Add Flux transformer export for easier use outside of tests (#700)

Adapt the model to accept parameters as structured in the HF repo.

Make Punet parameters importation from HF more general to serve other
models as well.

When downloading a dataset from Hugging Face make it return the local
location of all downloaded files including extras, not just the
"leading" file.

Add sample_inputs method to the BaseLayer interface to help standardize
exportation.

Introduce a standard export function for static-sized models.

Author: sogartar

.github/workflows/ci-sharktank.yml

@@ -136,15 +136,19 @@ jobs:
           pip install --no-compile -r requirements.txt -r sharktank/requirements-tests.txt -e sharktank/
 
       - name: Run tests
-      # TODO: unify with-t5-data and with-clip-data flags into a single flag
-      # and make it possible to run only tests that require data.
+      # TODO: unify with-*-data flags into a single flag and make it possible to run
+      # only tests that require data.
+      # We would still want the separate flags as we may endup with data being
+      # scattered on different CI machines.
         run: |
           source ${VENV_DIR}/bin/activate
           pytest \
-          --with-clip-data \
+            --with-clip-data \
+            --with-flux-data \
             --with-t5-data \
             sharktank/tests/models/clip/clip_test.py \
             sharktank/tests/models/t5/t5_test.py \
+            sharktank/tests/models/flux/flux_test.py \
             --durations=0
 
 

sharktank/conftest.py

@@ -97,6 +97,15 @@ def pytest_addoption(parser):
             "code. The user is expected to provide the data"
         ),
     )
+    parser.addoption(
+        "--with-flux-data",
+        action="store_true",
+        default=False,
+        help=(
+            "Enable tests that use Flux data like models that is not a part of the source "
+            "code. The user is expected to provide the data"
+        ),
+    )
     parser.addoption(
         "--with-t5-data",
         action="store_true",

sharktank/integration/models/punet/integration_test.py

@@ -67,7 +67,7 @@ def download(filename):
 
 @pytest.fixture(scope="module")
 def sdxl_fp16_dataset(sdxl_fp16_base_files, temp_dir):
-    from sharktank.models.punet.tools import import_hf_dataset
+    from sharktank.tools import import_hf_dataset
 
     dataset = temp_dir / "sdxl_fp16_dataset.irpa"
     import_hf_dataset.main(

sharktank/sharktank/export.py

@@ -4,11 +4,14 @@
 # See https://llvm.org/LICENSE.txt for license information.
 # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
-from typing import Callable, Any
+from typing import Callable, Optional, Any
 import torch
+from os import PathLike
+import iree.turbine.aot as aot
 from iree.turbine.aot import DeviceAffinity, FxProgramsBuilder
 from torch.utils._pytree import tree_structure, tree_unflatten, tree_flatten
 from .types.tensors import ShardedTensor
+from .layers import BaseLayer
 from torch.utils._pytree import PyTree, _is_leaf
 import functools
 
@@ -172,3 +175,51 @@ def flat_fn(*args, **kwargs):
         )
 
     assert False, "TODO: implement the case when not using an FxProgramsBuilder"
+
+
+def export_static_model_mlir(
+    model: BaseLayer,
+    output_path: PathLike,
+    function_batch_size_pairs: Optional[dict[Optional[str], list[int]]] = None,
+    batch_sizes: Optional[list[int]] = None,
+):
+    """Export a model with no dynamic dimensions.
+
+    For the set of provided function name batch sizes pair, the resulting MLIR will
+    have function names with the below format.
+    ```
+    <function_name>_bs<batch_size>
+    ```
+
+    If `batch_sizes` is given then it defaults to a single function with named
+    "forward".
+
+    The model is required to implement method `sample_inputs`.
+    """
+
+    assert not (function_batch_size_pairs is not None and batch_sizes is not None)
+
+    if batch_sizes is not None:
+        function_batch_size_pairs = {None: batch_sizes}
+
+    if function_batch_size_pairs is None and batch_sizes is None:
+        function_batch_size_pairs = {None: batch_sizes}
+
+    fxb = FxProgramsBuilder(model)
+
+    for function, batch_sizes in function_batch_size_pairs.items():
+        for batch_size in batch_sizes:
+            args, kwargs = model.sample_inputs(batch_size, function)
+
+            @fxb.export_program(
+                name=f"{function or 'forward'}_bs{batch_size}",
+                args=args,
+                kwargs=kwargs,
+                dynamic_shapes=None,
+                strict=False,
+            )
+            def _(model, **kwargs):
+                return model(**kwargs)
+
+    output = aot.export(fxb)
+    output.save_mlir(output_path)

sharktank/sharktank/layers/base.py

@@ -4,15 +4,12 @@
 # See https://llvm.org/LICENSE.txt for license information.
 # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
-from typing import Dict
-
+from typing import Dict, Optional
+from collections import OrderedDict
 import torch
 import torch.nn as nn
 
