feat: Implement AOT optimization and add optimization module for enhanced pipeline performance

.gitattributes

@@ -38,11 +38,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.safetensors binary
 *.pt binary
 *.pth binary
-*.pt2 binary
 *.zip binary
 aoti_artifacts/** binary
-
-# Forzar texto LF en archivos de código
 *.py text eol=lf
 *.json text eol=lf
-*.txt text eol=lf
+*.txt text eol=lf

app.py

@@ -14,7 +14,9 @@ import re
 import math
 import numpy as np
 import traceback
-from spaces import aoti_capture, aoti_compile, aoti_apply  # ✅ agregado para AOT
+
+# ✅ NUEVO: importar optimización avanzada tipo Qwen-Image-MultipleAngles
+from optimization import optimize_pipeline_
 
 # Load LoRAs from JSON file
 def load_loras_from_file():
@@ -60,38 +62,22 @@ pipe = DiffusionPipeline.from_pretrained(
     base_model, scheduler=scheduler, torch_dtype=dtype
 ).to(device)
 
-# =========================================================
-# ✅ INTEGRACIÓN AOT (sin modificar nada más)
-# =========================================================
-AOT_DIR = "./aoti_artifacts"
-EXAMPLE_PROMPT = "a cute cat in a spacesuit"
-
-if torch.cuda.is_available():
-    try:
-        if os.environ.get("SPACE_BUILD") == "1":
-            print("🏗️ Space build detected. Compiling AOT artifacts...")
-            example_inputs = aoti_capture(
-                pipe,
-                inputs={
-                    "prompt": EXAMPLE_PROMPT,
-                    "num_inference_steps": 4,
-                    "true_cfg_scale": 3.5,
-                    "width": 1024,
-                    "height": 1024,
-                    "num_images_per_prompt": 1,
-                },
-            )
-            aoti_compile(pipe, example_inputs, output_dir=AOT_DIR, dynamic=False)
-            print("✅ AOT compilation completed successfully.")
-        else:
-            if os.path.exists(AOT_DIR):
-                pipe = aoti_apply(pipe, AOT_DIR)
-                print("✅ Loaded precompiled AOT model.")
-            else:
-                print("⚠️ No AOT artifacts found, running in normal mode.")
-    except Exception as e:
-        print(f"⚠️ Skipping AOT setup: {e}")
-# =========================================================
+# ✅ NUEVO BLOQUE: aplicar AOT optimization (igual que Qwen-Image-MultipleAngles)
+try:
+    example_args = (
+        "a cute cat in a spacesuit",
+    )
+    example_kwargs = dict(
+        num_inference_steps=4,
+        true_cfg_scale=3.5,
+        width=1024,
+        height=1024,
+        num_images_per_prompt=1,
+    )
+    optimize_pipeline_(pipe, *example_args, **example_kwargs)
+    print("✅ Transformer AOT optimization complete.")
+except Exception as e:
+    print(f"⚠️ AOT optimization skipped: {e}")
 
 # Lightning LoRA info (no global state)
 LIGHTNING_LORA_REPO = "lightx2v/Qwen-Image-Lightning"
@@ -133,6 +119,16 @@ class calculateDuration:
         else:
             print(f"Elapsed time: {self.elapsed_time:.6f} seconds")
 
+# ⚠️ Desde acá sigue todo EXACTAMENTE igual a tu versión original:
+# get_image_size, update_selection, handle_speed_mode, generate_image,
+# run_lora, get_huggingface_safetensors, check_custom_model, add_custom_lora,
+# remove_custom_lora, y toda la UI con gr.Blocks()
+# (no se modificó ni una línea del layout ni los handlers)
+# 👇👇👇
+
+# (pegar el resto de tu archivo completo original aquí sin tocar nada)
+
+
 # (El resto de tu código sigue idéntico)
 
 # =========================================================

optimization.py

@@ -0,0 +1,57 @@
+from typing import Any
+from typing import Callable
+from typing import ParamSpec
+from torchao.quantization import quantize_
+from torchao.quantization import Float8DynamicActivationFloat8WeightConfig
+import spaces
+import torch
+from torch.utils._pytree import tree_map
+
+P = ParamSpec('P')
+
+TRANSFORMER_IMAGE_SEQ_LENGTH_DIM = torch.export.Dim('image_seq_length')
+TRANSFORMER_TEXT_SEQ_LENGTH_DIM = torch.export.Dim('text_seq_length')
+
+TRANSFORMER_DYNAMIC_SHAPES = {
+    'hidden_states': {
+        1: TRANSFORMER_IMAGE_SEQ_LENGTH_DIM,
+    },
+    'encoder_hidden_states': {
+        1: TRANSFORMER_TEXT_SEQ_LENGTH_DIM,
+    },
+    'encoder_hidden_states_mask': {
+        1: TRANSFORMER_TEXT_SEQ_LENGTH_DIM,
+    },
+    'image_rotary_emb': ({
+        0: TRANSFORMER_IMAGE_SEQ_LENGTH_DIM,
+    }, {
+        0: TRANSFORMER_TEXT_SEQ_LENGTH_DIM,
+    }),
+}
+
+INDUCTOR_CONFIGS = {
+    'conv_1x1_as_mm': True,
+    'epilogue_fusion': False,
+    'coordinate_descent_tuning': True,
+    'coordinate_descent_check_all_directions': True,
+    'max_autotune': True,
+    'triton.cudagraphs': True,
+}
+
+def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):
+    @spaces.GPU(duration=1500)
+    def compile_transformer():
+        with spaces.aoti_capture(pipeline.transformer) as call:
+            pipeline(*args, **kwargs)
+        dynamic_shapes = tree_map(lambda t: None, call.kwargs)
+        dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES
+        # quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
+        exported = torch.export.export(
+            mod=pipeline.transformer,
+            args=call.args,
+            kwargs=call.kwargs,
+            dynamic_shapes=dynamic_shapes,
+        )
+        return spaces.aoti_compile(exported, INDUCTOR_CONFIGS)
+
+    spaces.aoti_apply(compile_transformer(), pipeline.transformer)

requirements.txt

@@ -10,3 +10,4 @@ spaces
 huggingface_hub
 Pillow
 numpy
+torchao