diff --git a/README.md b/README.md
index cb64c84..6780481 100644
--- a/README.md
+++ b/README.md
@@ -4,7 +4,7 @@
 
 Official release of [SeedVR2](https://github.com/ByteDance-Seed/SeedVR) for ComfyUI that enables high-quality video and image upscaling.
 
-Can run as **Multi-GPU standalone CLI** too, see [🖥️ Run as Standalone](#️-run-as-standalone-cli) section.
+Can run as **Multi-GPU standalone CLI** too, see [🖥️ Run as Standalone](#-run-as-standalone-cli) section.
 
 [![SeedVR2 v2.5 Deep Dive Tutorial](https://img.youtube.com/vi/MBtWYXq_r60/maxresdefault.jpg)](https://youtu.be/MBtWYXq_r60)
 
@@ -14,8 +14,8 @@ Can run as **Multi-GPU standalone CLI** too, see [🖥️ Run as Standalone](#
 
 ## 📋 Quick Access
 
-- [🆙 Future Releases](#-future-releases)
-- [🚀 Updates](#-updates)
+- [🆙 Future Work](#-future-work)
+- [🚀 Release Notes](#-release-notes)
 - [🎯 Features](#-features)
 - [🔧 Requirements](#-requirements)
 - [📦 Installation](#-installation)
@@ -26,7 +26,7 @@ Can run as **Multi-GPU standalone CLI** too, see [🖥️ Run as Standalone](#
 - [🙏 Credits](#-credits)
 - [📜 License](#-license)
 
-## 🆙 Future Releases
+## 🆙 Future Work
 
 We're actively working on improvements and new features. To stay informed:
 
@@ -34,13 +34,31 @@ We're actively working on improvements and new features. To stay informed:
 - **💬 Join the Community**: Learn from others, share your workflows, and get help in the [Discussions](https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler/discussions)
 - **🔮 Next Model Survey**: We're looking for community input on the next open-source super-powerful generic restoration model. Share your suggestions in [Issue #164](https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler/issues/164)
 
-## 🚀 Updates
+## 🚀 Release Notes
 
-**2025.12.08 - Version 2.5.18**
+**2025.12.12 - Version 2.5.20**
 
-- **🚀 Feature: SageAttention Support** - Added `sd2` and `sd3` attention modes for experimental support of SageAttention v2 and v3 kernels.
-- **⚡ Feature: Precision Control** - Added `precision` selector (auto/fp16/bf16/bf32) to main upscaler node for fine-grained control over computation dtype.
-- **🔧 Optimization: High-end GPU Tuning** - Improved auto-detection defaults for Blackwell (50-series) and high-VRAM GPUs.
+- **⚡ Expanded attention backends** - Full support for Flash Attention 2 (Ampere+), Flash Attention 3 (Hopper+), SageAttention 2, and SageAttention 3 (Blackwell/RTX 50xx), with automatic fallback chains to PyTorch SDPA when unavailable *(based on PR by [@naxci1](https://github.com/naxci1) - thank you!)*
+- **🍎 macOS/Apple Silicon compatibility** - Replaced MPS autocast with explicit dtype conversion throughout VAE and DiT pipelines, resolving hangs and crashes on M-series Macs. BlockSwap now auto-disables with warning (unified memory makes it meaningless)
+- **🛡️ Flash Attention graceful fallback** - Added compatibility shims for corrupted or partially installed flash_attn/xformers DLLs, preventing startup crashes
+- **🛡️ AMD ROCm: bitsandbytes conflict fix** - Prevent kernel registration errors when diffusers attempts to re-import broken bitsandbytes installations
+- **📦 ComfyUI Manager: macOS classifier fix** - Removed NVIDIA CUDA classifier causing false "GPU not supported" warnings on macOS
+- **📚 Documentation updates** - Updated README with attention backend details, BlockSwap macOS notes, and clarified model caching descriptions
+
+**2025.12.10 - Version 2.5.19**
+
+- **🎨 New header logo design** - Refreshed ASCII art banner *(thanks [@naxci1](https://github.com/naxci1))*
+- **🧹 Remove dead flash attention wrapper** - Removed legacy code from FP8CompatibleDiT; FlashAttentionVarlen already handles backend switching via its `attention_mode` attribute
+- **🛡️ Fix graceful fallback from flash-attn** - Add compatibility shims for corrupted flash_attn/xformers DLLs, preventing startup crashes when CUDA extensions are broken
+- **📊 Improved VRAM tracking** - Separate allocated vs reserved memory tracking, Windows-only overflow detection (WDDM paging behavior)
+- **♻️ Centralize backend detection** - Unified `is_mps_available()`, `is_cuda_available()`, `get_gpu_backend()` helpers across codebase
+- **🔄 Revert 2.5.14 VRAM limit enforcement** - Removed `set_per_process_memory_fraction` call; Overflow detection and warnings remain.
+
+**2025.12.09 - Version 2.5.18**
+
+- **🚀 CLI: Streaming mode for long videos** - New `--chunk_size` flag processes videos in memory-bounded chunks, enabling arbitrarily long videos without RAM limits. Works with model caching (`--cache_dit`/`--cache_vae`) for chunk-to-chunk reuse *(inspired by [disk02](https://github.com/disk02) PR contribution)*
+- **⚡ CLI: Multi-GPU streaming** - Each GPU now streams its segment internally with independent model caching, improving memory efficiency and enabling `--temporal_overlap` blending at GPU boundaries
+- **🔧 CLI: Fix large video MemoryError** - Shared memory transfer replaces numpy pickling, preventing crashes on high-resolution/long video outputs *(inspired by  [FurkanGozukara](https://github.com/FurkanGozukara) PR contribution)*
 
 **2025.12.05 - Version 2.5.17**
 
@@ -223,7 +241,7 @@ We're actively working on improvements and new features. To stay informed:
 
 **2025.07.03**
 
-- 🛠️ Can run as **standalone mode** with **Multi GPU** see [🖥️ Run as Standalone](#️-run-as-standalone-cli)
+- 🛠️ Can run as **standalone mode** with **Multi GPU** see [🖥️ Run as Standalone](#run-as-standalone-cli)
 
 **2025.06.30**
 
@@ -270,8 +288,8 @@ We're actively working on improvements and new features. To stay informed:
 ### Performance Features
 - **torch.compile Integration**: Optional 20-40% DiT speedup and 15-25% VAE speedup with PyTorch 2.0+ compilation
 - **Multi-GPU CLI**: Distribute workload across multiple GPUs with automatic temporal overlap blending
-- **Model Caching**: Keep models loaded in memory for faster batch processing
-- **Flexible Attention Backends**: Choose between PyTorch SDPA (stable, always available) or Flash Attention 2 (faster on supported hardware)
+- **Model Caching**: Keep models loaded between generations for single-GPU directory processing or multi-GPU streaming
+- **Flexible Attention Backends**: Choose between PyTorch SDPA (stable, always available), Flash Attention 2/3, or SageAttention 2/3 for faster computation on supported hardware
 
 ### Quality Control
 - **Advanced Color Correction**: Five methods including LAB (recommended for highest fidelity), wavelet, wavelet adaptive, HSV, and AdaIN
@@ -300,7 +318,7 @@ With the current optimizations (tiling, BlockSwap, GGUF quantization), SeedVR2 c
 - **Python**: 3.12+ (Python 3.12 and 3.13 tested and recommended)
 - **PyTorch**: 2.0+ for torch.compile support (optional but recommended)
 - **Triton**: Required for torch.compile with inductor backend (optional)
-- **Flash Attention 2**: Provides faster attention computation on supported hardware (optional, falls back to PyTorch SDPA)
+- **Flash Attention / SageAttention**: Flash Attention 2 (Ampere+), Flash Attention 3 (Hopper+), SageAttention 2 or SageAttention 3 (Blackwell) provide faster attention computation on supported hardware (optional, falls back to PyTorch SDPA)
 
 ## 📦 Installation
 
@@ -415,14 +433,21 @@ Configure the DiT (Diffusion Transformer) model for video upscaling.
   - Requires offload_device to be set and different from device
 
 - **attention_mode**: Attention computation backend
-  - `sdpa`: PyTorch scaled_dot_product_attention (default, stable, always available)
-  - `flash_attn`: Flash Attention 2 (faster on supported hardware, requires flash-attn package)
+  - `sdpa`: PyTorch scaled_dot_product_attention (default, always available)
+  - `flash_attn_2`: Flash Attention 2 (Ampere+, requires flash-attn package)
+  - `flash_attn_3`: Flash Attention 3 (Hopper+, requires flash-attn with FA3 support)
+  - `sageattn_2`: SageAttention 2 (requires sageattention package)
+  - `sageattn_3`: SageAttention 3 (Blackwell/RTX 50xx, requires sageattn3 package)
 
 - **torch_compile_args**: Connect to SeedVR2 Torch Compile Settings node for 20-40% speedup
 
 **BlockSwap Explained:**
 
-BlockSwap enables running large models on GPUs with limited VRAM by dynamically swapping transformer blocks between GPU and CPU memory during inference. Here's how it works:
+BlockSwap enables running large models on GPUs with limited VRAM by dynamically swapping transformer blocks between GPU and CPU memory during inference.
+
+> **Note:** BlockSwap is not available on macOS. Apple Silicon Macs use unified memory architecture where GPU and CPU share the same memory pool, making BlockSwap meaningless. The option will be automatically disabled with a warning if requested on macOS.
+
+Here's how it works:
 
 - **What it does**: Keeps only the currently-needed transformer blocks on the GPU, while storing the rest on CPU or another device
 - **When to use it**: When you get OOM (Out of Memory) errors during the upscaling phase
@@ -777,9 +802,18 @@ The CLI provides comprehensive options for single-GPU, multi-GPU, and batch proc
 # Basic image upscaling
 python inference_cli.py image.jpg
 
-# Basic video video upscaling with temporal consistency
+# Basic video upscaling with temporal consistency
 python inference_cli.py video.mp4 --resolution 720 --batch_size 33
 
+# Streaming mode for long videos (memory-efficient)
+# Processes video in chunks of 330 frames to avoid loading entire video into RAM
+# Use --temporal_overlap to ensure smooth transitions between chunks
+python inference_cli.py long_video.mp4 \
+    --resolution 1080 \
+    --batch_size 33 \
+    --chunk_size 330 \
+    --temporal_overlap 3
+
 # Multi-GPU processing with temporal overlap
 python inference_cli.py video.mp4 \
     --cuda_device 0,1 \
@@ -836,7 +870,8 @@ python inference_cli.py media_folder/ \
 - `--batch_size`: Frames per batch (must follow 4n+1: 1, 5, 9, 13, 17, 21...). Ideally matches shot length for best temporal consistency (default: 5)
 - `--seed`: Random seed for reproducibility (default: 42)
 - `--skip_first_frames`: Skip N initial frames (default: 0)
-- `--load_cap`: Load maximum N frames from video. 0 = load all (default: 0)
+- `--load_cap`: Maximum total frames to load from video. 0 = load all (default: 0)
+- `--chunk_size`: Frames per chunk for streaming mode. When > 0, processes video in memory-bounded chunks of N frames, writing each chunk before loading the next. Essential for long videos that would otherwise exceed RAM. Use with `--temporal_overlap` for seamless chunk transitions. 0 = load all frames at once (default: 0)
 - `--prepend_frames`: Prepend N reversed frames to reduce start artifacts (auto-removed) (default: 0)
 - `--temporal_overlap`: Frames to overlap between batches/GPUs for smooth blending (default: 0)
 
@@ -848,9 +883,8 @@ python inference_cli.py media_folder/ \
 **Memory Management:**
 - `--dit_offload_device`: Device to offload DiT model: 'none' (keep on GPU), 'cpu', or 'cuda:X' (default: none)
 - `--vae_offload_device`: Device to offload VAE model: 'none', 'cpu', or 'cuda:X' (default: none)
-- `--blocks_to_swap`: Number of transformer blocks to swap (0=disabled, 3B: 0-32, 7B: 0-36). Requires dit_offload_device (default: 0)
-- `--swap_io_components`: Offload I/O components for additional VRAM savings. Requires dit_offload_device
-- `--use_non_blocking`: Use non-blocking memory transfers for BlockSwap (recommended)
+- `--blocks_to_swap`: Number of transformer blocks to swap (0=disabled, 3B: 0-32, 7B: 0-36). Requires dit_offload_device (default: 0). Not available on macOS.
+- `--swap_io_components`: Offload I/O components for additional VRAM savings. Requires dit_offload_device. Not available on macOS.
 
 **VAE Tiling:**
 - `--vae_encode_tiled`: Enable VAE encode tiling to reduce VRAM during encoding
@@ -862,7 +896,8 @@ python inference_cli.py media_folder/ \
 - `--tile_debug`: Visualize tiles: 'false' (default), 'encode', or 'decode'
 
 **Performance Optimization:**
-- `--attention_mode`: Attention backend: 'sdpa' (default, stable) or 'flash_attn' (faster, requires package)
+- `--allow_vram_overflow`: Allow VRAM overflow to system RAM. Prevents OOM but may cause severe slowdown
+- `--attention_mode`: Attention backend: 'sdpa' (default), 'flash_attn_2' (Ampere+), 'flash_attn_3' (Hopper+), 'sageattn_2', or 'sageattn_3' (Blackwell)
 - `--compile_dit`: Enable torch.compile for DiT model (20-40% speedup, requires PyTorch 2.0+ and Triton)
 - `--compile_vae`: Enable torch.compile for VAE model (15-25% speedup, requires PyTorch 2.0+ and Triton)
 - `--compile_backend`: Compilation backend: 'inductor' (full optimization) or 'cudagraphs' (lightweight) (default: inductor)
@@ -873,8 +908,8 @@ python inference_cli.py media_folder/ \
 - `--compile_dynamo_recompile_limit`: Max recompilation attempts before fallback (default: 128)
 
 **Model Caching (batch processing):**
-- `--cache_dit`: Cache DiT model between files (single GPU only, speeds up directory processing)
-- `--cache_vae`: Cache VAE model between files (single GPU only, speeds up directory processing)
+- `--cache_dit`: Keep DiT model in memory between generations. Works with single-GPU directory processing or multi-GPU streaming (`--chunk_size`). Requires `--dit_offload_device`
+- `--cache_vae`: Keep VAE model in memory between generations. Works with single-GPU directory processing or multi-GPU streaming (`--chunk_size`). Requires `--vae_offload_device`
 
 **Multi-GPU:**
 - `--cuda_device`: CUDA device id(s). Single id (e.g., '0') or comma-separated list '0,1' for multi-GPU
@@ -977,7 +1012,7 @@ For detailed contribution guidelines, see [CONTRIBUTING.md](CONTRIBUTING.md).
 
 This ComfyUI implementation is a collaborative project by **[NumZ](https://github.com/numz)** and **[AInVFX](https://www.youtube.com/@AInVFX)** (Adrien Toupet), based on the original [SeedVR2](https://github.com/ByteDance-Seed/SeedVR) by ByteDance Seed Team.
 
