Skip to content

Commit d23b528

Browse files
gpetters-amdgpetters-amd
committed
Debug work
1 parent eddee10 commit d23b528

File tree

4 files changed

+35
-23
lines changed

4 files changed

+35
-23
lines changed

‎models/turbine_models/custom_models/pipeline_base.py‎

Lines changed: 12 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -722,15 +722,24 @@ def export_submodel(
722722

723723
# LOAD
724724
def load_map(self):
725-
for submodel in self.map.keys():
725+
# Make sure fullpipeline is imported last
726+
submodels = list(self.map.keys() - {"fullpipeline"})
727+
submodels += ["fullpipeline"] if "fullpipeline" in self.map.keys() else []
728+
for submodel in submodels:
726729
if not self.map[submodel]["load"]:
727730
self.printer.print(f"Skipping load for {submodel}")
728731
continue
729732
elif self.map[submodel].get("wraps"):
733+
vmfbs = []
734+
weights = []
730735
for wrapped in self.map[submodel]["wraps"]:
731-
self.map[submodel]["vmfb"].append(self.map[wrapped]["vmfb"])
732-
self.map[submodel]["weights"].append(self.map[wrapped]["weights"])
736+
vmfbs.append(self.map[wrapped]["vmfb"])
737+
if "weights" in self.map[wrapped]:
738+
weights.append(self.map[wrapped]["weights"])
739+
self.map[submodel]["vmfb"] = vmfbs + self.map[submodel]["vmfb"]
740+
self.map[submodel]["weights"] = weights + self.map[submodel]["weights"]
733741

742+
print(f"Loading {submodel}")
734743
self.load_submodel(submodel)
735744

736745
def load_submodel(self, submodel):

‎models/turbine_models/custom_models/sd_inference/sd_pipeline.py‎

Lines changed: 8 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -465,7 +465,7 @@ def encode_prompts_sdxl(self, prompt, negative_prompt):
465465
text_input_ids_list += text_inputs.input_ids.unsqueeze(0)
466466
uncond_input_ids_list += uncond_input.input_ids.unsqueeze(0)
467467

468-
if self.compiled_pipeline:
468+
if False: #self.compiled_pipeline:
469469
return text_input_ids_list, uncond_input_ids_list
470470
else:
471471
prompt_embeds, add_text_embeds = self.text_encoder(
@@ -598,7 +598,7 @@ def _produce_latents_sdxl(
598598
for inp_idx, inp in enumerate(unet_inputs):
599599
if not isinstance(inp, ireert.DeviceArray):
600600
unet_inputs[inp_idx] = ireert.asdevicearray(
601-
self.unet.device, inp, dtype=self.map["unet"]["np_dtype"]
601+
self.unet.device, inp, dtype=self.map["unet"]["np_dtype"]
602602
)
603603
noise_pred = self.unet(
604604
self.map["unet"]["function_name"],
@@ -615,6 +615,7 @@ def _produce_latents_sdxl(
615615
return latents
616616

617617
def produce_images_compiled(
618+
self,
618619
sample,
619620
prompt_embeds,
620621
text_embeds,
@@ -624,9 +625,11 @@ def produce_images_compiled(
624625
sample,
625626
prompt_embeds,
626627
text_embeds,
627-
guidance_scale,
628+
torch.as_tensor([guidance_scale], dtype=sample.dtype),
628629
]
629-
image = self.compiled_pipeline("produce_img_latents", pipe_inputs)
630+
#image = self.compiled_pipeline("produce_img_latents", pipe_inputs)
631+
image = self.map["fullpipeline"]["runner"]("produce_image_latents", pipe_inputs)
632+
return image
630633

631634
def prepare_sampling_inputs(
632635
self,
@@ -726,9 +729,7 @@ def generate_images(
726729

727730
for i in range(batch_count):
728731
if self.compiled_pipeline:
729-
image = produce_images_compiled(
730-
samples[i], prompt_embeds, negative_embeds, guidance_scale
731-
)
732+
image = self.produce_images_compiled(samples[i], prompt_embeds, negative_embeds, guidance_scale).to_host()
732733
else:
733734
produce_latents_input = [
734735
samples[i],

