
 void finalize(Stream s) {
+  // Surface any prior async Metal error before sealing the stream.
+  throw_if_captured(s);
   auto pool = metal::new_scoped_memory_pool();
   auto& encoder = metal::get_command_encoder(s);
   auto* cb = encoder.get_command_buffer();
   encoder.end_encoding();
-  cb->addCompletedHandler([](MTL::CommandBuffer* cbuf) { check_error(cbuf); });
+  cb->addCompletedHandler(
+      [s](MTL::CommandBuffer* cbuf) { capture_async_error(s, cbuf); });
   encoder.commit();
 }
 
 void synchronize(Stream s) {
   metal::get_command_encoder(s).synchronize();
+  // CommandEncoder::synchronize only checks the FINAL command buffer.
+  // An async failure on an earlier (already-completed) command buffer
+  // landed in the captured-error slot — surface it here too so
+  // synchronize() is a true "all my queued work succeeded" guarantee.
+  throw_if_captured(s);
 }
 
 void clear_streams() {
   metal::get_command_encoders().clear();
+  // After tearing down encoder state, also clear any poisoned-stream
+  // flags so the streams are usable again. (The flags would otherwise
+  // outlive the encoders they refer to and refuse all subsequent
+  // work — see scheduler::StreamThread.)
+  scheduler::reset_all_errors();
 }
 
 } // namespace mlx::core::gpu
diff --git a/mlx/scheduler.h b/mlx/scheduler.h
index c84ab62855..3988a1d025 100644
--- a/mlx/scheduler.h
+++ b/mlx/scheduler.h
@@ -23,6 +23,58 @@ struct StreamThread {
   bool stop;
   std::thread thread;
 
+  // Errors raised from async GPU completion handlers can't safely
+  // propagate as C++ exceptions through Metal's dispatch infrastructure
+  // (an exception unwinding through Objective-C frames hits
+  // _objc_terminate -> abort()). The handler captures the message
+  // here and the next user-thread eval()/finalize() call re-throws
+  // it on a thread the language runtime can handle.
+  //
+  // To address the state-safety concern raised in upstream issue
+  // #2670 ("we don't have guarantees on the state being in a
+  // reasonable condition if there is an exception during eval"),
+  // the stream is also POISONED on capture. Once poisoned, every
+  // subsequent user-thread entry refuses to queue more work and
+  // throws — the caller must explicitly reset the stream
+  // (mx.clear_streams()) to resume. This ensures no further
+  // operations execute against potentially-corrupt encoder state.
+  std::mutex error_mtx;
+  std::string captured_error;
+  bool poisoned{false};
+
+  void capture_error(std::string msg) {
+    std::lock_guard<std::mutex> lk(error_mtx);
+    if (captured_error.empty()) {
+      captured_error = std::move(msg);
+    }
+    poisoned = true;
+  }
+
+  // Returns the captured error string (and clears it) on the first
+  // call after poisoning. Subsequent calls return a generic
+  // "stream poisoned" message until the stream is explicitly reset
+  // via reset_error().
+  std::string take_error() {
+    std::lock_guard<std::mutex> lk(error_mtx);
+    if (!poisoned) {
+      return {};
+    }
+    if (!captured_error.empty()) {
+      auto out = std::move(captured_error);
+      captured_error.clear();
+      return out;
+    }
+    return "[METAL] Stream is in error state from a prior failure. "
+           "Call mx.clear_streams() (or destroy this Stream) before "
+           "queuing more work.";
+  }
+
+  void reset_error() {
+    std::lock_guard<std::mutex> lk(error_mtx);
+    captured_error.clear();
+    poisoned = false;
+  }
+
   StreamThread() : stop(false), thread(&StreamThread::thread_fn, this) {}
 
   ~StreamThread() {
@@ -93,6 +145,48 @@ class MLX_API Scheduler {
     completion_cv.notify_all();
   }
 
+  // Capture an error from an async callback for the given stream's
+  // worker thread. Safe to call from Metal completion handlers — never
+  // throws, never blocks on user threads. Silently no-ops if the
+  // stream's thread doesn't exist (already torn down).
+  void capture_error(const Stream& stream, std::string msg) {
+    std::shared_lock lock(threads_mtx_);
+    auto it = threads_.find(stream.index);
+    if (it != threads_.end()) {
+      it->second->capture_error(std::move(msg));
+    }
+  }
+
+  // Take + clear the captured error for a stream. Returns empty string
+  // when no error is pending. Stream stays poisoned until reset_error.
+  std::string take_error(const Stream& stream) {
+    std::shared_lock lock(threads_mtx_);
+    auto it = threads_.find(stream.index);
+    if (it != threads_.end()) {
+      return it->second->take_error();
+    }
+    return {};
+  }
+
+  // Clear the poisoned flag for a stream. Called by mx.clear_streams()
+  // (and any future reset API) so the stream is usable again. The
+  // caller is expected to also have torn down the underlying GPU
+  // encoder state before clearing the flag.
+  void reset_error(const Stream& stream) {
+    std::shared_lock lock(threads_mtx_);
+    auto it = threads_.find(stream.index);
+    if (it != threads_.end()) {
+      it->second->reset_error();
+    }
+  }
+
+  void reset_all_errors() {
+    std::shared_lock lock(threads_mtx_);
+    for (auto& [_, st] : threads_) {
+      st->reset_error();
+    }
+  }
+
   int n_active_tasks() const {
     return n_active_tasks_;
   }
@@ -136,6 +230,22 @@ inline void notify_task_completion(const Stream& stream) {
   scheduler().notify_task_completion(stream);
 }
 
+inline void capture_error(const Stream& stream, std::string msg) {
+  scheduler().capture_error(stream, std::move(msg));
+}
+
+inline std::string take_error(const Stream& stream) {
+  return scheduler().take_error(stream);
+}
+
+inline void reset_error(const Stream& stream) {
+  scheduler().reset_error(stream);
+}
+
+inline void reset_all_errors() {
+  scheduler().reset_all_errors();
+}
+
 inline void wait_for_one() {
   scheduler().wait_for_one();
 }