introduce torch_xla2.compile API, make sdxl to use it (#8269)

Author: qihqi

experimental/reference_models/sdxl_inference/README.md

@@ -0,0 +1,5 @@
+# How to run:
+
+```
+python sdxl.py
+```

experimental/reference_models/sdxl_inference/astronaut_rides_horse.png

@@ -0,0 +1,73 @@
+import time
+import functools
+import jax
+import torch
+import torch_xla2
+from torch_xla2 import interop
+from torch_xla2.interop import JittableModule
+
+from transformers.modeling_outputs import BaseModelOutputWithPooling
+
+from jax.tree_util import register_pytree_node
+import jax
+
+def base_model_output_with_pooling_flatten(v):
+  return (v.last_hidden_state, v.pooler_output, v.hidden_states, v.attentions), None
+
+def base_model_output_with_pooling_unflatten(aux_data, children):
+  return BaseModelOutputWithPooling(*children)
+
+register_pytree_node(
+  BaseModelOutputWithPooling,
+  base_model_output_with_pooling_flatten,
+  base_model_output_with_pooling_unflatten
+)
+
+
+from diffusers import StableDiffusionPipeline
+pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-base")
+
+prompt = "a photograph of an astronaut riding a horse"
+# image = pipe(prompt).images[0]
+
+
+env = torch_xla2.default_env()
+jax.config.update('jax_enable_x64', False)
+
+def move_scheduler(scheduler):
+  for k, v in scheduler.__dict__.items():
+    if isinstance(v, torch.Tensor):
+      setattr(scheduler, k, v.to('jax'))
+
+
+with env:
+  pipe.to('jax:1')
+  move_scheduler(pipe.scheduler)
+  pipe.unet = torch_xla2.compile(
+    pipe.unet, torch_xla2.CompileOptions(
+      jax_jit_kwargs={'static_argnames': ('return_dict',)}
+    )
+  )
+  import pdb; pdb.set_trace()
+  pipe.text_encoder = torch_xla2.compile(pipe.text_encoder)
+
+  BS = 4
+  prompt = [prompt] * BS 
+  pipe.vae = torch_xla2.compile(
+    pipe.vae, torch_xla2.CompileOptions(
+      jax_jit_kwargs={'static_argnames': ('return_dict',)},
+      methods_to_compile=['decode'],
+    )
+  )
+  image = pipe(prompt).images[0]
+
+  jax.profiler.start_trace('/tmp/sdxl')
+  start = time.perf_counter()
+  image = pipe(prompt, num_inference_steps=20).images[0]
+  end = time.perf_counter()
+  jax.profiler.stop_trace()
+  print('Total time is ', end - start, 'bs = ', BS)
+  image.save(f"astronaut_rides_horse.png")
+
+
+

experimental/reference_models/sdxl_inference/sdxl.py

@@ -0,0 +1,14 @@
+import torch
+from diffusers import StableDiffusionPipeline
+
+import torch_xla2
+env = torch_xla2.default_env()
+
+# this is now contains torhc.Tensor
+pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-base")
+
+with env:
+  pipe.to('jax')
+  prompt = "a photograph of an astronaut riding a horse"
+  image = pipe(prompt, num_inference_steps=10).images[0]
+  image.save(f"astronaut_rides_horse_orig.png")

experimental/reference_models/sdxl_inference/sdxl_beginning.py

@@ -1,3 +1,5 @@
+from typing import List, Dict, Any, Optional
+import dataclasses
 import jax
 import os
 import torch
@@ -91,4 +93,30 @@ def enable_accuracy_mode():
 def enable_performance_mode():
   jax.config.update('jax_enable_x64', False)
   jax.config.update('jax_default_matmul_precision', 'default')
-  default_env().config.internal_respect_torch_return_dtypes = False
+  default_env().config.internal_respect_torch_return_dtypes = False
+
+
+
+@dataclasses.dataclass
+class CompileOptions:
+  # only valid if compiling nn.Module
+  methods_to_compile: List[str] = dataclasses.field(default_factory=lambda: ['forward'])  
+  jax_jit_kwargs: Dict[str, Any] = dataclasses.field(default_factory=dict)
+  mode: str = 'jax' # or dynamo or export
+
+
+def compile(fn, options: Optional[CompileOptions] = None):
+  options = options or CompileOptions()
+  if options.mode == 'jax':
+    from torch_xla2 import interop
+    if isinstance(fn, torch.nn.Module):
+      module = interop.JittableModule(fn, extra_jit_args=options.jax_jit_kwargs)
+      for n in options.methods_to_compile:
+        module.make_jitted(n)
+      return module
+    else:
+      return interop.jax_jit(fn)
+  elif options.mode == 'dynamo':
+    raise RuntimeError('dynamo mode is not supported yet')
+  elif options.mode == 'export':
+    raise RuntimeError('export mode is not supported yet')

experimental/torch_xla2/torch_xla2/__init__.py

@@ -49,8 +49,6 @@ def set_one(module, prefix):
 
 class JittableModule(torch.nn.Module):
 
-    # TODO: add statedict loading hook
-
     def __init__(self, m: torch.nn.Module, extra_jit_args={}):
         super().__init__()
         self.params, self.buffers = extract_all_buffers(m)