From b12cbcb5235bb43629d39364eb15a11b995bc72c Mon Sep 17 00:00:00 2001
From: CalamitousFelicitousness <iowasovereign@gmail.com>
Date: Sat, 18 Jul 2026 20:12:23 +0100
Subject: [PATCH] feat(api): report model memory placement in memory endpoint

Add a model section to /sdapi/v1/memory with loaded-model bytes summed
per pipeline component and device, so clients can tell resident weights
from offloaded ones and loop-critical components from edge ones.

- walk components over parameters and buffers, dedupe shared storages,
  key by component name then device type
- read the raw model slot so a memory poll never triggers a model load
- section is exception-isolated like ram and cuda; reports an error
  string if the walk races a reload
---
 modules/api/models.py |  1 +
 modules/api/server.py | 32 +++++++++++++++++++++++++++++++-
 2 files changed, 32 insertions(+), 1 deletion(-)

diff --git a/modules/api/models.py b/modules/api/models.py
index 7bb0f8f27..bbe5a5771 100644
--- a/modules/api/models.py
+++ b/modules/api/models.py
@@ -552,6 +552,7 @@ class ResEmbeddings(BaseModel):
 class ResMemory(BaseModel):
     ram: dict = Field(title="RAM", description="System memory stats")
     cuda: dict = Field(title="CUDA", description="nVidia CUDA memory stats")
+    model: dict = Field(default={}, title="Model", description="Loaded model bytes per component and device")
 
 class ResScripts(BaseModel):
     txt2img: list[str] = Field(title="Txt2img", description="Titles of scripts (txt2img)")
diff --git a/modules/api/server.py b/modules/api/server.py
index 3145ac61a..aee51ae4a 100644
--- a/modules/api/server.py
+++ b/modules/api/server.py
@@ -165,6 +165,32 @@ def post_skip():
     shared.state.skip()
     return Response(status_code=204)
 
+def get_model_placement():
+    import torch
+    from modules.modeldata import model_data
+    pipe = model_data.sd_model # raw slot: the shared.sd_model property can trigger a model load
+    if pipe is None:
+        return {}
+    components = getattr(pipe, 'components', None) or ({ 'model': pipe } if isinstance(pipe, torch.nn.Module) else {})
+    placement = {}
+    seen = set()
+    for name, component in components.items():
+        if not isinstance(component, torch.nn.Module):
+            continue
+        devmap = {}
+        for tensors in (component.parameters(), component.buffers()):
+            for t in tensors:
+                ptr = 0 if t.is_meta else t.untyped_storage().data_ptr()
+                if ptr:
+                    if ptr in seen:
+                        continue
+                    seen.add(ptr)
+                devmap[t.device.type] = devmap.get(t.device.type, 0) + t.numel() * t.element_size()
+        if devmap:
+            placement[name] = devmap
+    return placement
+
+
 def get_memory():
     try:
         import psutil
@@ -197,4 +223,8 @@ def get_memory():
             cuda = { 'error': 'unavailable' }
     except Exception as err:
         cuda = { 'error': f'{err}' }
-    return models.ResMemory(ram = ram, cuda = cuda)
+    try:
+        model = get_model_placement() # walk can race a model reload or offload rewrap; report the error and keep the endpoint alive
+    except Exception as err:
+        model = { 'error': f'{err}' }
+    return models.ResMemory(ram = ram, cuda = cuda, model = model)