Skip to content

downcast MiniMax-H3 position_ids at the device transfer on fp64-less backends #14767

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
SammyTourani wants to merge 1 commit into
base: main
Choose a base branch
from
+9 −3
Open

downcast MiniMax-H3 position_ids at the device transfer on fp64-less backends #14767

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 9 additions & 3 deletions src/diffusers/modular_pipelines/minimax_h3/before_denoise.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@

from ...schedulers import MiniMaxH3Scheduler
from ...utils import logging
from ...utils.torch_utils import randn_tensor
from ...utils.torch_utils import maybe_adjust_dtype_for_device, randn_tensor
from ..modular_pipeline import ModularPipelineBlocks, PipelineState
from ..modular_pipeline_utils import ComponentSpec, ConfigSpec, InputParam, OutputParam
from .modular_pipeline import (
Expand Down Expand Up @@ -441,7 +441,10 @@ def __call__(self, components: MiniMaxH3ModularPipeline, state: PipelineState) -
components.video_tag,
block_state.keyframe_anchors,
)
block_state.position_ids = position_ids.to(device)
# The grid is built in fp64 to reproduce the released coordinates exactly, but MPS, NPU and Neuron have no
# fp64; the transformer's rope casts to fp32 anyway, so downcasting here is all the transfer needs.
position_ids_dtype = maybe_adjust_dtype_for_device(position_ids.dtype, device)
block_state.position_ids = position_ids.to(device, position_ids_dtype)
block_state.token_tags = token_tags.to(device)
block_state.video_indices = video_indices.to(device)
block_state.audio_indices = audio_indices.to(device)
Expand Down Expand Up @@ -765,7 +768,10 @@ def __call__(self, components: MiniMaxH3ModularPipeline, state: PipelineState) -
components.audio_tag,
components.video_tag,
)
block_state.position_ids = position_ids.to(device)
# The grid is built in fp64 to reproduce the released coordinates exactly, but MPS, NPU and Neuron have no
# fp64; the transformer's rope casts to fp32 anyway, so downcasting here is all the transfer needs.
position_ids_dtype = maybe_adjust_dtype_for_device(position_ids.dtype, device)
block_state.position_ids = position_ids.to(device, position_ids_dtype)
block_state.token_tags = token_tags.to(device)
block_state.video_indices = video_indices.to(device)
block_state.audio_indices = audio_indices.to(device)
Expand Down
Loading