horde-engine 3.9.2__tar.gz → 3.9.3__tar.gz

{horde_engine-3.9.2 → horde_engine-3.9.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: horde-engine
-Version: 3.9.2
+Version: 3.9.3
 Summary: A wrapper around ComfyUI to allow use by AI Horde.
 Project-URL: Homepage, https://github.com/Haidra-Org/hordelib
 Project-URL: Bug Tracker, https://github.com/Haidra-Org/hordelib/issues

{horde_engine-3.9.2 → horde_engine-3.9.3}/hordelib/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '3.9.2'
-__version_tuple__ = version_tuple = (3, 9, 2)
+__version__ = version = '3.9.3'
+__version_tuple__ = version_tuple = (3, 9, 3)
 
 __commit_id__ = commit_id = None

{horde_engine-3.9.2 → horde_engine-3.9.3}/hordelib/comfy_horde.py

@@ -1015,6 +1015,8 @@ class Comfy_Horde:
         pipeline: GraphDict,
         params: dict[str, typing.Any],
         comfyui_progress_callback: typing.Callable[[ComfyUIProgress, str], None] | None = None,
+        *,
+        defer_vram_unload: bool = False,
     ) -> list[dict[str, typing.Any]] | None:
         start_time = time.time()
         _t0 = time.perf_counter()
@@ -1134,7 +1136,12 @@ class Comfy_Horde:
                 _execute_seconds = _t_post_execute - _t_pre_execute
                 if use_native_progress:
                     set_run_progress_callback(None)
-                if self.aggressive_unloading:
+                # ``aggressive_unloading`` evicts the just-used model from VRAM after every job so N
+                # sibling ComfyUI instances sharing one GPU never collectively over-commit. ``defer_vram_unload``
+                # lets the host keep the model resident across this job when it knows the same model runs next
+                # and the VRAM budget allows it, so the back-to-back force-reload (the dominant non-sampling cost
+                # on small jobs) is skipped. The host owns the safety decision; here we only honor it.
+                if self.aggressive_unloading and not defer_vram_unload:
                     global _comfy_cleanup_models
                     logger.debug("Cleaning up models")
                     with logfire.span("comfy.cleanup"):
@@ -1191,6 +1198,8 @@ class Comfy_Horde:
         pipeline: GraphDict,
         params: dict[str, typing.Any],
         comfyui_progress_callback: typing.Callable[[ComfyUIProgress, str], None] | None = None,
+        *,
+        defer_vram_unload: bool = False,
     ) -> list[dict[str, typing.Any]]:
         # From the horde point of view, let us assume the output we are interested in
         # is always in a HordeImageOutput node named "output_image". This is an array of
@@ -1219,7 +1228,7 @@ class Comfy_Horde:
             if idle_time > 1 and UserSettings.enable_idle_time_warning.active:
                 logger.warning("No job ran recently", idle_seconds=round(idle_time, 3))
 
-        result = self._run_pipeline(pipeline, params, comfyui_progress_callback)
+        result = self._run_pipeline(pipeline, params, comfyui_progress_callback, defer_vram_unload=defer_vram_unload)
 
         if result:
             return result

{horde_engine-3.9.2 → horde_engine-3.9.3}/hordelib/execution/in_process.py

@@ -87,11 +87,17 @@ class InProcessComfyBackend:
         graph: dict[str, Any],
         *,
         progress_callback: ProgressCallback | None = None,
+        defer_vram_unload: bool = False,
     ) -> list[OutputArtifact]:
         self._ensure_started()
         assert self._comfy is not None
 
-        results = self._comfy.run_image_pipeline(graph, {}, progress_callback)
+        results = self._comfy.run_image_pipeline(
+            graph,
+            {},
+            progress_callback,
+            defer_vram_unload=defer_vram_unload,
+        )
         return self._to_artifacts(results)
 
     @staticmethod

{horde_engine-3.9.2 → horde_engine-3.9.3}/hordelib/execution/interface.py

@@ -59,12 +59,17 @@ class ExecutionBackend(Protocol):
         graph: dict[str, Any],
         *,
         progress_callback: ProgressCallback | None = None,
+        defer_vram_unload: bool = False,
     ) -> list[OutputArtifact]:
         """Execute a fully materialized API-format graph and return its outputs.
 
         Args:
             graph: The pipeline graph in ComfyUI API format, with all parameters already set.
             progress_callback: Optionally called with progress updates during execution.
+            defer_vram_unload: When True, keep the model resident in VRAM after this run instead of
+                evicting it, so a following job that reuses it skips the RAM->VRAM reload. The caller
+                owns the VRAM-safety decision (it must know the model fits alongside the live set);
+                backends that never evict between runs ignore this. Defaults to False.
 
         Returns:
             list[OutputArtifact]: The outputs produced by the run.

{horde_engine-3.9.2 → horde_engine-3.9.3}/hordelib/horde.py

@@ -201,6 +201,7 @@ class HordeLib:
         *,
         single_image_expected: bool = True,
         comfyui_progress_callback: Callable[[ComfyUIProgress, str], None] | None = None,
+        defer_vram_unload: bool = False,
     ) -> list[ResultingImageReturn] | ResultingImageReturn:
         start_time = time.time()
 
@@ -236,6 +237,7 @@ class HordeLib:
         artifacts = self.backend.run_pipeline(
             graph.to_api_dict(),
             progress_callback=comfyui_progress_callback,
+            defer_vram_unload=defer_vram_unload,
         )
 
         ret_results = []
@@ -266,6 +268,7 @@ class HordeLib:
         payload: dict | ImageGenerateJobPopResponse,
         *,
         progress_callback: Callable[[ProgressReport], None] | None = None,
+        defer_vram_unload: bool = False,
     ) -> list[ResultingImageReturn]:
         post_processing_requested: list[str] | None = None
         if isinstance(payload, dict):
@@ -447,6 +450,7 @@ class HordeLib:
             payload,
             single_image_expected=False,
             comfyui_progress_callback=_default_progress_callback,
+            defer_vram_unload=defer_vram_unload,
         )
 
         if not isinstance(result, list):