Flux2TTRTrainer

Flux2TTRTrainer

advanced/attentionexperimental

Phase-1: distill Flux TTR linear attention modules from native attention.

Pack: ComfyUI-Taylor-Attention

custom_nodes.ComfyUI-Taylor-Attention

Inputs (31)

Name	Type	Required
model	MODEL	required
latents	LATENT	required
conditioning	CONDITIONING	required
steps	INT	required
training	BOOLEAN	required
training_preview_ttr	BOOLEAN	required
checkpoint_path	STRING	required
feature_dim	INT	required
query_chunk_size	INT	required
key_chunk_size	INT	required
landmark_fraction	FLOAT	required
landmark_min	INT	required
landmark_max	INT	required
text_tokens_guess	INT	required
alpha_init	FLOAT	required
training_query_token_cap	INT	required
replay_buffer_size	INT	required
replay_offload_cpu	BOOLEAN	required
replay_max_mb	INT	required
train_steps_per_call	INT	required
readiness_threshold	FLOAT	required
readiness_min_updates	INT	required
enable_memory_reserve	BOOLEAN	required
layer_start	INT	required
layer_end	INT	required
cfg_scale	FLOAT	required
min_swap_layers	INT	required
max_swap_layers	INT	required
inference_mixed_precision	BOOLEAN	required
controller_checkpoint_path	STRING	required
training_config	TTR_TRAINING_CONFIG	optional

Outputs (2)

Name	Type
MODEL	MODEL
loss_value	FLOAT