-from ..types import (
-    InferenceTensor,
-    Theta,
-)
+from ..types import InferenceTensor, Theta, AnyTensor
 from ..utils import debugging
 
 __all__ = [
@@ -56,6 +53,21 @@ def assert_not_nan(self, *ts: torch.Tensor):
                 if torch.isnan(t).any():
                     raise AssertionError(f"Tensor contains nans! {t}")
 
+    def sample_inputs(
+        self, batch_size: int = 1, function: Optional[str] = None
+    ) -> tuple[tuple[AnyTensor], OrderedDict[str, AnyTensor]]:
+        """Return sample inputs that can be used to run the function from the model.
+        If function is None then layer is treated as the callable.
+        E.g.
+        ```
+        args, kwargs = model.sample_inputs()
+        model(*args, **kwargs)
+        ```
+
+        One purpose of this method is to standardize exportation of models to MLIR.
+        """
+        raise NotImplementedError()
+
 
 class ThetaLayer(BaseLayer):
     "Base class for layers that derive parameters from a Theta object."

sharktank/sharktank/layers/mmdit.py

@@ -55,11 +55,15 @@ def __init__(self, theta, num_heads: int):
         self.add_module("img_attn_qkv", LinearLayer(theta("img_attn.qkv")))
         self.add_module(
             "img_attn_norm_q",
-            RMSNormLayer(theta("img_attn.norm.query_norm"), epsilon=1e-6),
+            RMSNormLayer(
+                theta("img_attn.norm.query_norm"), weight_name="scale", epsilon=1e-6
+            ),
         )
         self.add_module(
             "img_attn_norm_k",
-            RMSNormLayer(theta("img_attn.norm.key_norm"), epsilon=1e-6),
+            RMSNormLayer(
+                theta("img_attn.norm.key_norm"), weight_name="scale", epsilon=1e-6
+            ),
         )
         self.add_module("img_attn_proj", LinearLayer(theta("img_attn.proj")))
 
@@ -70,11 +74,15 @@ def __init__(self, theta, num_heads: int):
         self.add_module("txt_attn_qkv", LinearLayer(theta("txt_attn.qkv")))
         self.add_module(
             "txt_attn_norm_q",
-            RMSNormLayer(theta("txt_attn.norm.query_norm"), epsilon=1e-6),
+            RMSNormLayer(
+                theta("txt_attn.norm.query_norm"), weight_name="scale", epsilon=1e-6
+            ),
         )
         self.add_module(
             "txt_attn_norm_k",
-            RMSNormLayer(theta("txt_attn.norm.key_norm"), epsilon=1e-6),
+            RMSNormLayer(
+                theta("txt_attn.norm.key_norm"), weight_name="scale", epsilon=1e-6
+            ),
         )
         self.add_module("txt_attn_proj", LinearLayer(theta("txt_attn.proj")))
 
@@ -151,14 +159,15 @@ def __init__(self, theta, num_heads: int):
         super().__init__(theta)
 
         self.num_heads = num_heads
-        self.add_module("mod", ModulationLayer(theta("mod"), double=False))
+        self.add_module("mod", ModulationLayer(theta("modulation"), double=False))
         self.add_module(
-            "attn_norm_q", RMSNormLayer(theta("attn.norm.query_norm"), epsilon=1e-6)
+            "attn_norm_q",
+            RMSNormLayer(theta("norm.query_norm"), weight_name="scale", epsilon=1e-6),
         )
         self.add_module(
-            "attn_norm_k", RMSNormLayer(theta("attn.norm.key_norm"), epsilon=1e-6)
+            "attn_norm_k",
+            RMSNormLayer(theta("norm.key_norm"), weight_name="scale", epsilon=1e-6),
         )
-        self.add_module("attn_proj", LinearLayer(theta("attn.proj")))
 
         self.add_module("linear1", LinearLayer(theta("linear1")))
         self.add_module("linear2", LinearLayer(theta("linear2")))