-Special thanks to our community contributors including [benjaminherb](https://github.com/benjaminherb), [cmeka](https://github.com/cmeka), [FurkanGozukara](https://github.com/FurkanGozukara), [JohnAlcatraz](https://github.com/JohnAlcatraz), [lihaoyun6](https://github.com/lihaoyun6), [Luchuanzhao](https://github.com/Luchuanzhao), [Luke2642](https://github.com/Luke2642), [naxci1](https://github.com/naxci1), [q5sys](https://github.com/q5sys), and many others for their improvements, bug fixes, and testing.
+Special thanks to our community contributors including [naxci1](https://github.com/naxci1), [benjaminherb](https://github.com/benjaminherb), [cmeka](https://github.com/cmeka), [FurkanGozukara](https://github.com/FurkanGozukara), [JohnAlcatraz](https://github.com/JohnAlcatraz), [lihaoyun6](https://github.com/lihaoyun6), [Luchuanzhao](https://github.com/Luchuanzhao), [Luke2642](https://github.com/Luke2642), [proxyid](https://github.com/proxyid), [q5sys](https://github.com/q5sys), and many others for their improvements, bug fixes, and testing.
 
 ## 📜 License
 
diff --git a/inference_cli.py b/inference_cli.py
index 4701ae8..7c80200 100644
--- a/inference_cli.py
+++ b/inference_cli.py
@@ -8,10 +8,12 @@ Supports single and multi-GPU processing with advanced memory optimization.
 Key Features:
     • Multi-GPU Processing: Automatic workload distribution across multiple GPUs with
       temporal overlap blending for seamless transitions
+    • Streaming Mode: Memory-efficient processing of long videos in chunks, avoiding
+      full video loading into RAM while maintaining temporal consistency
     • Memory Optimization: BlockSwap for limited VRAM, VAE tiling for large resolutions,
       intelligent tensor offloading between processing phases
     • Performance: Torch.compile integration, BFloat16 compute pipeline,
-      efficient model caching for batch processing
+      efficient model caching for batch and streaming processing
     • Flexibility: Multiple output formats (MP4/PNG), advanced color correction methods,
       directory batch processing with auto-format detection
     • Quality Control: Temporal overlap blending, frame prepending for artifact reduction,
@@ -48,7 +50,7 @@ import argparse
 import time
 import platform
 import multiprocessing as mp
-from typing import Dict, Any, List, Optional, Tuple, Literal
+from typing import Dict, Any, List, Optional, Tuple, Literal, Generator
 from datetime import datetime
 from pathlib import Path
 
@@ -74,7 +76,7 @@ if platform.system() == "Darwin":
 else:
     os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "backend:cudaMallocAsync")
 
-    # Pre-parse CUDA device argument for validation and environment setup
+    # Pre-parse arguments that must be handled before torch import
     _pre_parser = argparse.ArgumentParser(add_help=False)
     _pre_parser.add_argument("--cuda_device", type=str, default=None)
     _pre_args, _ = _pre_parser.parse_known_args()
@@ -125,23 +127,13 @@ from src.core.generation_phases import (
     postprocess_all_batches
 )
 from src.utils.debug import Debug
+from src.optimization.memory_manager import clear_memory, get_gpu_backend, is_cuda_available
 debug = Debug(enabled=False)  # Will be enabled via --debug CLI flag
 
-
 # =============================================================================
 # Device Management Helpers
 # =============================================================================
 
-def _get_platform_type() -> str:
-    """Determine the platform device type (cuda/mps/cpu)."""
-    if platform.system() == "Darwin":
-        return "mps"
-    elif torch.cuda.is_available():
-        return "cuda"
-    else:
-        return "cpu"
-
-
 def _device_id_to_name(device_id: str, platform_type: str = None) -> str:
     """
     Convert device ID to full device name.
@@ -157,7 +149,7 @@ def _device_id_to_name(device_id: str, platform_type: str = None) -> str:
         return device_id
     
     if platform_type is None:
-        platform_type = _get_platform_type()
+        platform_type = get_gpu_backend()
     
     # MPS typically doesn't use indices
     if platform_type == "mps":
@@ -356,6 +348,9 @@ def process_single_file(input_path: str, args: argparse.Namespace, device_list:
     """
     Process a single video or image file with optional model caching.
     
+    For videos, supports streaming mode (chunk_size > 0) which processes in memory-bounded
+    chunks with temporal overlap for seamless transitions between chunks.
+    
     Args:
         input_path: Path to input file
         args: Command-line arguments with all processing settings
@@ -365,7 +360,7 @@ def process_single_file(input_path: str, args: argparse.Namespace, device_list:
         runner_cache: Optional cache dict for model reuse across multiple files
     
     Returns:
-        Number of frames processed from the input
+        Number of frames written to output
     """
     input_type = get_input_type(input_path)
     
@@ -375,19 +370,6 @@ def process_single_file(input_path: str, args: argparse.Namespace, device_list:
     
     debug.log(f"Processing {input_type}: {Path(input_path).name}", category="generation", force=True)
     
-    # Extract frames
-    if input_type == "video":
-        start_time = time.time()
-        frames_tensor, original_fps = extract_frames_from_video(
-            input_path, args.skip_first_frames, args.load_cap
-        )
-        debug.log(f"Frame extraction time: {time.time() - start_time:.2f}s", category="timing")
-    else:
-        frames_tensor, original_fps = extract_frames_from_image(input_path)
-    
-    # Track frames before processing (for FPS calculation)
-    input_frame_count = len(frames_tensor)
-    
     # Generate or validate output path
     if output_path is None:
         output_path = generate_output_path(input_path, args.output_format, input_type=input_type)
@@ -400,237 +382,360 @@ def process_single_file(input_path: str, args: argparse.Namespace, device_list:
     format_prefix = "Auto-detected" if format_auto_detected else "Requested"
     debug.log(f"{format_prefix} output format: {args.output_format}", category="info", force=True, indent_level=1)
     
-    # Process frames
-    processing_start = time.time()
-    # Use direct processing if caching enabled OR on Mac (MPS doesn't support multiprocessing well)
-    if runner_cache is not None or platform.system() == "Darwin":
-        # Direct single-GPU processing (required for Mac MPS, optional for caching)
-        result = _single_gpu_direct_processing(frames_tensor, args, device_list[0], runner_cache)
-    else:
-        # Multi-GPU or non-cached processing via worker processes
-        result = _gpu_processing(frames_tensor, device_list, args)
-    debug.log(f"Processing time: {time.time() - processing_start:.2f}s", category="timing")
-
-    # Save results
-    is_png_format = args.output_format == "png"
-    is_single_image = input_type == "image"
-    
-    if is_png_format and is_single_image:
-        # Single PNG file
-        os.makedirs(Path(output_path).parent, exist_ok=True)
-        frame_np = (result[0].cpu().numpy() * 255.0).astype(np.uint8)
-        # Convert RGB(A) to BGR(A) based on channel count
-        if frame_np.shape[2] == 4:
-            frame_save = cv2.cvtColor(frame_np, cv2.COLOR_RGBA2BGRA)
+    # === VIDEO PROCESSING ===
+    if input_type == "video":
+        if not os.path.exists(input_path):
+            raise FileNotFoundError(f"Video file not found: {input_path}")
+        
+        cap = cv2.VideoCapture(input_path)
+        if not cap.isOpened():
+            raise ValueError(f"Cannot open video file: {input_path}")
+        
+        fps = cap.get(cv2.CAP_PROP_FPS) or 30.0
+        total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+        width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+        height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+        
+        debug.log(f"Video info: {total_frames} frames, {width}x{height}, {fps:.2f} FPS", category="info")
+        
+       # Skip initial frames
+        if args.skip_first_frames > 0:
+            debug.log(f"Skipping first {args.skip_first_frames} frames", category="info")
+            cap.set(cv2.CAP_PROP_POS_FRAMES, args.skip_first_frames)
+        
+        # Calculate frames to process (apply load_cap if set)
+        frames_to_process = total_frames - args.skip_first_frames
+        if args.load_cap > 0:
+            frames_to_process = min(frames_to_process, args.load_cap)
+        
+        # Early exit for empty/exhausted video
+        if frames_to_process <= 0:
+            debug.log(f"No frames to process after skipping {args.skip_first_frames} of {total_frames}", 
+                     level="WARNING", category="file", force=True)
+            cap.release()
+            return 0
+        
+        # Streaming mode: process in chunks
+        chunk_size = args.chunk_size if args.chunk_size > 0 else frames_to_process
+        streaming = args.chunk_size > 0
+        total_chunks = (frames_to_process + chunk_size - 1) // chunk_size  # ceiling division
+        
+        if streaming:
+            debug.log(f"Streaming mode: chunks of {chunk_size} frames, overlap={args.temporal_overlap}", 
+                     category="info", force=True, indent_level=1)
+        
+        is_png = args.output_format == "png"
+        video_writer = None
+        overlap = args.temporal_overlap
+        frames_written = 0
+        chunk_idx = 0
+        base_name = Path(input_path).stem
+        
+        # Multi-GPU: workers stream their own segments
+        if len(device_list) > 1:
+            cap.release()  # Workers will reopen
+            video_info = {
+                'video_path': input_path,
+                'start_frame': args.skip_first_frames,
+                'frames_to_process': frames_to_process,
+            }
+            result = _gpu_processing(None, device_list, args, video_info=video_info)
+            
+            # Save result
+            if is_png:
+                save_frames_to_image(result, output_path, base_name)
+            else:
+                video_writer = save_frames_to_video(result, output_path, fps)
+                if video_writer is not None:
+                    video_writer.release()
+            
+            frames_written = result.shape[0]
+        
+        # Single GPU: stream in main process
         else:
-            frame_save = cv2.cvtColor(frame_np, cv2.COLOR_RGB2BGR)
-        cv2.imwrite(output_path, frame_save)
-    
-    elif is_png_format:
-        # PNG sequence (save_frames_to_png creates directory internally)
-        save_frames_to_png(result, output_path, base_name=Path(input_path).stem)
-    
-    else:
-        # Video file
-        os.makedirs(Path(output_path).parent, exist_ok=True)
-        save_frames_to_video(result, output_path, original_fps)
-    
-    # Log appropriate save message based on format
-    if is_png_format and not is_single_image:
-        debug.log(f"PNG frames saved in directory: {output_path}", category="file", force=True)
-    else:
+            chunk_count = 0
+            for result in _stream_video_chunks(
+                cap=cap,
+                frames_to_process=frames_to_process,
+                chunk_size=chunk_size,
+                overlap=overlap,
+                args=args,
+                device_id=device_list[0],
+                debug=debug,
+                runner_cache=runner_cache,
+                log_progress=streaming,
+                total_chunks=total_chunks,
+                cleanup_timer_name="chunk_cleanup"
+            ):
+                chunk_count += 1
+                
+                # Save output
+                if is_png:
+                    save_frames_to_image(result, output_path, base_name, start_index=frames_written)
+                else:
+                    video_writer = save_frames_to_video(result, output_path, fps, writer=video_writer)
+                
+                frames_written += result.shape[0]
+                del result
+            
+            chunk_idx = chunk_count
+            cap.release()
+            if video_writer is not None:
+                video_writer.release()
+        
+        if streaming:
+            debug.log("", category="none", force=True)
+            if len(device_list) > 1:
+                debug.log(f"Streaming complete: {frames_written} frames across {len(device_list)} GPUs", category="success", force=True)
+            else:
+                debug.log(f"Streaming complete: {frames_written} frames in {chunk_idx} chunks", category="success", force=True)
+        
         debug.log(f"Output saved to: {output_path}", category="file", force=True)
+        return frames_written
     
-    return input_frame_count
+    # === IMAGE PROCESSING ===
+    frames_tensor, _ = extract_frames_from_image(input_path)
+    
+    processing_start = time.time()
+    # Process frames (multiprocessing only for multi-GPU)
+    if len(device_list) > 1:
+        result = _gpu_processing(frames_tensor, device_list, args)
+    else:
+        result = _single_gpu_direct_processing(frames_tensor, args, device_list[0], runner_cache)
+    debug.log(f"Processing time: {time.time() - processing_start:.2f}s", category="timing")
+    
+    # Save single image
+    os.makedirs(Path(output_path).parent, exist_ok=True)
+    frame_np = (result[0].cpu().numpy() * 255.0).astype(np.uint8)
+    _save_image_bgr(frame_np, output_path)
+    
+    debug.log(f"Output saved to: {output_path}", category="file", force=True)
+    return 1
 
 
-def extract_frames_from_video(
-    video_path: str, 
-    skip_first_frames: int = 0, 
-    load_cap: Optional[int] = None
-) -> Tuple[torch.Tensor, float]:
+def _read_frames_from_cap(cap: cv2.VideoCapture, max_frames: int) -> Optional[torch.Tensor]:
     """
-    Extract frames from video file and convert to tensor format.
-    
-    Reads video using OpenCV, converts BGR to RGB, normalizes to [0,1] range.
-    Note: Frame prepending is handled later in the processing pipeline via
-    compute_generation_info(), not in this function.
+    Read up to max_frames from an already-open VideoCapture.
     
