ollamadiffuser 2.0.1__tar.gz → 2.0.3__tar.gz

{ollamadiffuser-2.0.1/ollamadiffuser.egg-info → ollamadiffuser-2.0.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ollamadiffuser
-Version: 2.0.1
+Version: 2.0.3
 Summary: Local AI Image Generation with Ollama-style CLI for Stable Diffusion, FLUX, and LoRA support
 Home-page: https://github.com/ollamadiffuser/ollamadiffuser
 Author: OllamaDiffuser Team

{ollamadiffuser-2.0.1 → ollamadiffuser-2.0.3}/ollamadiffuser/__init__.py

@@ -4,7 +4,7 @@ OllamaDiffuser - Local AI Image Generation with Ollama-style CLI
 A tool for managing and running Stable Diffusion, FLUX.1, and other AI image generation models locally.
 """
 
-__version__ = "2.0.1"
+__version__ = "2.0.3"
 __author__ = "OllamaDiffuser Team"
 __email__ = "ollamadiffuser@gmail.com"
 __description__ = "🎨 Local AI Image Generation with Ollama-style CLI for Stable Diffusion, FLUX.1, and LoRA support"

{ollamadiffuser-2.0.1 → ollamadiffuser-2.0.3}/ollamadiffuser/core/inference/strategies/flux_strategy.py

@@ -38,6 +38,11 @@ class FluxStrategy(InferenceStrategy):
             if device == "cpu":
                 load_kwargs["torch_dtype"] = torch.float32
                 logger.warning("FLUX on CPU will be very slow for this 12B parameter model")
+            elif device == "mps":
+                # MPS has limited bfloat16 support; float16 avoids VAE decode crashes
+                load_kwargs["torch_dtype"] = torch.float16
+                load_kwargs["use_safetensors"] = True
+                load_kwargs["low_cpu_mem_usage"] = True
             else:
                 load_kwargs["torch_dtype"] = torch.bfloat16
                 load_kwargs["use_safetensors"] = True
@@ -46,11 +51,11 @@ class FluxStrategy(InferenceStrategy):
                 model_config.path, **load_kwargs
             )
 
-            if device in ("cuda", "mps") and hasattr(self.pipeline, "enable_model_cpu_offload"):
-                # CPU offloading manages device placement itself — don't call _move_to_device
+            if device == "cuda" and hasattr(self.pipeline, "enable_model_cpu_offload"):
                 self.pipeline.enable_model_cpu_offload(device=device)
                 logger.info(f"Enabled CPU offloading for FLUX on {device}")
             else:
+                # MPS: unified memory means CPU offload adds overhead without saving memory
                 self._move_to_device(device)
             self._apply_memory_optimizations()
 
@@ -102,7 +107,9 @@ class FluxStrategy(InferenceStrategy):
 
         max_seq_len = kwargs.get("max_sequence_length", params.get("max_sequence_length", 512))
 
-        generator, used_seed = self._make_generator(seed, self.device)
+        # CPU offload moves tensors between CPU/device; use CPU generator to avoid device mismatches
+        gen_device = "cpu" if self.device == "mps" else self.device
+        generator, used_seed = self._make_generator(seed, gen_device)
 
         gen_kwargs = {
             "prompt": prompt,

{ollamadiffuser-2.0.1 → ollamadiffuser-2.0.3}/ollamadiffuser/core/inference/strategies/generic_strategy.py

@@ -72,20 +72,8 @@ class GenericPipelineStrategy(InferenceStrategy):
 
             # Device placement
             enable_offload = params.get("enable_cpu_offload", False)
-            # Auto-enable CPU offload on MPS to avoid OOM on unified memory
-            if device == "mps":
-                enable_offload = True
-
-            if enable_offload and device in ("cuda", "mps"):
-                if device == "mps" and hasattr(self.pipeline, "enable_model_cpu_offload"):
-                    # MPS/unified memory: model-level offload is more effective than
-                    # sequential offload because it fully deallocates entire components
-                    # (T5 encoder, transformer, VAE) between stages, reducing peak
-                    # memory pressure on the MPS allocator.
-                    self.pipeline.enable_model_cpu_offload(device=device)
-                    logger.info(f"Enabled model CPU offloading on {device}")
-                elif hasattr(self.pipeline, "enable_sequential_cpu_offload"):
-                    # CUDA: sequential offload moves individual layers, lowest VRAM usage
+            if enable_offload and device == "cuda":
+                if hasattr(self.pipeline, "enable_sequential_cpu_offload"):
                     self.pipeline.enable_sequential_cpu_offload(device=device)
                     logger.info(f"Enabled sequential CPU offloading on {device}")
                 elif hasattr(self.pipeline, "enable_model_cpu_offload"):
@@ -94,6 +82,7 @@ class GenericPipelineStrategy(InferenceStrategy):
                 else:
                     self._move_to_device(device)
             else:
+                # MPS: unified memory means CPU offload adds overhead without saving memory
                 self._move_to_device(device)
 
             self._apply_memory_optimizations()