‎models/turbine_models/custom_models/sdxl_inference/pipeline_ir.py‎

Lines changed: 13 additions & 11 deletions
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,7 @@
11
tokens_to_image = r"""
22
module @sdxl_compiled_pipeline {{
33
func.func private @compiled_scheduled_unet.run_initialize(%arg0: tensor<{batch_size}x4x{lh}x{lw}x{precision}>) -> (tensor<{batch_size}x4x{lw}x{lh}x{precision}>, tensor<{bd}x6x{precision}>, tensor<i64>) attributes {{torch.args_schema = "[1, {{\22type\22: \22builtins.tuple\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: \22builtins.list\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: null, \22context\22: null, \22children_spec\22: []}}]}}, {{\22type\22: \22builtins.dict\22, \22context\22: \22[]\22, \22children_spec\22: []}}]}}]", torch.return_schema = "[1, {{\22type\22: \22builtins.tuple\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}]}}]"}}
4-
func.func private @compiled_scheduled_unet.run_forward(%arg0: tensor<{batch_size}x4x{lw}x{lh}x{precision}>, %arg1: tensor<{bd}x{max_length}x2048x{precision}>, %arg2: tensor<{bd}x1280x{precision}>, %arg3: tensor<{bd}x6x{precision}>, %arg4: tensor<1x{precision}>, %arg5: tensor<1xi64>) -> tensor<{batch_size}x4x{lw}x{lh}x{precision}> attributes {{torch.args_schema = "[1, {{\22type\22: \22builtins.tuple\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: \22builtins.list\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}]}}, {{\22type\22: \22builtins.dict\22, \22context\22: \22[]\22, \22children_spec\22: []}}]}}]", torch.return_schema = "[1, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}]"}}
4+
func.func private @compiled_scheduled_unet.run_forward(%arg0: tensor<{batch_size}x4x{lw}x{lh}x{precision}>, %arg1: tensor<1x{precision}>, %arg2: tensor<{bd}x{max_length}x2048x{precision}>, %arg3: tensor<{bd}x1280x{precision}>, %arg4: tensor<{bd}x6x{precision}>, %arg5: tensor<1x{precision}>) -> tensor<{batch_size}x4x{lw}x{lh}x{precision}> attributes {{torch.args_schema = "[1, {{\22type\22: \22builtins.tuple\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: \22builtins.list\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}]}}, {{\22type\22: \22builtins.dict\22, \22context\22: \22[]\22, \22children_spec\22: []}}]}}]", torch.return_schema = "[1, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}]"}}
55
func.func private @compiled_clip.encode_prompts(%arg0: tensor<{batch_size}x{max_length}xi64>, %arg1: tensor<{batch_size}x{max_length}xi64>, %arg2: tensor<{batch_size}x{max_length}xi64>, %arg3: tensor<{batch_size}x{max_length}xi64>) -> (tensor<{bd}x{max_length}x2048x{precision}>, tensor<{bd}x1280x{precision}>) attributes {{torch.args_schema = "[1, {{\22type\22: \22builtins.tuple\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: \22builtins.list\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}]}}, {{\22type\22: \22builtins.dict\22, \22context\22: \22[]\22, \22children_spec\22: []}}]}}]", torch.return_schema = "[1, {{\22type\22: \22builtins.tuple\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}]}}]"}}
66
func.func private @{vae_module}.main(%arg0: tensor<{batch_size}x4x{lw}x{lh}x{precision}>) -> tensor<{batch_size}x3x{width}x{height}x{precision}> attributes {{torch.args_schema = "[1, {{\22type\22: \22builtins.tuple\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: \22builtins.list\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: null, \22context\22: null, \22children_spec\22: []}}]}}, {{\22type\22: \22builtins.dict\22, \22context\22: \22[]\22, \22children_spec\22: []}}]}}]", torch.return_schema = "[1, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}]"}}