     Args:
-        video_path: Path to input video file
-        skip_first_frames: Number of initial frames to skip (default: 0)
-        load_cap: Maximum number of frames to load, None loads all (default: None)
-        
+        cap: An already opened cv2.VideoCapture instance
+        max_frames: Maximum number of frames to read in this call
+    
     Returns:
-        Tuple containing:
-            - frames_tensor: Frames in format [T, H, W, C], Float32, range [0,1]
-            - fps: Original video frames per second
-    
-    Raises:
-        FileNotFoundError: If video file doesn't exist
-        ValueError: If video cannot be opened or no frames extracted
+        Tensor [T, H, W, C] float32 [0,1], or None if no frames available
     """
-    debug.log(f"Extracting frames from video: {video_path}", category="file")
-    
-    if not os.path.exists(video_path):
-        raise FileNotFoundError(f"Video file not found: {video_path}")
-    
-    # Open video
-    cap = cv2.VideoCapture(video_path)
-    if not cap.isOpened():
-        raise ValueError(f"Cannot open video file: {video_path}")
-    
-    # Get video properties
-    fps = cap.get(cv2.CAP_PROP_FPS)
-    frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
-    width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
-    height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
-    
-    debug.log(f"Video info: {frame_count} frames, {width}x{height}, {fps:.2f} FPS", category="info")
-    if skip_first_frames:
-        debug.log(f"Will skip first {skip_first_frames} frames", category="info")
-    if load_cap:
-        debug.log(f"Will load maximum {load_cap} frames", category="info")
-    
     frames = []
-    frame_idx = 0
-    frames_loaded = 0
-    
-    while True:
+    for _ in range(max_frames):
         ret, frame = cap.read()
         if not ret:
             break
-        
-        # Skip first frame if requested
-        if frame_idx < skip_first_frames:
-            frame_idx += 1
-            continue
-
-        if skip_first_frames > 0 and frame_idx == skip_first_frames:
-            debug.log(f"Skipped first {skip_first_frames} frames", category="info") 
-
-        # Check load cap
-        if load_cap is not None and load_cap > 0 and frames_loaded >= load_cap:
-            debug.log(f"Reached load cap of {load_cap} frames", category="info")
-            break
-        
-        # Convert BGR to RGB
-        frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
-        
-        # Convert to float32 and normalize to 0-1
-        frame = frame.astype(np.float32) / 255.0
-        
+        frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB).astype(np.float32) / 255.0
         frames.append(frame)
-        frame_idx += 1
-        frames_loaded += 1
+    
+    if not frames:
+        return None
+    return torch.from_numpy(np.stack(frames)).to(torch.float32)
+
+
+def _stream_video_chunks(
+    cap: cv2.VideoCapture,
+    frames_to_process: int,
+    chunk_size: int,
+    overlap: int,
+    args: argparse.Namespace,
+    device_id: str,
+    debug: 'Debug',
+    runner_cache: Optional[Dict[str, Any]],
+    log_progress: bool = False,
+    total_chunks: int = 0,
+    cleanup_timer_name: Optional[str] = None,
+    log_prefix: str = ""
+) -> Generator[torch.Tensor, None, None]:
+    """
+    Generator that streams and processes video chunks.
+    
+    Handles frame reading, temporal context prepending, processing via
+    _process_frames_core, context removal from output, and memory cleanup.
+    Caller is responsible for VideoCapture lifecycle and result handling.
+    
+    Args:
+        cap: Open VideoCapture positioned at start frame
+        frames_to_process: Total frames to read and process
+        chunk_size: Frames per chunk (use frames_to_process for single chunk)
+        overlap: Temporal overlap frames between chunks for blending
+        args: Processing arguments (copied internally, prepend_frames zeroed after first chunk)
+        device_id: GPU device ID for processing
+        debug: Debug instance for logging
+        runner_cache: Optional model cache dict for reuse across chunks
+        log_progress: If True, log chunk progress with separators
+        total_chunks: Total chunks for progress display (used if log_progress=True)
+        cleanup_timer_name: Optional timer name for memory cleanup logging
+        log_prefix: Optional prefix for log messages (e.g., "[GPU 0] " for worker identification)
+    
+    Yields:
+        Processed frames tensor [T, H, W, C] for each chunk, context frames removed
+    """
+    chunk_args = argparse.Namespace(**vars(args))
+    frames_read = 0
+    prev_raw_tail = None
+    chunk_idx = 0
+    streaming = chunk_size < frames_to_process
+    
+    while frames_read < frames_to_process:
+        read_count = min(chunk_size, frames_to_process - frames_read)
+        new_frames = _read_frames_from_cap(cap, read_count)
+        if new_frames is None:
+            break
+        frames_read += new_frames.shape[0]
+        chunk_idx += 1
         
-        if debug.enabled and frames_loaded % 100 == 0:
-            total_to_load = min(frame_count, load_cap) if load_cap else frame_count
-            debug.log(f"Extracted {frames_loaded}/{total_to_load} frames", category="file")
-    
-    cap.release()
-    
-    if len(frames) == 0:
-        raise ValueError(f"No frames extracted from video: {video_path}")
-    
-    debug.log(f"Extracted {len(frames)} frames", category="success")
+        # Disable prepend_frames after first chunk
+        if chunk_idx > 1:
+            chunk_args.prepend_frames = 0
+        
+        # Prepend context from previous chunk
+        if prev_raw_tail is not None and overlap > 0:
+            context_count = min(overlap, prev_raw_tail.shape[0])
+            frames = torch.cat([prev_raw_tail[-context_count:], new_frames], dim=0)
+        else:
+            frames = new_frames
+            context_count = 0
+        
+        # Log progress if enabled
+        if log_progress and streaming:
+            if chunk_idx > 1:
+                debug.log("", category="none", force=True)
+                debug.log("━" * 60, category="none", force=True)
+            debug.log("", category="none", force=True)
+            debug.log(f"{log_prefix}Chunk {chunk_idx}/{total_chunks}: {new_frames.shape[0]} new + {context_count} context frames", 
+                     category="generation", force=True)
+            debug.log("", category="none", force=True)
+        
+        # Process chunk
+        result = _process_frames_core(
+            frames_tensor=frames.to(torch.float16),
+            args=chunk_args,
+            device_id=device_id,
+            debug=debug,
+            runner_cache=runner_cache
+        )
+        
+        # Remove context frames from output
+        if context_count > 0:
+            result = result[context_count:]
+        
+        # Save tail for next chunk context
+        prev_raw_tail = new_frames[-overlap:].clone() if overlap > 0 else None
+        
+        # Cleanup before yield
+        del frames
+        
+        yield result
+        
+        # Memory cleanup between chunks
+        if streaming:
+            clear_memory(debug=debug, deep=True, force=True, timer_name=cleanup_timer_name)
 
-    # Convert to tensor (will be cast to compute_dtype in worker process)
-    frames_tensor = torch.from_numpy(np.stack(frames)).to(torch.float32)
-    
-    debug.log(f"Frames tensor shape: {frames_tensor.shape}, dtype: {frames_tensor.dtype}", category="memory")
 
-    return frames_tensor, fps
+def _save_image_bgr(frame_np: np.ndarray, file_path: str) -> None:
+    """
+    Save a single RGB(A) uint8 frame to disk, converting to BGR(A) for OpenCV.
+    
+    Args:
+        frame_np: Frame as uint8 numpy array [H, W, C] where C is 3 (RGB) or 4 (RGBA)
+        file_path: Output file path
+    """
+    if frame_np.shape[2] == 4:
+        frame_bgr = cv2.cvtColor(frame_np, cv2.COLOR_RGBA2BGRA)
+    else:
+        frame_bgr = cv2.cvtColor(frame_np, cv2.COLOR_RGB2BGR)
+    cv2.imwrite(file_path, frame_bgr)
 
 
 def save_frames_to_video(
     frames_tensor: torch.Tensor, 
     output_path: str, 
-    fps: float = 30.0
-) -> None:
+    fps: float = 30.0,
+    writer: Optional[cv2.VideoWriter] = None
+) -> Optional[cv2.VideoWriter]:
     """
     Save frames tensor to MP4 video file.
     
     Converts tensor from Float32 [0,1] to uint8 [0,255], RGB to BGR for OpenCV,
-    and writes to video file using mp4v codec.
+    and writes to video file using mp4v codec. Supports streaming mode where
+    an existing writer is passed and kept open for subsequent chunks.
     
     Args:
         frames_tensor: Frames in format [T, H, W, C], Float32, range [0,1]
-        output_path: Output video file path (will be created if doesn't exist)
+        output_path: Output video file path (directory created if doesn't exist)
         fps: Frames per second for output video (default: 30.0)
+        writer: Existing VideoWriter for streaming (if None, creates new one)
+    
+    Returns:
+        VideoWriter if streaming mode (caller must close), None if standalone mode
     
     Raises:
         ValueError: If video writer cannot be initialized
     """
-    debug.log(f"Saving {frames_tensor.shape[0]} frames to video: {output_path}", category="file")
-    
-    # Convert tensor to numpy and denormalize
-    frames_np = frames_tensor.cpu().numpy()
-    frames_np = (frames_np * 255.0).astype(np.uint8)
-    
-    # Get video properties
+    frames_np = (frames_tensor.cpu().numpy() * 255.0).astype(np.uint8)
     T, H, W, C = frames_np.shape
     
-    # Initialize video writer
-    fourcc = cv2.VideoWriter_fourcc(*'mp4v')
-    out = cv2.VideoWriter(output_path, fourcc, fps, (W, H))
+    if writer is None:
+        debug.log(f"Saving {T} frames to video: {output_path}", category="file")
+        os.makedirs(Path(output_path).parent, exist_ok=True)
+        fourcc = cv2.VideoWriter_fourcc(*'mp4v')
+        writer = cv2.VideoWriter(output_path, fourcc, fps, (W, H))
+        if not writer.isOpened():
+            raise ValueError(f"Cannot create video writer for: {output_path}")
     
-    if not out.isOpened():
-        raise ValueError(f"Cannot create video writer for: {output_path}")
-    
-    # Write frames
     for i, frame in enumerate(frames_np):
-        # Convert RGB to BGR for OpenCV
         frame_bgr = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
-        out.write(frame_bgr)
-
+        writer.write(frame_bgr)
         if debug.enabled and (i + 1) % 100 == 0:
-            debug.log(f"Saved {i + 1}/{T} frames", category="file")
-
-    out.release()
+            debug.log(f"Written {i + 1}/{T} frames", category="file")
     
-    debug.log(f"Video saved successfully: {output_path}", category="success")
+    return writer  # Caller always closes
 
 
-def save_frames_to_png(
+def save_frames_to_image(
     frames_tensor: torch.Tensor, 
     output_dir: str, 
-    base_name: str
-) -> None:
+    base_name: str,
+    start_index: int = 0
+) -> int:
     """
     Save frames tensor as sequential PNG image files.
     
-    Each frame saved as {base_name}_{index:05d}.png with zero-padded indices.
+    Each frame saved as {base_name}_{index:0Nd}.png with zero-padded indices.
     Converts Float32 [0,1] to uint8 [0,255] and RGB(A) to BGR(A) for OpenCV.
     
     Args:
         frames_tensor: Frames in format [T, H, W, C], Float32, range [0,1]
         output_dir: Directory to save PNG files (created if doesn't exist)
         base_name: Base name for output files (e.g., "frame" → "frame_00000.png")
+        start_index: Starting index for filenames (for streaming continuation)
+    
+    Returns:
+        Number of frames saved
     """
-    debug.log(f"Saving {frames_tensor.shape[0]} frames as PNGs to directory: {output_dir}", category="file")
-
-    # Ensure output directory exists
     os.makedirs(output_dir, exist_ok=True)
-
-    # Convert to numpy uint8 RGB
+    
     frames_np = (frames_tensor.cpu().numpy() * 255.0).astype(np.uint8)
     total = frames_np.shape[0]
-    digits = max(5, len(str(total)))  # at least 5 digits
+    
+    if start_index == 0:
+        debug.log(f"Saving {total} frames as PNGs to directory: {output_dir}", category="file")
+    digits = 6  # Supports up to 999,999 frames (~11.5 hours at 24fps)
 
     for idx, frame in enumerate(frames_np):
-        filename = f"{base_name}_{idx:0{digits}d}.png"
+        filename = f"{base_name}_{start_index + idx:0{digits}d}.png"
         file_path = os.path.join(output_dir, filename)
-        # Convert RGB(A) to BGR(A) for cv2 based on channel count
-        if frame.shape[2] == 4:
-            frame_save = cv2.cvtColor(frame, cv2.COLOR_RGBA2BGRA)
-        else:
-            frame_save = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
-        cv2.imwrite(file_path, frame_save)
+        _save_image_bgr(frame, file_path)
         if debug.enabled and (idx + 1) % 100 == 0:
-            debug.log(f"Saved {idx + 1}/{total} PNGs", category="file")
+            debug.log(f"Saved {idx + 1}/{total} images", category="file")
 
-    debug.log(f"PNG saving completed: {total} files in '{output_dir}'", category="success")
+    debug.log(f"Saved {total} images to '{output_dir}'", category="success")
+    return total
 
 
 # =============================================================================
@@ -661,7 +766,7 @@ def _process_frames_core(
         Upscaled frames tensor [T', H', W', C], Float32, range [0,1]
     """    
     # Determine platform and convert device IDs to full names
-    platform_type = _get_platform_type()
+    platform_type = get_gpu_backend()
     inference_device = _device_id_to_name(device_id, platform_type)
     
     # Parse offload devices (with caching defaults)
@@ -820,46 +925,100 @@ def _process_frames_core(
 def _worker_process(
     proc_idx: int, 
     device_id: str, 
-    frames_np: np.ndarray, 
+    frames_np: Optional[np.ndarray],
     shared_args: Dict[str, Any], 
-    return_queue: mp.Queue
+    return_queue: mp.Queue,
+    done_barrier: mp.Barrier,
+    video_info: Optional[Dict[str, Any]] = None
 ) -> None:
     """
     Worker process for multi-GPU upscaling.
     
-    CUDA_VISIBLE_DEVICES is set by parent before spawn, so this worker
-    only sees its assigned GPU. Results returned via queue as numpy arrays.
-    """
-    # Note: CUDA_VISIBLE_DEVICES and PYTORCH_CUDA_ALLOC_CONF are inherited
-    # from parent (set before spawn). torch is imported at module level.
+    Supports two modes:
+    1. frames_np provided: Process pre-loaded frames (for images)
+    2. video_info provided: Stream video segment internally (for videos)
+       - Each worker opens the video, seeks to its assigned range, and streams
+         with internal chunking and model caching for memory efficiency
     
+    Args:
+        proc_idx: Worker index for result ordering
+        device_id: GPU device ID (used for CUDA_VISIBLE_DEVICES inheritance)
+        frames_np: Pre-loaded frames as numpy array, or None for video streaming
+        shared_args: Serialized args namespace as dict
+        return_queue: Queue for returning results to parent
+        done_barrier: Barrier for synchronizing shared memory handoff
+        video_info: Optional dict with 'video_path', 'start_frame', 'end_frame'
+                   for video streaming mode
+    """
     # Create debug instance for this worker
     worker_debug = Debug(enabled=shared_args["debug"])
     
-    # Convert numpy back to tensor
-    frames_tensor = torch.from_numpy(frames_np).to(torch.float16)
-    
-    # Create args namespace from shared_args
     args = argparse.Namespace(**shared_args)
     
-    # Process frames (no caching in worker mode)
-    result_tensor = _process_frames_core(
-        frames_tensor=frames_tensor,
-        args=args,
-        device_id="0",  # Worker sees only 1 GPU (index 0) due to CUDA_VISIBLE_DEVICES
-        debug=worker_debug,
-        runner_cache=None  # No caching in multiprocessing mode
-    )
+    # Video streaming mode: worker reads and processes its assigned segment
+    if video_info is not None:
+        cap = cv2.VideoCapture(video_info['video_path'])
+        cap.set(cv2.CAP_PROP_POS_FRAMES, video_info['start_frame'])
+        
+        segment_frames = video_info['end_frame'] - video_info['start_frame']
+        chunk_size = args.chunk_size if args.chunk_size > 0 else segment_frames
+        
+        worker_debug.log(f"GPU {proc_idx}: frames {video_info['start_frame']}-{video_info['end_frame']} "
+                        f"({segment_frames} frames, chunks of {chunk_size})",
+                        category="generation", force=True)
+        
+        # Only GPU 0 uses prepend_frames (applies to video start only)
+        worker_args = argparse.Namespace(**vars(args))
+        if proc_idx != 0:
+            worker_args.prepend_frames = 0
+        
+        # Enable model caching within worker only if requested
+        runner_cache = {} if (args.cache_dit or args.cache_vae) else None
+        
+        total_chunks = (segment_frames + chunk_size - 1) // chunk_size
+        results = []
+        for result in _stream_video_chunks(
+            cap=cap,
+            frames_to_process=segment_frames,
+            chunk_size=chunk_size,
+            overlap=args.temporal_overlap,
+            args=worker_args,
+            device_id="0",
+            debug=worker_debug,
+            runner_cache=runner_cache,
+            log_progress=total_chunks > 1,
+            total_chunks=total_chunks,
+            log_prefix=f"[GPU {proc_idx}] "
+        ):
+            results.append(result.cpu())
+        
+        cap.release()
+        result_tensor = torch.cat(results, dim=0) if results else torch.empty(0, dtype=torch.float32)
     
-    # Send back result as numpy array
-    return_queue.put((proc_idx, result_tensor.numpy()))
+    # Pre-loaded frames mode (original behavior)
+    else:
+        frames_tensor = torch.from_numpy(frames_np).to(torch.float16)
+        result_tensor = _process_frames_core(
+            frames_tensor=frames_tensor,
+            args=args,
+            device_id="0",
+            debug=worker_debug,
+            runner_cache=None
+        )
+    
+    # Share tensor memory for efficient cross-process transfer (avoids pickling large arrays)
+    return_queue.put((proc_idx, result_tensor.share_memory_()))
+    
+    # Wait for parent to copy shared tensors before exiting
+    # (shared memory requires creating process to stay alive during access)
+    done_barrier.wait()
 
 
 def _single_gpu_direct_processing(
     frames_tensor: torch.Tensor,
     args: argparse.Namespace,
     device_id: str,
-    runner_cache: Dict[str, Any]
+    runner_cache: Optional[Dict[str, Any]]
 ) -> torch.Tensor:
     """
     Direct single-GPU processing with model caching support.
@@ -876,88 +1035,115 @@ def _single_gpu_direct_processing(
 
 
 def _gpu_processing(
-    frames_tensor: torch.Tensor, 
+    frames_tensor: Optional[torch.Tensor],
     device_list: List[str], 
-    args: argparse.Namespace
+    args: argparse.Namespace,
+    video_info: Optional[Dict[str, Any]] = None
 ) -> torch.Tensor:
     """
     Orchestrate multi-GPU parallel video upscaling with temporal overlap blending.
     
