Upload src/pipeline.py with huggingface_hub

src/pipeline.py

@@ -1,4 +1,4 @@
-import gc
+# onediff.compile
 import os
 from typing import TypeAlias
 
@@ -10,6 +10,7 @@ from pipelines.models import TextToImageRequest
 from torch import Generator
 from torchao.quantization import quantize_, int8_weight_only
 from transformers import T5EncoderModel, CLIPTextModel
+from functools import partial
 
 
 Pipeline: TypeAlias = FluxPipeline
@@ -28,37 +29,42 @@ TinyVAE = "madebyollin/taef1"
 TinyVAE_REV = "2d552378e58c9c94201075708d7de4e1163b2689"
 
 
-def load_pipeline() -> Pipeline:
-    path = os.path.join(HF_HUB_CACHE, "models--jokerbit--flux.1-schnell-Robert-int8wo/snapshots/5ef0012f11a863e5111ec56540302a023bc8587b/transformer")
-    transformer = FluxTransformer2DModel.from_pretrained(
-            path, 
-            use_safetensors=False,
-            local_files_only=True,
-            torch_dtype=torch.bfloat16)
-    vae = AutoencoderTiny.from_pretrained(
-            TinyVAE,
-            revision=TinyVAE_REV,
-            local_files_only=True,
-            torch_dtype=torch.bfloat16
-            )    
-    pipeline = FluxPipeline.from_pretrained(
-        CHECKPOINT,
-        revision=REVISION,
-        transformer=transformer,
-        vae=vae,
+my_quantize = partial(quantize_, apply_tensor_subclass=int8_weight_only())
+
+
+path = os.path.join(HF_HUB_CACHE, "models--jokerbit--flux.1-schnell-Robert-int8wo/snapshots/5ef0012f11a863e5111ec56540302a023bc8587b/transformer")
+transformer = FluxTransformer2DModel.from_pretrained(
+        path, 
+        use_safetensors=False,
         local_files_only=True,
-        torch_dtype=torch.bfloat16,
-    )
+        torch_dtype=torch.bfloat16)
+vae = AutoencoderTiny.from_pretrained(
+        TinyVAE,
+        revision=TinyVAE_REV,
+        local_files_only=True,
+        torch_dtype=torch.bfloat16
+        )    
+pipeline = FluxPipeline.from_pretrained(
+    CHECKPOINT,
+    revision=REVISION,
+    transformer=transformer,
+    vae=vae,
+    local_files_only=True,
+    torch_dtype=torch.bfloat16,
+)
+
+pipeline.transformer.to(memory_format=torch.channels_last)
+pipeline.set_progress_bar_config(disable=True)
+pipeline.to("cuda")
+my_quantize(pipeline.vae)
 
-    pipeline.transformer.to(memory_format=torch.channels_last)
-    pipeline.set_progress_bar_config(disable=True)
-    pipeline.to("cuda")
-    quantize_(pipeline.vae, int8_weight_only())
+
+def load_pipeline():
     for _ in range(4):
         pipeline("cat", num_inference_steps=4)
-
     return pipeline
 
+
 @torch.inference_mode()
 def infer(request: TextToImageRequest, pipeline: Pipeline, generator: torch.Generator) -> Image: