refactor the env setup and install of fastvideo

[PorridgeSwim]

Support optional lint install.
Remove sta_attn dependency.
Debug env_setup.py (flash_attn should has wheel preinstalled).

[SolitaryThinker]

Thank you for the effort!
Quick question, is it not possible to completely remove the env_setup.sh script? why?

[SolitaryThinker]

I don't like the use of scripts to install flash-attn or sta. I am going to take a look at whether we just put these two dependencies as optional under the .toml file. I will directly modify this PR and work on this.

[SolitaryThinker]

Okay, I checked xDit/xfuers, which depends on flash-attn package, and their instructions in README actually fails if you start with a clean conda env.

The best way forward is to add flash-attn instruction in our README if people want to install it and fall back to torch sdpa backend if flash-attn package is not available (already being done)

cc @jzhang38 @PorridgeSwim

[zhisbug]

Okay, I checked xDit/xfuers, which depends on flash-attn package, and their instructions in README actually fails if you start with a clean conda env.

The best way forward is to add flash-attn instruction in our README if people want to install it and fall back to torch sdpa backend if flash-attn package is not available (already being done)

cc @jzhang38 @PorridgeSwim

agree. Let's do that then. To summarize:

1. README should be explict in suggesting to pip install flash-attn
2. The attention backend will check fa existence and warn if not existed
3. if fa is not installed, fall back to use torch attention.

[jzhang38]

Yes, I will set the default backend to SDPA in my PR: #312

[SolitaryThinker]

merging to unblock ci/cd work