Fix CogVideoX support #261
Conversation
…ion _init_sync_pipeline
| ], | ||
| extras_require={ | ||
| "all": [ | ||
| "flash_attn>=2.6.3", |
There was a problem hiding this comment.
we should update readme.
pip install xfuser [flash_attn]?
There was a problem hiding this comment.
Ok I changed all to flash_attn and modified the readme.
| @@ -12,40 +12,34 @@ def get_cuda_version(): | |||
| except Exception as e: | |||
| return 'no_cuda' | |||
|
|
|||
| def get_install_requires(cuda_version): | |||
There was a problem hiding this comment.
The func was introduced in #259, why remove it?
There was a problem hiding this comment.
This package (xdit) does not require a specific CUDA version to build or run.
So this warning is meaningless.
There was a problem hiding this comment.
And even this commit #259 is buggy. It actually wants to print a warning when the version is not 12.4. but actually it prints a warning when the version is equal to 12.4...😢
There was a problem hiding this comment.
The way it checks the cuda version is also incorrect. It checks the version of the system cuda rather than the cuda bundled with the PyTorch installation. So the get_cuda_version func could be removed in a future commit.
| def torch_compile_disable_if_v100(func): | ||
| if is_v100(): | ||
| return torch.compiler.disable(func) | ||
| return func | ||
|
|
||
|
|
||
| def apply_rotary_emb( |
There was a problem hiding this comment.
This change result in the following error:
torchrun --nproc_per_node=2 ./examples/flux_example.py --model /cfs/dit/FLUX.1-dev --ulysses_degree 2 --prompt "A snowy mountain" --num_inference_steps 20
[rank0]: Traceback (most recent call last):
[rank0]: File "~/xDiT/./examples/flux_example.py", line 77, in <module>
[rank0]: main()
[rank0]: File "~/xDiT/./examples/flux_example.py", line 35, in main
[rank0]: pipe.prepare_run(input_config)
[rank0]: File "~/xDiT/xfuser/model_executor/pipelines/pipeline_flux.py", line 69, in prepare_run
[rank0]: self.__call__(
[rank0]: File "~/miniconda3/envs/long_ctx_attn/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
[rank0]: return func(*args, **kwargs)
[rank0]: File "~/xDiT/xfuser/model_executor/pipelines/base_pipeline.py", line 181, in wrapper
[rank0]: return func(*args, **kwargs)
[rank0]: File "~/xDiT/xfuser/model_executor/pipelines/base_pipeline.py", line 133, in data_parallel_fn
[rank0]: return func(self, *args, **kwargs)
[rank0]: File "~/xDiT/xfuser/model_executor/pipelines/base_pipeline.py", line 149, in check_naive_forward_fn
[rank0]: return func(self, *args, **kwargs)
[rank0]: File "~/xDiT/xfuser/model_executor/pipelines/pipeline_flux.py", line 297, in __call__
[rank0]: latents = self._sync_pipeline(
[rank0]: File "~/xDiT/xfuser/model_executor/pipelines/pipeline_flux.py", line 399, in _sync_pipeline
[rank0]: latents, encoder_hidden_states = self._backbone_forward(
[rank0]: File "~/xDiT/xfuser/model_executor/pipelines/pipeline_flux.py", line 484, in _backbone_forward
[rank0]: noise_pred, encoder_hidden_states = self.transformer(
[rank0]: File "~/miniconda3/envs/long_ctx_attn/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
[rank0]: return self._call_impl(*args, **kwargs)
[rank0]: File "~/miniconda3/envs/long_ctx_attn/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl
[rank0]: return forward_call(*args, **kwargs)
[rank0]: File "~/xDiT/xfuser/model_executor/models/transformers/transformer_flux.py", line 147, in forward
[rank0]: encoder_hidden_states, hidden_states = block(
[rank0]: File "~/miniconda3/envs/long_ctx_attn/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
[rank0]: return self._call_impl(*args, **kwargs)
[rank0]: File "~/miniconda3/envs/long_ctx_attn/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl
[rank0]: return forward_call(*args, **kwargs)
[rank0]: File "~/miniconda3/envs/long_ctx_attn/lib/python3.10/site-packages/diffusers/models/transformers/transformer_flux.py", line 200, in forward
[rank0]: attn_output, context_attn_output = self.attn(
[rank0]: File "~/miniconda3/envs/long_ctx_attn/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
[rank0]: return self._call_impl(*args, **kwargs)
[rank0]: File "~/miniconda3/envs/long_ctx_attn/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl
[rank0]: return forward_call(*args, **kwargs)
[rank0]: File "~/xDiT/xfuser/model_executor/layers/attention_processor.py", line 223, in forward
[rank0]: return self.processor(
[rank0]: File "~/xDiT/xfuser/model_executor/layers/attention_processor.py", line 705, in __call__
[rank0]: query = apply_rotary_emb(query, image_rotary_emb)
[rank0]: File "~/xDiT/xfuser/model_executor/layers/attention_processor.py", line 76, in apply_rotary_emb
[rank0]: cos, sin = freqs_cis # [S, D]
[rank0]: ValueError: not enough values to unpack (expected 2, got 1)
We can merge this pr again after the problem is fixed.
There was a problem hiding this comment.
Oh, looks like the latest diffusers has changed a lot and make itself incompatible with the flux implementation in xdit anymore.
I guess a refactor could be needed.
There was a problem hiding this comment.
After modifying the attention processor and remove resuing encoder_hidden_states it now works with diffusers==0.30.2.
Anyway it is still only for reference for you. I guess the frequent changes in diffusers and lack of some automatic code patching mechanism could be challenging to fit. So those changes need to be trated carefully.🙂
This reverts commit 9484590.
This reverts commit 9484590.
This reverts commit 9484590.
This reverts commit 9484590.
xFuserCogVideoXAttnProcessor2_0to supportCogVideoXAttnProcessor2_0in newer version ofdiffusers.diffusers(use a bundledapply_rotary_emb).Current status
torchrun --nproc_per_node=2 examples/cogvideox_example.py --ulysses_degree 2 \ --model THUDM/CogVideoX-5b --height 480 --width 720 --num_frames 30 \ --prompt "a panda playing piano"Works perfectly now.