From 0648f4426b4fcf11f326cc87030bd4727f44666e Mon Sep 17 00:00:00 2001
From: fszontagh <51741446+fszontagh@users.noreply.github.com>
Date: Sat, 6 Jun 2026 10:32:03 +0200
Subject: [PATCH] perf: ratchet streaming budget so plan stops re-merging every
 step (#1611)

---
 src/ggml_extend.hpp    | 36 +++++++++++++++++++++++++++---------
 src/ggml_graph_cut.cpp |  6 +++---
 2 files changed, 30 insertions(+), 12 deletions(-)

diff --git a/src/ggml_extend.hpp b/src/ggml_extend.hpp
index 9dada344..f6ac1d65 100644
--- a/src/ggml_extend.hpp
+++ b/src/ggml_extend.hpp
@@ -1706,8 +1706,9 @@ protected:
     std::unordered_set<ggml_tensor*> resident_param_set;
     uint64_t resident_state_token = 0;
 
-    size_t max_graph_vram_bytes = 0;
-    bool stream_layers_enabled  = false;
+    size_t max_graph_vram_bytes           = 0;
+    bool   stream_layers_enabled          = false;
+    size_t observed_max_effective_budget_ = 0;
 
     sd::layer_registry::LayerRegistry layer_registry_;
 
@@ -2446,15 +2447,25 @@ protected:
                 constexpr size_t safety_margin = 512ull * 1024 * 1024;
                 size_t free_clamp              = (free_vram > safety_margin) ? (free_vram - safety_margin) : 0;
                 if (free_clamp < effective_budget) {
-                    LOG_INFO("%s clamping streaming budget: actual free VRAM %.2f MB < user cap %.2f MB",
-                             get_desc().c_str(),
-                             free_clamp / (1024.0 * 1024.0),
-                             effective_budget / (1024.0 * 1024.0));
+                    LOG_DEBUG("%s clamping streaming budget: actual free VRAM %.2f MB < user cap %.2f MB",
+                              get_desc().c_str(),
+                              free_clamp / (1024.0 * 1024.0),
+                              effective_budget / (1024.0 * 1024.0));
                     effective_budget = free_clamp;
                 }
             }
         }
 
+        bool budget_increased = false;
+        if (stream_layers_enabled) {
+            if (effective_budget > observed_max_effective_budget_) {
+                observed_max_effective_budget_ = effective_budget;
+                budget_increased               = true;
+            } else {
+                effective_budget = observed_max_effective_budget_;
+            }
+        }
+
         if (effective_budget_out != nullptr) {
             *effective_budget_out = effective_budget;
         }
@@ -2466,9 +2477,15 @@ protected:
                                                      params_tensor_set_,
                                                      get_desc().c_str());
         if (stream_layers_enabled) {
-            LOG_INFO("%s streaming budget = %.2f MB",
-                     get_desc().c_str(),
-                     effective_budget / (1024.0 * 1024.0));
+            if (budget_increased) {
+                LOG_INFO("%s streaming budget = %.2f MB",
+                         get_desc().c_str(),
+                         effective_budget / (1024.0 * 1024.0));
+            } else {
+                LOG_DEBUG("%s streaming budget = %.2f MB",
+                          get_desc().c_str(),
+                          effective_budget / (1024.0 * 1024.0));
+            }
         }
         return true;
     }
@@ -3053,6 +3070,7 @@ public:
             ggml_backend_buffer_free(params_buffer);
             params_buffer = nullptr;
         }
+        observed_max_effective_budget_ = 0;
     }
 
     size_t get_params_buffer_size() {
diff --git a/src/ggml_graph_cut.cpp b/src/ggml_graph_cut.cpp
index 61234eaf..2009eaf0 100644
--- a/src/ggml_graph_cut.cpp
+++ b/src/ggml_graph_cut.cpp
@@ -699,9 +699,9 @@ namespace sd::ggml_graph_cut {
         }
 
         if (log_desc != nullptr) {
-            LOG_INFO("%s graph cut max_vram budget merge took %lld ms",
-                     log_desc,
-                     ggml_time_ms() - t_budget_begin);
+            LOG_DEBUG("%s graph cut max_vram budget merge took %lld ms",
+                      log_desc,
+                      ggml_time_ms() - t_budget_begin);
         }
 
         return merged_plan;