Extend VaeDecoderModel for flux compatibility (#750)

Co-authored-by: Kyle Herndon <kyle.herndon@amd.com>
Co-authored-by: Kyle Herndon <kherndon@amd.com>

sharktank/requirements.txt

@@ -5,6 +5,11 @@ gguf>=0.11.0
 numpy
 
 # Model deps.
-huggingface-hub==0.22.2
+huggingface-hub
 transformers==4.40.0
 datasets
+einops
+
+# Serving deps.
+fastapi>=0.112.2
+uvicorn>=0.30.6

sharktank/sharktank/models/vae/config.py

@@ -38,6 +38,8 @@ class HParams:
     layers_per_block: int = 2
     norm_num_groups: int = 32
     scaling_factor: float = 0.13025
+    use_post_quant_conv: bool = True
+    shift_factor: float = 0.0
 
     def assert_default_values(self, attr_names: Sequence[str]):
         for name in attr_names:

sharktank/sharktank/models/vae/layers.py

@@ -17,7 +17,6 @@
     ResnetBlock2D,
     Upsample2D,
     GroupNormLayer,
-    AttentionLayer,
 )
 from .config import *
 
@@ -84,7 +83,6 @@ def forward(
             hidden_states = self.group_norm(hidden_states.transpose(1, 2)).transpose(
                 1, 2
             )
-
         query = self.to_q(hidden_states)
 
         if encoder_hidden_states is None:
@@ -110,6 +108,7 @@ def forward(
         hidden_states = hidden_states.transpose(1, 2).reshape(
             batch_size, -1, self.heads * head_dim
         )
+        hidden_states = hidden_states.to(query.dtype)
 
         # linear proj
         hidden_states = self.to_out(hidden_states)

sharktank/sharktank/models/vae/model.py

@@ -15,6 +15,8 @@
 from .layers import *
 from sharktank.models.punet.layers import UpDownBlock2D, GroupNormLayer
 from typing import Optional
+from einops import rearrange
+import math
 
 
 class VaeDecoderModel(ThetaLayer):
@@ -23,12 +25,13 @@ def from_dataset(cls, ds: Dataset) -> "VaeDecoderModel":
         hp = HParams.from_dict(ds.properties["hparams"])
         return cls(hp, ds.root_theta)
 
-    def __init__(self, hp: HParams, theta: Theta):
+    def __init__(self, hp, theta: Theta):
         super().__init__(theta)
         self.hp = hp
 
         # input conv
-        self.post_quant_conv = Conv2DLayer(theta("post_quant_conv"), padding=(0, 0))
+        if hp.use_post_quant_conv:
+            self.post_quant_conv = Conv2DLayer(theta("post_quant_conv"), padding=(0, 0))
         self.conv_in = Conv2DLayer(theta("decoder")("conv_in"), padding=(1, 1))
         # Mid
         self.mid_block = self._create_mid_block(theta("decoder")("mid_block"))
@@ -71,9 +74,20 @@ def forward(
                 "latent_embeds": latent_embeds,
             },
         )
-        sample = 1 / self.hp.scaling_factor * sample
+        if not self.hp.use_post_quant_conv:
+            sample = rearrange(
+                sample,
+                "b (h w) (c ph pw) -> b c (h ph) (w pw)",
+                h=math.ceil(1024 / 16),
+                w=math.ceil(1024 / 16),
+                ph=2,
+                pw=2,
+            )
+        sample = sample / self.hp.scaling_factor + self.hp.shift_factor
+
+        if self.hp.use_post_quant_conv:
+            sample = self.post_quant_conv(sample)
 
-        sample = self.post_quant_conv(sample)
         sample = self.conv_in(sample)
         self.trace_golden("conv_in", sample)
         # TODO add training and gradient checkpointing support
@@ -90,7 +104,11 @@ def forward(
 
         sample = self.conv_act(sample)
         sample = self.conv_out(sample)
-        sample = (sample / 2 + 0.5).clamp(0, 1)
+
+        if not self.hp.use_post_quant_conv:
+            sample = sample.clamp(-1, 1)
+        else:
+            sample = (sample / 2 + 0.5).clamp(0, 1)
         return sample
 
     def _create_mid_block(self, mid_block_theta: Theta) -> nn.Module:

sharktank/sharktank/models/vae/tools/diffuser_ref.py

@@ -7,6 +7,8 @@
 import argparse
 import torch
 from diffusers import AutoencoderKL
+from einops import rearrange
+import math
 
 
 class VaeModel(torch.nn.Module):
@@ -51,6 +53,34 @@ def decode(self, inp):
 
