Added support for full finetuning flux with randomized param activation. Examples coming soon

toolkit/config_modules.py

@@ -389,6 +389,10 @@ class TrainConfig:
         # will cache a blank prompt or the trigger word, and unload the text encoder to cpu
         # will make training faster and use less vram
         self.unload_text_encoder = kwargs.get('unload_text_encoder', False)
+        # for swapping which parameters are trained during training
+        self.do_paramiter_swapping = kwargs.get('do_paramiter_swapping', False)
+        # 0.1 is 10% of the parameters active at a time lower is less vram, higher is more
+        self.paramiter_swapping_factor = kwargs.get('paramiter_swapping_factor', 0.1)
 
 
 class ModelConfig:
@@ -898,4 +902,16 @@ class GenerateImageConfig:
         if self.logger is None:
             return
 
-        self.logger.log_image(image, count, self.prompt)
+        self.logger.log_image(image, count, self.prompt)
+        
+        
+def validate_configs(
+    train_config: TrainConfig,
+    model_config: ModelConfig,
+    save_config: SaveConfig,
+):
+    if model_config.is_flux:
+        if save_config.save_format != 'diffusers':
+            # make it diffusers
+            save_config.save_format = 'diffusers'
+        

toolkit/models/flux_sage_attn.py

@@ -0,0 +1,94 @@
+from typing import Optional
+from diffusers.models.attention_processor import Attention
+import torch
+import torch.nn.functional as F
+
+
+class FluxSageAttnProcessor2_0:
+    """Attention processor used typically in processing the SD3-like self-attention projections."""
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("FluxAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.FloatTensor:
+        from sageattention import sageattn
+        
+        batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+
+        # `sample` projections.
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
+        if encoder_hidden_states is not None:
+            # `context` projections.
+            encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
+            encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+            encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+
+            encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
+                batch_size, -1, attn.heads, head_dim
+            ).transpose(1, 2)
+            encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
+                batch_size, -1, attn.heads, head_dim
+            ).transpose(1, 2)
+            encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
+                batch_size, -1, attn.heads, head_dim
+            ).transpose(1, 2)
+
+            if attn.norm_added_q is not None:
+                encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
+            if attn.norm_added_k is not None:
+                encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
+
+            # attention
+            query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
+            key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
+            value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
+
+        if image_rotary_emb is not None:
+            from diffusers.models.embeddings import apply_rotary_emb
+
+            query = apply_rotary_emb(query, image_rotary_emb)
+            key = apply_rotary_emb(key, image_rotary_emb)
+
+        hidden_states = sageattn(query, key, value, dropout_p=0.0, is_causal=False)
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        if encoder_hidden_states is not None:
+            encoder_hidden_states, hidden_states = (
+                hidden_states[:, : encoder_hidden_states.shape[1]],
+                hidden_states[:, encoder_hidden_states.shape[1] :],
+            )
+
+            # linear proj
+            hidden_states = attn.to_out[0](hidden_states)
+            # dropout
+            hidden_states = attn.to_out[1](hidden_states)
+            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+            return hidden_states, encoder_hidden_states
+        else:
+            return hidden_states

toolkit/optimizers/adafactor.py

@@ -3,6 +3,7 @@ from typing import List
 import torch
 from toolkit.optimizers.optimizer_utils import copy_stochastic, stochastic_grad_accummulation
 from optimum.quanto import QBytesTensor
+import random
 
 
 class Adafactor(torch.optim.Optimizer):
@@ -105,6 +106,8 @@ class Adafactor(torch.optim.Optimizer):
         scale_parameter=True,
         relative_step=True,
         warmup_init=False,
+        do_paramiter_swapping=False,
+        paramiter_swapping_factor=0.1,
     ):
         if lr is not None and relative_step:
             raise ValueError(
@@ -140,6 +143,49 @@ class Adafactor(torch.optim.Optimizer):
                     param.register_post_accumulate_grad_hook(
                         stochastic_grad_accummulation
                     )
+    
+        self.do_paramiter_swapping = do_paramiter_swapping
+        self.paramiter_swapping_factor = paramiter_swapping_factor
+        self._total_paramiter_size = 0
+        # count total paramiters
+        for group in self.param_groups:
+            for param in group['params']:
+                self._total_paramiter_size += torch.numel(param)
+        # pretty print total paramiters with comma seperation
+        print(f"Total training paramiters: {self._total_paramiter_size:,}")
+        
+        # needs to be enabled to count paramiters
+        if self.do_paramiter_swapping:
+            self.enable_paramiter_swapping(self.paramiter_swapping_factor)
+        
+    
+    def enable_paramiter_swapping(self, paramiter_swapping_factor=0.1):
+        self.do_paramiter_swapping = True
+        self.paramiter_swapping_factor = paramiter_swapping_factor
+        # call it an initial time
+        self.swap_paramiters()
+                    
+    def swap_paramiters(self):
+        all_params = []
+        # deactivate all paramiters
+        for group in self.param_groups:
+            for param in group['params']:
+                param.requires_grad_(False)
+                # remove any grad
+                param.grad = None
+                all_params.append(param)
+        # shuffle all paramiters
+        random.shuffle(all_params)
+        
+        # keep activating paramiters until we are going to go over the target paramiters
+        target_paramiters = int(self._total_paramiter_size * self.paramiter_swapping_factor)
+        total_paramiters = 0
+        for param in all_params:
+            total_paramiters += torch.numel(param)
+            if total_paramiters >= target_paramiters:
+                break
+            else:
+                param.requires_grad_(True)
 
     @staticmethod
     def _get_lr(param_group, param_state):
@@ -209,7 +255,7 @@ class Adafactor(torch.optim.Optimizer):
 
         for group in self.param_groups:
             for p in group["params"]:
-                if p.grad is None:
+                if p.grad is None or not p.requires_grad:
                     continue
 
                 grad = p.grad