sharktank/sharktank/layers/testing.py

@@ -65,10 +65,10 @@ def make_mmdit_double_block_random_theta(
     mlp_hidden_size3 = int(2 * (mlp_ratio - 1) * hidden_size)
     return Theta(
         {
-            "img_attn.norm.key_norm.weight": DefaultPrimitiveTensor(  #
+            "img_attn.norm.key_norm.scale": DefaultPrimitiveTensor(  #
                 data=make_rand_torch((in_channels,), dtype=dtype)
             ),
-            "img_attn.norm.query_norm.weight": DefaultPrimitiveTensor(  #
+            "img_attn.norm.query_norm.scale": DefaultPrimitiveTensor(  #
                 data=make_rand_torch((in_channels,), dtype=dtype)
             ),
             "img_attn.proj.bias": DefaultPrimitiveTensor(
@@ -101,10 +101,10 @@ def make_mmdit_double_block_random_theta(
             "img_mod.lin.weight": DefaultPrimitiveTensor(
                 data=make_rand_torch((mlp_hidden_size3, hidden_size), dtype=dtype)
             ),
-            "txt_attn.norm.key_norm.weight": DefaultPrimitiveTensor(  #
+            "txt_attn.norm.key_norm.scale": DefaultPrimitiveTensor(  #
                 data=make_rand_torch((in_channels,), dtype=dtype)
             ),
-            "txt_attn.norm.query_norm.weight": DefaultPrimitiveTensor(  #
+            "txt_attn.norm.query_norm.scale": DefaultPrimitiveTensor(  #
                 data=make_rand_torch((in_channels,), dtype=dtype)
             ),
             "txt_attn.proj.bias": DefaultPrimitiveTensor(
@@ -155,10 +155,10 @@ def make_mmdit_single_block_random_theta(
     mlp_hidden_size3 = int((2 * mlp_ratio - 1) * hidden_size)
     return Theta(
         {
-            "attn.norm.key_norm.weight": DefaultPrimitiveTensor(  #
+            "norm.key_norm.scale": DefaultPrimitiveTensor(  #
                 data=make_rand_torch((in_channels,), dtype=dtype)
             ),
-            "attn.norm.query_norm.weight": DefaultPrimitiveTensor(  #
+            "norm.query_norm.scale": DefaultPrimitiveTensor(  #
                 data=make_rand_torch((in_channels,), dtype=dtype)
             ),
             "attn.proj.bias": DefaultPrimitiveTensor(
@@ -179,10 +179,10 @@ def make_mmdit_single_block_random_theta(
             "linear2.weight": DefaultPrimitiveTensor(
                 data=make_rand_torch((hidden_size, mlp_hidden_size2), dtype=dtype)
             ),
-            "mod.lin.bias": DefaultPrimitiveTensor(
+            "modulation.lin.bias": DefaultPrimitiveTensor(
                 data=make_rand_torch((mlp_hidden_size,), dtype=dtype)
             ),
-            "mod.lin.weight": DefaultPrimitiveTensor(
+            "modulation.lin.weight": DefaultPrimitiveTensor(
                 data=make_rand_torch((mlp_hidden_size, hidden_size), dtype=dtype)
             ),
         }

sharktank/sharktank/models/flux/export.py

@@ -0,0 +1,49 @@
+# Copyright 2024 Advanced Micro Devices, Inc.
+#
+# Licensed under the Apache License v2.0 with LLVM Exceptions.
+# See https://llvm.org/LICENSE.txt for license information.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+from os import PathLike
+
+from ...export import export_static_model_mlir
+from ...tools.import_hf_dataset import import_hf_dataset
+from .flux import FluxModelV1, FluxParams
+from ...types import Dataset
+from ...utils.hf_datasets import get_dataset
+
+flux_transformer_default_batch_sizes = [4]
+
+
+def export_flux_transformer_model_mlir(
+    model: FluxModelV1,
+    output_path: PathLike,
+    batch_sizes: list[int] = flux_transformer_default_batch_sizes,
+):
+    export_static_model_mlir(model, output_path=output_path, batch_sizes=batch_sizes)
+
+
+def export_flux_transformer_from_hugging_face(
+    repo_id: str,
+    mlir_output_path: PathLike,
+    parameters_output_path: PathLike,
+    batch_sizes: list[int] = flux_transformer_default_batch_sizes,
+):
+    hf_dataset = get_dataset(
+        repo_id,
+    ).download()
+
+    import_hf_dataset(
+        config_json_path=hf_dataset["config"][0],
+        param_paths=hf_dataset["parameters"],
+        output_irpa_file=parameters_output_path,
+    )
+
+    dataset = Dataset.load(parameters_output_path)
+    model = FluxModelV1(
+        theta=dataset.root_theta,
+        params=FluxParams.from_hugging_face_properties(dataset.properties),
+    )
+    export_flux_transformer_model_mlir(
+        model, output_path=mlir_output_path, batch_sizes=batch_sizes
+    )