-    Splits input frames across multiple GPUs with optional temporal overlap,
-    spawns worker processes for parallel processing, and reassembles results
-    with smooth blending of overlapping regions.
-    
-    Processing flow:
-        1. Split frames into chunks (with overlap if enabled)
-        2. Spawn worker processes on each GPU
-        3. Wait for all workers to complete
-        4. Blend overlapping regions using Hann window crossfade
-        5. Remove prepended frames from final result
+    Supports two modes:
+    1. video_info provided: Workers stream their assigned video segments internally
+       (each GPU reads and processes its frame range with internal chunking)
+    2. frames_tensor provided: Workers process pre-loaded frame chunks
+       (non streaming behavior for images or pre-loaded videos)
     
     Args:
-        frames_tensor: Input frames [T, H, W, C], Float32, range [0,1]
+        frames_tensor: Input frames [T, H, W, C] or None if using video_info mode
         device_list: List of device IDs as strings (e.g., ["0", "1"])
         args: Parsed command-line arguments containing all processing settings
+        video_info: Optional dict with 'video_path', 'start_frame', 'frames_to_process'
+                   for streaming mode where workers read video directly
     
     Returns:
         Upscaled frames tensor [T', H', W', C], Float32, range [0,1]
-        where T' may be less than T if prepend_frames were removed
-    
-    Note:
-        - Single GPU: Can use multiprocessing or direct processing
-        - Multi-GPU with overlap: Chunks sized to multiples of batch_size for
-          proper temporal blending
-        - Prepended frames removed after all GPU workers complete (multi-GPU safe)
     """
     num_devices = len(device_list)
-    total_frames = frames_tensor.shape[0]
+    overlap = args.temporal_overlap
     
-    # Create overlapping chunks (for multi GPU); ensures every chunk is 
-    # a multiple of batch_size (except last one) to avoid blending issues
-    if args.temporal_overlap > 0 and num_devices > 1:
-        chunk_with_overlap = total_frames // num_devices + args.temporal_overlap
-        if args.batch_size > 1:
-            chunk_with_overlap = ((chunk_with_overlap + args.batch_size - 1) // args.batch_size) * args.batch_size
-        base_chunk_size = chunk_with_overlap - args.temporal_overlap
-
-        chunks = []
-        for i in range(num_devices):
-            start_idx = i * base_chunk_size
-            if i == num_devices - 1: # last chunk/device
-                end_idx = total_frames
-            else:
-                end_idx = min(start_idx + chunk_with_overlap, total_frames)
-            chunks.append(frames_tensor[start_idx:end_idx])
-    else:
-        chunks = torch.chunk(frames_tensor, num_devices, dim=0)
-
-    # Use direct Queue with explicit unlimited size for large video chunks
-    return_queue = mp.Queue(maxsize=0)  # 0 = unlimited (explicit)
+    return_queue = mp.Queue(maxsize=0)
+    done_barrier = mp.Barrier(num_devices + 1)
     workers = []
-
-    # Convert args namespace to dict for serialization
     shared_args = vars(args).copy()
-
-    # Start all workers
-    for idx, (device_id, chunk_tensor) in enumerate(zip(device_list, chunks)):
-        # Set CUDA_VISIBLE_DEVICES before spawning so child inherits it
-        os.environ["CUDA_VISIBLE_DEVICES"] = device_id
+    
+    # Video streaming mode: distribute frame ranges to workers
+    if video_info is not None:
+        total_frames = video_info['frames_to_process']
+        start_frame = video_info['start_frame']
+        video_path = video_info['video_path']
         
-        p = mp.Process(
-            target=_worker_process,
-            args=(idx, device_id, chunk_tensor.cpu().numpy(), shared_args, return_queue),
-        )
-        p.start()
-        workers.append(p)
+        base_per_gpu = total_frames // num_devices
+        remainder = total_frames % num_devices
+        
+        current_start = start_frame
+        for idx, device_id in enumerate(device_list):
+            gpu_frames = base_per_gpu + (1 if idx < remainder else 0)
+            gpu_end = current_start + gpu_frames
+            
+            # Add overlap frames for blending (except last GPU)
+            if idx < num_devices - 1 and overlap > 0:
+                gpu_end = min(gpu_end + overlap, start_frame + total_frames)
+            
+            worker_video_info = {
+                'video_path': video_path,
+                'start_frame': current_start,
+                'end_frame': gpu_end,
+            }
+            
+            os.environ["CUDA_VISIBLE_DEVICES"] = device_id
+            p = mp.Process(
+                target=_worker_process,
+                args=(idx, device_id, None, shared_args, return_queue, done_barrier),
+                kwargs={'video_info': worker_video_info}
+            )
+            p.start()
+            workers.append(p)
+            
+            current_start += gpu_frames
+    
+    # Pre-loaded frames mode (original behavior for images or non-streaming)
+    else:
+        total_frames = frames_tensor.shape[0]
+        
+        if overlap > 0 and num_devices > 1:
+            chunk_with_overlap = total_frames // num_devices + overlap
+            if args.batch_size > 1:
+                chunk_with_overlap = ((chunk_with_overlap + args.batch_size - 1) // args.batch_size) * args.batch_size
+            base_chunk_size = chunk_with_overlap - overlap
+
+            chunks = []
+            for i in range(num_devices):
+                start_idx = i * base_chunk_size
+                if i == num_devices - 1:
+                    end_idx = total_frames
+                else:
+                    end_idx = min(start_idx + chunk_with_overlap, total_frames)
+                chunks.append(frames_tensor[start_idx:end_idx])
+        else:
+            chunks = torch.chunk(frames_tensor, num_devices, dim=0)
+
+        for idx, (device_id, chunk_tensor) in enumerate(zip(device_list, chunks)):
+            os.environ["CUDA_VISIBLE_DEVICES"] = device_id
+            p = mp.Process(
+                target=_worker_process,
+                args=(idx, device_id, chunk_tensor.cpu().numpy(), shared_args, return_queue, done_barrier),
+            )
+            p.start()
+            workers.append(p)
 
     # Collect results before joining to prevent deadlock
+    # Tensors arrive via shared memory - copy to numpy while workers still alive
     results_np = [None] * num_devices
     collected = 0
     while collected < num_devices:
-        proc_idx, res_np = return_queue.get()
-        results_np[proc_idx] = res_np
+        proc_idx, result_tensor = return_queue.get()
+        results_np[proc_idx] = result_tensor.numpy()
         collected += 1
     
+    # Release workers now that shared tensors are copied
+    done_barrier.wait()
+    
     # Now safe to join
     for p in workers:
         p.join()
@@ -1041,9 +1227,12 @@ Examples:
   Basic image upscaling:
     python {invocation} image.jpg
 
-  Basic video video upscaling with temporal consistency
+  Basic video upscaling with temporal consistency:
     python {invocation} video.mp4 --resolution 720 --batch_size 33
     
+  Streaming mode for long videos:
+    python {invocation} long_video.mp4 --resolution 1080 --batch_size 33 --chunk_size 330 --temporal_overlap 3
+
   Multi-GPU processing with temporal overlap:
     python {invocation} video.mp4 --cuda_device 0,1 --resolution 1080 --batch_size 81 --uniform_batch_size --temporal_overlap 3 --prepend_frames 4 
 
@@ -1101,6 +1290,9 @@ Examples:
                         help="Skip N initial frames (default: 0)")
     process_group.add_argument("--load_cap", type=int, default=0,
                         help="Load maximum N frames from video. 0 = load all (default: 0)")
+    process_group.add_argument("--chunk_size", type=int, default=0,
+                        help="Frames per chunk for streaming mode. When > 0, processes video in "
+                             "memory-bounded chunks of N frames. 0 = load all frames at once (default: 0)")
     process_group.add_argument("--prepend_frames", type=int, default=0,
                         help="Prepend N reversed frames to reduce start artifacts (auto-removed). Default: 0")
     process_group.add_argument("--temporal_overlap", type=int, default=0,
@@ -1135,9 +1327,10 @@ Examples:
     blockswap_group = parser.add_argument_group('Memory optimization (BlockSwap)')
     blockswap_group.add_argument("--blocks_to_swap", type=int, default=0,
                         help="Transformer blocks to swap for VRAM savings. 0-32 (3B) or 0-36 (7B). "
-                             "Requires --dit_offload_device. Default: 0 (disabled)")
+                             "Requires --dit_offload_device. Not available on macOS. Default: 0 (disabled)")
     blockswap_group.add_argument("--swap_io_components", action="store_true",
-                        help="Offload DiT I/O layers for extra VRAM savings. Requires --dit_offload_device")
+                        help="Offload DiT I/O layers for extra VRAM savings. Requires --dit_offload_device. "
+                             "Not available on macOS")
     
     # VAE Tiling
     vae_group = parser.add_argument_group('VAE tiling (for high resolution upscale)')
@@ -1159,11 +1352,8 @@ Examples:
     # Performance
     perf_group = parser.add_argument_group('Performance optimization')
     perf_group.add_argument("--attention_mode", type=str, default="sdpa",
-                        choices=["sdpa", "flash_attn", "sa2", "sa3"],
-                        help="Attention backend: 'sdpa' (default), 'flash_attn' (faster), 'sa2' (SageAttention v2), 'sa3' (SageAttention v3)")
-    perf_group.add_argument("--precision", type=str, default="auto",
-                        choices=["auto", "fp16", "bf16", "bf32"],
-                        help="Compute precision: 'auto' (default), 'fp16', 'bf16', or 'bf32' (TF32)")
+                        choices=["sdpa", "flash_attn_2", "flash_attn_3", "sageattn_2", "sageattn_3"],
+                        help="Attention backend: 'sdpa' (default), 'flash_attn_2', 'flash_attn_3', 'sageattn_2', or 'sageattn_3' (Blackwell GPUs)")
     perf_group.add_argument("--compile_dit", action="store_true", 
                         help="Enable torch.compile for DiT model (20-40%% speedup, requires PyTorch 2.0+ and Triton)")
     perf_group.add_argument("--compile_vae", action="store_true",
@@ -1185,9 +1375,11 @@ Examples:
     # Model Caching (for batch processing)
     cache_group = parser.add_argument_group('Model caching (batch processing)')
     cache_group.add_argument("--cache_dit", action="store_true",
-                        help="Cache DiT model between files (single GPU only, speeds up directory processing)")
+                        help="Keep DiT model in memory between generations. Works with single-GPU directory processing "
+                             "or multi-GPU streaming (--chunk_size). Requires --dit_offload_device")
     cache_group.add_argument("--cache_vae", action="store_true",
-                        help="Cache VAE model between files (single GPU only, speeds up directory processing)")
+                        help="Keep VAE model in memory between generations. Works with single-GPU directory processing "
+                             "or multi-GPU streaming (--chunk_size). Requires --vae_offload_device")
     
     # Debugging
     debug_group = parser.add_argument_group('Debugging')
@@ -1246,27 +1438,9 @@ def main() -> None:
         debug.log(f"VAE decode tile overlap ({args.vae_decode_tile_overlap}) must be smaller than tile size ({args.vae_decode_tile_size})", level="ERROR", category="vae", force=True)
         sys.exit(1)
     
-    # Validate BlockSwap configuration - either blocks_to_swap or swap_io_components requires dit_offload_device
-    blockswap_enabled = args.blocks_to_swap > 0 or args.swap_io_components
-    if blockswap_enabled and args.dit_offload_device == "none":
-        config_details = []
-        if args.blocks_to_swap > 0:
-            config_details.append(f"blocks_to_swap={args.blocks_to_swap}")
-        if args.swap_io_components:
-            config_details.append("swap_io_components=True")
-        
-        debug.log(
-            f"BlockSwap enabled ({', '.join(config_details)}) but dit_offload_device='none'. "
-            "BlockSwap requires dit_offload_device to be set (typically 'cpu'). "
-            "Either set --dit_offload_device cpu or disable BlockSwap "
-            "(--blocks_to_swap 0 and do not use --swap_io_components)",
-            level="ERROR", category="blockswap", force=True
-        )
-        sys.exit(1)
-    
     # Inform about caching defaults
     if args.cache_dit and args.dit_offload_device == "none":
-        offload_target = "system memory (CPU)" if _get_platform_type() != "mps" else "unified memory"
+        offload_target = "system memory (CPU)" if get_gpu_backend() != "mps" else "unified memory"
         debug.log(
             f"DiT caching enabled: Using default {offload_target} for offload. "
             "Set --dit_offload_device explicitly to use a different device.",
@@ -1274,7 +1448,7 @@ def main() -> None:
         )
     
     if args.cache_vae and args.vae_offload_device == "none":
-        offload_target = "system memory (CPU)" if _get_platform_type() != "mps" else "unified memory"
+        offload_target = "system memory (CPU)" if get_gpu_backend() != "mps" else "unified memory"
         debug.log(
             f"VAE caching enabled: Using default {offload_target} for offload. "
             "Set --vae_offload_device explicitly to use a different device.",
@@ -1287,7 +1461,7 @@ def main() -> None:
         else:
             # Show actual CUDA device visibility
             debug.log(f"CUDA_VISIBLE_DEVICES: {os.environ.get('CUDA_VISIBLE_DEVICES', 'Not set (all)')}", category="device")
-            if torch.cuda.is_available():
+            if is_cuda_available():
                 debug.log(f"torch.cuda.device_count(): {torch.cuda.device_count()}", category="device")
                 debug.log(f"Using device index 0 inside script (mapped to selected GPU)", category="device")
     