{ollamadiffuser-2.0.1 → ollamadiffuser-2.0.3}/ollamadiffuser/core/inference/strategies/hidream_strategy.py

@@ -34,6 +34,9 @@ class HiDreamStrategy(InferenceStrategy):
             load_kwargs = {**SAFETY_DISABLED_KWARGS}
             if device == "cpu":
                 load_kwargs["torch_dtype"] = torch.float32
+            elif device == "mps":
+                load_kwargs["torch_dtype"] = torch.float16
+                load_kwargs["low_cpu_mem_usage"] = True
             else:
                 load_kwargs["torch_dtype"] = torch.bfloat16
 
@@ -41,8 +44,7 @@ class HiDreamStrategy(InferenceStrategy):
                 model_config.path, **load_kwargs
             )
 
-            if device in ("cuda", "mps") and hasattr(self.pipeline, "enable_model_cpu_offload"):
-                # CPU offloading manages device placement itself — don't call _move_to_device
+            if device == "cuda" and hasattr(self.pipeline, "enable_model_cpu_offload"):
                 self.pipeline.enable_model_cpu_offload(device=device)
             else:
                 self._move_to_device(device)
@@ -77,7 +79,8 @@ class HiDreamStrategy(InferenceStrategy):
         guidance = guidance_scale if guidance_scale is not None else params.get("guidance_scale", 5.0)
         max_seq_len = kwargs.get("max_sequence_length", params.get("max_sequence_length", 128))
 
-        generator, used_seed = self._make_generator(seed, self.device)
+        gen_device = "cpu" if self.device == "mps" else self.device
+        generator, used_seed = self._make_generator(seed, gen_device)
 
         gen_kwargs = {
             "prompt": prompt,

{ollamadiffuser-2.0.1 → ollamadiffuser-2.0.3}/ollamadiffuser/core/inference/strategies/video_strategy.py

@@ -46,8 +46,7 @@ class VideoStrategy(InferenceStrategy):
                 steps_offset=1,
             )
 
-            if device in ("cuda", "mps") and hasattr(self.pipeline, "enable_model_cpu_offload"):
-                # CPU offloading manages device placement itself — don't call _move_to_device
+            if device == "cuda" and hasattr(self.pipeline, "enable_model_cpu_offload"):
                 self.pipeline.enable_model_cpu_offload(device=device)
             else:
                 self._move_to_device(device)

{ollamadiffuser-2.0.1 → ollamadiffuser-2.0.3/ollamadiffuser.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ollamadiffuser
-Version: 2.0.1
+Version: 2.0.3
 Summary: Local AI Image Generation with Ollama-style CLI for Stable Diffusion, FLUX, and LoRA support
 Home-page: https://github.com/ollamadiffuser/ollamadiffuser
 Author: OllamaDiffuser Team

{ollamadiffuser-2.0.1 → ollamadiffuser-2.0.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "ollamadiffuser"
-version = "2.0.1"
+version = "2.0.3"
 authors = [
     {name = "OllamaDiffuser Team", email = "ollamadiffuser@gmail.com"}
 ]

{ollamadiffuser-2.0.1 → ollamadiffuser-2.0.3}/tests/test_engine.py

@@ -242,9 +242,9 @@ class TestGenericStrategy:
     def test_generic_strategy_cpu_offload_mps(self):
         result, mock_pipe = self._load_generic_with_offload("mps")
         assert result is True
-        # MPS prefers model-level offload (more effective on unified memory)
-        mock_pipe.enable_model_cpu_offload.assert_called_once()
-        mock_pipe.to.assert_not_called()
+        # MPS: unified memory means CPU offload adds overhead, so load directly to device
+        mock_pipe.enable_model_cpu_offload.assert_not_called()
+        mock_pipe.to.assert_called_once()
 
 
 class TestInferenceStrategyBase: