refactor Flux transformer to use scanned blocks, dynamic checkpointing, and decoupled projections#417
Open
prishajain1 wants to merge 1 commit into
Open
refactor Flux transformer to use scanned blocks, dynamic checkpointing, and decoupled projections#417prishajain1 wants to merge 1 commit into
prishajain1 wants to merge 1 commit into
Conversation
prishajain1
force-pushed
the
prisha/flux_training
branch
from
6dfd5ea to
4696256
Compare
…ng, and weight loading improvements
prishajain1
force-pushed
the
prisha/flux_training
branch
from
4696256 to
11ddfef
Compare
|
🤖 Hi @prishajain1, I've received your request, and I'm working on it now! You can track my progress in the logs for more details. |
|
🤖 I'm sorry @prishajain1, but I was unable to process your request. Please see the logs for more details. |
Perseus14
requested changes
Jun 12, 2026
Comment on lines
+220
to
+221
| names_which_can_be_saved=self.config.names_which_can_be_saved, | ||
| names_which_can_be_offloaded=self.config.names_which_can_be_offloaded, |
Collaborator
There was a problem hiding this comment.
Can we make these variable names a little more explicit? Maybe something like saved_transformer_layer_names or savable_transformer_layer_names and offloaded_transformer_layer_names or offloadable_transformer_layer_names? I will leave it upto you
Comment on lines
+290
to
292
| names_which_can_be_saved=self.config.names_which_can_be_saved, | ||
| names_which_can_be_offloaded=self.config.names_which_can_be_offloaded, | ||
| ) |
Comment on lines
+236
to
+237
| names_which_can_be_saved: [] | ||
| names_which_can_be_offloaded: [] |
Collaborator
There was a problem hiding this comment.
Can we rename the file to transformer_flux.py?
Comment on lines
-154
to
-166
| r""" | ||
| A Transformer block following the MMDiT architecture, introduced in Stable Diffusion 3. | ||
|
|
||
| Reference: https://arxiv.org/abs/2403.03206 | ||
|
|
||
| Parameters: | ||
| dim (`int`): The number of channels in the input and output. | ||
| num_attention_heads (`int`): The number of heads to use for multi-head attention. | ||
| attention_head_dim (`int`): The number of channels in each head. | ||
| context_pre_only (`bool`): Boolean to determine if we should add some blocks associated with the | ||
| processing of `context` conditions. | ||
| """ | ||
|
|
Collaborator
There was a problem hiding this comment.
Let's keep the comment block. Update if required
| vec_shape = ( | ||
| batch_size, | ||
| 768, # Sequence length of clip, how to get this programmatically? | ||
| 768, |
Comment on lines
-54
to
-61
| r""" | ||
| Norm layer adaptive layer norm zero (adaLN-Zero). | ||
|
|
||
| Parameters: | ||
| embedding_dim (`int`): The size of each embedding vector. | ||
| num_embeddings (`int`): The size of the embeddings dictionary. | ||
| """ | ||
|
|
Collaborator
There was a problem hiding this comment.
Let's keep the comment block
Comment on lines
-104
to
-111
| r""" | ||
| Norm layer adaptive layer norm zero (adaLN-Zero). | ||
|
|
||
| Parameters: | ||
| embedding_dim (`int`): The size of each embedding vector. | ||
| num_embeddings (`int`): The size of the embeddings dictionary. | ||
| """ | ||
|
|
Collaborator
There was a problem hiding this comment.
Let's keep the comment block
|
|
||
| x = (x - mean) * inv_std * (1.0 + scale_msa) + shift_msa | ||
| else: | ||
| raise ValueError(f"Unsupported `norm_type` ({self.norm_type}) provided. Supported ones are: 'layer_norm'.") |
Collaborator
There was a problem hiding this comment.
Is there a need to change this text comment?
Comment on lines
+318
to
+319
| names_which_can_be_saved=config.names_which_can_be_saved, | ||
| names_which_can_be_offloaded=config.names_which_can_be_offloaded, |
Collaborator
There was a problem hiding this comment.
Please change variable names
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Overview
This PR refactors the Flux model architecture in MaxDiffusion to support scanned blocks (nn.scan) for double and single blocks, implements configurable gradient checkpointing (rematerialization) policies, and updates the weights loader to support loading pretrained checkpoints under the scanned format.
Key Changes
MlpAndOutputBlockwrapper) to eliminate redundant recomputation of attention and projection outputs.jnp.splitacross Flux transformer blocks for cleaner layout constraints.nn.scanto optimize compiler tracing and step execution speed on TPUs.FLUX_OPTIMIZEDtoGradientCheckpointTypeto allow configuring block-specific rematerialization policies dynamically via configuration files instead of being hardcoded.util.py) to slice, group, and stack PyTorch checkpoint weights along axis 0 to match the expected format ofnn.scanlayers.