Add xfail tests for single layer Flux transformer that verify IREE re…

…sults against Torch (#741)

The test accuracy is not of sufficient quality and needs further
investigation.
The tests compare IREE bf16 and f32 against Torch f32.

Refactor the sample input generation and make it produce noise images
for final size of 1024x1024 instead of 512x512.

Remove unused duplicated function for random Theta generation.

sharktank/sharktank/models/flux/flux.py

@@ -21,6 +21,7 @@
 from ...layers import *
 from ...types import *
 from ...utils.create_cache import *
+from ...utils.testing import make_rand_torch
 from ... import ops
 
 __all__ = [
@@ -196,13 +197,37 @@ def sample_inputs(
         if not (function is None or function == "forward"):
             raise ValueError(f'Only function "forward" is supported. Got "{function}"')
 
-        # TODO: do not hardcode these but derive the required shapes from the config.
-        img = torch.rand([batch_size, 1024, 64], dtype=self.dtype)
-        img_ids = torch.rand([batch_size, 1024, 3], dtype=torch.float32)
-        txt = torch.rand([batch_size, 512, 4096], dtype=self.dtype)
-        txt_ids = torch.rand([batch_size, 512, 3], dtype=torch.float32)
+        # The allowed range of these values is dependent on the model size.
+        # They will not work for all variants, specifically toy-sized models.
+        output_img_height = 1024
+        output_img_width = 1024
+        output_img_channels = 3
+
+        img = self._get_noise(
+            batch_size, output_img_height, output_img_width, self.dtype
+        )
+
+        _, c, h, w = img.shape
+        img = img.reshape(batch_size, h * w // 4, c * 4)
+
+        img_ids = torch.zeros(h // 2, w // 2, output_img_channels)
+        img_ids[..., 1] = img_ids[..., 1] + torch.arange(h // 2)[:, None]
+        img_ids[..., 2] = img_ids[..., 2] + torch.arange(w // 2)[None, :]
+        img_ids = img_ids.reshape(1, h * w // 4, output_img_channels)
+        img_ids = img_ids.repeat(batch_size, 1, 1)
+
+        # T5 encoder output
+        txt_context_length = 512
+        txt_dims_per_token = 4096
+        txt = torch.rand([1, txt_context_length, txt_dims_per_token], dtype=self.dtype)
+        txt = txt.repeat(batch_size, 1, 1)
+        txt_ids = torch.zeros(batch_size, txt.shape[1], output_img_channels)
+
         timesteps = torch.rand([batch_size], dtype=self.dtype)
-        y = torch.rand([batch_size, 768], dtype=self.dtype)
+
+        # CLIP text model output
+        y = make_rand_torch([1, 768], dtype=self.dtype)
+        y = y.repeat(batch_size, 1)
 
         args = tuple()
         kwargs = OrderedDict(
@@ -217,10 +242,26 @@ def sample_inputs(
         )
 
         if self.guidance:
-            kwargs["guidance"] = torch.rand([batch_size], dtype=self.dtype)
+            kwargs["guidance"] = torch.full([batch_size], 3.5, dtype=self.dtype)
 
         return args, kwargs
 
+    def _get_noise(
+        self,
+        batch_size: int,
+        height: int,
+        width: int,
+        dtype: torch.dtype,
+    ):
+        return torch.randn(
+            batch_size,
+            16,
+            # allow for packing
+            2 * math.ceil(height / 16),
+            2 * math.ceil(width / 16),
+            dtype=dtype,
+        )
+
     def _deduce_dtype(self) -> torch.dtype:
         dtype = self.theta("img_in.weight").dtype
         assert (

sharktank/sharktank/models/flux/testing.py

@@ -216,17 +216,8 @@ def make_random_theta(config: FluxParams, dtype: torch.dtype):
     return Theta(tensor_dict)
 
