EasyTeacache

a easy library to use teacache for speedup dit models, contain auto params solve. support different sequence length use special params to minimum loss.

Introduction

you can use tea-cache to speedup your DIT models, and easily solve the parameters associated with the model weights.

Usage

Speedup mode

if you already have params, you can use this mode

from EasyTeaCache import TeaCache
cache = TeaCache(
    min_skip_step=2,        # teacache can skip first step is index==1 (start from 0)
    max_skip_step=48, # teacache can skip first step is index==48 (start from 0)
    threshold=0.04,
    model_keys=["mymodel","function"],  # any strings to sign your model-weight, support any depth
    cache_path="config/teacache/cache.json", # load config from here 
)

# in transformer.forward
skip_blocks = False
if teacache is not None:
    skip_blocks = teacache.check(
        step=time_stemp_index,
        t_mod=timestep_proj,
        sequence_length=hidden_states.size(1),
    )

if skip_blocks:
    hidden_states = teacache.update(timestep_proj, hidden_states)
else:
    input_hidden_states = hidden_states
    # 4. Transformer blocks
    for block in self.blocks:
        hidden_states = block(
            hidden_states,
            encoder_hidden_states,
            timestep_proj,
            rotary_emb,
            encoder_hidden_states_mask,
        )

    if teacache is not None:
        teacache.store_truth(
            step=time_stemp_index,
            t_mod=timestep_proj,
            input_latent=input_hidden_states,
            output_latent=hidden_states,
            sequence_length=hidden_states.size(1),
        )

Solve Params

you can not use sp-parallel in this mode

from EasyTeaCache import TeaCache
cache = TeaCache(
    min_skip_step=0,
    max_skip_step=-1, 
    threshold=0.04,
    model_keys=["mymodel","function"],  # any strings to sign your model-weight, support any depth
    cache_path="config/teacache/cache.json",   # save config here 
    speedup_mode=False,
)

# in transformer.forward, same with speedup mode
skip_blocks = False
if teacache is not None:
    skip_blocks = teacache.check(
        step=time_stemp_index,
        t_mod=timestep_proj,
        sequence_length=hidden_states.size(1),
    )

if skip_blocks:
    hidden_states = teacache.update(timestep_proj, hidden_states)
else:
    input_hidden_states = hidden_states
    # 4. Transformer blocks
    for block in self.blocks:
        hidden_states = block(
            hidden_states,
            encoder_hidden_states,
            timestep_proj,
            rotary_emb,
            encoder_hidden_states_mask,
        )

    if teacache is not None:
        teacache.store_truth(
            step=time_stemp_index,
            t_mod=timestep_proj,
            input_latent=input_hidden_states,
            output_latent=hidden_states,
            sequence_length=hidden_states.size(1),
        )

Reference

During the implementation of this project, we referred to the methods in the papers as well as possible open-source projects (either officially released or by third parties). However, there might still be some modifications. This could be due to the unclear or non-public nature of some methods, or because I considered them to be effective improvements.

• TeaCache: https://arxiv.org/html/2411.19108v2
• FoCa: https://arxiv.org/html/2508.16211v1