[Training Node] algo support, grad acc, optional grad ckpt (#9015)

* Add factorization utils for lokr

* Add lokr train impl

* Add loha train impl

* Add adapter map for algo selection

* Add optional grad ckpt and algo selection

* Update __init__.py

* correct key name for loha

* Use custom fwd/bwd func and better init for loha

* Support gradient accumulation

* Fix bugs of loha

* use more stable init

* Add OFT training

* linting

comfy/weight_adapter/oft.py

@@ -3,7 +3,58 @@ from typing import Optional
 
 import torch
 import comfy.model_management
-from .base import WeightAdapterBase, weight_decompose
+from .base import WeightAdapterBase, WeightAdapterTrainBase, weight_decompose, factorization
+
+
+class OFTDiff(WeightAdapterTrainBase):
+    def __init__(self, weights):
+        super().__init__()
+        # Unpack weights tuple from LoHaAdapter
+        blocks, rescale, alpha, _ = weights
+
+        # Create trainable parameters
+        self.oft_blocks = torch.nn.Parameter(blocks)
+        if rescale is not None:
+            self.rescale = torch.nn.Parameter(rescale)
+            self.rescaled = True
+        else:
+            self.rescaled = False
+        self.block_num, self.block_size, _ = blocks.shape
+        self.constraint = float(alpha)
+        self.alpha = torch.nn.Parameter(torch.tensor(alpha), requires_grad=False)
+
+    def __call__(self, w):
+        org_dtype = w.dtype
+        I = torch.eye(self.block_size, device=self.oft_blocks.device)
+
+        ## generate r
+        # for Q = -Q^T
+        q = self.oft_blocks - self.oft_blocks.transpose(1, 2)
+        normed_q = q
+        if self.constraint:
+            q_norm = torch.norm(q) + 1e-8
+            if q_norm > self.constraint:
+                normed_q = q * self.constraint / q_norm
+        # use float() to prevent unsupported type
+        r = (I + normed_q) @ (I - normed_q).float().inverse()
+
+        ## Apply chunked matmul on weight
+        _, *shape = w.shape
+        org_weight = w.to(dtype=r.dtype)
+        org_weight = org_weight.unflatten(0, (self.block_num, self.block_size))
+        # Init R=0, so add I on it to ensure the output of step0 is original model output
+        weight = torch.einsum(
+            "k n m, k n ... -> k m ...",
+            r,
+            org_weight,
+        ).flatten(0, 1)
+        if self.rescaled:
+            weight = self.rescale * weight
+        return weight.to(org_dtype)
+
+    def passive_memory_usage(self):
+        """Calculates memory usage of the trainable parameters."""
+        return sum(param.numel() * param.element_size() for param in self.parameters())
 
 
 class OFTAdapter(WeightAdapterBase):
@@ -13,6 +64,18 @@ class OFTAdapter(WeightAdapterBase):
         self.loaded_keys = loaded_keys
         self.weights = weights
 
+    @classmethod
+    def create_train(cls, weight, rank=1, alpha=1.0):
+        out_dim = weight.shape[0]
+        block_size, block_num = factorization(out_dim, rank)
+        block = torch.zeros(block_num, block_size, block_size, device=weight.device, dtype=weight.dtype)
+        return OFTDiff(
+            (block, None, alpha, None)
+        )
+
+    def to_train(self):
+        return OFTDiff(self.weights)
+
     @classmethod
     def load(
         cls,
@@ -60,6 +123,8 @@ class OFTAdapter(WeightAdapterBase):
         blocks = v[0]
         rescale = v[1]
         alpha = v[2]
+        if alpha is None:
+            alpha = 0
         dora_scale = v[3]
 
         blocks = comfy.model_management.cast_to_device(blocks, weight.device, intermediate_dtype)