@@ -1328,18 +1502,18 @@ def main() -> None:
             
             debug.log(f"Found {len(media_files)} media files to process", category="file", force=True)
             
-            # Validate caching with multi-GPU (not supported in CLI - would need shared memory)
-            if (args.cache_dit or args.cache_vae) and len(device_list) > 1:
+            # Multi-GPU caching requires streaming (workers cache within their chunk loops)
+            if (args.cache_dit or args.cache_vae) and len(device_list) > 1 and args.chunk_size <= 0:
                 debug.log(
-                    "Model caching requires single GPU selection (you selected multiple GPUs). "
-                    "Disabling caching for this run.", 
+                    "Model caching requires streaming mode (--chunk_size > 0) for multi-GPU. "
+                    "Disabling caching for this run.",
                     level="WARNING", category="cache", force=True
                 )
                 args.cache_dit = False
                 args.cache_vae = False
             
-            # Initialize runner cache if caching enabled
-            runner_cache = {} if (args.cache_dit or args.cache_vae) else None
+            # Single-GPU: runner_cache persists across files; multi-GPU: workers cache internally
+            runner_cache = {} if (args.cache_dit or args.cache_vae) and len(device_list) == 1 else None
             
             for idx, file_path in enumerate(media_files, 1):
                 # Visual separation between files (except before first file)
@@ -1379,27 +1553,32 @@ def main() -> None:
             if format_auto_detected:
                 args.output_format = "mp4" if input_type == "video" else "png"
             
-            # Validate caching for single file (would provide no benefit but shouldn't error)
-            if (args.cache_dit or args.cache_vae):
+            # Caching: single-GPU streaming uses runner_cache, multi-GPU streaming workers cache internally
+            runner_cache = None
+            streaming = args.chunk_size > 0
+            
+            if args.cache_dit or args.cache_vae:
                 if len(device_list) > 1:
-                    debug.log(
-                        "Model caching requires single GPU selection (you selected multiple GPUs). "
-                        "Disabling caching for this run.",
-                        level="WARNING", category="cache", force=True
-                    )
-                    args.cache_dit = False
-                    args.cache_vae = False
+                    if not streaming:
+                        debug.log(
+                            "Model caching requires streaming mode (--chunk_size > 0) for multi-GPU. "
+                            "Disabling caching for this run.",
+                            level="WARNING", category="cache", force=True
+                        )
+                        args.cache_dit = False
+                        args.cache_vae = False
+                elif streaming:
+                    runner_cache = {}
                 else:
                     debug.log(
-                        "Model caching has no benefit for single file processing (only useful for directories). "
+                        "Model caching has no benefit for single file processing (only useful for directories or streaming mode). "
                         "Consider removing --cache_dit/--cache_vae for single files.",
                         category="tip", force=True
                     )
             
-            # No caching for single file (no benefit)
             frames = process_single_file(args.input, args, device_list, args.output,
                                         format_auto_detected=format_auto_detected,
-                                        runner_cache=None)
+                                        runner_cache=runner_cache)
             total_frames_processed += frames
         
         else:
diff --git a/pyproject.toml b/pyproject.toml
index 8c6f1fd..61410dd 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,15 +1,14 @@
 [project]
 name = "seedvr2_videoupscaler"
 description = "SeedVR2 official ComfyUI integration: ByteDance-Seed's one-step diffusion-based video/image upscaling with memory-efficient inference"
-version = "2.5.17"
+version = "2.5.20"
 authors = [
     {name = "numz"},
     {name = "adrientoupet"}
 ]
 license = {file = "LICENSE"} 
 classifiers = [
-     "Operating System :: OS Independent",
-     "Environment :: GPU :: NVIDIA CUDA"
+     "Operating System :: OS Independent"
 ]
 dependencies = [
     "torch",
diff --git a/src/common/distributed/basic.py b/src/common/distributed/basic.py
index 4615967..d92610e 100644
--- a/src/common/distributed/basic.py
+++ b/src/common/distributed/basic.py
@@ -21,6 +21,7 @@ from datetime import timedelta
 import torch
 import torch.distributed as dist
 from torch.nn.parallel import DistributedDataParallel
+from ...optimization.memory_manager import is_mps_available
 
 def get_global_rank() -> int:
     """
@@ -47,7 +48,7 @@ def get_device() -> torch.device:
     """
     Get current rank device.
     """
-    if hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
+    if is_mps_available():
         return torch.device("mps")
     return torch.device("cuda", get_local_rank())
 
diff --git a/src/core/generation_phases.py b/src/core/generation_phases.py
index 7fe1f33..909b3cf 100644
--- a/src/core/generation_phases.py
+++ b/src/core/generation_phases.py
@@ -700,7 +700,7 @@ def upscale_all_batches(
             )
             conditions = [condition]
             
-            # Detect DiT model dtype (handle FP8CompatibleDiT wrapper)
+            # Detect DiT model dtype (handle CompatibleDiT wrapper)
             dit_model = runner.dit.dit_model if hasattr(runner.dit, 'dit_model') else runner.dit
             try:
                 dit_dtype = next(dit_model.parameters()).dtype
@@ -708,9 +708,10 @@ def upscale_all_batches(
                 dit_dtype = ctx['compute_dtype']  # Fallback for meta device or empty model
             
             # Use autocast if DiT dtype differs from compute dtype
+            # Skip autocast on MPS (CompatibleDiT already handles dtype conversion)
             debug.start_timer(f"dit_inference_{upscale_idx+1}")
             with torch.no_grad():
-                if dit_dtype != ctx['compute_dtype']:
+                if dit_dtype != ctx['compute_dtype'] and ctx['dit_device'].type != 'mps':
                     with torch.autocast(ctx['dit_device'].type, ctx['compute_dtype'], enabled=True):
                         upscaled_latents = runner.inference(
                             noises=noises,
@@ -1414,7 +1415,7 @@ def postprocess_all_batches(
                     total_computed += (num_valid_samples - 1) * actual_overlap
                 frame_info += f" ({total_computed} computed with {' + '.join(adjustments)} removed)"
             
-            debug.log(f"Final output assembled: {frame_info}, Resolution: {Wf}x{Hf}px, Channels: {channels_str}", 
+            debug.log(f"Output assembled: {frame_info}, Resolution: {Wf}x{Hf}px, Channels: {channels_str}", 
                     category="generation", force=True)
         else:
             ctx['final_video'] = torch.empty((0, 0, 0, 0), dtype=ctx['compute_dtype'])
diff --git a/src/core/generation_utils.py b/src/core/generation_utils.py
index 5b14790..f89642e 100644
--- a/src/core/generation_utils.py
+++ b/src/core/generation_utils.py
@@ -365,10 +365,6 @@ def setup_generation_context(
         interrupt_fn = None
         comfyui_available = False
     
-    # Determine compute dtype (allow override in context setup if needed, but standard flow uses COMPUTE_DTYPE)
-    # The precision selection in UI overrides this later in model_configuration.
-    compute_dtype = COMPUTE_DTYPE
-
     # Create generation context
     ctx = {
         'dit_device': dit_device,
@@ -376,7 +372,7 @@ def setup_generation_context(
         'dit_offload_device': dit_offload_device,
         'vae_offload_device': vae_offload_device,
         'tensor_offload_device': tensor_offload_device,
-        'compute_dtype': compute_dtype,
+        'compute_dtype': COMPUTE_DTYPE,
         'interrupt_fn': interrupt_fn,
         'video_transform': None,
         'text_embeds': None,
@@ -436,7 +432,6 @@ def prepare_runner(
     decode_tile_overlap: Optional[Tuple[int, int]] = None,
     tile_debug: str = "false",
     attention_mode: str = 'sdpa',
-    precision: str = 'auto',
     torch_compile_args_dit: Optional[Dict[str, Any]] = None,
     torch_compile_args_vae: Optional[Dict[str, Any]] = None
 ) -> Tuple['VideoDiffusionInfer', Dict[str, Any]]:
@@ -462,7 +457,7 @@ def prepare_runner(
         decode_tile_size: Tile size for decoding (height, width)
         decode_tile_overlap: Tile overlap for decoding (height, width)
         tile_debug: Tile visualization mode (false/encode/decode)
-        attention_mode: Attention computation backend ('sdpa' or 'flash_attn')
+        attention_mode: Attention computation backend ('sdpa', 'flash_attn_2', 'flash_attn_3', 'sageattn_2', or 'sageattn_3')
         torch_compile_args_dit: Optional torch.compile configuration for DiT model
         torch_compile_args_vae: Optional torch.compile configuration for VAE model
         
@@ -507,7 +502,6 @@ def prepare_runner(
         decode_tile_overlap=decode_tile_overlap,
         tile_debug=tile_debug,
         attention_mode=attention_mode,
-        precision=precision,
         torch_compile_args_dit=torch_compile_args_dit,
         torch_compile_args_vae=torch_compile_args_vae
     )
@@ -535,8 +529,8 @@ def load_text_embeddings(script_directory: str, device: torch.device,
         - Memory-efficient embedding preparation
         - Consistent movement logging
     """
-    text_pos_embeds = torch.load(os.path.join(script_directory, 'pos_emb.pt'))
-    text_neg_embeds = torch.load(os.path.join(script_directory, 'neg_emb.pt'))
+    text_pos_embeds = torch.load(os.path.join(script_directory, 'pos_emb.pt'), weights_only=True)
+    text_neg_embeds = torch.load(os.path.join(script_directory, 'neg_emb.pt'), weights_only=True)
     
     text_pos_embeds = manage_tensor(
         tensor=text_pos_embeds,
@@ -825,4 +819,4 @@ def ensure_precision_initialized(
             debug.log(f"Model precision: {', '.join(parts)}", category="precision")
             
     except Exception as e:
-        debug.log(f"Could not log model dtypes: {e}", level="WARNING", category="precision", force=True)
+        debug.log(f"Could not log model dtypes: {e}", level="WARNING", category="precision", force=True)
\ No newline at end of file
diff --git a/src/core/infer.py b/src/core/infer.py
index 2566daf..a0869ca 100644
--- a/src/core/infer.py
+++ b/src/core/infer.py
@@ -154,15 +154,26 @@ class VideoDiffusionInfer():
                     vae_dtype = dtype  # Fallback
 
                 # Use autocast if VAE dtype differs from input dtype
+                # Skip autocast on MPS (only supports bf16, unified memory = no benefit)
+                # Instead, explicitly convert input to model dtype
                 if vae_dtype != sample.dtype:
-                    with torch.autocast(device.type, sample.dtype, enabled=True):
+                    if device.type == 'mps':
+                        # MPS: explicit dtype conversion instead of autocast
+                        sample = sample.to(vae_dtype)
                         if use_sample:
                             latent = self.vae.encode(sample, tiled=self.encode_tiled, tile_size=self.encode_tile_size, 
                                                     tile_overlap=self.encode_tile_overlap).latent
                         else:
-                            # Deterministic vae encode, only used for i2v inference (optionally)
                             latent = self.vae.encode(sample, tiled=self.encode_tiled, tile_size=self.encode_tile_size,
                                                 tile_overlap=self.encode_tile_overlap).posterior.mode().squeeze(2)
+                    else:
+                        with torch.autocast(device.type, sample.dtype, enabled=True):
+                            if use_sample:
+                                latent = self.vae.encode(sample, tiled=self.encode_tiled, tile_size=self.encode_tile_size, 
+                                                        tile_overlap=self.encode_tile_overlap).latent
+                            else:
+                                latent = self.vae.encode(sample, tiled=self.encode_tiled, tile_size=self.encode_tile_size,
+                                                    tile_overlap=self.encode_tile_overlap).posterior.mode().squeeze(2)
                 else:
                     if use_sample:
                         latent = self.vae.encode(sample, tiled=self.encode_tiled, tile_size=self.encode_tile_size, 
@@ -230,13 +241,23 @@ class VideoDiffusionInfer():
                     vae_dtype = dtype  # Fallback
 
                 # Use autocast if VAE dtype differs from latent dtype
+                # Skip autocast on MPS (only supports bf16, unified memory = no benefit)
                 if vae_dtype != latent.dtype:
-                    with torch.autocast(device.type, latent.dtype, enabled=True):
+                    if device.type == 'mps':
+                        # MPS: explicit dtype conversion instead of autocast
+                        latent = latent.to(vae_dtype)
                         sample = self.vae.decode(
                             latent,
                             tiled=self.decode_tiled, tile_size=self.decode_tile_size,
                             tile_overlap=self.decode_tile_overlap
                         ).sample
+                    else:
+                        with torch.autocast(device.type, latent.dtype, enabled=True):
+                            sample = self.vae.decode(
+                                latent,
+                                tiled=self.decode_tiled, tile_size=self.decode_tile_size,
+                                tile_overlap=self.decode_tile_overlap
+                            ).sample
                 else:
                     sample = self.vae.decode(
                         latent,
diff --git a/src/core/model_configuration.py b/src/core/model_configuration.py
index 5e2e808..6129762 100644
--- a/src/core/model_configuration.py
+++ b/src/core/model_configuration.py
@@ -70,13 +70,11 @@ from .model_cache import get_global_cache
 from ..common.config import load_config
 from ..models.video_vae_v3.modules.causal_inflation_lib import InflatedCausalConv3d
 from ..optimization.compatibility import (
-    FP8CompatibleDiT,
+    CompatibleDiT,
     TRITON_AVAILABLE,
-    validate_flash_attention_availability,
-    detect_high_end_system,
-    log_system_capabilities
+    validate_attention_mode
 )
-from ..optimization.blockswap import is_blockswap_enabled, apply_block_swap_to_dit, cleanup_blockswap
+from ..optimization.blockswap import is_blockswap_enabled, validate_blockswap_config, apply_block_swap_to_dit, cleanup_blockswap
 from ..optimization.memory_manager import cleanup_dit, cleanup_vae
 from ..utils.constants import find_model_file
 
@@ -173,7 +171,7 @@ def _describe_attention_mode(attention_mode: Optional[str]) -> str:
     Generate human-readable description of attention mode configuration.
     
