Skip to content

README.md

Latest commit

 

History

History
76 lines (41 loc) · 2.67 KB

README.md

File metadata and controls

76 lines (41 loc) · 2.67 KB

flux-poke

An effort to make the Flux-dev model run in less memory...

Step by step

Gather analysis data

  • Enable the Enable to run with analysis block, but then bypass the two nodes referring to Hidden State (unless you are happy to spend a lot of disc space on data!)
  • Use the prompt generator, or put your own prompts into CLIPTextEncodeFlux
  • Mess with the settings all you like
  • Each time you run, internal statistics will be saved into internals.safetensors
  • These stats record how often each internal state (there are 466,944 of them) gets activated (out of typically 80,000 possible triggers per image)
  • The more runs, the better the stats will be

Play with a pruned version

  • Bypass Enable to run with analysis and enable the Enable to run with pruned model block.
  • Try generating images with the model loaded using one of the Pruned nodes.
  • You can choose different ranges of layers (they are numbered 0-18 inclusive)
  • Either set a number of lines to cut ()

Notes

DoubleStreamBlocks (19) - 2.8B mlp parameters, 2.15B attn-mod parameters, 1.4B self-attn parameters

Flux has 19 DoubleStreamBlocks, each of which has two mlp (img and txt), each consisting of two Linear matrices. These have internal sizes of 12288 and hidden_size of 3072, so a total of 151M parameters each layer (12288.3072.2.2), or 2.8B parameters.

img_mod and txt_mod are 3072x18432 each - 2.15B

attn is 3072x3072x4 for each of txt and img - 1.4B

SingleStreamBlocks (38) - 2.8B mlp parameters, 2.15B attn-mod parameters, no self attn

The SingleStreamBlocks (of which there are 38) have
linear1 (in_features=3072, out_features=21504, bias=True) and linear2 (in_features=15360, out_features=3072, bias=True)

linear1 maps the hidden state (3072) to (qkvi), where i is the 12288 intermediate states

linear2 maps (ai) where a is the post-attention value of qkv onto hidden state.

So each of this includes a submatrix, linear1[9216:,:] and linear2[:,3072:] which are mlp-like. In these layers both the img and the txt data streams are multiplied by the same matrices.

Download prompts

Optionally, run python download_prompts.py in the custom node's directory to download prompts

Workflow

Normal Run

Analysis

Updates internals.safetensors and adds to the hidden_states directory

Done for 100 images from the whole dataset, then another 100 requiring there to be a " in the prompt

Pruned Run

Just throw stuff away

layer retrain

Uses internals.safetensors and the hidden_states directory. Prunes a layer, retrains it, and saves it in retrained_layers

Replaced run

Loads the model, replaces the specified layers with retrained versions