Fix models getting stuck in memory when merging

expand on functionality and reduce restrictions when extracting LoRA

nodes/lora_extract_svd.py

@@ -46,14 +46,14 @@ def _index_sv_ratio(S: torch.Tensor, ratio: float) -> int:
 
 def _index_sv_cumulative(S: torch.Tensor, target: float, max_rank: int = None) -> int:
     """Cumulative mode - keep enough SVs to reach target % of total.
-    
+
     Calculates relative to max_rank if provided, otherwise relative to full.
     """
     if max_rank is not None and max_rank < len(S):
         total = torch.sum(S[:max_rank])
     else:
         total = torch.sum(S)
-    
+
     if total < 1e-8:
         return 1
     cumsum = torch.cumsum(S, dim=0) / total
@@ -63,7 +63,7 @@ def _index_sv_cumulative(S: torch.Tensor, target: float, max_rank: int = None) -
 
 def _index_sv_fro(S: torch.Tensor, target: float, max_rank: int = None) -> int:
     """Frobenius norm mode - preserve target fraction of Frobenius norm.
-    
+
     Calculates relative to max_rank if provided, otherwise relative to full.
     This means "retain target% of what's achievable within max_rank".
     """
@@ -75,10 +75,10 @@ def _index_sv_fro(S: torch.Tensor, target: float, max_rank: int = None) -> int:
     else:
         S_sq = S.pow(2)
         total_sq = torch.sum(S_sq)
-    
+
     if total_sq < 1e-8:
         return 1
-    
+
     # Cumsum of all S (not capped) to find where we reach target
     cumsum = torch.cumsum(S.pow(2), dim=0) / total_sq
     rank = int(torch.searchsorted(cumsum, target ** 2).item()) + 1
@@ -172,27 +172,27 @@ def _svd_extract_linear_lowrank(
 ) -> tuple:
     """
     Low-rank SVD decomposition using svd_lowrank for fixed rank extraction.
-    
+
     Much faster than full SVD when rank << min(m, n) because it only computes
     the top-k singular values using randomized algorithms.
-    
+
     Args:
         niter: Power iterations for SVD accuracy (higher = more accurate but slower)
-    
+
     Returns:
         (lora_down, lora_up, diff), "low rank" or (weight, "full")
     """
     weight = weight.to(device=device, dtype=torch.float32)
     out_dim, in_dim = weight.shape
-    
+
     # Clamp rank to valid range
     max_possible_rank = min(out_dim, in_dim)
-    rank = max(1, min(rank, max_possible_rank - 1))
-    
+    rank = max(1, min(rank, max_possible_rank))
+
     # Check if decomposition is worthwhile
-    if rank >= max_possible_rank // 2:
+    if rank >= max_possible_rank:
         return weight, "full"
-    
+
     # Use svd_lowrank for faster low-rank approximation
     try:
         U, S, V = torch.svd_lowrank(weight, q=rank, niter=niter)
@@ -200,7 +200,7 @@ def _svd_extract_linear_lowrank(
         Vh = V.T  # [rank, in_dim]
     except Exception as e:
         raise RuntimeError(f"svd_lowrank failed: {e}")
-    
+
     # Clamp outliers
     if clamp_quantile < 1.0 and len(S) > 0:
         try:
@@ -208,14 +208,14 @@ def _svd_extract_linear_lowrank(
             S = S.clamp(max=max_val)
         except RuntimeError:
             pass
-    
+
     # Construct LoRA matrices
     lora_up = U @ torch.diag(S)  # [out_dim, rank]
     lora_down = Vh               # [rank, in_dim]
-    
+
     # Compute reconstruction diff
     diff = weight - (lora_up @ lora_down)
-    
+
     return (lora_down, lora_up, diff), "low rank"
 
 
@@ -239,7 +239,7 @@ def _svd_extract_linear(
     """
     weight = weight.to(device=device, dtype=torch.float32)
     out_dim, in_dim = weight.shape
-    
+
     # For fixed rank, use the faster svd_lowrank
     if mode == "fixed" and mode_param > 0:
         target_rank = int(mode_param)
@@ -257,7 +257,7 @@ def _svd_extract_linear(
     rank = _compute_rank(S, mode, mode_param, max_rank)
 
     # Check if decomposition is worthwhile
-    if rank >= min(out_dim, in_dim) // 2:
+    if rank >= min(out_dim, in_dim):
         return weight, "full"
 
     # Truncate to rank
@@ -303,7 +303,7 @@ def _svd_extract_conv(
 
     # Flatten for SVD
     if is_1x1:
-        mat = weight.squeeze()  # [out_ch, in_ch]
+        mat = weight.view(out_ch, in_ch)  # [out_ch, in_ch]
     else:
         mat = weight.reshape(out_ch, -1)  # [out_ch, in_ch*k*k]
 
@@ -382,8 +382,11 @@ def _extract_chunked_layer(
             return None, None, 0
 
     # Combine chunks
+    # Note: Using the first chunk's down matrix as the basis for all chunks
+    # is a better approximation than a simple mean if the chunks share the same input space (like QKV).
+    # Ideally, we would use a more sophisticated joint SVD, but this is a reasonable fallback.
     combined_lora_up = torch.cat(all_lora_up, dim=0)
-    combined_lora_down = torch.stack(all_lora_down, dim=0).mean(dim=0)
+    combined_lora_down = all_lora_down[0]
     rank = combined_lora_down.shape[0]
 
     return combined_lora_up, combined_lora_down, rank
@@ -476,39 +479,40 @@ def extract_lora_from_files(
     try:
         keys_a = set(handler_a.keys())
         keys_b = set(handler_b.keys())
-
         weight_keys = [k for k in keys_a if k.endswith(".weight")]
-
         pbar = comfy.utils.ProgressBar(len(weight_keys))
-
         stats = {"extracted": 0, "full": 0, "skipped": 0, "chunked": 0}
 
-        for key in tqdm(weight_keys, desc="Extracting LoRA", unit="layers"):
+        def _process_layer(key):
             lora_name = _format_lora_key(key)
 
             if _matches_any_pattern(key, skip_patterns):
-                stats["skipped"] += 1
-                pbar.update(1)
-                continue
+                return "skipped", None
 
             # Load tensors with pinned memory for CUDA
             use_pinned = device == 'cuda'
-            
+
             if key not in keys_b:
                 if mismatch_mode == "skip":
-                    stats["skipped"] += 1
-                    pbar.update(1)
-                    continue
+                    return "skipped", None
+                if mismatch_mode == "error":
+                    raise ValueError(f"Key {key} not found in model B")
+
                 cpu_a = handler_a.get_tensor(key)
                 if use_pinned:
                     weight_diff = transfer_to_gpu_pinned(cpu_a, device, torch.float32)
                 else:
                     weight_diff = cpu_a.to(device=device, dtype=torch.float32)
                 del cpu_a
+
+                if mismatch_mode == "zeros":
+                    # For zeros mode, we treat the missing base as a zeroed tensor of same shape
+                    # which is already captured by weight_diff = tensor_a
+                    pass
             else:
                 cpu_a = handler_a.get_tensor(key)
                 cpu_b = handler_b.get_tensor(key)
-                
+
                 if use_pinned:
                     tensor_a = transfer_to_gpu_pinned(cpu_a, device, torch.float32)
                     tensor_b = transfer_to_gpu_pinned(cpu_b, device, torch.float32)
@@ -519,10 +523,12 @@ def extract_lora_from_files(
 
                 if tensor_a.shape != tensor_b.shape:
                     if mismatch_mode == "skip":
-                        stats["skipped"] += 1
-                        pbar.update(1)
                         del tensor_a, tensor_b
-                        continue
+                        return "skipped", None
+                    if mismatch_mode == "error":
+                        raise ValueError(f"Shape mismatch for {key}: {tensor_a.shape} vs {tensor_b.shape}")
+
+                    # For zeros/fallback, use tensor_a as the difference
                     weight_diff = tensor_a
                     del tensor_b
                 else:
@@ -531,19 +537,16 @@ def extract_lora_from_files(
 
             # Skip small differences
             if min_diff > 0 and weight_diff.abs().max() < min_diff:
-                stats["skipped"] += 1
                 del weight_diff
-                pbar.update(1)
-                continue
+                return "skipped", None
 
             # Skip 1D tensors
             if weight_diff.ndim < 2:
-                stats["skipped"] += 1
                 del weight_diff
-                pbar.update(1)
-                continue
+                return "skipped", None
 
             is_conv = weight_diff.ndim == 4
+            layer_results = {}
 
             try:
                 if is_conv:
@@ -564,32 +567,36 @@ def extract_lora_from_files(
                             weight_diff, num_chunks, mode, linear_param, device, linear_max_rank
                         )
                         if lora_up is not None:
-                            output_sd[f"{lora_name}.lora_up.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
-                            output_sd[f"{lora_name}.lora_down.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
-                            output_sd[f"{lora_name}.alpha"] = torch.tensor(rank).to(save_torch_dtype)
-                            stats["chunked"] += 1
+                            layer_results[f"{lora_name}.lora_up.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
+                            layer_results[f"{lora_name}.lora_down.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
+                            layer_results[f"{lora_name}.alpha"] = torch.tensor(rank).to(save_torch_dtype)
                             del weight_diff
-                            pbar.update(1)
-                            continue
+                            return "chunked", layer_results
 
                 print(f"[LoRA Extract] Failed: {key}: {e}")
-                stats["skipped"] += 1
                 del weight_diff
-                pbar.update(1)
-                continue
+                return "skipped", None
 
             # Store result
             if mode_str == "full":
-                output_sd[f"{lora_name}.diff"] = weight_diff.to(save_torch_dtype).cpu().contiguous()
-                stats["full"] += 1
+                layer_results[f"{lora_name}.diff"] = weight_diff.to(save_torch_dtype).cpu().contiguous()
+                status = "full"
             else:
                 lora_down, lora_up, _ = result
-                output_sd[f"{lora_name}.lora_down.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
-                output_sd[f"{lora_name}.lora_up.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
-                output_sd[f"{lora_name}.alpha"] = torch.tensor(lora_down.shape[0]).to(save_torch_dtype)
-                stats["extracted"] += 1
+                layer_results[f"{lora_name}.lora_down.weight"] = lora_down.to(save_torch_dtype).cpu().contiguous()
+                layer_results[f"{lora_name}.lora_up.weight"] = lora_up.to(save_torch_dtype).cpu().contiguous()
+                layer_results[f"{lora_name}.alpha"] = torch.tensor(lora_down.shape[0]).to(save_torch_dtype)
+                status = "extracted"
 
             del weight_diff
+            return status, layer_results
+
+        for key in tqdm(weight_keys, desc="Extracting LoRA", unit="layers"):
+            status, layer_sd = _process_layer(key)
+            stats[status] += 1
+            if layer_sd:
+                output_sd.update(layer_sd)
+
             if force_clear_cache:
                 import gc
                 gc.collect()
@@ -706,7 +713,7 @@ def _format_lora_key(key: str) -> str:
     # Handle already correct prefixes
     if key.startswith("diffusion_model.") or key.startswith("transformer."):
         return key
-    
+
     # Handle known Unet blocks (Diffusers format UNet)
     if any(key.startswith(p) for p in ["down_blocks", "up_blocks", "mid_block", "conv_in", "conv_out", "time_embedding", "class_embedding"]):
         return f"diffusion_model.{key}"
@@ -746,9 +753,9 @@ class LoRAExtractFixed(io.ComfyNode):
             description="Extract LoRA with specified fixed rank for each layer type.",
             inputs=[
                 *_get_model_inputs(),
-                io.Int.Input("linear_dim", default=64, min=1, max=4096,
+                io.Int.Input("linear_dim", default=64, min=1, max=16384,
                             tooltip="Rank for linear/attention layers"),
-                io.Int.Input("conv_dim", default=32, min=1, max=4096,
+                io.Int.Input("conv_dim", default=32, min=1, max=16384,
                             tooltip="Rank for conv layers"),
                 io.Int.Input("svd_niter", default=2, min=0, max=10,
                             tooltip="SVD power iterations (higher = more accurate but slower)"),
@@ -768,7 +775,7 @@ class LoRAExtractFixed(io.ComfyNode):
 
         output_sd = extract_lora_from_files(
             model_a_path, model_b_path, "fixed", linear_dim, conv_dim,
-            "lora_unet_", device, save_dtype, linear_dim, conv_dim,
+            device, save_dtype, linear_dim, conv_dim,
             clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers, svd_niter,
             lazy_load, force_clear_cache
         )
@@ -792,9 +799,9 @@ class LoRAExtractRatio(io.ComfyNode):
                               tooltip="Ratio threshold for linear layers (higher = more SVs kept)"),
                 io.Float.Input("conv_ratio", default=2.0, min=1.0, max=100.0, step=0.1,
                               tooltip="Ratio threshold for conv layers (higher = more SVs kept)"),
-                io.Int.Input("linear_max_rank", default=128, min=1, max=3072,
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for linear layers"),
-                io.Int.Input("conv_max_rank", default=128, min=1, max=3072,
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for conv layers"),
                 *_get_common_inputs(),
             ],
@@ -812,7 +819,7 @@ class LoRAExtractRatio(io.ComfyNode):
 
         output_sd = extract_lora_from_files(
             model_a_path, model_b_path, "ratio", linear_ratio, conv_ratio,
-            "lora_unet_", device, save_dtype, linear_max_rank, conv_max_rank,
+            device, save_dtype, linear_max_rank, conv_max_rank,
             clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
             lazy_load=lazy_load, force_clear_cache=force_clear_cache
         )
@@ -836,9 +843,9 @@ class LoRAExtractQuantile(io.ComfyNode):
                               tooltip="Target cumulative % for linear layers"),
                 io.Float.Input("conv_quantile", default=0.9, min=0.0, max=1.0, step=0.01,
                               tooltip="Target cumulative % for conv layers"),
-                io.Int.Input("linear_max_rank", default=128, min=1, max=3072,
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for linear layers"),
-                io.Int.Input("conv_max_rank", default=128, min=1, max=3072,
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for conv layers"),
                 *_get_common_inputs(),
             ],
@@ -856,7 +863,7 @@ class LoRAExtractQuantile(io.ComfyNode):
 
         output_sd = extract_lora_from_files(
             model_a_path, model_b_path, "quantile", linear_quantile, conv_quantile,
-            "lora_unet_", device, save_dtype, linear_max_rank, conv_max_rank,
+            device, save_dtype, linear_max_rank, conv_max_rank,
             clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
             lazy_load=lazy_load, force_clear_cache=force_clear_cache
         )
@@ -878,9 +885,9 @@ class LoRAExtractKnee(io.ComfyNode):
                 *_get_model_inputs(),
                 io.Combo.Input("knee_method", options=["sv_knee", "sv_cumulative_knee"],
                               default="sv_knee", tooltip="Knee detection method"),
-                io.Int.Input("linear_max_rank", default=128, min=1, max=3072,
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for linear layers"),
-                io.Int.Input("conv_max_rank", default=128, min=1, max=3072,
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for conv layers"),
                 *_get_common_inputs(),
             ],
@@ -898,7 +905,7 @@ class LoRAExtractKnee(io.ComfyNode):
 
