Merge pull request #243 from mgz-dev/dynamic-dim-lora-resize

Enable ability to resize lora dim based off sv ratios

networks/resize_lora.py

@@ -3,12 +3,13 @@
 # Thanks to cloneofsimo and kohya
 
 import argparse
-import os
 import torch
 from safetensors.torch import load_file, save_file, safe_open
 from tqdm import tqdm
 from library import train_util, model_util
+import numpy as np
 
+MIN_SV = 1e-6
 
 def load_state_dict(file_name, dtype):
   if model_util.is_safetensors(file_name):
@@ -38,12 +39,149 @@ def save_to_file(file_name, model, state_dict, dtype, metadata):
     torch.save(model, file_name)
 
 
-def resize_lora_model(lora_sd, new_rank, save_dtype, device, verbose):
+def index_sv_cumulative(S, target):
+  original_sum = float(torch.sum(S))
+  cumulative_sums = torch.cumsum(S, dim=0)/original_sum
+  index = int(torch.searchsorted(cumulative_sums, target)) + 1
+  if index >= len(S):
+    index = len(S) - 1
+
+  return index
+
+
+def index_sv_fro(S, target):
+  S_squared = S.pow(2)
+  s_fro_sq = float(torch.sum(S_squared))
+  sum_S_squared = torch.cumsum(S_squared, dim=0)/s_fro_sq
+  index = int(torch.searchsorted(sum_S_squared, target**2)) + 1
+  if index >= len(S):
+    index = len(S) - 1
+
+  return index
+
+
+# Modified from Kohaku-blueleaf's extract/merge functions
+def extract_conv(weight, lora_rank, dynamic_method, dynamic_param, device, scale=1):
+    out_size, in_size, kernel_size, _ = weight.size()
+    U, S, Vh = torch.linalg.svd(weight.reshape(out_size, -1).to(device))
+
+    param_dict = rank_resize(S, lora_rank, dynamic_method, dynamic_param, scale)
+    lora_rank = param_dict["new_rank"]
+
+    U = U[:, :lora_rank]
+    S = S[:lora_rank]
+    U = U @ torch.diag(S)
+    Vh = Vh[:lora_rank, :]
+
+    param_dict["lora_down"] = Vh.reshape(lora_rank, in_size, kernel_size, kernel_size).cpu()
+    param_dict["lora_up"] = U.reshape(out_size, lora_rank, 1, 1).cpu()
+    del U, S, Vh, weight
+    return param_dict
+
+
+def extract_linear(weight, lora_rank, dynamic_method, dynamic_param, device, scale=1):
+    out_size, in_size = weight.size()
+
+    U, S, Vh = torch.linalg.svd(weight.to(device))
+
+    param_dict = rank_resize(S, lora_rank, dynamic_method, dynamic_param, scale)
+    lora_rank = param_dict["new_rank"]
+
+    U = U[:, :lora_rank]
+    S = S[:lora_rank]
+    U = U @ torch.diag(S)
+    Vh = Vh[:lora_rank, :]
+
+    param_dict["lora_down"] = Vh.reshape(lora_rank, in_size).cpu()
+    param_dict["lora_up"] = U.reshape(out_size, lora_rank).cpu()
+    del U, S, Vh, weight
+    return param_dict
+
+
+def merge_conv(lora_down, lora_up, device):
+    in_rank, in_size, kernel_size, k_ = lora_down.shape
+    out_size, out_rank, _, _ = lora_up.shape
+    assert in_rank == out_rank and kernel_size == k_, f"rank {in_rank} {out_rank} or kernel {kernel_size} {k_} mismatch"
+
+    lora_down = lora_down.to(device)
+    lora_up = lora_up.to(device)
+
+    merged = lora_up.reshape(out_size, -1) @ lora_down.reshape(in_rank, -1)
+    weight = merged.reshape(out_size, in_size, kernel_size, kernel_size)
+    del lora_up, lora_down
+    return weight
+
+
+def merge_linear(lora_down, lora_up, device):
+    in_rank, in_size = lora_down.shape
+    out_size, out_rank = lora_up.shape
+    assert in_rank == out_rank, f"rank {in_rank} {out_rank} mismatch"
+
+    lora_down = lora_down.to(device)
+    lora_up = lora_up.to(device)
+
+    weight = lora_up @ lora_down
+    del lora_up, lora_down
+    return weight
+
+
+def rank_resize(S, rank, dynamic_method, dynamic_param, scale=1):
+    param_dict = {}
+
+    if dynamic_method=="sv_ratio":
+        # Calculate new dim and alpha based off ratio
+        max_sv = S[0]
+        min_sv = max_sv/dynamic_param
+        new_rank = max(torch.sum(S > min_sv).item(),1)
+        new_alpha = float(scale*new_rank)
+
+    elif dynamic_method=="sv_cumulative":
+        # Calculate new dim and alpha based off cumulative sum
+        new_rank = index_sv_cumulative(S, dynamic_param)
+        new_rank = max(new_rank, 1)
+        new_alpha = float(scale*new_rank)
+
+    elif dynamic_method=="sv_fro":
+        # Calculate new dim and alpha based off sqrt sum of squares
+        new_rank = index_sv_fro(S, dynamic_param)
+        new_rank = min(max(new_rank, 1), len(S)-1)
+        new_alpha = float(scale*new_rank)
+    else:
+        new_rank = rank
+        new_alpha = float(scale*new_rank)
+
+
+    if S[0] <= MIN_SV: # Zero matrix, set dim to 1
+        new_rank = 1
+        new_alpha = float(scale*new_rank)
+    elif new_rank > rank: # cap max rank at rank
+        new_rank = rank
+        new_alpha = float(scale*new_rank)
+
+
+    # Calculate resize info
+    s_sum = torch.sum(torch.abs(S))
+    s_rank = torch.sum(torch.abs(S[:new_rank]))
+
+    S_squared = S.pow(2)
+    s_fro = torch.sqrt(torch.sum(S_squared))
+    s_red_fro = torch.sqrt(torch.sum(S_squared[:new_rank]))
+    fro_percent = float(s_red_fro/s_fro)
+
+    param_dict["new_rank"] = new_rank
+    param_dict["new_alpha"] = new_alpha
+    param_dict["sum_retained"] = (s_rank)/s_sum
+    param_dict["fro_retained"] = fro_percent
+    param_dict["max_ratio"] = S[0]/S[new_rank]
+
+    return param_dict
+
+
+def resize_lora_model(lora_sd, new_rank, save_dtype, device, dynamic_method, dynamic_param, verbose):
   network_alpha = None
   network_dim = None
   verbose_str = "\n"
-
-  CLAMP_QUANTILE = 0.99
+  fro_list = []
 