     Args:
-        attention_mode: Attention mode string ('sdpa' or 'flash_attn' or 'sa2' or 'sa3')
+        attention_mode: Attention mode string ('sdpa', 'flash_attn_2', 'flash_attn_3', 'sageattn_2', or 'sageattn_3')
         
     Returns:
         Human-readable description string
@@ -183,9 +181,10 @@ def _describe_attention_mode(attention_mode: Optional[str]) -> str:
     
     mode_descriptions = {
         'sdpa': 'PyTorch SDPA',
-        'flash_attn': 'Flash Attention 2',
-        'sa2': 'SageAttention v2',
-        'sa3': 'SageAttention v3'
+        'flash_attn_2': 'Flash Attention 2',
+        'flash_attn_3': 'Flash Attention 3',
+        'sageattn_2': 'SageAttention 2',
+        'sageattn_3': 'SageAttention 3 (Blackwell)'
     }
     
     return mode_descriptions.get(attention_mode, attention_mode)
@@ -440,7 +439,7 @@ def _update_dit_config(
             - dynamic: bool - Enable dynamic shapes
             - dynamo_cache_size_limit: int - Cache size limit
             - dynamo_recompile_limit: int - Recompilation limit
-        attention_mode: Attention computation backend ('sdpa' or 'flash_attn')
+        attention_mode: Attention computation backend ('sdpa', 'flash_attn_2', 'flash_attn_3', 'sageattn_2', or 'sageattn_3')
         debug: Debug instance for logging
         
     Returns:
@@ -749,7 +748,6 @@ def configure_runner(
     decode_tile_overlap: Optional[Tuple[int, int]] = None,
     tile_debug: str = "false",
     attention_mode: str = 'sdpa',
-    precision: str = 'auto',
     torch_compile_args_dit: Optional[Dict[str, Any]] = None,
     torch_compile_args_vae: Optional[Dict[str, Any]] = None
 ) -> Tuple[VideoDiffusionInfer, Dict[str, Any]]:
@@ -776,7 +774,7 @@ def configure_runner(
         decode_tile_size: Tile size for decoding (height, width)
         decode_tile_overlap: Tile overlap for decoding (height, width)
         tile_debug: Tile visualization mode (false/encode/decode)
-        attention_mode: Attention computation backend ('sdpa' or 'flash_attn')
+        attention_mode: Attention computation backend ('sdpa', 'flash_attn_2', 'flash_attn_3', 'sageattn_2', or 'sageattn_3')
         torch_compile_args_dit: Optional torch.compile configuration for DiT model
         torch_compile_args_vae: Optional torch.compile configuration for VAE model
         
@@ -797,9 +795,14 @@ def configure_runner(
     if debug is None:
         raise ValueError("Debug instance must be provided to configure_runner")
     
-    # Log installed attention backends and versions
-    log_system_capabilities(debug)
-
+    # Validate BlockSwap configuration early (before any model loading)
+    block_swap_config = validate_blockswap_config(
+        block_swap_config=block_swap_config,
+        dit_device=ctx['dit_device'],
+        dit_offload_device=ctx.get('dit_offload_device'),
+        debug=debug
+    )
+    
     # Phase 1: Initialize cache and get cached models
     cache_context = _initialize_cache_context(
         dit_cache, vae_cache, dit_id, vae_id, 
@@ -822,9 +825,6 @@ def configure_runner(
         block_swap_config, debug
     )
     
-    # Store precision setting
-    runner._precision = precision
-
     # Phase 4: Setup models (load from cache or create new)
     _setup_models(
         runner, cache_context, dit_model, vae_model, 
@@ -868,7 +868,7 @@ def _configure_runner_settings(
         decode_tile_size: Tile dimensions (height, width) for decoding in pixels
         decode_tile_overlap: Overlap dimensions (height, width) between decoding tiles
         tile_debug: Tile visualization mode (false/encode/decode)
-        attention_mode: Attention computation backend ('sdpa' or 'flash_attn')
+        attention_mode: Attention computation backend ('sdpa', 'flash_attn_2', 'flash_attn_3', 'sageattn_2', or 'sageattn_3')
         torch_compile_args_dit: torch.compile configuration for DiT model or None
         torch_compile_args_vae: torch.compile configuration for VAE model or None
         block_swap_config: BlockSwap configuration for DiT model or None
@@ -906,16 +906,6 @@ def _configure_runner_settings(
     runner._tensor_offload_device = ctx['tensor_offload_device']
     runner._compute_dtype = ctx['compute_dtype']
 
-    # Auto-detection for 5070ti/similar hardware
-    system_opts = detect_high_end_system()
-    if system_opts.get('high_vram', False):
-        if debug:
-            debug.log(f"Detected high-end system optimizations: {system_opts}", category="setup")
-        # Apply recommended settings if not overridden
-        # For example, we might favor speed/quality trade-offs differently
-        # Here we just log it as the user has control via UI, but we could set defaults if they were None
-        pass
-
     runner.debug = debug
 
 
@@ -1186,43 +1176,27 @@ def apply_model_specific_config(model: torch.nn.Module, runner: VideoDiffusionIn
     """
     if is_dit:
         # DiT-specific
-        # Apply FP8 compatibility wrapper with compute_dtype
-        if not isinstance(model, FP8CompatibleDiT):
-            debug.log("Applying FP8/RoPE compatibility wrapper to DiT model", category="setup")
-            debug.start_timer("FP8CompatibleDiT")
+        # Apply compatibility wrapper with compute_dtype
+        if not isinstance(model, CompatibleDiT):
+            debug.log("Applying DiT compatibility wrapper", category="setup")
+            debug.start_timer("CompatibleDiT")
             # Get compute_dtype from runner if available, fallback to bfloat16
             compute_dtype = getattr(runner, '_compute_dtype', torch.bfloat16)
-            model = FP8CompatibleDiT(model, debug, compute_dtype=compute_dtype, skip_conversion=False)
-            debug.end_timer("FP8CompatibleDiT", "FP8/RoPE compatibility wrapper application")
+            model = CompatibleDiT(model, debug, compute_dtype=compute_dtype, skip_conversion=False)
+            debug.end_timer("CompatibleDiT", "Compatibility wrapper application")
         else:
-            debug.log("Reusing existing FP8/RoPE compatibility wrapper", category="reuse")
+            debug.log("Reusing existing DiT compatibility wrapper", category="reuse")
         
         # Apply attention mode and compute_dtype to all FlashAttentionVarlen modules
         if hasattr(runner, '_dit_attention_mode'):
             requested_attention_mode = runner._dit_attention_mode or 'sdpa'
             
             # Validate and get final attention_mode (with warning if fallback needed)
-            attention_mode = validate_flash_attention_availability(requested_attention_mode, debug)
+            attention_mode = validate_attention_mode(requested_attention_mode, debug)
             
-            # Get compute_dtype from runner, override if precision set
-            compute_dtype = getattr(runner, '_compute_dtype', torch.bfloat16)
-
-            # Apply precision override if set
-            precision_override = getattr(runner, '_precision', 'auto')
-            if precision_override == 'fp16':
-                compute_dtype = torch.float16
-            elif precision_override == 'bf16':
-                compute_dtype = torch.bfloat16
-            elif precision_override == 'bf32':
-                 # TF32 is an attribute of the context, not dtype, but we can respect it here or in setup
-                 # For compute dtype, usually bf32 implies tf32 or float32 with tf32
-                 compute_dtype = torch.float32
-                 # We will handle TF32 setting globally elsewhere or here if needed
-
-            # Log final decision prominently
-            mode_desc = _describe_attention_mode(attention_mode)
-            debug.log(f"Using Attention Mode: {mode_desc}", category="info", force=True)
-            debug.log(f"Using Compute Dtype: {compute_dtype}", category="info", force=True)
+            # Get compute_dtype from runner
+            compute_dtype = getattr(runner, '_compute_dtype', torch.bfloat16)            
+            debug.log(f"Applying {attention_mode} attention mode and {compute_dtype} compute dtype to model", category="setup")
             
             # Get the actual model (unwrap if needed)
             actual_model = model.dit_model if hasattr(model, 'dit_model') else model
@@ -1503,4 +1477,4 @@ def _propagate_debug_to_modules(module: torch.nn.Module, debug: 'Debug') -> None
     for name, submodule in module.named_modules():
         if submodule.__class__.__name__ in target_modules:
             if not hasattr(submodule, 'debug'):  # Only set if not already present
-                submodule.debug = debug
+                submodule.debug = debug
\ No newline at end of file
diff --git a/src/data/image/transforms/area_resize.py b/src/data/image/transforms/area_resize.py
index 5873b85..fc025da 100644
--- a/src/data/image/transforms/area_resize.py
+++ b/src/data/image/transforms/area_resize.py
@@ -19,6 +19,7 @@ import torch
 from PIL import Image
 from torchvision.transforms import functional as TVF
 from torchvision.transforms.functional import InterpolationMode
+from ....optimization.memory_manager import is_mps_available
 
 
 class AreaResize:
@@ -31,7 +32,7 @@ class AreaResize:
         self.max_area = max_area
         self.downsample_only = downsample_only
         self.interpolation = interpolation
-        if hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
+        if is_mps_available():
             self.interpolation = InterpolationMode.BILINEAR
 
     def __call__(self, image: Union[torch.Tensor, Image.Image]):
diff --git a/src/data/image/transforms/na_resize.py b/src/data/image/transforms/na_resize.py
index 61a186a..e1111c7 100644
--- a/src/data/image/transforms/na_resize.py
+++ b/src/data/image/transforms/na_resize.py
@@ -18,6 +18,7 @@ from torchvision.transforms import CenterCrop, Compose, InterpolationMode, Resiz
 
 from .area_resize import AreaResize
 from .side_resize import SideResize
+from ....optimization.memory_manager import is_mps_available
 
 def NaResize(
     resolution: int,
@@ -26,7 +27,7 @@ def NaResize(
     max_resolution: int = 0,
     interpolation: InterpolationMode = InterpolationMode.BICUBIC,
 ):
-    Interpolation = InterpolationMode.BILINEAR if (hasattr(torch.backends, 'mps') and torch.backends.mps.is_available()) else interpolation
+    Interpolation = InterpolationMode.BILINEAR if is_mps_available() else interpolation
     if mode == "area":
         return AreaResize(
             max_area=resolution**2,
diff --git a/src/data/image/transforms/side_resize.py b/src/data/image/transforms/side_resize.py
index 6d5273f..01362ae 100644
--- a/src/data/image/transforms/side_resize.py
+++ b/src/data/image/transforms/side_resize.py
@@ -17,6 +17,7 @@ import torch
 from PIL import Image
 from torchvision.transforms import InterpolationMode
 from torchvision.transforms import functional as TVF
+from ....optimization.memory_manager import is_mps_available
 
 class SideResize:
     def __init__(
@@ -30,7 +31,7 @@ class SideResize:
         self.max_size = max_size
         self.downsample_only = downsample_only
         self.interpolation = interpolation
-        if hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
+        if is_mps_available():
             self.interpolation = InterpolationMode.BILINEAR
 
     def __call__(self, image: Union[torch.Tensor, Image.Image]):
diff --git a/src/interfaces/dit_model_loader.py b/src/interfaces/dit_model_loader.py
index 526754d..9d8204e 100644
--- a/src/interfaces/dit_model_loader.py
+++ b/src/interfaces/dit_model_loader.py
@@ -66,7 +66,8 @@ class SeedVR2LoadDiTModel(io.ComfyNode):
                         "• 3B model: 0-32 blocks\n"
                         "• 7B model: 0-36 blocks\n"
                         "\n"
-                        "Requires offload_device to be set and different from device."
+                        "Requires offload_device to be set and different from device.\n"
+                        "Not available on macOS (unified memory architecture)."
                     )
                 ),
                 io.Boolean.Input("swap_io_components",
@@ -74,7 +75,8 @@ class SeedVR2LoadDiTModel(io.ComfyNode):
                     optional=True,
                     tooltip=(
                         "Offload input/output embeddings and normalization layers to reduce VRAM.\n"
-                        "Requires offload_device to be set and different from device."
+                        "Requires offload_device to be set and different from device.\n"
+                        "Not available on macOS (unified memory architecture)."
                     )
                 ),
                 io.Combo.Input("offload_device",
@@ -100,15 +102,16 @@ class SeedVR2LoadDiTModel(io.ComfyNode):
                     )
                 ),
                 io.Combo.Input("attention_mode",
-                    options=["sdpa", "flash_attn", "sa2", "sa3"],
+                    options=["sdpa", "flash_attn_2", "flash_attn_3", "sageattn_2", "sageattn_3"],
                     default="sdpa",
                     optional=True,
                     tooltip=(
                         "Attention computation backend:\n"
                         "• sdpa: PyTorch scaled_dot_product_attention (default, stable, always available)\n"
-                        "• flash_attn: Flash Attention 2 (faster on supported hardware, requires flash-attn package)\n"
-                        "• sa2: SageAttention v2 (requires sageattention package)\n"
-                        "• sa3: SageAttention v3 (requires sageattention package)\n"
+                        "• flash_attn_2: Flash Attention 2 (Ampere+, requires flash-attn package)\n"
+                        "• flash_attn_3: Flash Attention 3 (Hopper+, requires flash-attn with FA3 support)\n"
+                        "• sageattn_2: SageAttention 2 (requires sageattention package)\n"
+                        "• sageattn_3: SageAttention 3 (Blackwell/RTX 50xx only, requires sageattn3 package)\n"
                         "\n"
                         "SDPA is recommended - stable and works everywhere.\n"
                         "Flash Attention and SageAttention provide speedup through optimized CUDA kernels on compatible GPUs."
@@ -144,23 +147,15 @@ class SeedVR2LoadDiTModel(io.ComfyNode):
             cache_model: Whether to keep model loaded between runs
             blocks_to_swap: Number of transformer blocks to swap (requires offload_device != device)
             swap_io_components: Whether to offload I/O components (requires offload_device != device)
-            attention_mode: Attention computation backend ('sdpa' or 'flash_attn')
+            attention_mode: Attention computation backend ('sdpa', 'flash_attn_2', 'flash_attn_3', 'sageattn_2', or 'sageattn_3')
             torch_compile_args: Optional torch.compile configuration from settings node
             
         Returns:
             NodeOutput containing configuration dictionary for SeedVR2 main node
             
         Raises:
-            ValueError: If BlockSwap is enabled but offload_device is invalid
+            ValueError: If cache_model is enabled but offload_device is not set
         """
-        # Validate BlockSwap configuration
-        if (blocks_to_swap > 0 or swap_io_components) and (offload_device == "none" or offload_device == device):
-            raise ValueError(
-                "BlockSwap requires offload_device to be set and different from device. "
-                f"Current: device='{device}', offload_device='{offload_device}'. "
-                "Please set offload_device to a different device (e.g., 'cpu' or another GPU)."
-            )
-        
         # Validate cache_model configuration
         if cache_model and offload_device == "none":
             raise ValueError(
diff --git a/src/interfaces/video_upscaler.py b/src/interfaces/video_upscaler.py
index 2116c91..54d6399 100644
--- a/src/interfaces/video_upscaler.py
+++ b/src/interfaces/video_upscaler.py
@@ -204,18 +204,6 @@ class SeedVR2VideoUpscaler(io.ComfyNode):
                         "• 'cuda:X': Offload to another GPU (good balance if available, faster than CPU)"
                     )
                 ),
-                io.Combo.Input("precision",
-                    options=["auto", "fp16", "bf16", "bf32"],
-                    default="auto",
-                    optional=True,
-                    tooltip=(
-                        "Precision for main generation process (default: auto).\n"
-                        "• auto: Automatically select based on device capabilities\n"
-                        "• fp16: Half precision (fastest, standard)\n"
-                        "• bf16: BFloat16 (better dynamic range, requires Ampere+ GPU)\n"
-                        "• bf32: Float32 with TF32 enabled (Ampere+ GPU)"
-                    )
-                ),
                 io.Boolean.Input("enable_debug",
                     default=False,
                     optional=True,
@@ -239,7 +227,7 @@ class SeedVR2VideoUpscaler(io.ComfyNode):
                 uniform_batch_size: bool = False, temporal_overlap: int = 0, prepend_frames: int = 0,
                 color_correction: str = "wavelet", input_noise_scale: float = 0.0,
                 latent_noise_scale: float = 0.0, offload_device: str = "none", 
-                precision: str = "auto", enable_debug: bool = False) -> io.NodeOutput:
+                enable_debug: bool = False) -> io.NodeOutput:
         """
         Execute SeedVR2 video upscaling with progress reporting
         