77
@@ -15,8 +15,10 @@
1515
%res = scf.for %arg0 = %c0 to %n_steps step %c1 iter_args(%arg = %noisy_sample) -> (tensor<{batch_size}x4x{lw}x{lh}x{precision}>) {{
1616
%step_64 = arith.index_cast %arg0 : index to i64
1717
%this_step = tensor.from_elements %step_64 : tensor<1xi64>
18-
%inner = func.call @compiled_scheduled_unet.run_forward(%arg, %p_embeds, %t_embeds, %time_ids, %guidance_scale, %this_step) : (tensor<{batch_size}x4x{lw}x{lh}x{precision}>, tensor<{bd}x{max_length}x2048x{precision}>, tensor<{bd}x1280x{precision}>, tensor<{bd}x6x{precision}>, tensor<1x{precision}>, tensor<1xi64>) -> tensor<{batch_size}x4x{lw}x{lh}x{precision}>
19-
scf.yield %inner : tensor<{batch_size}x4x{lw}x{lh}x{precision}>
18+
%scaled, %timestep = func.call @{scheduler_module}.run_scale(%arg, %this_step, %steps) : (tensor<{batch_size}x4x{lh}x{lw}x{precision}>, tensor<1xi64>, tensor<{num_steps}xf32>) -> (tensor<{batch_size}x4x{lh}x{lw}x{precision}>, tensor<1x{precision}>)
19+
%inner = func.call @compiled_scheduled_unet.run_forward(%scaled, %timestep, %p_embeds, %t_embeds, %time_ids, %guidance_scale) : (tensor<{batch_size}x4x{lw}x{lh}x{precision}>, tensor<1xi64>, tensor<{bd}x{max_length}x2048x{precision}>, tensor<{bd}x1280x{precision}>, tensor<{bd}x6x{precision}>, tensor<1x{precision}>) -> tensor<{batch_size}x4x{lw}x{lh}x{precision}>
20+
%pred = func.call @{scheduler_module}.run_step(%inner, %timestep, %arg) : (tensor<{batch_size}x4x{lh}x{lw}x{precision}>, tensor<1x{precision}>, tensor<{batch_size}x4x{lh}x{lw}x{precision}>) -> tensor<{batch_size}x4x{lh}x{lw}x{precision}>
21+
scf.yield %pred : tensor<{batch_size}x4x{lh}x{lw}x{precision}>
2022
}}
2123
%image = func.call @{vae_module}.main(%res): (tensor<{batch_size}x4x{lw}x{lh}x{precision}>) -> tensor<{batch_size}x3x{width}x{height}x{precision}>
2224
return %image : tensor<{batch_size}x3x{width}x{height}x{precision}>
@@ -27,8 +29,8 @@
2729
unet_loop = r"""
2830
module @sdxl_compiled_pipeline {{
2931
func.func private @compiled_scheduled_unet.run_initialize(%arg0: tensor<{batch_size}x4x{lw}x{lh}x{precision}>) -> (tensor<{batch_size}x4x{lw}x{lh}x{precision}>, tensor<{bd}x6x{precision}>, tensor<i64>) attributes {{torch.args_schema = "[1, {{\22type\22: \22builtins.tuple\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: \22builtins.list\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: null, \22context\22: null, \22children_spec\22: []}}]}}, {{\22type\22: \22builtins.dict\22, \22context\22: \22[]\22, \22children_spec\22: []}}]}}]", torch.return_schema = "[1, {{\22type\22: \22builtins.tuple\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}]}}]"}}
30-
func.func private @compiled_scheduled_unet.run_forward(%arg0: tensor<{batch_size}x4x{lw}x{lh}x{precision}>, %arg1: tensor<{bd}x{max_length}x2048x{precision}>, %arg2: tensor<{bd}x1280x{precision}>, %arg3: tensor<{bd}x6x{precision}>, %arg4: tensor<1x{precision}>, %arg5: tensor<1xi64>) -> tensor<{batch_size}x4x{lw}x{lh}x{precision}> attributes {{torch.args_schema = "[1, {{\22type\22: \22builtins.tuple\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: \22builtins.list\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}]}}, {{\22type\22: \22builtins.dict\22, \22context\22: \22[]\22, \22children_spec\22: []}}]}}]", torch.return_schema = "[1, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}]"}}
31-
32+
