MODEL: SDXL issues tracking.

[monorimet]

We need a centralized resource for SDXL bringup.

This will serve as a high-level map of the pending tasks and their assignees.

We do not want to split into different IR paths in the future, so first priority is making sure we can decompose sdpfa properly in IREE. To reproduce any issues tracked here, please for the time being use the ean-sd-fp16 branch of SHARK-Turbine

Model Files: https://github.com/nod-ai/playbook?tab=readme-ov-file#ai-model-download-links

Compile Commands: https://github.com/nod-ai/playbook/blob/main/HOWTO/sd-commands.md

Following are tables for each backend that track the current state of SDXL compilation and numerics validation, for each submodel.

By default these tables track : stabilityai/stable-diffusion-xl-base-1.0 output size 1024x1024, f16 precision, BS1, max length 64.

CPU:

Variant	Submodel	Compile	Runtime	Numerics	Assignee
SDXL
	UNet + attn	Good	Good	Fail (NaN)	@hanhanW
	Clip1	Good	Good	Good
	Clip2	Good	Good	Good
	VAE decode + attn	Good	Good	Fail (NaN)	@hanhanW
	VAE encode + attn	Good	Good	Fail (NaN)	@hanhanW

Vulkan-SPIRV:

Variant	Submodel	Compile	Runtime	Numerics	Assignee
SDXL
	UNet + attn	Fail (attn tile+decomp)			@erman-gurses
	Clip1	Fail			@Eliasj42
	Clip2	Fail			@Eliasj42
	VAE decode + attn	Fail (attn tile+decomp)			@erman-gurses
	VAE encode + attn	Fail (attn tile+decomp)			@erman-gurses

ROCM:

Variant	Submodel	Compile	Runtime	Numerics	Assignee
SDXL
	UNet + attn	Good	Fail (shared memory exceeds limit by ~800 bytes		@erman-gurses
	Clip1				@Eliasj42
	Clip2				@Eliasj42
	VAE decode + attn	Fail (shared memory)			@erman-gurses
	VAE encode + attn	Fail (shared memory)			@erman-gurses

Additionally, we track testing for SDXL models and what needs done for them @jinchen62 :

• Benchmarks
• Parametrized/configurable tests (dont hardcode all the args as is currently done in turbine_models/tests/sd(xl)_test.py)
• Actual testing functionalities e.g. xfails, reproducer automation, etc.
• optional tear-down (artifacts cleanup) that works on windows+linux
• multiple entry-points (i.e. don't recompile vmfbs or re-import .mlir if they already exist, unless the user asks for recompile/import)