@@ -354,10 +342,6 @@ class SeedVR2VideoUpscaler(io.ComfyNode):
         attention_mode = dit.get("attention_mode", "sdpa")
         vae_cache = vae.get("cache_model", False)
 
-        # Override attention mode if specified in dit config but allow validation later
-        if "attention_mode" in dit:
-             attention_mode = dit["attention_mode"]
-
         # BlockSwap configuration - construct from individual values
         blocks_to_swap = dit.get("blocks_to_swap", 0)
         swap_io_components = dit.get("swap_io_components", False)
@@ -365,13 +349,12 @@ class SeedVR2VideoUpscaler(io.ComfyNode):
 
         block_swap_config = None
         if blocks_to_swap > 0 or swap_io_components:
-            # Convert offload device string to torch.device for BlockSwap
+            block_swap_config = {
+                "blocks_to_swap": blocks_to_swap,
+                "swap_io_components": swap_io_components,
+            }
             if dit_offload_str != "none":
-                block_swap_config = {
-                    "blocks_to_swap": blocks_to_swap,
-                    "swap_io_components": swap_io_components,
-                    "offload_device": torch.device(dit_offload_str)
-                }
+                block_swap_config["offload_device"] = torch.device(dit_offload_str)
 
         # Device configuration for offloading - convert "none" to None, else torch.device
         vae_offload_str = vae.get("offload_device", "none")
@@ -447,7 +430,6 @@ class SeedVR2VideoUpscaler(io.ComfyNode):
                 decode_tile_overlap=(decode_tile_overlap, decode_tile_overlap),
                 tile_debug=tile_debug,
                 attention_mode=attention_mode,
-                precision=precision,
                 torch_compile_args_dit=dit_torch_compile_args,
                 torch_compile_args_vae=vae_torch_compile_args
             )
diff --git a/src/models/dit_3b/attention.py b/src/models/dit_3b/attention.py
index cd6741b..03e3c63 100644
--- a/src/models/dit_3b/attention.py
+++ b/src/models/dit_3b/attention.py
@@ -15,16 +15,14 @@
 import torch
 import torch.nn.functional as F
 
-# Import flash_attn with automatic fallback from compatibility layer
-from ...optimization.compatibility import flash_attn_varlen_func, FLASH_ATTN_AVAILABLE, SAGE_ATTN_AVAILABLE
+# Import flash/sage attn with automatic fallback from compatibility layer
+from ...optimization.compatibility import (
+    call_flash_attn_2_varlen, call_flash_attn_3_varlen,
+    call_sage_attn_2_varlen, call_sage_attn_3_varlen
+)
 
 from torch import nn
 
-# Safe import for SageAttention
-try:
-    import sageattention
-except ImportError:
-    sageattention = None
 
 def pytorch_varlen_attention(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q=None, max_seqlen_k=None, dropout_p=0.0, softmax_scale=None, causal=False, deterministic=False):
     """
@@ -66,81 +64,6 @@ def pytorch_varlen_attention(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q=N
     return torch.cat(output_splits, dim=0)
 
 
-@torch._dynamo.disable
-def _call_flash_attn_varlen_func(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, **kwargs):
-    """
-    Wrapper for flash_attn_varlen_func that handles tensor-to-scalar conversion.
-    
-    This function is excluded from torch.compile because:
-    1. flash_attn is a C++ extension that can't be compiled anyway
-    2. It requires Python int scalars for max_seqlen parameters
-    3. Disabling compilation here keeps the rest of the model compilable
-    """
-    if not FLASH_ATTN_AVAILABLE:
-        raise ImportError("flash_attn is not available")
-    
-    # Convert tensor max_seqlen to Python int if needed
-    if torch.is_tensor(max_seqlen_q):
-        max_seqlen_q = int(max_seqlen_q.item())
-    if torch.is_tensor(max_seqlen_k):
-        max_seqlen_k = int(max_seqlen_k.item())
-    
-    return flash_attn_varlen_func(
-        q=q,
-        k=k,
-        v=v,
-        cu_seqlens_q=cu_seqlens_q,
-        cu_seqlens_k=cu_seqlens_k,
-        max_seqlen_q=max_seqlen_q,
-        max_seqlen_k=max_seqlen_k,
-        **kwargs
-    )
-
-@torch._dynamo.disable
-def _call_sage_attn_varlen_func(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, is_causal=False, implementation="sa2"):
-    """
-    Wrapper for SageAttention variable length function.
-
-    Args:
-        implementation: "sa2" (SageAttention v2) or "sa3" (SageAttention v3)
-    """
-    if not SAGE_ATTN_AVAILABLE:
-        raise ImportError("SageAttention is not available")
-
-    # SageAttention expects q, k, v as (total_tokens, heads, head_dim)
-    # The input q, k, v here are (total_tokens, heads, head_dim)
-
-    # Convert tensor max_seqlen to Python int if needed
-    if torch.is_tensor(max_seqlen_q):
-        max_seqlen_q = int(max_seqlen_q.item())
-    if torch.is_tensor(max_seqlen_k):
-        max_seqlen_k = int(max_seqlen_k.item())
-
-    # Ensure tensors are contiguous
-    q = q.contiguous()
-    k = k.contiguous()
-    v = v.contiguous()
-
-    # SageAttention API usage
-    # sageattn_varlen(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, is_causal, sm_scale)
-    try:
-         from sageattention import sageattn_varlen
-    except ImportError:
-         # Fallback or error
-         raise ImportError("sageattn_varlen not found in sageattention package")
-
-    # Check if sm_scale is needed (usually 1/sqrt(head_dim))
-    sm_scale = 1.0 / (q.shape[-1] ** 0.5)
-
-    # Calling sageattn_varlen
-    # Signature assumptions: q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, is_causal, sm_scale
-    if not hasattr(_call_sage_attn_varlen_func, "_logged"):
-        print(f"🚀 Executing SageAttention ({implementation}) kernel for the first time")
-        _call_sage_attn_varlen_func._logged = True
-
-    return sageattn_varlen(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, is_causal, sm_scale)
-
-
 class TorchAttention(nn.Module):
     def tflops(self, args, kwargs, output) -> float:
         assert len(args) == 0 or len(args) > 2, "query, key should both provided by args / kwargs"
@@ -156,12 +79,16 @@ class TorchAttention(nn.Module):
 
 class FlashAttentionVarlen(nn.Module):
     """
-    Variable-length attention with configurable backend (Flash Attention or PyTorch SDPA).
+    Variable-length attention with configurable backend.
     
-    Backend selection is validated during model configuration.
-    Compilation behavior:
-    - SDPA: Fully compilable, optimal performance
-    - Flash Attention: Uses @torch._dynamo.disable wrapper (C++ extension)
+    Supported backends:
+    - sdpa: PyTorch SDPA (fully compilable, always available)
+    - flash_attn_2: Flash Attention 2 (Ampere+)
+    - flash_attn_3: Flash Attention 3 (Hopper+)
+    - sageattn_2: SageAttention 2
+    - sageattn_3: SageAttention 3 (Blackwell/RTX 50xx)
+    
+    All non-SDPA backends use @torch._dynamo.disable wrapper (C++ extensions).
     """
 
     def __init__(self, attention_mode: str = 'sdpa', compute_dtype: torch.dtype = None):
@@ -169,7 +96,7 @@ class FlashAttentionVarlen(nn.Module):
         Initialize with specified attention backend.
         
         Args:
-            attention_mode: 'flash_attn' or 'sdpa' (validated externally by validate_flash_attention_availability)
+            attention_mode: 'sdpa', 'flash_attn_2', 'flash_attn_3', 'sageattn_2', or 'sageattn_3'
             compute_dtype: Compute dtype for attention (set by pipeline, defaults to None for auto-detection)
         """
         super().__init__()
@@ -193,20 +120,25 @@ class FlashAttentionVarlen(nn.Module):
             k = k.to(self.compute_dtype)
             v = v.to(self.compute_dtype)
         
-        if self.attention_mode == 'flash_attn':
-            return _call_flash_attn_varlen_func(
+        if self.attention_mode == 'flash_attn_3':
+            return call_flash_attn_3_varlen(
                 q, k, v, cu_seqlens_q, cu_seqlens_k, 
                 max_seqlen_q, max_seqlen_k, **kwargs
             )
-        elif self.attention_mode in ['sa2', 'sa3']:
-            # Use SageAttention
-            # Extract causal flag if present in kwargs, default to False
-            is_causal = kwargs.get('causal', False)
-            return _call_sage_attn_varlen_func(
+        elif self.attention_mode == 'flash_attn_2':
+            return call_flash_attn_2_varlen(
+                q, k, v, cu_seqlens_q, cu_seqlens_k, 
+                max_seqlen_q, max_seqlen_k, **kwargs
+            )
+        elif self.attention_mode == 'sageattn_3':
+            return call_sage_attn_3_varlen(
                 q, k, v, cu_seqlens_q, cu_seqlens_k,
-                max_seqlen_q, max_seqlen_k,
-                is_causal=is_causal,
-                implementation=self.attention_mode
+                max_seqlen_q, max_seqlen_k, **kwargs
+            )
+        elif self.attention_mode == 'sageattn_2':
+            return call_sage_attn_2_varlen(
+                q, k, v, cu_seqlens_q, cu_seqlens_k,
+                max_seqlen_q, max_seqlen_k, **kwargs
             )
         else:
             # PyTorch SDPA
diff --git a/src/models/dit_7b/attention.py b/src/models/dit_7b/attention.py
index cd6741b..03e3c63 100644
--- a/src/models/dit_7b/attention.py
+++ b/src/models/dit_7b/attention.py
@@ -15,16 +15,14 @@
 import torch
 import torch.nn.functional as F
 
-# Import flash_attn with automatic fallback from compatibility layer
-from ...optimization.compatibility import flash_attn_varlen_func, FLASH_ATTN_AVAILABLE, SAGE_ATTN_AVAILABLE
+# Import flash/sage attn with automatic fallback from compatibility layer
+from ...optimization.compatibility import (
+    call_flash_attn_2_varlen, call_flash_attn_3_varlen,
+    call_sage_attn_2_varlen, call_sage_attn_3_varlen
+)
 
 from torch import nn
 
-# Safe import for SageAttention
-try:
-    import sageattention
-except ImportError:
-    sageattention = None
 
 def pytorch_varlen_attention(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q=None, max_seqlen_k=None, dropout_p=0.0, softmax_scale=None, causal=False, deterministic=False):
     """
@@ -66,81 +64,6 @@ def pytorch_varlen_attention(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q=N
     return torch.cat(output_splits, dim=0)
 
 
-@torch._dynamo.disable
-def _call_flash_attn_varlen_func(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, **kwargs):
-    """
-    Wrapper for flash_attn_varlen_func that handles tensor-to-scalar conversion.
-    
-    This function is excluded from torch.compile because:
-    1. flash_attn is a C++ extension that can't be compiled anyway
-    2. It requires Python int scalars for max_seqlen parameters
-    3. Disabling compilation here keeps the rest of the model compilable
-    """
-    if not FLASH_ATTN_AVAILABLE:
-        raise ImportError("flash_attn is not available")
-    
-    # Convert tensor max_seqlen to Python int if needed
-    if torch.is_tensor(max_seqlen_q):
-        max_seqlen_q = int(max_seqlen_q.item())
-    if torch.is_tensor(max_seqlen_k):
-        max_seqlen_k = int(max_seqlen_k.item())
-    
-    return flash_attn_varlen_func(
-        q=q,
-        k=k,
-        v=v,
-        cu_seqlens_q=cu_seqlens_q,
-        cu_seqlens_k=cu_seqlens_k,
-        max_seqlen_q=max_seqlen_q,
-        max_seqlen_k=max_seqlen_k,
-        **kwargs
-    )
-
-@torch._dynamo.disable
-def _call_sage_attn_varlen_func(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, is_causal=False, implementation="sa2"):
-    """
-    Wrapper for SageAttention variable length function.
-
-    Args:
-        implementation: "sa2" (SageAttention v2) or "sa3" (SageAttention v3)
-    """
-    if not SAGE_ATTN_AVAILABLE:
-        raise ImportError("SageAttention is not available")
-
-    # SageAttention expects q, k, v as (total_tokens, heads, head_dim)
-    # The input q, k, v here are (total_tokens, heads, head_dim)
-
-    # Convert tensor max_seqlen to Python int if needed
-    if torch.is_tensor(max_seqlen_q):
-        max_seqlen_q = int(max_seqlen_q.item())
-    if torch.is_tensor(max_seqlen_k):
-        max_seqlen_k = int(max_seqlen_k.item())
-
-    # Ensure tensors are contiguous
-    q = q.contiguous()
-    k = k.contiguous()
-    v = v.contiguous()
-
-    # SageAttention API usage
-    # sageattn_varlen(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, is_causal, sm_scale)
-    try:
-         from sageattention import sageattn_varlen
-    except ImportError:
-         # Fallback or error
-         raise ImportError("sageattn_varlen not found in sageattention package")
-
-    # Check if sm_scale is needed (usually 1/sqrt(head_dim))
-    sm_scale = 1.0 / (q.shape[-1] ** 0.5)
-
-    # Calling sageattn_varlen
-    # Signature assumptions: q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, is_causal, sm_scale
-    if not hasattr(_call_sage_attn_varlen_func, "_logged"):
-        print(f"🚀 Executing SageAttention ({implementation}) kernel for the first time")
-        _call_sage_attn_varlen_func._logged = True
-
-    return sageattn_varlen(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, is_causal, sm_scale)
-
-
 class TorchAttention(nn.Module):
     def tflops(self, args, kwargs, output) -> float:
         assert len(args) == 0 or len(args) > 2, "query, key should both provided by args / kwargs"
@@ -156,12 +79,16 @@ class TorchAttention(nn.Module):
 
 class FlashAttentionVarlen(nn.Module):
     """
-    Variable-length attention with configurable backend (Flash Attention or PyTorch SDPA).
+    Variable-length attention with configurable backend.
     
-    Backend selection is validated during model configuration.
-    Compilation behavior:
-    - SDPA: Fully compilable, optimal performance
-    - Flash Attention: Uses @torch._dynamo.disable wrapper (C++ extension)
+    Supported backends:
+    - sdpa: PyTorch SDPA (fully compilable, always available)
+    - flash_attn_2: Flash Attention 2 (Ampere+)
+    - flash_attn_3: Flash Attention 3 (Hopper+)
+    - sageattn_2: SageAttention 2
+    - sageattn_3: SageAttention 3 (Blackwell/RTX 50xx)
+    
+    All non-SDPA backends use @torch._dynamo.disable wrapper (C++ extensions).
     """
 
     def __init__(self, attention_mode: str = 'sdpa', compute_dtype: torch.dtype = None):
@@ -169,7 +96,7 @@ class FlashAttentionVarlen(nn.Module):
         Initialize with specified attention backend.
         
