From a3fe4a1b2d72ee539b70a3455c08e13937eca843 Mon Sep 17 00:00:00 2001
From: Fei Chen <feich@microsoft.com>
Date: Wed, 2 Jul 2025 14:47:44 +0800
Subject: [PATCH] release or cache buffer early in Release not OnRefresh

---
 .../core/providers/webgpu/buffer_manager.cc   | 28 ++++++-------------
 1 file changed, 9 insertions(+), 19 deletions(-)

diff --git a/onnxruntime/core/providers/webgpu/buffer_manager.cc b/onnxruntime/core/providers/webgpu/buffer_manager.cc
index 1d8c689cbd909..985df3b332d4f 100644
--- a/onnxruntime/core/providers/webgpu/buffer_manager.cc
+++ b/onnxruntime/core/providers/webgpu/buffer_manager.cc
@@ -193,30 +193,21 @@ class BucketCacheManager : public IBufferCacheManager {
   }
 
   void ReleaseBuffer(WGPUBuffer buffer) override {
-    pending_buffers_.emplace_back(buffer);
-  }
-
-  void OnRefresh() override {
-    // TODO: consider graph capture. currently not supported
-
-    for (auto& buffer : pending_buffers_) {
-      auto buffer_size = static_cast<size_t>(wgpuBufferGetSize(buffer));
+    auto buffer_size = static_cast<size_t>(wgpuBufferGetSize(buffer));
 
-      auto it = buckets_.find(buffer_size);
-      if (it != buckets_.end() && it->second.size() < buckets_limit_[buffer_size]) {
-        it->second.emplace_back(buffer);
-      } else {
-        wgpuBufferRelease(buffer);
-      }
+    auto it = buckets_.find(buffer_size);
+    if (it != buckets_.end() && it->second.size() < buckets_limit_[buffer_size]) {
+      it->second.emplace_back(buffer);
+    } else {
+      wgpuBufferRelease(buffer);
     }
+  }
 
-    pending_buffers_.clear();
+  void OnRefresh() override {
+    // no-op
   }
 
   ~BucketCacheManager() {
-    for (auto& buffer : pending_buffers_) {
-      wgpuBufferRelease(buffer);
-    }
     for (auto& pair : buckets_) {
       for (auto& buffer : pair.second) {
         wgpuBufferRelease(buffer);
@@ -245,7 +236,6 @@ class BucketCacheManager : public IBufferCacheManager {
   }
   std::unordered_map<size_t, size_t> buckets_limit_;
   std::unordered_map<size_t, std::vector<WGPUBuffer>> buckets_;
-  std::vector<WGPUBuffer> pending_buffers_;
   std::vector<size_t> buckets_keys_;
 };