From f5bdfcf0fe9d7ac27d4c2e1863bc930aea0387f6 Mon Sep 17 00:00:00 2001
From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com>
Date: Tue, 22 Jul 2025 16:29:24 +0800
Subject: [PATCH] Add loha train impl

---
 comfy/weight_adapter/loha.py | 77 +++++++++++++++++++++++++++++++++++-
 1 file changed, 76 insertions(+), 1 deletion(-)

diff --git a/comfy/weight_adapter/loha.py b/comfy/weight_adapter/loha.py
index ce79abad5..1a73768a7 100644
--- a/comfy/weight_adapter/loha.py
+++ b/comfy/weight_adapter/loha.py
@@ -3,7 +3,63 @@ from typing import Optional
 
 import torch
 import comfy.model_management
-from .base import WeightAdapterBase, weight_decompose
+from .base import WeightAdapterBase, WeightAdapterTrainBase, weight_decompose
+
+
+class LohaDiff(WeightAdapterTrainBase):
+    def __init__(self, weights):
+        super().__init__()
+        # Unpack weights tuple from LoHaAdapter
+        w1a, w1b, alpha, w2a, w2b, t1, t2, dora_scale = weights
+
+        # Create trainable parameters
+        self.w1a = torch.nn.Parameter(w1a)
+        self.w1b = torch.nn.Parameter(w1b)
+        self.w2a = torch.nn.Parameter(w2a)
+        self.w2b = torch.nn.Parameter(w2b)
+
+        self.use_tucker = False
+        if t1 is not None and t2 is not None:
+            self.use_tucker = True
+            self.t1 = torch.nn.Parameter(t1)
+            self.t2 = torch.nn.Parameter(t2)
+        else:
+            # Keep the attributes for consistent access
+            self.t1 = None
+            self.t2 = None
+
+        # Store rank and non-trainable alpha
+        self.rank = w1b.shape[0]
+        self.alpha = torch.nn.Parameter(torch.tensor(alpha), requires_grad=False)
+        # dora_scale is not used in the training forward pass
+
+    def __call__(self, w):
+        org_dtype = w.dtype
+
+        # Reconstruct the two matrices m1 and m2
+        if self.use_tucker:
+            # CP/Tucker decomposition case
+            m1 = torch.einsum('i j k l, j r, i p -> p r k l', self.t1, self.w1b, self.w1a)
+            m2 = torch.einsum('i j k l, j r, i p -> p r k l', self.t2, self.w2b, self.w2a)
+        else:
+            # Standard Hadmard product case
+            m1 = self.w1a @ self.w1b
+            m2 = self.w2a @ self.w2b
+
+        # Calculate the final difference via element-wise product
+        diff = m1 * m2
+
+        # Apply scaling
+        scale = self.alpha / self.rank
+
+        # Add the scaled difference to the original weight
+        weight = w + scale * diff.reshape(w.shape)
+
+        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 LoHaAdapter(WeightAdapterBase):
@@ -13,6 +69,25 @@ class LoHaAdapter(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]
+        in_dim = weight.shape[1:].numel()
+        mat1 = torch.empty(out_dim, rank, device=weight.device, dtype=weight.dtype)
+        mat2 = torch.empty(rank, in_dim, device=weight.device, dtype=weight.dtype)
+        torch.nn.init.kaiming_uniform_(mat1, a=5**0.5)
+        torch.nn.init.constant_(mat2, 0.0)
+        mat3 = torch.empty(out_dim, rank, device=weight.device, dtype=weight.dtype)
+        mat4 = torch.empty(rank, in_dim, device=weight.device, dtype=weight.dtype)
+        torch.nn.init.kaiming_uniform_(mat1, a=5**0.5)
+        torch.nn.init.kaiming_uniform_(mat2, a=5**0.5)
+        return LohaDiff(
+            (mat1, mat2, alpha, mat3, mat4, None, None, None)
+        )
+
+    def to_train(self):
+        return LohaDiff(self.weights)
+
     @classmethod
     def load(
         cls,