         output_sd = extract_lora_from_files(
             model_a_path, model_b_path, knee_method, 0, 0,
-            "lora_unet_", device, save_dtype, linear_max_rank, conv_max_rank,
+            device, save_dtype, linear_max_rank, conv_max_rank,
             clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
             lazy_load=lazy_load, force_clear_cache=force_clear_cache
         )
@@ -922,9 +929,9 @@ class LoRAExtractFrobenius(io.ComfyNode):
                               tooltip="Target Frobenius norm fraction for linear"),
                 io.Float.Input("conv_target", default=0.9, min=0.0, max=1.0, step=0.01,
                               tooltip="Target Frobenius norm fraction for conv"),
-                io.Int.Input("linear_max_rank", default=128, min=1, max=3072,
+                io.Int.Input("linear_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for linear layers"),
-                io.Int.Input("conv_max_rank", default=128, min=1, max=3072,
+                io.Int.Input("conv_max_rank", default=128, min=1, max=16384,
                             tooltip="Maximum rank for conv layers"),
                 *_get_common_inputs(),
             ],
@@ -942,7 +949,7 @@ class LoRAExtractFrobenius(io.ComfyNode):
 
         output_sd = extract_lora_from_files(
             model_a_path, model_b_path, "sv_fro", linear_target, conv_target,
-            "lora_unet_", device, save_dtype, linear_max_rank, conv_max_rank,
+            device, save_dtype, linear_max_rank, conv_max_rank,
             clamp_quantile, min_diff, skip_patterns, mismatch_mode, chunk_large_layers,
             lazy_load=lazy_load, force_clear_cache=force_clear_cache
         )

nodes/lora_merger.py

@@ -13,8 +13,8 @@ from typing import List, Dict, Tuple, Optional
 from .merger_utils import MemoryEfficientSafeOpen, transfer_to_gpu_pinned
 from .device_utils import estimate_model_size, prepare_for_large_operation, cleanup_after_operation
 from .lora_resize import (
-    detect_lora_format, 
-    extract_lora_pairs, 
+    detect_lora_format,
+    extract_lora_pairs,
     detect_lora_rank,
     _compile_patterns,
     _matches_any_pattern,
@@ -32,7 +32,7 @@ def merge_multi_loras(
 ) -> str:
     """
     Merge multiple LoRAs into a single LoRA file.
-    
+
     Modes:
     - concatenate: Mathematically sound merge by stacking ranks.
                    New rank = sum(input ranks). Alpha set to new rank.
@@ -45,22 +45,22 @@ def merge_multi_loras(
     if verbose:
         print(f"[LoRA Multi-Merge] Preparing memory for {total_size_gb:.2f}GB operation...")
         print(f"[LoRA Multi-Merge] Mode: {merge_mode}")
-    
+
     prepare_for_large_operation(total_size_gb * 2.5, torch.device(device))
 
-    handlers = [MemoryEfficientSafeOpen(p) for p in lora_paths]
-    
+    handlers = [MemoryEfficientSafeOpen(p, low_memory=True) for p in lora_paths]
+
     try:
         # 1. Analyze all LoRAs
         layer_map = {}
         lora_infos = []
-        
+
         for i, handler in enumerate(handlers):
             keys = handler.keys()
             fmt = detect_lora_format(keys)
             pairs = extract_lora_pairs(keys, fmt)
             rank, alpha = detect_lora_rank(handler, pairs)
-            
+
             info = {
                 "name": os.path.basename(lora_paths[i]),
                 "format": fmt,
@@ -71,10 +71,10 @@ def merge_multi_loras(
                 "scale": alpha / rank if rank > 0 else 1.0
             }
             lora_infos.append(info)
-            
+
             if verbose:
                 print(f"[LoRA Multi-Merge] LoRA {i+1} ({info['name']}): {len(pairs)} layers, dim={rank}, format={fmt['format']}")
-            
+
             for block_name in pairs.keys():
                 core = _format_lora_key(block_name)
                 if core not in layer_map:
@@ -94,37 +94,37 @@ def merge_multi_loras(
                 ups = []
                 is_conv = False
                 max_rank = 0
-                
+
                 # First pass: load and determine max rank
                 for idx in indices:
                     info = lora_infos[idx]
                     handler = handlers[idx]
                     orig_block = next(b for b in info["pairs"].keys() if _format_lora_key(b) == core)
                     block_keys = info["pairs"][orig_block]
-                    
+
                     if "down" not in block_keys or "up" not in block_keys:
                         continue
-                        
+
                     t_down = handler.get_tensor(block_keys["down"])
                     t_up = handler.get_tensor(block_keys["up"])
-                    
+
                     if device == 'cuda':
                         t_down = transfer_to_gpu_pinned(t_down, device, torch.float32)
                         t_up = transfer_to_gpu_pinned(t_up, device, torch.float32)
                     else:
                         t_down = t_down.to(device=device, dtype=torch.float32)
                         t_up = t_up.to(device=device, dtype=torch.float32)
-                    
+
                     # Store with original info for padding/weighting
                     is_conv = len(t_down.shape) == 4
                     current_rank = t_down.shape[0]
                     max_rank = max(max_rank, current_rank)
-                    
+
                     # Apply scale immediately for concatenate mode, or keep for later
                     if merge_mode == "concatenate":
                         effective_weight = info["weight"] * info["scale"]
                         t_up = t_up * effective_weight
-                    
+
                     downs.append((t_down, info))
                     ups.append((t_up, info))
 
@@ -148,10 +148,10 @@ def merge_multi_loras(
                     target_down_shape[0] = max_rank
                     target_up_shape = list(ups[0][0].shape)
                     target_up_shape[1] = max_rank
-                    
+
                     merged_down = torch.zeros(target_down_shape, device=device, dtype=torch.float32)
                     merged_up = torch.zeros(target_up_shape, device=device, dtype=torch.float32)
-                    
+
                     for (t_d, info_d), (t_u, info_u) in zip(downs, ups):
                         r = t_d.shape[0]
                         w = info_d["weight"]
@@ -159,7 +159,7 @@ def merge_multi_loras(
                         # Kohya just uses weights. But to be safe with different alphas,
                         # we apply scale to the final delta.
                         # For direct weight merge, we'll just use weights.
-                        
+
                         if r < max_rank:
                             # Pad down: [r, in...] -> [max_rank, in...]
                             pad_d = [0] * (len(t_d.shape) * 2)
@@ -168,15 +168,15 @@ def merge_multi_loras(
                             # For [r, in], padding is (0,0, 0, max_rank-r)
                             padding_d = [0, 0] * (len(t_d.shape) - 1) + [0, max_rank - r]
                             t_d = torch.nn.functional.pad(t_d, tuple(padding_d))
-                            
+
                             # Pad up: [out, r...] -> [out, max_rank...]
                             # For [out, r], padding is (0, max_rank-r, 0, 0)
                             padding_u = [0, max_rank - r] + [0, 0] * (len(t_u.shape) - 1)
                             t_u = torch.nn.functional.pad(t_u, tuple(padding_u))
-                        
+
                         merged_down += w * t_d
                         merged_up += w * t_u
-                    
+
                     new_rank = max_rank
                     # Alpha is usually max of alphas or same as rank
                     new_alpha = float(max_rank)
@@ -185,12 +185,18 @@ def merge_multi_loras(
                 output_sd[f"{core}.lora_down.weight"] = merged_down.to(save_dtype).cpu().contiguous()
                 output_sd[f"{core}.lora_up.weight"] = merged_up.to(save_dtype).cpu().contiguous()
                 output_sd[f"{core}.alpha"] = torch.tensor(new_alpha, dtype=save_dtype)
-                
+
                 # Cleanup
                 del merged_down, merged_up
                 for d, _ in downs: del d
                 for u, _ in ups: del u
-                
+                downs.clear()
+                ups.clear()
+                import gc
+                gc.collect()
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+
                 pbar.update(1)
 
