Fix AoT (#1)

- Fix AoT (35ee5b0267f6846d0c86392c55dd3bf67deb512c)
- Update app.py (152019b5b68dd7404d072590a79c42c29bd14814)


Co-authored-by: Charles Bensimon <[email protected]>

app.py

@@ -2,7 +2,6 @@ import os
 import subprocess
 import sys
 import io
-from kernels import get_kernel
 import gradio as gr
 import numpy as np
 import random
@@ -38,14 +37,11 @@ def remote_text_encoder(prompts):
     return prompt_embeds
 
 # Load model
-fa3_kernel = get_kernel("kernels-community/flash-attn3", revision="fake-ops-return-probs")
-
 repo_id = "black-forest-labs/FLUX.2-dev"
 
 dit = Flux2Transformer2DModel.from_pretrained(
     repo_id,
     subfolder="transformer",
-    attn_implementation=fa3_kernel,
     torch_dtype=torch.bfloat16
 )
 
@@ -56,16 +52,18 @@ pipe = Flux2Pipeline.from_pretrained(
     torch_dtype=torch.bfloat16
 )
 pipe.to("cuda")
-# pipe.transformer.compile_repeated_blocks(dynamic=True)
-
-#optimize_pipeline_(
-#    pipe,
-#    prompt_embeds=remote_text_encoder("prompt").to("cuda"),
-#    guidance_scale=2.5,
-#    width=1024,
-#    height=1024,
-#    num_inference_steps=1,
-#)
+
+pipe.transformer.set_attention_backend("_flash_3_hub")
+
+optimize_pipeline_(
+    pipe,
+    image=[Image.new("RGB", (1024, 1024))],
+    prompt_embeds = remote_text_encoder("prompt").to("cuda"),
+    guidance_scale=2.5,
+    width=1024,
+    height=1024,
+    num_inference_steps=1
+)
 
 @spaces.GPU(duration=180)
 def infer(prompt, input_images, seed=42, randomize_seed=False, width=1024, height=1024, num_inference_steps=50, guidance_scale=2.5, progress=gr.Progress(track_tqdm=True)):

optimization.py

@@ -6,25 +6,32 @@ from typing import Callable
 from typing import ParamSpec
 import spaces
 import torch
+from spaces.zero.torch.aoti import ZeroGPUCompiledModel
+from spaces.zero.torch.aoti import ZeroGPUWeights
 from torch.utils._pytree import tree_map
 
 P = ParamSpec('P')
 
-TRANSFORMER_IMAGE_SEQ_LENGTH_DIM = torch.export.Dim('image_seq_length', min=64, max=16384)
-TRANSFORMER_TEXT_SEQ_LENGTH_DIM = torch.export.Dim('text_seq_length', min=64, max=512)
+TRANSFORMER_IMAGE_DIM = torch.export.Dim('image_seq_length', min=4096, max=16384) # min: 0 images, max: 3 (1024x1024) images
 
 TRANSFORMER_DYNAMIC_SHAPES = {
-    'hidden_states': {
-        1: TRANSFORMER_IMAGE_SEQ_LENGTH_DIM,
+    'double': {
+        'hidden_states': {
+            1: TRANSFORMER_IMAGE_DIM,
+        },
+        'image_rotary_emb': (
+            {0: TRANSFORMER_IMAGE_DIM + 512},
+            {0: TRANSFORMER_IMAGE_DIM + 512},
+        ),
     },
-    'encoder_hidden_states': {
-        1: TRANSFORMER_TEXT_SEQ_LENGTH_DIM,
-    },
-    'img_ids': {
-        1: TRANSFORMER_IMAGE_SEQ_LENGTH_DIM,
-    },
-    'txt_ids': {
-        1: TRANSFORMER_TEXT_SEQ_LENGTH_DIM, 
+    'single': {
+        'hidden_states': {
+            1: TRANSFORMER_IMAGE_DIM + 512,
+        },
+        'image_rotary_emb': (
+            {0: TRANSFORMER_IMAGE_DIM + 512},
+            {0: TRANSFORMER_IMAGE_DIM + 512},
+        ),
     },
 }
 
@@ -33,28 +40,38 @@ INDUCTOR_CONFIGS = {
     'epilogue_fusion': False,
     'coordinate_descent_tuning': True,
     'coordinate_descent_check_all_directions': True,
-    #'max_autotune': True,
-    #'triton.cudagraphs': 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():
+    blocks = {
+        'double': pipeline.transformer.transformer_blocks,
+        'single': pipeline.transformer.single_transformer_blocks,
+    }
 
-        with spaces.aoti_capture(pipeline.transformer) as call:
+    @spaces.GPU(duration=1200)
+    def compile_block(blocks_kind: str):
+        block = blocks[blocks_kind][0]
+        with spaces.aoti_capture(block) as call:
             pipeline(*args, **kwargs)
 
         dynamic_shapes = tree_map(lambda t: None, call.kwargs)
-        dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES
-        
-        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)
+        dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES[blocks_kind]
+
+        with torch.no_grad():      
+            exported = torch.export.export(
+                mod=block,
+                args=call.args,
+                kwargs=call.kwargs,
+                dynamic_shapes=dynamic_shapes,
+            )
+
+        return spaces.aoti_compile(exported, INDUCTOR_CONFIGS).archive_file
+
+    for blocks_kind in ('double', 'single'):
+        archive_file = compile_block(blocks_kind)
+        for block in blocks[blocks_kind]:
+            weights = ZeroGPUWeights(block.state_dict())
+            block.forward = ZeroGPUCompiledModel(archive_file, weights)