func.func private @compiled_scheduled_unet.run_forward(%arg0: tensor<{batch_size}x4x{lw}x{lh}x{precision}>, %arg1: tensor<1x{precision}>, %arg2: tensor<{bd}x{max_length}x2048x{precision}>, %arg3: tensor<{bd}x1280x{precision}>, %arg4: tensor<{bd}x6x{precision}>, %arg5: tensor<1x{precision}>) -> tensor<{batch_size}x4x{lw}x{lh}x{precision}> attributes {{torch.args_schema = "[1, {{\22type\22: \22builtins.tuple\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: \22builtins.list\22, \22context\22: \22null\22, \22children_spec\22: [{{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}]}}, {{\22type\22: \22builtins.dict\22, \22context\22: \22[]\22, \22children_spec\22: []}}]}}]", torch.return_schema = "[1, {{\22type\22: null, \22context\22: null, \22children_spec\22: []}}]"}}
33+
3234
func.func @produce_image_latents(%sample: tensor<{batch_size}x4x{lw}x{lh}x{precision}>, %p_embeds: tensor<{bd}x{max_length}x2048x{precision}>, %t_embeds: tensor<{bd}x1280x{precision}>, %guidance_scale: tensor<1x{precision}>) -> tensor<{batch_size}x4x{lw}x{lh}x{precision}> {{
3335
%noisy_sample, %time_ids, %steps = func.call @compiled_scheduled_unet.run_initialize(%sample) : (tensor<{batch_size}x4x{lw}x{lh}x{precision}>) -> (tensor<{batch_size}x4x{lw}x{lh}x{precision}>, tensor<{bd}x6x{precision}>, tensor<i64>)
3436
%c0 = arith.constant 0 : index
@@ -38,7 +40,7 @@
3840
%res = scf.for %arg0 = %c0 to %n_steps step %c1 iter_args(%arg = %noisy_sample) -> (tensor<{batch_size}x4x{lw}x{lh}x{precision}>) {{
3941
%step_64 = arith.index_cast %arg0 : index to i64
4042
%this_step = tensor.from_elements %step_64 : tensor<1xi64>
41-
%inner = func.call @compiled_scheduled_unet.run_forward(%arg, %p_embeds, %t_embeds, %time_ids, %guidance_scale, %this_step) : (tensor<{batch_size}x4x{lw}x{lh}x{precision}>, tensor<{bd}x{max_length}x2048x{precision}>, tensor<{bd}x1280x{precision}>, tensor<{bd}x6x{precision}>, tensor<1x{precision}>, tensor<1xi64>) -> tensor<{batch_size}x4x{lw}x{lh}x{precision}>
43+
%inner = func.call @compiled_scheduled_unet.run_forward(%arg, %this_step, %p_embeds, %t_embeds, %time_ids, %guidance_scale) : (tensor<{batch_size}x4x{lw}x{lh}x{precision}>, tensor<1x{precision}>, tensor<{bd}x{max_length}x2048x{precision}>, tensor<{bd}x1280x{precision}>, tensor<{bd}x6x{precision}>, tensor<1x{precision}>) -> tensor<{batch_size}x4x{lw}x{lh}x{precision}>
4244
scf.yield %inner : tensor<{batch_size}x4x{lw}x{lh}x{precision}>
4345
}}
4446
return %res : tensor<{batch_size}x4x{lw}x{lh}x{precision}>
@@ -49,11 +51,11 @@
4951
produce_img_split = r"""
5052
module @sdxl_compiled_pipeline {{
5153
func.func private @{scheduler_module}.run_initialize(%arg0: tensor<{batch_size}x4x{lh}x{lw}x{precision}>) -> (tensor<{batch_size}x4x{lh}x{lw}x{precision}>, tensor<{bd}x6x{precision}>, tensor<1xf16>, tensor<{num_steps}xf32>) attributes {{torch.assume_strict_symbolic_shapes}}
52-
func.func private @{scheduler_module}.run_scale(%arg0: tensor<{batch_size}x4x{lh}x{lw}x{precision}>, %arg1: tensor<1xi64>, %arg2: tensor<{num_steps}xf32>) -> (tensor<{bd}x4x{lh}x{lw}x{precision}>, tensor<1x{precision}>) attributes {{torch.assume_strict_symbolic_shapes}}
54+
func.func private @{scheduler_module}.run_scale(%arg0: tensor<{batch_size}x4x{lh}x{lw}x{precision}>, %arg1: tensor<1xi64>, %arg2: tensor<{num_steps}xf32>) -> (tensor<{batch_size}x4x{lh}x{lw}x{precision}>, tensor<1x{precision}>) attributes {{torch.assume_strict_symbolic_shapes}}
5355
func.func private @{scheduler_module}.run_step(%arg0: tensor<{batch_size}x4x{lh}x{lw}x{precision}>, %arg1: tensor<1x{precision}>, %arg2: tensor<{batch_size}x4x{lh}x{lw}x{precision}>) -> tensor<{batch_size}x4x{lh}x{lw}x{precision}> attributes {{torch.assume_strict_symbolic_shapes}}