         # 3. Final Summary
@@ -199,7 +205,7 @@ def merge_multi_loras(
             for indices in layer_map.values():
                 for idx in indices:
                     matched_stats[idx] += 1
-            
+
             print(f"[LoRA Multi-Merge] --- Merge Summary ---")
             for i, info in enumerate(lora_infos):
                 print(f"[LoRA Multi-Merge] LoRA {i+1}: {matched_stats[i]}/{len(info['pairs'])} layers used in merge")
@@ -210,16 +216,16 @@ def merge_multi_loras(
             "ss_training_comment": f"Merged {len(lora_paths)} LoRAs via concatenation",
             "ss_network_module": "networks.lora",
         }
-        
+
         # 4. Save
         output_dir = folder_paths.get_folder_paths("loras")[0]
         os.makedirs(output_dir, exist_ok=True)
         output_path = os.path.join(output_dir, f"{output_filename.strip()}.safetensors")
-        
+
         save_file(output_sd, output_path, metadata)
         if verbose:
             print(f"[LoRA Multi-Merge] Saved merged LoRA to {output_path}")
-            
+
         return output_path
 
     finally:
@@ -265,7 +271,7 @@ class LoRAMultiMerge(io.ComfyNode):
                 # LoRA 8
                 io.Combo.Input("lora_8", options=["None"] + folder_paths.get_filename_list("loras"), default="None"),
                 io.Float.Input("weight_8", default=1.0, min=-10.0, max=10.0, step=0.01),
-                
+
                 io.Combo.Input("merge_mode", options=["concatenate", "weighted_sum"], default="concatenate",
                               tooltip="concatenate: safe, increases rank. weighted_sum: fixed rank (to max input rank), mathematically lossy but standard in Kohya."),
                 io.String.Input("output_filename", default="merged_lora"),
@@ -281,25 +287,25 @@ class LoRAMultiMerge(io.ComfyNode):
                 lora_1, weight_1, lora_2, weight_2, lora_3, weight_3, lora_4, weight_4,
                 lora_5, weight_5, lora_6, weight_6, lora_7, weight_7, lora_8, weight_8,
                 merge_mode, output_filename, save_dtype, device) -> io.NodeOutput:
-        
+
         count = int(lora_count)
         names = [lora_1, lora_2, lora_3, lora_4, lora_5, lora_6, lora_7, lora_8]
         weights = [weight_1, weight_2, weight_3, weight_4, weight_5, weight_6, weight_7, weight_8]
-        
+
         valid_paths = []
         valid_weights = []
-        
+
         for i in range(count):
             if names[i] and names[i] != "None":
                 path = folder_paths.get_full_path_or_raise("loras", names[i])
                 valid_paths.append(path)
                 valid_weights.append(weights[i])
-                
+
         if not valid_paths:
             raise ValueError("No LoRAs selected for merging.")
-            
+
         dtype = {"fp16": torch.float16, "bf16": torch.bfloat16, "fp32": torch.float32}[save_dtype]
-        
+
         path = merge_multi_loras(
             lora_paths=valid_paths,
             lora_weights=valid_weights,
@@ -308,7 +314,7 @@ class LoRAMultiMerge(io.ComfyNode):
             save_dtype=dtype,
             output_filename=output_filename
         )
-        
+
         return io.NodeOutput(path)
 
 
@@ -359,25 +365,25 @@ class LoRAMultiMergeDARE(io.ComfyNode):
                 lora_1, weight_1, lora_2, weight_2, lora_3, weight_3, lora_4, weight_4,
                 lora_5, weight_5, lora_6, weight_6, lora_7, weight_7, lora_8, weight_8,
                 drop_rate, trim_quantile, seed, output_filename, save_dtype, device) -> io.NodeOutput:
-        
+
         count = int(lora_count)
         names = [lora_1, lora_2, lora_3, lora_4, lora_5, lora_6, lora_7, lora_8]
         weights = [weight_1, weight_2, weight_3, weight_4, weight_5, weight_6, weight_7, weight_8]
-        
+
         valid_paths = []
         valid_weights = []
-        
+
         for i in range(count):
             if names[i] and names[i] != "None":
                 path = folder_paths.get_full_path_or_raise("loras", names[i])
                 valid_paths.append(path)
                 valid_weights.append(weights[i])
-                
+
         if not valid_paths:
             raise ValueError("No LoRAs selected for merging.")
-            
+
         dtype = {"fp16": torch.float16, "bf16": torch.bfloat16, "fp32": torch.float32}[save_dtype]
-        
+
         path = merge_multi_loras_dare(
             lora_paths=valid_paths,
             lora_weights=valid_weights,
@@ -388,7 +394,7 @@ class LoRAMultiMergeDARE(io.ComfyNode):
             save_dtype=dtype,
             output_filename=output_filename
         )
-        
+
         return io.NodeOutput(path)
 
 def merge_multi_loras_dare(
@@ -412,8 +418,8 @@ def merge_multi_loras_dare(
         print(f"[LoRA Multi-Merge DARE] Drop rate: {drop_rate}, Trim quantile: {trim_quantile}")
 
     prepare_for_large_operation(total_size_gb * 2.5, torch.device(device))
-    handlers = [MemoryEfficientSafeOpen(p) for p in lora_paths]
-    
+    handlers = [MemoryEfficientSafeOpen(p, low_memory=True) for p in lora_paths]
+
     rng = torch.Generator(device=device).manual_seed(seed)
 
     try:
@@ -433,7 +439,7 @@ def merge_multi_loras_dare(
                 "weight": lora_weights[i]
             }
             lora_infos.append(info)
-            
+
             if verbose:
                 print(f"[LoRA Multi-Merge DARE] LoRA {i+1} ({info['name']}): {len(pairs)} layers, dim={rank}")
 
