ollamadiffuser 2.0.1__py3-none-any.whl → 2.0.3__py3-none-any.whl

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/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/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/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/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.dist-info → ollamadiffuser-2.0.3.dist-info}/METADATA

@@ -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.dist-info → ollamadiffuser-2.0.3.dist-info}/RECORD

@@ -1,4 +1,4 @@
-ollamadiffuser/__init__.py,sha256=WNSenduGIDUa8FBGS-GSR0yS3430KCofd102wjN7K7E,1127
+ollamadiffuser/__init__.py,sha256=DrDVJ1zOBxtKD04fUKIidYLBgLORzagrUcTPw4zDqZM,1127
 ollamadiffuser/__main__.py,sha256=tNWMvEHq4ddtKLp7DrhIoOdnFw3F8RNrETC_u5xpkFI,141
 ollamadiffuser/api/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 ollamadiffuser/api/server.py,sha256=MqEKjckz8x8pU9dhZlLXET5XWt7ERfVcAfvdVfDLFWw,15464
@@ -19,14 +19,14 @@ ollamadiffuser/core/inference/base.py,sha256=AyoM6j37nhhh5RXQeH9Ycn9x1_eRPQQfC5n
 ollamadiffuser/core/inference/engine.py,sha256=xUCSQmGke9yAIoKGzh9jRaH1XE3GP9-uGM2hseuUzao,8139
 ollamadiffuser/core/inference/strategies/__init__.py,sha256=5LQgTeS5JVin-HiGX7UvjlPzd0awombKrNhvAVr3SSw,53
 ollamadiffuser/core/inference/strategies/controlnet_strategy.py,sha256=_lGlCaYhrrdn7N6Aw0X9a4L90wKPjCrr6EBfQqPVH2E,6712
-ollamadiffuser/core/inference/strategies/flux_strategy.py,sha256=NPlwoKC9TsoVOkCLP0Gzf33D2cWg-77McO6_hYWWbsU,5216
-ollamadiffuser/core/inference/strategies/generic_strategy.py,sha256=fGD4nnUspSlOtK7XWTjDJRRN5uDfQIQIuefBJSJhCEA,6843
+ollamadiffuser/core/inference/strategies/flux_strategy.py,sha256=E5OIWlylL886V2lOxGSxHS-LULlBMj-kjWRpdFnFxdQ,5660
+ollamadiffuser/core/inference/strategies/generic_strategy.py,sha256=IyCuIDi-MyscUD4LVvbrF6ZcpJRsLbDfQ6juHFU7JHU,6129
 ollamadiffuser/core/inference/strategies/gguf_strategy.py,sha256=kIGT85tDCcSsliXdaxEJoQz4Gm7Xt7TfEcu6xcmTvJg,3893
-ollamadiffuser/core/inference/strategies/hidream_strategy.py,sha256=D1BeqEXiMRzJER5SEPAGJAGm9B_lnczMM94wu6sVrHE,3707
+ollamadiffuser/core/inference/strategies/hidream_strategy.py,sha256=qVCI1Z5S0_p7A5mlHvs2ueZeGTiR6FpMN-K2xM33brE,3818
 ollamadiffuser/core/inference/strategies/sd15_strategy.py,sha256=qz5eGA2xkcA_3oNywP-rCliXzP7jYpH60728QmOT5fw,4966
 ollamadiffuser/core/inference/strategies/sd3_strategy.py,sha256=6DjWebeyjaH7jiRm8hf2ismkJ3Gth69u71enVgMMPi8,2772
 ollamadiffuser/core/inference/strategies/sdxl_strategy.py,sha256=tslfENJIvEhDuj1D6aClFF6hv8i0JO2PukFQZsTCwQY,5137
-ollamadiffuser/core/inference/strategies/video_strategy.py,sha256=xJJU5GbHol8SMpNBllga8AkjQRTgZ0sZUkAwFIoJqk8,3939
+ollamadiffuser/core/inference/strategies/video_strategy.py,sha256=WnzTkWY_b70kZQa0j4w6iSEnv0gwkqKG8IaAYrP3bRI,3834
 ollamadiffuser/core/models/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 ollamadiffuser/core/models/gguf_loader.py,sha256=ocfl3_MDVXC9nSjW8YJdz4kX1Q-Qe2ltu6w4fbqhxVY,35724
 ollamadiffuser/core/models/manager.py,sha256=rTEAameGih3wPcVG_Y-4k_brBeEqEoBjoI7fjggNtiY,16799
@@ -53,9 +53,9 @@ ollamadiffuser/ui/samples/scribble/face_sketch.png,sha256=MVVYy_aS48xoS_RnIDzLUa
 ollamadiffuser/ui/samples/scribble/tree_sketch.png,sha256=3P-NGgW25xRwreDxiBYKcDhd2oHZAwKSkjNVM5oPTWY,3017
 ollamadiffuser/ui/templates/index.html,sha256=XcrYZqtDR65dAiu959Ea19t3MbtYmXl9PVyMnR1Telk,42358
 ollamadiffuser/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-ollamadiffuser-2.0.1.dist-info/licenses/LICENSE,sha256=cnGL9l2P510Uk3TCnv62kot6vAfdSawhOZh7Y-oYoIE,1071
-ollamadiffuser-2.0.1.dist-info/METADATA,sha256=9MSH7COrR7xKmnSCL_1hK9OPmiJgZcS5LdeEWrYZlNo,31097
-ollamadiffuser-2.0.1.dist-info/WHEEL,sha256=wUyA8OaulRlbfwMtmQsvNngGrxQHAvkKcvRmdizlJi0,92
-ollamadiffuser-2.0.1.dist-info/entry_points.txt,sha256=Bp-ZzV3F7QpQu02Mcafeza-oTMjDslomz9qrhvfcQUA,116
-ollamadiffuser-2.0.1.dist-info/top_level.txt,sha256=97wOGgTCxDE765Nr_o7B4Kwr_M_jy8fCCeQ81sMKlC4,15
-ollamadiffuser-2.0.1.dist-info/RECORD,,
+ollamadiffuser-2.0.3.dist-info/licenses/LICENSE,sha256=cnGL9l2P510Uk3TCnv62kot6vAfdSawhOZh7Y-oYoIE,1071
+ollamadiffuser-2.0.3.dist-info/METADATA,sha256=64JcTKGPRp343mivzk_Y7Je06KZ7N0KB-96-9UcWQZo,31097
+ollamadiffuser-2.0.3.dist-info/WHEEL,sha256=wUyA8OaulRlbfwMtmQsvNngGrxQHAvkKcvRmdizlJi0,92
+ollamadiffuser-2.0.3.dist-info/entry_points.txt,sha256=Bp-ZzV3F7QpQu02Mcafeza-oTMjDslomz9qrhvfcQUA,116
+ollamadiffuser-2.0.3.dist-info/top_level.txt,sha256=97wOGgTCxDE765Nr_o7B4Kwr_M_jy8fCCeQ81sMKlC4,15
+ollamadiffuser-2.0.3.dist-info/RECORD,,