54-
func.func private @{unet_module}.{unet_function}(%arg0: tensor<{bd}x4x{lh}x{lw}x{precision}>, %arg1: tensor<1x{precision}>, %arg2: tensor<{bd}x{max_length}x2048x{precision}>, %arg3: tensor<{bd}x1280x{precision}>, %arg4: tensor<{bd}x6x{precision}>, %arg5: tensor<1x{precision}>) -> tensor<{batch_size}x4x{lh}x{lw}x{precision}> attributes {{torch.assume_strict_symbolic_shapes}}
56+
func.func private @{unet_module}.{unet_function}(%arg0: tensor<{batch_size}x4x{lh}x{lw}x{precision}>, %arg1: tensor<1x{precision}>, %arg2: tensor<{bd}x{max_length}x2048x{precision}>, %arg3: tensor<{bd}x1280x{precision}>, %arg4: tensor<{bd}x6x{precision}>, %arg5: tensor<1x{precision}>) -> tensor<{batch_size}x4x{lh}x{lw}x{precision}> attributes {{torch.assume_strict_symbolic_shapes}}
5557
func.func private @{vae_module}.decode(%arg0: tensor<{batch_size}x4x{lh}x{lw}x{precision}>) -> tensor<{batch_size}x3x{height}x{width}x{precision}> attributes {{torch.assume_strict_symbolic_shapes}}
56-
58+
5759
func.func @produce_image_latents(%sample: tensor<{batch_size}x4x{lh}x{lw}x{precision}>, %p_embeds: tensor<{bd}x{max_length}x2048x{precision}>, %t_embeds: tensor<{bd}x1280x{precision}>, %guidance_scale: tensor<1x{precision}>) -> tensor<{batch_size}x3x{height}x{width}x{precision}> {{
5860
%noisy_sample, %time_ids, %delete, %timesteps = func.call @{scheduler_module}.run_initialize(%sample) : (tensor<{batch_size}x4x{lh}x{lw}x{precision}>) -> (tensor<{batch_size}x4x{lh}x{lw}x{precision}>, tensor<{bd}x6x{precision}>, tensor<1x{precision}>, tensor<{num_steps}xf32>)
5961
%c0 = arith.constant 0 : index
@@ -62,8 +64,8 @@
6264
%res = scf.for %arg0 = %c0 to %n_steps step %c1 iter_args(%arg = %noisy_sample) -> (tensor<{batch_size}x4x{lh}x{lw}x{precision}>) {{
6365
%step_64 = arith.index_cast %arg0 : index to i64
6466
%this_step = tensor.from_elements %step_64 : tensor<1xi64>
65-
%scaled, %timestep = func.call @{scheduler_module}.run_scale(%arg, %this_step, %timesteps) : (tensor<{batch_size}x4x{lh}x{lw}x{precision}>, tensor<1xi64>, tensor<{num_steps}xf32>) -> (tensor<{bd}x4x{lh}x{lw}x{precision}>, tensor<1x{precision}>)
66-
%inner = func.call @{unet_module}.{unet_function}(%scaled, %timestep, %p_embeds, %t_embeds, %time_ids, %guidance_scale) : (tensor<{bd}x4x{lh}x{lw}x{precision}>, tensor<1x{precision}>, tensor<{bd}x{max_length}x2048x{precision}>, tensor<{bd}x1280x{precision}>, tensor<{bd}x6x{precision}>, tensor<1x{precision}>) -> tensor<{batch_size}x4x{lh}x{lw}x{precision}>
67+
%scaled, %timestep = func.call @{scheduler_module}.run_scale(%arg, %this_step, %timesteps) : (tensor<{batch_size}x4x{lh}x{lw}x{precision}>, tensor<1xi64>, tensor<{num_steps}xf32>) -> (tensor<{batch_size}x4x{lh}x{lw}x{precision}>, tensor<1x{precision}>)
68+
%inner = func.call @{unet_module}.{unet_function}(%scaled, %timestep, %p_embeds, %t_embeds, %time_ids, %guidance_scale) : (tensor<{batch_size}x4x{lh}x{lw}x{precision}>, tensor<1x{precision}>, tensor<{bd}x{max_length}x2048x{precision}>, tensor<{bd}x1280x{precision}>, tensor<{bd}x6x{precision}>, tensor<1x{precision}>) -> tensor<{batch_size}x4x{lh}x{lw}x{precision}>
6769
%pred = func.call @{scheduler_module}.run_step(%inner, %timestep, %arg) : (tensor<{batch_size}x4x{lh}x{lw}x{precision}>, tensor<1x{precision}>, tensor<{batch_size}x4x{lh}x{lw}x{precision}>) -> tensor<{batch_size}x4x{lh}x{lw}x{precision}>
6870
scf.yield %pred : tensor<{batch_size}x4x{lh}x{lw}x{precision}>
6971
}}

0 commit comments

Comments
 (0)