@@ -454,16 +460,16 @@ def merge_multi_loras_dare(
                 downs = []
                 ups = []
                 max_rank = 0
-                
+
                 for idx in indices:
                     info = lora_infos[idx]
                     orig_block = next(b for b in info["pairs"].keys() if _format_lora_key(b) == core)
                     block_keys = info["pairs"][orig_block]
                     if "down" not in block_keys or "up" not in block_keys: continue
-                    
+
                     t_d = handlers[idx].get_tensor(block_keys["down"]).to(device=device, dtype=torch.float32)
                     t_u = handlers[idx].get_tensor(block_keys["up"]).to(device=device, dtype=torch.float32)
-                    
+
                     max_rank = max(max_rank, t_d.shape[0])
                     downs.append((t_d, info["weight"]))
                     ups.append((t_u, info["weight"]))
@@ -484,13 +490,13 @@ def merge_multi_loras_dare(
                             padding[rev_dim*2 + 1] = max_rank - t.shape[dim_to_pad]
                             t = torch.nn.functional.pad(t, tuple(padding))
                         padded.append(t * w)
-                    
+
                     # DARE
                     if drop_rate > 0:
                         for i in range(len(padded)):
                             mask = (torch.rand(padded[i].shape, generator=rng, device=device) > drop_rate).float()
                             padded[i] = (padded[i] * mask) / (1 - drop_rate)
-                    
+
                     # TIES
                     # 1. Trim
                     if trim_quantile > 0:
@@ -500,17 +506,17 @@ def merge_multi_loras_dare(
                             if k > 0:
                                 threshold = torch.kthvalue(flat, k).values
                                 padded[i] = torch.where(padded[i].abs() < threshold, torch.zeros_like(padded[i]), padded[i])
-                    
+
                     # 2. Elect & Merge
                     stacked = torch.stack(padded) # [N, ...]
                     signs = torch.sign(stacked)
                     sum_signs = signs.sum(dim=0)
                     dominant_sign = torch.sign(sum_signs)
-                    
+
                     # Filter those matching dominant sign
                     mask = (signs == dominant_sign) & (dominant_sign != 0)
                     filtered = torch.where(mask, stacked, torch.zeros_like(stacked))
-                    
+
                     # Average matching signs
                     count = mask.sum(dim=0)
                     result = filtered.sum(dim=0) / torch.clamp(count, min=1.0)
@@ -522,7 +528,17 @@ def merge_multi_loras_dare(
                 output_sd[f"{core}.lora_down.weight"] = merged_down.to(save_dtype).cpu().contiguous()
                 output_sd[f"{core}.lora_up.weight"] = merged_up.to(save_dtype).cpu().contiguous()
                 output_sd[f"{core}.alpha"] = torch.tensor(float(max_rank), dtype=save_dtype)
-                
+
+                del merged_down, merged_up
+                for d, _ in downs: del d
+                for u, _ in ups: del u
+                downs.clear()
+                ups.clear()
+                import gc
+                gc.collect()
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+
                 pbar.update(1)
 
         # Final Summary
@@ -531,7 +547,7 @@ def merge_multi_loras_dare(
             for indices in layer_map.values():
                 for idx in indices:
                     matched_stats[idx] += 1
-            
+
             print(f"[LoRA Multi-Merge DARE] --- Merge Summary ---")
             for i, info in enumerate(lora_infos):
                 print(f"[LoRA Multi-Merge DARE] LoRA {i+1}: {matched_stats[i]}/{len(info['pairs'])} layers used in merge")

nodes/lora_resize.py

@@ -461,6 +461,12 @@ def resize_lora_file(
                 if alpha_suffix:
                     output_sd[f"{new_block_name}{alpha_suffix}"] = torch.tensor(result["new_alpha"]).to(save_dtype)
 
+                del result
+                import gc
+                gc.collect()
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+
                 pbar.update(1)
 
         if verbose and fro_list:
@@ -854,6 +860,10 @@ def merge_loras_to_model(
                     del cpu_base
                     stats["copied"] += 1
 
+                import gc
+                gc.collect()
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
 
                 pbar.update(1)
 
@@ -886,11 +896,11 @@ class LoRAMergeToModel(io.ComfyNode):
             node_id="LoRAMergeToModel",
             display_name="LoRA Merge To Model",
             category="ModelUtils/LoRA/Merge",
-            description="Merge 1-4 LoRAs into a base model and save the result. Saves to base model directory.",
+            description="Merge 1-8 LoRAs into a base model and save the result. Saves to base model directory.",
             inputs=[
                 io.Combo.Input("base_model", options=folder_paths.get_filename_list("diffusion_models"),
                               tooltip="Base model the LoRAs were trained on"),
-                io.Combo.Input("lora_count", options=["1", "2", "3", "4"], default="2",
+                io.Combo.Input("lora_count", options=["1", "2", "3", "4", "5", "6", "7", "8"], default="2",
                               tooltip="Number of LoRAs to merge"),
                 # LoRA 1
                 io.Combo.Input("lora_1", options=folder_paths.get_filename_list("loras"),
@@ -912,6 +922,26 @@ class LoRAMergeToModel(io.ComfyNode):
                               default="None", tooltip="Fourth LoRA"),
                 io.Float.Input("weight_4", default=1.0, min=-10.0, max=10.0, step=0.01,
                               tooltip="Weight strength for LoRA 4"),
+                # LoRA 5
+                io.Combo.Input("lora_5", options=["None"] + folder_paths.get_filename_list("loras"),
+                              default="None", tooltip="Fifth LoRA"),
+                io.Float.Input("weight_5", default=1.0, min=-10.0, max=10.0, step=0.01,
+                              tooltip="Weight strength for LoRA 5"),
+                # LoRA 6
+                io.Combo.Input("lora_6", options=["None"] + folder_paths.get_filename_list("loras"),
+                              default="None", tooltip="Sixth LoRA"),
+                io.Float.Input("weight_6", default=1.0, min=-10.0, max=10.0, step=0.01,
+                              tooltip="Weight strength for LoRA 6"),
+                # LoRA 7
+                io.Combo.Input("lora_7", options=["None"] + folder_paths.get_filename_list("loras"),
+                              default="None", tooltip="Seventh LoRA"),
+                io.Float.Input("weight_7", default=1.0, min=-10.0, max=10.0, step=0.01,
+                              tooltip="Weight strength for LoRA 7"),
+                # LoRA 8
+                io.Combo.Input("lora_8", options=["None"] + folder_paths.get_filename_list("loras"),
+                              default="None", tooltip="Eighth LoRA"),
+                io.Float.Input("weight_8", default=1.0, min=-10.0, max=10.0, step=0.01,
+                              tooltip="Weight strength for LoRA 8"),
                 # Settings
                 io.String.Input("skip_patterns", default="", multiline=True,
                                tooltip="Regex patterns for layers to skip"),
@@ -926,12 +956,13 @@ class LoRAMergeToModel(io.ComfyNode):
     @classmethod
     def execute(cls, base_model, lora_count,
                 lora_1, weight_1, lora_2, weight_2, lora_3, weight_3, lora_4, weight_4,
+                lora_5, weight_5, lora_6, weight_6, lora_7, weight_7, lora_8, weight_8,
                 skip_patterns, output_filename, save_dtype, device) -> io.NodeOutput:
 
         # Build LoRA list based on count
         count = int(lora_count)
-        lora_names = [lora_1, lora_2, lora_3, lora_4][:count]
-        lora_weights = [weight_1, weight_2, weight_3, weight_4][:count]
+        lora_names = [lora_1, lora_2, lora_3, lora_4, lora_5, lora_6, lora_7, lora_8][:count]
+        lora_weights = [weight_1, weight_2, weight_3, weight_4, weight_5, weight_6, weight_7, weight_8][:count]
 
         # Filter out "None" entries
         valid_loras = [(name, weight) for name, weight in zip(lora_names, lora_weights) if name != "None"]

nodes/merger.py

@@ -77,18 +77,17 @@ class MergerLogic:
             print(f"[Merger] Preparing memory for {total_size_gb:.2f}GB merge operation...")
             prepare_for_large_operation(total_size_gb * 1.2, torch.device(process_device))
 
-        lazy_load = recipe_params.get('lazy_load', True)
         handlers = {}
         for name in model_names.values():
             if name and name != "None":
                 path = folder_paths.get_full_path(model_type, name)
                 if not path:
                     raise FileNotFoundError(f"Model '{name}' not found.")
-                handlers[name] = MemoryEfficientSafeOpen(path, low_memory=lazy_load)
+                handlers[name] = MemoryEfficientSafeOpen(path, low_memory=True)
 
         primary_handler = handlers[primary_model_name]
         all_keys = primary_handler.keys()
-        metadata = primary_handler.header.get("__metadata__", {})
+        metadata = primary_handler.metadata()
 