   # Extract loaded lora dim and alpha
   for key, value in lora_sd.items():
@@ -57,9 +195,9 @@ def resize_lora_model(lora_sd, new_rank, save_dtype, device, verbose):
       network_alpha = network_dim
 
   scale = network_alpha/network_dim
-  new_alpha = float(scale*new_rank)  # calculate new alpha from scale
 
-  print(f"old dimension: {network_dim}, old alpha: {network_alpha}, new alpha: {new_alpha}")
+  if dynamic_method:
+    print(f"Dynamically determining new alphas and dims based off {dynamic_method}: {dynamic_param}, max rank is {new_rank}")
 
   lora_down_weight = None
   lora_up_weight = None
@@ -68,7 +206,6 @@ def resize_lora_model(lora_sd, new_rank, save_dtype, device, verbose):
   block_down_name = None
   block_up_name = None
 
-  print("resizing lora...")
   with torch.no_grad():
     for key, value in tqdm(lora_sd.items()):
       if 'lora_down' in key:
@@ -85,57 +222,43 @@ def resize_lora_model(lora_sd, new_rank, save_dtype, device, verbose):
         conv2d = (len(lora_down_weight.size()) == 4)
 
         if conv2d:
-          lora_down_weight = lora_down_weight.squeeze()
-          lora_up_weight = lora_up_weight.squeeze()
-
-        if device:
-          org_device = lora_up_weight.device
-          lora_up_weight = lora_up_weight.to(args.device)
-          lora_down_weight = lora_down_weight.to(args.device)
-
-        full_weight_matrix = torch.matmul(lora_up_weight, lora_down_weight)
-
-        U, S, Vh = torch.linalg.svd(full_weight_matrix)
+          full_weight_matrix = merge_conv(lora_down_weight, lora_up_weight, device)
+          param_dict = extract_conv(full_weight_matrix, new_rank, dynamic_method, dynamic_param, device, scale)
+        else:
+          full_weight_matrix = merge_linear(lora_down_weight, lora_up_weight, device)
+          param_dict = extract_linear(full_weight_matrix, new_rank, dynamic_method, dynamic_param, device, scale)
 
         if verbose:
-          s_sum = torch.sum(torch.abs(S))
-          s_rank = torch.sum(torch.abs(S[:new_rank]))
-          verbose_str+=f"{block_down_name:76} | "
-          verbose_str+=f"sum(S) retained: {(s_rank)/s_sum:.1%}, max(S) ratio: {S[0]/S[new_rank]:0.1f}\n"
+          max_ratio = param_dict['max_ratio']
+          sum_retained = param_dict['sum_retained']
+          fro_retained = param_dict['fro_retained']
+          if not np.isnan(fro_retained):
+            fro_list.append(float(fro_retained))
 
-        U = U[:, :new_rank]
-        S = S[:new_rank]
-        U = U @ torch.diag(S)
+          verbose_str+=f"{block_down_name:75} | "
+          verbose_str+=f"sum(S) retained: {sum_retained:.1%}, fro retained: {fro_retained:.1%}, max(S) ratio: {max_ratio:0.1f}"
 
