From 9fba0906eee536ec4ef358b93f128dbb4ed953e5 Mon Sep 17 00:00:00 2001
From: Tarjei Mandt <kernelpool@gmail.com>
Date: Fri, 7 Aug 2026 14:56:02 +1000
Subject: [PATCH] Remove the per-op completion handler in gpu::eval

---
 mlx/backend/metal/device.cpp |  3 +++
 mlx/backend/metal/device.h   | 10 ++++++++++
 mlx/backend/metal/eval.cpp   | 26 ++++++++++----------------
 3 files changed, 23 insertions(+), 16 deletions(-)

diff --git a/mlx/backend/metal/device.cpp b/mlx/backend/metal/device.cpp
index 197b919b1d..c7f0eaed24 100644
--- a/mlx/backend/metal/device.cpp
+++ b/mlx/backend/metal/device.cpp
@@ -520,6 +520,7 @@ void CommandEncoder::commit(std::function<void()> completion) {
       [&error_ = error_,
        wait_events = std::move(wait_events_),
        signal_events = std::move(signal_events_),
+       retained = std::move(retained_buffers_),
        completion = std::move(completion)](MTL::CommandBuffer* cbuf) {
         if (completion) {
           completion();
@@ -556,6 +557,8 @@ void CommandEncoder::commit(std::function<void()> completion) {
   buffer_ = NS::RetainPtr(queue_->commandBufferWithUnretainedReferences());
   buffer_ops_ = 0;
   buffer_sizes_ = 0;
+  retained_buffers_.clear();
+  retained_ptrs_.clear();
 }
 
 void CommandEncoder::synchronize() {
diff --git a/mlx/backend/metal/device.h b/mlx/backend/metal/device.h
index 871e95ccff..4ba79cebb6 100644
--- a/mlx/backend/metal/device.h
+++ b/mlx/backend/metal/device.h
@@ -54,6 +54,14 @@ class MLX_API CommandEncoder {
   void add_temporary(array arr);
   void add_temporaries(std::vector<array> arrays);
 
+  // Keep a buffer alive until the current command buffer completes. Deduped
+  // because a primitive may bind the same array in several slots.
+  void hold_buffer(const std::shared_ptr<array::Data>& buf) {
+    if (buf && retained_ptrs_.insert(buf.get()).second) {
+      retained_buffers_.push_back(buf);
+    }
+  }
+
   void dispatch_threadgroups(MTL::Size grid_dims, MTL::Size group_dims);
   void dispatch_threads(MTL::Size grid_dims, MTL::Size group_dims);
   void maybeInsertBarrier();
@@ -129,6 +137,8 @@ class MLX_API CommandEncoder {
   bool needs_barrier_{false};
   bool concurrent_{false};
   std::vector<array> temporaries_;
+  std::vector<std::shared_ptr<array::Data>> retained_buffers_;
+  std::unordered_set<const array::Data*> retained_ptrs_;
   std::unordered_set<MTL::Resource*> prev_inputs_;
   std::unordered_set<MTL::Resource*> prev_outputs_;
   std::unordered_set<MTL::Resource*> next_inputs_;
diff --git a/mlx/backend/metal/eval.cpp b/mlx/backend/metal/eval.cpp
index e6826253de..36c120bf70 100644
--- a/mlx/backend/metal/eval.cpp
+++ b/mlx/backend/metal/eval.cpp
@@ -1,6 +1,4 @@
 // Copyright © 2023-2024 Apple Inc.
-#include <memory>
-
 #include "mlx/backend/gpu/eval.h"
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/utils.h"
@@ -44,27 +42,23 @@ void eval(array& arr) {
     debug_set_primitive_buffer_label(command_buffer, arr.primitive());
     arr.primitive().eval_gpu(arr.inputs(), outputs);
   }
-  std::unordered_set<std::shared_ptr<array::Data>> buffers;
+  // Skip a donated output's buffer since holding it blocks allocator reuse.
+  const auto& out_data = arr.data_shared_ptr();
   for (auto& in : arr.inputs()) {
-    buffers.insert(in.data_shared_ptr());
-  }
-  for (auto& s : arr.siblings()) {
-    buffers.insert(s.data_shared_ptr());
+    if (in.data_shared_ptr() != out_data) {
+      encoder.hold_buffer(in.data_shared_ptr());
+    }
   }
-  // Remove the output if it was donated to by an input
-  if (auto it = buffers.find(arr.data_shared_ptr()); it != buffers.end()) {
-    buffers.erase(it);
+  for (auto& sib : arr.siblings()) {
+    if (sib.data_shared_ptr() != out_data) {
+      encoder.hold_buffer(sib.data_shared_ptr());
+    }
   }
 
   if (encoder.needs_commit()) {
     encoder.end_encoding();
     scheduler::notify_new_task(s);
-    encoder.commit([s, buffers = std::move(buffers)]() {
-      scheduler::notify_task_completion(s);
-    });
-  } else {
-    command_buffer->addCompletedHandler(
-        [buffers = std::move(buffers)](MTL::CommandBuffer* cbuf) {});
+    encoder.commit([s]() { scheduler::notify_task_completion(s); });
   }
 }