         # Convert mismatch_mode string to enum
         mismatch_mode_str = recipe_params.get('mismatch_mode', 'skip')
@@ -125,72 +124,78 @@ class MergerLogic:
         discarded_keys = 0
         error_keys = []
 
-        for key in tqdm(all_keys, desc="Merging layers", unit="layers"):
-            # Check discard patterns first - skip entirely
-            if _matches_any_pattern(key, discard_patterns):
-                discarded_keys += 1
+        with torch.no_grad():
+            for key in tqdm(all_keys, desc="Merging layers", unit="layers"):
+                # Check discard patterns first - skip entirely
+                if _matches_any_pattern(key, discard_patterns):
+                    discarded_keys += 1
+                    pbar.update(1)
+                    continue
+
+                # Pre-load Model A's tensor with pinned memory for CUDA
+                cpu_tensor = primary_handler.get_tensor(key)
+                if process_device == 'cuda':
+                    tensor_a = transfer_to_gpu_pinned(cpu_tensor, process_device, process_dtype)
+                else:
+                    tensor_a = cpu_tensor.to(device=process_device, dtype=process_dtype)
+                del cpu_tensor
+
+                # Check exclude patterns - use Model A only, no merge
+                if _matches_any_pattern(key, exclude_patterns):
+                    merged_state_dict[key] = tensor_a.detach().to(save_torch_dtype).cpu().clone()
+                    excluded_keys += 1
+                    pbar.update(1)
+                    continue
+
+                # Pass tensor_a metadata to recipes for zeros mode and fallback
+                recipe_params['_tensor_a'] = tensor_a
+                recipe_params['_tensor_a_shape'] = tensor_a.shape
+                recipe_params['_tensor_a_dtype'] = tensor_a.dtype
+
+                try:
+                    recipe = calc_mode_class.create_recipe(key=key, **recipe_params)
+                    result = recipe.merge()
+                except MissingTensorError as e:
+                    if mismatch_mode == MissingTensorBehavior.ERROR:
+                        raise ValueError(f"Layer mismatch error (mismatch_mode='error'): {e}")
+                    result = None
+                    error_keys.append(key)
+
+                # Handle None result (mismatch occurred with skip mode)
+                if result is None:
+                    result = tensor_a
+                    skipped_keys += 1
+
+                if isinstance(result, dict):
+                    for r_key, r_tensor in result.items():
+                        merged_state_dict[r_key] = r_tensor.detach().to(save_torch_dtype).cpu().clone()
+                else:
+                    # Ensure compatibility with Model A's architecture.
+                    # If alignment_mode is 'pad/crop', we crop results that were padded.
+                    # If alignment_mode is 'interpolate', resizing happened during operators.
+                    if alignment_mode == 'pad/crop':
+                        target_shape = recipe_params['_tensor_a_shape']
+                        if result.shape != target_shape:
+                            slices = tuple(slice(0, min(res_s, tgt_s)) for res_s, tgt_s in zip(result.shape, target_shape))
+                            result = result[slices]
+
+                    merged_state_dict[key] = result.detach().to(save_torch_dtype).cpu().clone()
+
+                # Clean up references to allow GC immediately.
+                # Local loop variables must be explicitly deleted to prevent PyTorch from keeping tensors in VRAM.
+                recipe.clean()
+                del recipe_params['_tensor_a']
+                del tensor_a
+                del recipe
+                del result
+
+                if recipe_params.get('force_clear_cache', True):
+                    import gc
+                    gc.collect()
+                    if torch.cuda.is_available():
+                        torch.cuda.empty_cache()
+
                 pbar.update(1)
-                continue
-
-            # Pre-load Model A's tensor with pinned memory for CUDA
-            cpu_tensor = primary_handler.get_tensor(key)
-            if process_device == 'cuda':
-                tensor_a = transfer_to_gpu_pinned(cpu_tensor, process_device, process_dtype)
-            else:
-                tensor_a = cpu_tensor.to(device=process_device, dtype=process_dtype)
-            del cpu_tensor
-
-            # Check exclude patterns - use Model A only, no merge
-            if _matches_any_pattern(key, exclude_patterns):
-                merged_state_dict[key] = tensor_a.to(save_torch_dtype).cpu().clone()
-                excluded_keys += 1
-                pbar.update(1)
-                continue
-
-            # Pass tensor_a metadata to recipes for zeros mode and fallback
-            recipe_params['_tensor_a'] = tensor_a
-            recipe_params['_tensor_a_shape'] = tensor_a.shape
-            recipe_params['_tensor_a_dtype'] = tensor_a.dtype
-
-            try:
-                recipe = calc_mode_class.create_recipe(key=key, **recipe_params)
-                result = recipe.merge()
-            except MissingTensorError as e:
-                if mismatch_mode == MissingTensorBehavior.ERROR:
-                    raise ValueError(f"Layer mismatch error (mismatch_mode='error'): {e}")
-                result = None
-                error_keys.append(key)
-
-            # Handle None result (mismatch occurred with skip mode)
-            if result is None:
-                result = tensor_a
-                skipped_keys += 1
-
-            if isinstance(result, dict):
-                for r_key, r_tensor in result.items():
-                    merged_state_dict[r_key] = r_tensor.to(save_torch_dtype).cpu().clone()
-            else:
-                # Ensure compatibility with Model A's architecture.
-                # If alignment_mode is 'pad/crop', we crop results that were padded.
-                # If alignment_mode is 'interpolate', resizing happened during operators.
-                if alignment_mode == 'pad/crop':
-                    target_shape = recipe_params['_tensor_a_shape']
-                    if result.shape != target_shape:
-                        slices = tuple(slice(0, min(res_s, tgt_s)) for res_s, tgt_s in zip(result.shape, target_shape))
-                        result = result[slices]
-
-                merged_state_dict[key] = result.to(save_torch_dtype).cpu().clone()
-
-            # Clean up tensor_a reference to allow GC
-            del recipe_params['_tensor_a']
-            
-            if recipe_params.get('force_clear_cache', True):
-                import gc
-                gc.collect()
-                if torch.cuda.is_available():
-                    torch.cuda.empty_cache()
-
-            pbar.update(1)
 
         # Log summary
         if excluded_keys > 0:

nodes/merger_ops.py

@@ -4,6 +4,8 @@ import torch.nn.functional as F
 from enum import Enum
 
 
+from .merger_utils import transfer_to_gpu_pinned
+
 class MissingTensorBehavior(Enum):
     """Controls behavior when a tensor key is missing from a model."""
     ERROR = "error"      # Raise exception (strict mode)
@@ -127,12 +129,21 @@ class Operation:
         raise NotImplementedError
 
     def recurse(self, operation):
-        source_tensors = [source_oper.merge() for source_oper in operation.sources]
-        return operation.oper(*source_tensors)
+        self._source_tensors = [source_oper.merge() for source_oper in operation.sources]
+        return operation.oper(*self._source_tensors)
 
     def merge(self):
         return self.merge_func(self)
 
+    def clean(self):
+        if hasattr(self, '_source_tensors'):
+            for t in self._source_tensors:
+                del t
+            del self._source_tensors
+        for source in self.sources:
+            if hasattr(source, 'clean'):
+                source.clean()
+
 class LoadTensor(Operation):
     def __init__(self, key, model_name, handlers, device, dtype,
                  on_missing=MissingTensorBehavior.ERROR, fallback_shape=None, fallback_dtype=None):
@@ -165,7 +176,17 @@ class LoadTensor(Operation):
                 dtype = self.fallback_dtype if self.fallback_dtype else self.dtype
                 return torch.zeros(self.fallback_shape, device=self.device, dtype=dtype)
 