-        Vh = Vh[:new_rank, :]
+        if verbose and dynamic_method:
+          verbose_str+=f", dynamic | dim: {param_dict['new_rank']}, alpha: {param_dict['new_alpha']}\n"
+        else:
+          verbose_str+=f"\n"
 
-        dist = torch.cat([U.flatten(), Vh.flatten()])
-        hi_val = torch.quantile(dist, CLAMP_QUANTILE)
-        low_val = -hi_val
-
-        U = U.clamp(low_val, hi_val)
-        Vh = Vh.clamp(low_val, hi_val)
-
-        if conv2d:
-          U = U.unsqueeze(2).unsqueeze(3)
-          Vh = Vh.unsqueeze(2).unsqueeze(3)
-
-        if device:
-          U = U.to(org_device)
-          Vh = Vh.to(org_device)
-
-        o_lora_sd[block_down_name + "." + "lora_down.weight"] = Vh.to(save_dtype).contiguous()
-        o_lora_sd[block_up_name + "." + "lora_up.weight"] = U.to(save_dtype).contiguous()
-        o_lora_sd[block_up_name + "." "alpha"] = torch.tensor(new_alpha).to(save_dtype)
+        new_alpha = param_dict['new_alpha']
+        o_lora_sd[block_down_name + "." + "lora_down.weight"] = param_dict["lora_down"].to(save_dtype).contiguous()
+        o_lora_sd[block_up_name + "." + "lora_up.weight"] = param_dict["lora_up"].to(save_dtype).contiguous()
+        o_lora_sd[block_up_name + "." "alpha"] = torch.tensor(param_dict['new_alpha']).to(save_dtype)
 
         block_down_name = None
         block_up_name = None
         lora_down_weight = None
         lora_up_weight = None
         weights_loaded = False
+        del param_dict
 
   if verbose:
     print(verbose_str)
+
+    print(f"Average Frobenius norm retention: {np.mean(fro_list):.2%} | std: {np.std(fro_list):0.3f}")
   print("resizing complete")
   return o_lora_sd, network_dim, new_alpha
 
@@ -151,6 +274,9 @@ def str_to_dtype(p):
       return torch.bfloat16
     return None
 
+  if args.dynamic_method and not args.dynamic_param:
+    raise Exception("If using dynamic_method, then dynamic_param is required")
+
   merge_dtype = str_to_dtype('float')  # matmul method above only seems to work in float32
   save_dtype = str_to_dtype(args.save_precision)
   if save_dtype is None:
@@ -159,17 +285,23 @@ def str_to_dtype(p):
   print("loading Model...")
   lora_sd, metadata = load_state_dict(args.model, merge_dtype)
 
-  print("resizing rank...")
-  state_dict, old_dim, new_alpha = resize_lora_model(lora_sd, args.new_rank, save_dtype, args.device, args.verbose)
+  print("Resizing Lora...")
+  state_dict, old_dim, new_alpha = resize_lora_model(lora_sd, args.new_rank, save_dtype, args.device, args.dynamic_method, args.dynamic_param, args.verbose)
 
   # update metadata
   if metadata is None:
     metadata = {}
 
   comment = metadata.get("ss_training_comment", "")
-  metadata["ss_training_comment"] = f"dimension is resized from {old_dim} to {args.new_rank}; {comment}"
-  metadata["ss_network_dim"] = str(args.new_rank)
-  metadata["ss_network_alpha"] = str(new_alpha)
+
+  if not args.dynamic_method:
+    metadata["ss_training_comment"] = f"dimension is resized from {old_dim} to {args.new_rank}; {comment}"
+    metadata["ss_network_dim"] = str(args.new_rank)
+    metadata["ss_network_alpha"] = str(new_alpha)
+  else:
+    metadata["ss_training_comment"] = f"Dynamic resize with {args.dynamic_method}: {args.dynamic_param} from {old_dim}; {comment}"
+    metadata["ss_network_dim"] = 'Dynamic'
+    metadata["ss_network_alpha"] = 'Dynamic'
 
   model_hash, legacy_hash = train_util.precalculate_safetensors_hashes(state_dict, metadata)
   metadata["sshs_model_hash"] = model_hash
@@ -193,6 +325,11 @@ def str_to_dtype(p):
   parser.add_argument("--device", type=str, default=None, help="device to use, cuda for GPU / 計算を行うデバイス、cuda でGPUを使う")
   parser.add_argument("--verbose", action="store_true", 
                       help="Display verbose resizing information / rank変更時の詳細情報を出力する")
+  parser.add_argument("--dynamic_method", type=str, default=None, choices=[None, "sv_ratio", "sv_fro", "sv_cumulative"],
+                      help="Specify dynamic resizing method, --new_rank is used as a hard limit for max rank")
+  parser.add_argument("--dynamic_param", type=float, default=None,
+                      help="Specify target for dynamic reduction")
+
 
   args = parser.parse_args()
   resize(args)