 
-def export_dev_random_single_layer(
-    dtype: torch.dtype,
-    mlir_output_path: PathLike,
-    parameters_output_path: PathLike,
-    batch_sizes: list[int] = flux_transformer_default_batch_sizes,
-):
-    rng_state = torch.get_rng_state()
-    torch.random.manual_seed(12345)
-
-    dtype = torch.bfloat16
-    params = FluxParams(
+def make_dev_single_layer_config():
+    return FluxParams(
         in_channels=64,
         out_channels=64,
         vec_in_dim=768,
@@ -241,6 +232,19 @@ def export_dev_random_single_layer(
         qkv_bias=True,
         guidance_embed=True,
     )
+
+
+def export_dev_random_single_layer(
+    dtype: torch.dtype,
+    mlir_output_path: PathLike,
+    parameters_output_path: PathLike,
+    batch_sizes: list[int] = flux_transformer_default_batch_sizes,
+):
+    rng_state = torch.get_rng_state()
+    torch.random.manual_seed(12345)
+
+    dtype = torch.bfloat16
+    params = make_dev_single_layer_config()
     theta = make_random_theta(params, dtype)
     flux = FluxModelV1(
         theta=theta,

sharktank/sharktank/utils/iree.py

@@ -119,6 +119,24 @@ def bfloat16_device_array_to_torch(
         return torch.tensor(device_array.to_host())
 
 
+def torch_tensor_to_device_array(
+    tensor: torch.Tensor, device: iree.runtime.HalDevice
+) -> iree.runtime.DeviceArray:
+    if tensor.dtype == torch.bfloat16:
+        tensor_as_int16 = tensor.view(dtype=torch.int16)
+        device_array_as_int16 = iree.runtime.asdevicearray(
+            device, unbox_tensor(tensor_as_int16).to("cpu").numpy()
+        )
+        buffer_view = iree.runtime.HalBufferView(
+            buffer=device_array_as_int16._buffer_view.get_buffer(),
+            shape=device_array_as_int16._buffer_view.shape,
+            element_type=iree.runtime.HalElementType.BFLOAT_16,
+        )
+        return iree.runtime.DeviceArray(device, buffer_view)
+
+    return iree.runtime.asdevicearray(device, unbox_tensor(tensor).to("cpu").numpy())
+
+
 def run_iree_module_function(
     module: iree.runtime.VmModule,
     vm_context: iree.runtime.VmContext,
@@ -180,11 +198,7 @@ def prepare_iree_module_function_args(
                 ]
             )
         elif isinstance(arg, (DefaultPrimitiveTensor, torch.Tensor)):
-            res.append(
-                iree.runtime.asdevicearray(
-                    devices[0], unbox_tensor(arg).to("cpu").numpy()
-                )
-            )
+            res.append(torch_tensor_to_device_array(arg, devices[0]))
         else:
             assert isinstance(arg, collections.abc.Sequence)
             res.extend(prepare_iree_module_function_args(arg, devices))
@@ -200,24 +214,27 @@ def flatten_for_iree_signature(tree: Tree) -> List[torch.Tensor]:
 def call_torch_module_function(
     module: torch.nn.Module,
     function_name: str,
-    kwargs: OrderedDict,
+    args: Optional[tuple[AnyTensor]] = None,
+    kwargs: Optional[OrderedDict] = None,
     trace_path_prefix: Optional[str] = None,
 ):
     """Call a torch module function with optional tracing.
     For tracing the arguments/results are flattened to match IREE's signature."""
+    args = args if args is not None else tuple()
+    kwargs = kwargs if kwargs is not None else OrderedDict()
     assert isinstance(
         kwargs, OrderedDict
     ), "Make sure when flattening the order is preserved"
     if trace_path_prefix is not None:
-        flat_args = flatten_for_iree_signature(kwargs)
+        flat_args = flatten_for_iree_signature([args, kwargs])
         for i, arg in enumerate(flat_args):
             np.save(
                 f"{trace_path_prefix}{function_name}_arg{i}.npy",
                 promote_bfloat16_to_float32(arg.to("cpu")).numpy(),
             )
-    res = getattr(module, function_name)(**kwargs)
+    res = getattr(module, function_name)(*args, **kwargs)
     if trace_path_prefix is not None:
-        flat_args = flatten_for_iree_signature(kwargs)
+        flat_args = flatten_for_iree_signature([args, kwargs])
         for i, arg in enumerate(flat_args):
             np.save(
                 f"{trace_path_prefix}{function_name}_arg{i}.npy",