 
 def run_torch_vae(hf_model_name, example_input):
-
     vae_model = VaeModel(hf_model_name)
     return vae_model.decode(example_input)
+
+
+# TODO Remove and integrate with VaeModel
+class FluxAEWrapper(torch.nn.Module):
+    def __init__(self, height=1024, width=1024):
+        super().__init__()
+        self.ae = AutoencoderKL.from_pretrained(
+            "black-forest-labs/FLUX.1-dev", subfolder="vae", torch_dtype=torch.bfloat16
+        )
+        self.height = height
+        self.width = width
+
+    def forward(self, z):
+        d_in = rearrange(
+            z,
+            "b (h w) (c ph pw) -> b c (h ph) (w pw)",
+            h=math.ceil(self.height / 16),
+            w=math.ceil(self.width / 16),
+            ph=2,
+            pw=2,
+        )
+        d_in = d_in / self.ae.config.scaling_factor + self.ae.config.shift_factor
+        return self.ae.decode(d_in, return_dict=False)[0].clamp(-1, 1)
+
+
+def run_flux_vae(example_input, dtype):
+    # TODO add support for other height/width sizes
+    vae_model = FluxAEWrapper(1024, 1024).to(dtype)
+    return vae_model.forward(example_input)

sharktank/sharktank/models/vae/tools/run_vae.py

@@ -21,6 +21,7 @@
 from iree.turbine.dynamo.passes import (
     DEFAULT_DECOMPOSITIONS,
 )
+import numpy as np
 
 
 def export_vae(model, sample_inputs, decomp_attn):
@@ -81,6 +82,19 @@ def main(argv):
         action="store_true",
         help="Compares results vs HF diffusers reference model",
     )
+
+    parser.add_argument(
+        "--torch_model",
+        default="stabilityai/stable-diffusion-xl-base-1.0",
+        help="HF reference model id",
+    )
+
+    parser.add_argument(
+        "--sharktank_config",
+        default="sdxl",
+        help="Sharktank config providing hyperparamters [sdxl or flux]",
+    )
+
     parser.add_argument(
         "--decomp_attn",
         action="store_true",
@@ -95,12 +109,13 @@ def main(argv):
     ds.to(device=device)
 
     mdl = VaeDecoderModel.from_dataset(ds)
-
     # Run a step for debugging.
     if args.inputs:
         inputs = load_inputs(args.inputs, dtype=dtype, device=device, bs=args.bs)
     else:
-        inputs = get_random_inputs(dtype=dtype, device=device, bs=args.bs)
+        inputs = get_random_inputs(
+            dtype=dtype, device=device, bs=args.bs, config=args.sharktank_config
+        )
 
     if args.export:
         # TODO move export from a run_vae file
@@ -126,11 +141,12 @@ def main(argv):
             intermediates_saver.save_file(args.save_intermediates_path)
 
         if args.compare_vs_torch:
-            from .diffuser_ref import run_torch_vae
+            from .diffuser_ref import run_torch_vae, run_flux_vae
 
-            diffusers_results = run_torch_vae(
-                "stabilityai/stable-diffusion-xl-base-1.0", inputs
-            )
+            if args.sharktank_config == "flux":
+                diffusers_results = run_flux_vae(inputs, torch.bfloat16)
+            elif args.sharktank_config == "sdxl":
+                run_torch_vae(args.torch_model, inputs)
             print("diffusers results:", diffusers_results)
             torch.testing.assert_close(diffusers_results, results)
 

sharktank/sharktank/models/vae/tools/sample_data.py

@@ -11,7 +11,15 @@
 import torch
 
 
-def get_random_inputs(dtype, device, bs: int = 2):
+def get_random_inputs(dtype, device, bs: int = 2, config: str = "sdxl"):
     height = 1024
     width = 1024
-    return torch.rand(bs, 4, width // 8, height // 8, dtype=dtype).to(device)
+    if config == "sdxl":
+        print("sdxl returning inputs")
+        return torch.rand(bs, 4, width // 8, height // 8, dtype=dtype).to(device)
+    elif config == "flux":
+        print("flux returning inputs")
+        return torch.rand(bs, int(width * height / 256), 64, dtype=dtype).to(device)
+    else:
+        print("config: ", config)
+        raise AssertionError(f"{config} config not implmented [sdxl, flux] implemented")

sharktank/sharktank/tools/import_hf_dataset.py

@@ -38,6 +38,7 @@ def import_hf_dataset(
     config_json_path: PathLike,
     param_paths: list[PathLike],
     output_irpa_file: Optional[PathLike] = None,
+    target_dtype=None,
 ) -> Optional[Dataset]:
     import safetensors
 
@@ -50,7 +51,9 @@ def import_hf_dataset(
     for params_path in param_paths:
         with safetensors.safe_open(params_path, framework="pt", device="cpu") as st:
             tensors = [
-                DefaultPrimitiveTensor(name=name, data=st.get_tensor(name))
+                DefaultPrimitiveTensor(
+                    name=name, data=st.get_tensor(name).to(target_dtype)
+                )
                 for name in st.keys()
             ]
 

sharktank/sharktank/types/tensors.py

@@ -388,6 +388,11 @@ def squeeze(self, dim: Optional[int] = None) -> "AnyTensor":
 
         return squeeze(self, dim)
 
+    def squeeze(self, dim: Optional[int] = None) -> "AnyTensor":
+        from ..ops import squeeze
+
+        return squeeze(self, dim)
+
     def transpose(self, dim0: int, dim1: int) -> "AnyTensor":
         from ..ops import transpose