-        return handler.get_tensor(self.key).to(device=self.device, dtype=self.dtype)
+        cpu_tensor = handler.get_tensor(self.key)
+        if self.device == 'cuda':
+            self._tensor = transfer_to_gpu_pinned(cpu_tensor, self.device, self.dtype)
+        else:
+            self._tensor = cpu_tensor.to(device=self.device, dtype=self.dtype)
+        del cpu_tensor
+        return self._tensor
+
+    def clean(self):
+        if hasattr(self, '_tensor'):
+            del self._tensor
 
 class Multiply(Operation):
     def __init__(self, key, alpha, *sources):
@@ -409,6 +430,10 @@ class PreloadedTensor(Operation):
     def merge(self) -> torch.Tensor:
         return self.tensor
 
+    def clean(self):
+        if hasattr(self, 'tensor'):
+            del self.tensor
+
 class WeightSum(CalcMode):
     name = 'Weight-Sum'
     description = 'A * (1 - α) + B * α. Simple linear interpolation.'

nodes/merger_utils.py

@@ -1,133 +1,89 @@
 """
-Memory-efficient safetensors utilities with parallel I/O and pinned memory support.
-
-Based on benchmarking, parallel reading with ThreadPoolExecutor provides 2-4x speedup
-for large models compared to sequential reads.
-
-Worker count is automatically optimized based on device capabilities when workers=None.
-Pinned memory enables faster CPU→GPU transfers (2-3x improvement).
+Memory-efficient safetensors utilities with pinned memory support.
+Forces direct file reads to minimize OS page cache usage and avoids VRAM accumulation.
 """
 import os
-import mmap
+import gc
 import json
 import struct
 import torch
-from concurrent.futures import ThreadPoolExecutor, as_completed
-from typing import Dict, List, Optional, Any, Tuple
+from safetensors import safe_open
+from typing import Dict, List, Optional, Any
 
-# Import ComfyUI's memory pinning functions if available
-try:
-    from comfy.model_management import pin_memory, unpin_memory, is_device_cuda
-    PINNING_AVAILABLE = True
-except ImportError:
-    PINNING_AVAILABLE = False
-    def pin_memory(tensor): return False
-    def unpin_memory(tensor): return False
-    def is_device_cuda(device): return str(device).startswith('cuda')
+# Module-level configuration for pinned transfer
+_verbose = False
+_pinned_transfer_stats = {"pinned": 0, "fallback": 0}
 
+def set_verbose(enabled: bool):
+    """Enable/disable verbose output for pinned transfers."""
+    global _verbose
+    _verbose = enabled
 
-class PinnedTensor:
-    """Context manager for automatic pinned memory management.
-
-    Usage:
-        with PinnedTensor(cpu_tensor) as pinned:
-            gpu_tensor = pinned.to('cuda')
-        # Memory automatically unpinned after transfer
-    """
-
-    def __init__(self, tensor: torch.Tensor, auto_pin: bool = True):
-        self.tensor = tensor
-        self.is_pinned = False
-        self.auto_pin = auto_pin and PINNING_AVAILABLE
-
-    def __enter__(self) -> torch.Tensor:
-        if self.auto_pin and self.tensor.device.type == 'cpu':
-            self.is_pinned = pin_memory(self.tensor)
-        return self.tensor
-
-    def __exit__(self, exc_type, exc_val, exc_tb):
-        if self.is_pinned:
-            unpin_memory(self.tensor)
-            self.is_pinned = False
+def get_pinned_transfer_stats():
+    """Return pinned transfer statistics."""
+    return _pinned_transfer_stats.copy()
 
+def reset_pinned_transfer_stats():
+    """Reset transfer statistics."""
+    global _pinned_transfer_stats
+    _pinned_transfer_stats = {"pinned": 0, "fallback": 0}
 
 def transfer_to_gpu_pinned(
     tensor: torch.Tensor,
     device: str = 'cuda',
     dtype: Optional[torch.dtype] = None
 ) -> torch.Tensor:
-    """Transfer tensor to GPU using pinned memory for faster transfer.
+    """Transfer tensor to GPU using pinned memory for faster transfer."""
+    global _pinned_transfer_stats
 
-    Args:
-        tensor: CPU tensor to transfer
-        device: Target GPU device
-        dtype: Optional dtype conversion
+    if tensor.device.type != 'cpu' or not torch.cuda.is_available():
+        if dtype is not None:
+            return tensor.to(device=device, dtype=dtype)
+        return tensor.to(device=device)
 
-    Returns:
-        Tensor on GPU device
-    """
-    if not PINNING_AVAILABLE or not is_device_cuda(torch.device(device)):
-        # Fall back to regular transfer
-        if dtype:
+    if not str(device).startswith('cuda'):
+        if dtype is not None:
             return tensor.to(device=device, dtype=dtype)
         return tensor.to(device=device)
 
-    # Pin memory for faster transfer
-    was_pinned = pin_memory(tensor)
     try:
-        if dtype:
-            result = tensor.to(device=device, dtype=dtype, non_blocking=True)
+        pinned = tensor.pin_memory()
+        if dtype is not None:
+            result = pinned.to(device=device, dtype=dtype, non_blocking=True)
         else:
-            result = tensor.to(device=device, non_blocking=True)
-        # Sync to ensure transfer is complete before unpinning
-        if was_pinned:
-            torch.cuda.current_stream().synchronize()
+            result = pinned.to(device=device, non_blocking=True)
+
+        torch.cuda.current_stream().synchronize()
+        _pinned_transfer_stats["pinned"] += 1
         return result
-    finally:
-        if was_pinned:
-            unpin_memory(tensor)
+    except Exception as e:
+        _pinned_transfer_stats["fallback"] += 1
+        if dtype is not None:
+            return tensor.to(device=device, dtype=dtype)
+        return tensor.to(device=device)
 
 
 class MemoryEfficientSafeOpen:
-    """Memory-efficient safetensors file reader with mmap, parallel I/O, and pinned memory support.
-
-    Features:
-    - mmap mode: Zero-copy tensor access via memory-mapped file
-    - low_memory mode: Direct file reads to minimize OS page cache usage
-    - Parallel loading: Multi-threaded tensor reads for 2-4x speedup
-    - Sorted batch reads: Keys sorted by file offset for sequential I/O
-    - Auto-optimized workers: Adjusts parallelism based on device capabilities
-    - Pinned memory: Faster CPU→GPU transfers when transferring to CUDA
-
-    Args:
-        filename: Path to safetensors file
-        device: Target device (default 'cpu')
-        mmap_mode: Use memory-mapped file for zero-copy (default True)
-        low_memory: Use direct file reads to minimize memory (overrides mmap_mode)
+    """Memory-efficient safetensors file reader with streaming and pinned memory support.
+    
+    Operates strictly in streaming mode (low memory). Tensors are read byte-by-byte from 
+    their exact offsets without using safe_open or memory mapping, ensuring tensors do not
+    remain persistently in OS page caches or RAM.
     """
 
