Add model torch settings patch node

README.md

@@ -71,6 +71,20 @@ Inputs:
 The CPU methods use `color-matcher`. The `reinhard_lab_gpu` method uses
 ComfyUI's Kornia dependency for Lab-space matching.
 
+### `Model Patch Torch Settings`
+
+Patches a `MODEL` with ComfyUI model callbacks for PyTorch backend settings.
+
+Inputs:
+
+- `model`: model to clone and patch
+- `enable_fp16_accumulation`: toggles
+  `torch.backends.cuda.matmul.allow_fp16_accumulation` while the model runs
+
+When enabled, the patched model turns full FP16 accumulation on before model
+execution and turns it back off during cleanup. When disabled, it forces the
+setting off before model execution.
+
 ### `LM Studio Unified (URL + API Key)`
 
 Sends text, an image, or both to an LM Studio OpenAI-compatible chat endpoint.

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "enviral-design-node-pack"
-version = "0.1.10"
+version = "0.1.11"
 description = "Small, dependency-light ComfyUI utility nodes from Enviral Design."
 readme = "README.md"
 requires-python = ">=3.10"

utility_nodes.py

@@ -478,3 +478,80 @@ an optional Kornia Lab-space Reinhard path for GPU-friendly batch work.
 
 NODE_CLASS_MAPPINGS["EnviralColorMatchV2"] = EnviralColorMatchV2
 NODE_DISPLAY_NAME_MAPPINGS["EnviralColorMatchV2"] = "Enviral Color Match V2"
+
+
+class EnviralModelPatchTorchSettings:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "model": ("MODEL",),
+                "enable_fp16_accumulation": (
+                    "BOOLEAN",
+                    {
+                        "default": False,
+                        "tooltip": (
+                            "Toggles torch.backends.cuda.matmul.allow_fp16_accumulation "
+                            "while the patched model runs."
+                        ),
+                    },
+                ),
+            },
+        }
+
+    RETURN_TYPES = ("MODEL",)
+    RETURN_NAMES = ("model",)
+    FUNCTION = "patch"
+    CATEGORY = "EnviralDesign/model"
+    DESCRIPTION = """
+Adds model callbacks that enable or disable PyTorch full FP16 accumulation for
+CUDA FP16 matmuls while the model runs.
+"""
+    EXPERIMENTAL = True
+
+    @staticmethod
+    def _ensure_supported():
+        if not hasattr(torch.backends.cuda.matmul, "allow_fp16_accumulation"):
+            raise RuntimeError(
+                "torch.backends.cuda.matmul.allow_fp16_accumulation is not "
+                "available in this PyTorch build."
+            )
+
+    @staticmethod
+    def _set_fp16_accumulation(enabled):
+        logging.info(
+            "Patching torch settings: "
+            "torch.backends.cuda.matmul.allow_fp16_accumulation = %s",
+            enabled,
+        )
+        torch.backends.cuda.matmul.allow_fp16_accumulation = enabled
+
+    def patch(self, model, enable_fp16_accumulation):
+        from comfy.patcher_extension import CallbacksMP
+
+        self._ensure_supported()
+        if not hasattr(model, "clone") or not hasattr(model, "add_callback"):
+            raise RuntimeError(
+                "Model Patch Torch Settings requires a ComfyUI MODEL patcher "
+                "with clone() and add_callback() support."
+            )
+
+        model_clone = model.clone()
+
+        def enable_callback(model_patcher):
+            self._set_fp16_accumulation(True)
+
+        def disable_callback(model_patcher):
+            self._set_fp16_accumulation(False)
+
+        if enable_fp16_accumulation:
+            model_clone.add_callback(CallbacksMP.ON_PRE_RUN, enable_callback)
+            model_clone.add_callback(CallbacksMP.ON_CLEANUP, disable_callback)
+        else:
+            model_clone.add_callback(CallbacksMP.ON_PRE_RUN, disable_callback)
+
+        return (model_clone,)
+
+
+NODE_CLASS_MAPPINGS["EnviralModelPatchTorchSettings"] = EnviralModelPatchTorchSettings
+NODE_DISPLAY_NAME_MAPPINGS["EnviralModelPatchTorchSettings"] = "Model Patch Torch Settings"