-    def __init__(self, filename: str, device: str = 'cpu', mmap_mode: bool = True, low_memory: bool = False):
+    def __init__(self, filename: str, device: str = 'cpu', **kwargs):
+        # Ignore mmap_mode, low_memory, or any other legacy kwargs. We always stream.
         self.filename = filename
         self.device = device
-        # low_memory mode forces mmap off to avoid OS page caching
-        self.low_memory = low_memory
-        self.mmap_mode = mmap_mode and not low_memory
         self.header, self.header_size = self._read_header()
         self.file = open(filename, "rb")
-        self.mmap_obj = None
-
-        if self.mmap_mode:
-            self.mmap_obj = mmap.mmap(self.file.fileno(), 0, access=mmap.ACCESS_READ)
 
     def __enter__(self):
         return self
 
     def __exit__(self, exc_type, exc_val, exc_tb):
-        if self.mmap_obj:
-            self.mmap_obj.close()
         self.file.close()
 
-
     def keys(self) -> List[str]:
         """Return all tensor keys (excluding metadata)."""
         return [k for k in self.header.keys() if k != "__metadata__"]
@@ -136,8 +92,17 @@ class MemoryEfficientSafeOpen:
         """Return file metadata."""
         return self.header.get("__metadata__", {})
 
+    def get_shape(self, key: str) -> tuple:
+        """Get tensor shape without loading tensor data."""
+        if key not in self.header:
+            raise KeyError(f"Tensor '{key}' not found in file")
+        return tuple(self.header[key]["shape"])
+
+    def get_ndim(self, key: str) -> int:
+        """Get tensor ndim without loading tensor data."""
+        return len(self.get_shape(key))
+
     def keys_sorted_by_offset(self) -> List[str]:
-        """Return keys sorted by file offset for optimal sequential I/O."""
         keys_with_offsets = []
         for key in self.keys():
             offset = self.header[key]["data_offsets"][0]
@@ -146,48 +111,28 @@ class MemoryEfficientSafeOpen:
         return [k for k, _ in keys_with_offsets]
 
     def get_tensor(self, key: str) -> torch.Tensor:
-        """Load a single tensor by key."""
+        """Load a single tensor by key directly from file byte offset."""
         if key not in self.header:
             raise KeyError(f"Tensor '{key}' not found in the file")
 
         metadata = self.header[key]
         offset_start, offset_end = metadata["data_offsets"]
 
-        if self.mmap_mode and self.mmap_obj:
-            if offset_start != offset_end:
-                start = self.header_size + 8 + offset_start
-                end = self.header_size + 8 + offset_end
-                tensor_bytes = memoryview(self.mmap_obj)[start:end]
-            else:
-                tensor_bytes = None
+        if offset_start != offset_end:
+            self.file.seek(self.header_size + 8 + offset_start)
+            tensor_bytes = bytearray(offset_end - offset_start)
+            self.file.readinto(tensor_bytes)
         else:
-            # Non-mmap mode: use pre-allocated bytearray with readinto for minimal copies
             tensor_bytes = None
-            if offset_start != offset_end:
-                self.file.seek(self.header_size + 8 + offset_start)
-                # Pre-allocate writable buffer and read directly into it
-                tensor_bytes = bytearray(offset_end - offset_start)
-                self.file.readinto(tensor_bytes)
 
         return self._deserialize_tensor(tensor_bytes, metadata)
 
-
     def get_tensor_to_gpu(
         self,
         key: str,
         device: str = 'cuda',
         dtype: Optional[torch.dtype] = None
     ) -> torch.Tensor:
-        """Load a tensor and transfer to GPU using pinned memory for faster transfer.
-
-        Args:
-            key: Tensor key to load
-            device: Target GPU device (default 'cuda')
-            dtype: Optional dtype conversion
-
-        Returns:
-            Tensor on GPU device
-        """
         cpu_tensor = self.get_tensor(key)
         return transfer_to_gpu_pinned(cpu_tensor, device, dtype)
 
@@ -197,20 +142,9 @@ class MemoryEfficientSafeOpen:
         workers: Optional[int] = None,
         sort_by_offset: bool = True
     ) -> Dict[str, torch.Tensor]:
-        """Load multiple tensors in parallel using ThreadPoolExecutor.
-
-        Args:
-            keys: List of keys to load (default: all keys)
-            workers: Number of worker threads (default: auto-calculated based on device)
-            sort_by_offset: Sort keys by file offset before loading (default: True)
-
-        Returns:
-            Dict mapping key -> tensor
-        """
         if keys is None:
             keys = self.keys()
 
-        # Auto-calculate workers based on device capabilities
         if workers is None:
             try:
                 from .device_utils import get_optimal_workers, estimate_model_size
@@ -220,13 +154,11 @@ class MemoryEfficientSafeOpen:
                 workers = min(4, os.cpu_count() or 4)
 
         if sort_by_offset:
-            # Sort by offset for better I/O patterns
             keys_with_offsets = [(k, self.header[k]["data_offsets"][0]) for k in keys]
             keys_with_offsets.sort(key=lambda x: x[1])
             keys = [k for k, _ in keys_with_offsets]
 
         result = {}
-
         with ThreadPoolExecutor(max_workers=workers) as executor:
             futures = {
                 executor.submit(
@@ -251,7 +183,6 @@ class MemoryEfficientSafeOpen:
         return result
 
     def get_tensor_as_dict(self, key: str) -> Dict[str, Any]:
-        """Load a tensor and decode as JSON dict (for config tensors)."""
         if key not in self.header:
             raise KeyError(f"Tensor '{key}' not found")
 
@@ -261,14 +192,8 @@ class MemoryEfficientSafeOpen:
         if offset_start == offset_end:
             return {}
 
-        if self.mmap_mode and self.mmap_obj:
-            start = self.header_size + 8 + offset_start
-            end = self.header_size + 8 + offset_end
-            tensor_bytes = bytes(self.mmap_obj[start:end])
-        else:
-            self.file.seek(self.header_size + 8 + offset_start)
-            tensor_bytes = self.file.read(offset_end - offset_start)
-
+        self.file.seek(self.header_size + 8 + offset_start)
+        tensor_bytes = self.file.read(offset_end - offset_start)
         return json.loads(tensor_bytes.decode("utf-8"))
 
     def _read_header(self):
@@ -285,19 +210,13 @@ class MemoryEfficientSafeOpen:
         if tensor_bytes is None:
             byte_tensor = torch.empty(0, dtype=torch.uint8)
         else:
-            # Avoid extra copy if already a bytearray (low_memory mode)
-            if isinstance(tensor_bytes, bytearray):
-                byte_tensor = torch.frombuffer(tensor_bytes, dtype=torch.uint8)
-            else:
-                # mmap memoryview needs copy to create writable tensor
-                byte_tensor = torch.frombuffer(bytearray(tensor_bytes), dtype=torch.uint8)
+            byte_tensor = torch.frombuffer(tensor_bytes, dtype=torch.uint8)
 
         if dtype_str in ["F8_E5M2", "F8_E4M3"]:
             return self._convert_float8(byte_tensor, dtype_str, shape)
 
         return byte_tensor.view(dtype).reshape(shape)
 
-
     @staticmethod
     def _get_torch_dtype(dtype_str):
         dtype_map = {
@@ -322,7 +241,7 @@ class MemoryEfficientSafeOpen:
         elif dtype_str == "F8_E4M3" and hasattr(torch, "float8_e4m3fn"):
             return byte_tensor.view(torch.float8_e4m3fn).reshape(shape)
         else:
-            raise ValueError(f"Unsupported float8 type: {dtype_str}. Your PyTorch version may be too old.")
+            raise ValueError(f"Unsupported float8 type: {dtype_str}")
 
 
 def _read_tensor_from_file(

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "ComfyUI-ModelUtils"
 description = "[WIP]Custom nodes for handling, inspecting, modifying and creating various model files."
-version = "0.2.7"
+version = "0.2.8"
 license = { file = "LICENSE" }
 
 [project.urls]