         Args:
-            attention_mode: 'flash_attn' or 'sdpa' (validated externally by validate_flash_attention_availability)
+            attention_mode: 'sdpa', 'flash_attn_2', 'flash_attn_3', 'sageattn_2', or 'sageattn_3'
             compute_dtype: Compute dtype for attention (set by pipeline, defaults to None for auto-detection)
         """
         super().__init__()
@@ -193,20 +120,25 @@ class FlashAttentionVarlen(nn.Module):
             k = k.to(self.compute_dtype)
             v = v.to(self.compute_dtype)
         
-        if self.attention_mode == 'flash_attn':
-            return _call_flash_attn_varlen_func(
+        if self.attention_mode == 'flash_attn_3':
+            return call_flash_attn_3_varlen(
                 q, k, v, cu_seqlens_q, cu_seqlens_k, 
                 max_seqlen_q, max_seqlen_k, **kwargs
             )
-        elif self.attention_mode in ['sa2', 'sa3']:
-            # Use SageAttention
-            # Extract causal flag if present in kwargs, default to False
-            is_causal = kwargs.get('causal', False)
-            return _call_sage_attn_varlen_func(
+        elif self.attention_mode == 'flash_attn_2':
+            return call_flash_attn_2_varlen(
+                q, k, v, cu_seqlens_q, cu_seqlens_k, 
+                max_seqlen_q, max_seqlen_k, **kwargs
+            )
+        elif self.attention_mode == 'sageattn_3':
+            return call_sage_attn_3_varlen(
                 q, k, v, cu_seqlens_q, cu_seqlens_k,
-                max_seqlen_q, max_seqlen_k,
-                is_causal=is_causal,
-                implementation=self.attention_mode
+                max_seqlen_q, max_seqlen_k, **kwargs
+            )
+        elif self.attention_mode == 'sageattn_2':
+            return call_sage_attn_2_varlen(
+                q, k, v, cu_seqlens_q, cu_seqlens_k,
+                max_seqlen_q, max_seqlen_k, **kwargs
             )
         else:
             # PyTorch SDPA
diff --git a/src/optimization/blockswap.py b/src/optimization/blockswap.py
index 33d5b75..0ecbdb6 100644
--- a/src/optimization/blockswap.py
+++ b/src/optimization/blockswap.py
@@ -47,6 +47,78 @@ def is_blockswap_enabled(config: Optional[Dict[str, Any]]) -> bool:
     return blocks_to_swap > 0 or swap_io_components
 
 
+def validate_blockswap_config(
+    block_swap_config: Optional[Dict[str, Any]],
+    dit_device: 'torch.device',
+    dit_offload_device: Optional['torch.device'],
+    debug: 'Debug'
+) -> Optional[Dict[str, Any]]:
+    """
+    Validate and potentially modify BlockSwap configuration.
+    
+    Performs platform-specific validation and configuration adjustment:
+    - On macOS (MPS): Auto-disables BlockSwap since unified memory makes it meaningless
+    - On other platforms: Validates that offload_device is properly configured
+    
+    This is the single authoritative validation point for BlockSwap configuration,
+    called early in configure_runner() before any model loading.
+    
+    Args:
+        block_swap_config: BlockSwap configuration dictionary (may be None)
+        dit_device: Target device for DiT model inference
+        dit_offload_device: Device for offloading DiT blocks (may be None)
+        debug: Debug instance for logging warnings/errors
+        
+    Returns:
+        Validated/modified block_swap_config (may be None or modified copy)
+        
+    Raises:
+        ValueError: If BlockSwap is enabled but offload_device is invalid (non-MPS only)
+    """
+    if not is_blockswap_enabled(block_swap_config):
+        return block_swap_config
+    
+    blocks_to_swap = block_swap_config.get("blocks_to_swap", 0)
+    swap_io_components = block_swap_config.get("swap_io_components", False)
+    
+    # Check for macOS unified memory - BlockSwap is meaningless there
+    if dit_device.type == "mps":
+        debug.log(
+            f"BlockSwap disabled: macOS uses unified memory (no separate VRAM/RAM). "
+            f"Ignoring blocks_to_swap={blocks_to_swap}, swap_io_components={swap_io_components}",
+            level="WARNING", category="blockswap", force=True
+        )
+        # Return disabled config
+        return {
+            **block_swap_config,
+            "blocks_to_swap": 0,
+            "swap_io_components": False
+        }
+    
+    # Validate offload_device is set and different from dit_device
+    offload_device_valid = (
+        dit_offload_device is not None and 
+        str(dit_offload_device) != str(dit_device)
+    )
+    
+    if not offload_device_valid:
+        config_details = []
+        if blocks_to_swap > 0:
+            config_details.append(f"blocks_to_swap={blocks_to_swap}")
+        if swap_io_components:
+            config_details.append("swap_io_components=True")
+        
+        offload_str = str(dit_offload_device) if dit_offload_device else "none"
+        raise ValueError(
+            f"BlockSwap enabled ({', '.join(config_details)}) but dit_offload_device is invalid. "
+            f"Current: device='{dit_device}', dit_offload_device='{offload_str}'. "
+            f"BlockSwap requires offload_device on the DiT Model to be set and different from device. "
+            f"Set --dit_offload_device cpu or disable BlockSwap."
+        )
+    
+    return block_swap_config
+
+
 # Timing helpers marked to skip torch.compile tracing
 # These functions are excluded from Dynamo's graph tracing to avoid warnings
 # about non-traceable builtins like time.time(), but they still execute normally
@@ -125,7 +197,7 @@ def apply_block_swap_to_dit(
     
     debug.start_timer("apply_blockswap")
 
-    # Get the actual model (handle FP8CompatibleDiT wrapper)
+    # Get the actual model (handle CompatibleDiT wrapper)
     model = runner.dit
     if hasattr(model, "dit_model"):
         model = model.dit_model
@@ -723,7 +795,7 @@ def set_blockswap_bypass(runner, bypass: bool, debug):
     if not hasattr(runner, "_blockswap_active") or not runner._blockswap_active:
         return
     
-    # Get the actual model (handle FP8CompatibleDiT wrapper)
+    # Get the actual model (handle CompatibleDiT wrapper)
     model = runner.dit
     if hasattr(model, "dit_model"):
         model = model.dit_model
@@ -758,7 +830,7 @@ def cleanup_blockswap(runner, keep_state_for_cache=False):
     
     debug = runner.debug
     
-    # Get the actual model (handle FP8CompatibleDiT wrapper)
+    # Get the actual model (handle CompatibleDiT wrapper)
     model = runner.dit
     if hasattr(model, "dit_model"):
         model = model.dit_model
@@ -786,7 +858,7 @@ def cleanup_blockswap(runner, keep_state_for_cache=False):
         return
 
     # Full cleanup when not caching
-    # Get the actual model (handle FP8CompatibleDiT wrapper)
+    # Get the actual model (handle CompatibleDiT wrapper)
     model = runner.dit
     if hasattr(model, "dit_model"):
         model = model.dit_model
diff --git a/src/optimization/compatibility.py b/src/optimization/compatibility.py
index 23509a9..cb386c8 100644
--- a/src/optimization/compatibility.py
+++ b/src/optimization/compatibility.py
@@ -5,9 +5,11 @@ Contains FP8/FP16 compatibility layers and wrappers for different model architec
 Extracted from: seedvr2.py (lines 1045-1630)
 """
 
-# Triton compatibility shim for bitsandbytes 0.45+ with triton 3.0+
-# Must be called before any diffusers import
+# Compatibility shims - Must run before any torch/diffusers import
 import sys
+import types
+import importlib.machinery
+
 
 def ensure_triton_compat():
     """Create minimal triton.ops stubs only if missing, to allow bitsandbytes import."""
@@ -20,8 +22,6 @@ def ensure_triton_compat():
     except (ImportError, ModuleNotFoundError, AttributeError):
         pass
     
-    import types
-    
     if 'triton.ops' not in sys.modules:
         sys.modules['triton.ops'] = types.ModuleType('triton.ops')
     
@@ -32,102 +32,519 @@ def ensure_triton_compat():
     sys.modules['triton.ops'].matmul_perf_model = matmul_perf
     sys.modules['triton.ops.matmul_perf_model'] = matmul_perf
 
-# Run immediately on import
+
+def ensure_flash_attn_safe():
+    """
+    Pre-test flash_attn package; stub if DLL is broken.
+    Prevents diffusers from crashing when flash_attn has broken DLLs.
+    """
+    if 'flash_attn' in sys.modules:
+        return  # Already loaded
+    
+    try:
+        import flash_attn
+    except (ImportError, OSError):
+        # DLL broken or not installed - create stub with proper __spec__
+        stub = types.ModuleType('flash_attn')
+        stub.__spec__ = importlib.machinery.ModuleSpec('flash_attn', None)
+        stub.__file__ = None
+        stub.__path__ = []
+        stub.__loader__ = None
+        # Provide attributes that diffusers/transformers import
+        stub.flash_attn_func = None
+        stub.flash_attn_varlen_func = None
+        sys.modules['flash_attn'] = stub
+
+
+def ensure_xformers_flash_compat():
+    """
+    Pre-test xformers._C_flashattention; stub if DLL is broken.
+    Prevents xformers.ops.fmha.flash from crashing on import.
+    """
+    if 'xformers._C_flashattention' in sys.modules:
+        return  # Already loaded
+    
+    try:
+        from xformers import _C_flashattention  # noqa: F401
+    except (ImportError, OSError):
+        # DLL broken or not installed - create stub with proper __spec__
+        class _FailingStub(types.ModuleType):
+            """Stub that lets xformers gracefully disable its flash backend."""
+            def __getattr__(self, name):
+                raise ImportError("_C_flashattention unavailable")
+        
+        stub = _FailingStub('xformers._C_flashattention')
+        stub.__spec__ = importlib.machinery.ModuleSpec('xformers._C_flashattention', None)
+        stub.__file__ = None
+        stub.__path__ = []
+        stub.__loader__ = None
+        sys.modules['xformers._C_flashattention'] = stub
+
+
+def ensure_bitsandbytes_safe():
+    """
+    Pre-test bitsandbytes; stub if broken to prevent import conflicts.
+    
+    On some systems (e.g., ROCm without proper binaries), bitsandbytes registers
+    PyTorch kernels during import then fails. If another node already triggered 
+    this partial load, re-importing causes kernel registration conflicts.
+    
+    This shim catches such failures and stubs the module so diffusers can load
+    gracefully without bitsandbytes quantization support.
+    """
+    if 'bitsandbytes' in sys.modules:
+        return  # Already loaded or stubbed
+    
+    try:
+        import bitsandbytes
+        # Success - bitsandbytes works, other nodes can use it
+    except (ImportError, OSError, RuntimeError):
+        # Installation broken or not present - create stub
+        stub = types.ModuleType('bitsandbytes')
+        stub.__spec__ = importlib.machinery.ModuleSpec('bitsandbytes', None)
+        stub.__file__ = None
+        stub.__path__ = []
+        stub.__version__ = "0.0.0"
+        sys.modules['bitsandbytes'] = stub
+
+
+# Run all shims immediately on import, before torch/diffusers
 ensure_triton_compat()
+ensure_flash_attn_safe()
+ensure_xformers_flash_compat()
+ensure_bitsandbytes_safe()
 
 
 import torch
-import types
 import os
-from typing import Dict, Any, Optional
 
 
-# Flash Attention & Triton Compatibility Layer
-# 1. Flash Attention - speedup for attention operations
+# Flash/Sage Attention & Triton Compatibility Layer
+
+# 1. Flash Attention 3 (Hopper+, faster, no dropout/window support)
+flash_attn_3_varlen_func = None
+FLASH_ATTN_3_AVAILABLE = False
 try:
-    from flash_attn import flash_attn_varlen_func
-    FLASH_ATTN_AVAILABLE = True
-except ImportError:
-    flash_attn_varlen_func = None
-    FLASH_ATTN_AVAILABLE = False
+    import flash_attn_interface
+    flash_attn_3_varlen_func = flash_attn_interface.flash_attn_varlen_func
+    FLASH_ATTN_3_AVAILABLE = True
+except (ImportError, AttributeError, OSError):
+    pass
 
-# 1.1 SageAttention - speedup for attention operations
+# 2. Flash Attention 2 (wider compatibility, supports dropout/window)
+flash_attn_2_varlen_func = None
+FLASH_ATTN_2_AVAILABLE = False
 try:
-    import sageattention
-    SAGE_ATTN_AVAILABLE = True
+    from flash_attn import flash_attn_varlen_func as _fa2_varlen
+    import flash_attn_2_cuda  # noqa: F401
+    flash_attn_2_varlen_func = _fa2_varlen
+    FLASH_ATTN_2_AVAILABLE = True
+except (ImportError, AttributeError, OSError):
+    pass
+
+FLASH_ATTN_AVAILABLE = FLASH_ATTN_2_AVAILABLE or FLASH_ATTN_3_AVAILABLE
+
+# 3. SageAttention 2 (varlen support)
+sageattn_varlen = None
+SAGE_ATTN_2_AVAILABLE = False
+try:
+    from sageattention import sageattn_varlen as _sa2_varlen
+    sageattn_varlen = _sa2_varlen
+    SAGE_ATTN_2_AVAILABLE = True
+except (ImportError, AttributeError, OSError):
+    pass
+
+# 4. SageAttention 3 / Blackwell (RTX 50xx only, batched attention)
+sageattn_blackwell = None
+SAGE_ATTN_3_AVAILABLE = False
+try:
+    from sageattn3 import sageattn3_blackwell as _sa3_blackwell
+    sageattn_blackwell = _sa3_blackwell
+    SAGE_ATTN_3_AVAILABLE = True
+except (ImportError, AttributeError, OSError):
     try:
-        from sageattention import sageattn_varlen
-        # Basic check to see if it's functional or mock
-        SAGE_ATTN_VARLEN_AVAILABLE = True
-    except ImportError:
-        SAGE_ATTN_VARLEN_AVAILABLE = False
-except ImportError:
-    SAGE_ATTN_AVAILABLE = False
-    SAGE_ATTN_VARLEN_AVAILABLE = False
+        from sageattention import sageattn_blackwell as _sa3_blackwell
+        sageattn_blackwell = _sa3_blackwell
+        SAGE_ATTN_3_AVAILABLE = True
+    except (ImportError, AttributeError, OSError):
+        pass
+
+SAGE_ATTN_AVAILABLE = SAGE_ATTN_2_AVAILABLE or SAGE_ATTN_3_AVAILABLE
 
 
-def validate_flash_attention_availability(requested_mode: str, debug=None) -> str:
+def validate_attention_mode(requested_mode: str, debug=None) -> str:
     """
-    Validate attention mode availability and warn if fallback needed.
+    Validate attention mode availability with automatic fallback.
     
     Args:
-        requested_mode: 'flash_attn', 'sdpa', 'sd2', or 'sd3'
+        requested_mode: 'sdpa', 'flash_attn_2', 'flash_attn_3', 'sageattn_2', or 'sageattn_3'
         debug: Optional debug instance for logging
         
     Returns:
-        Validated mode
+        Validated mode that is available
     """
-    if requested_mode == 'flash_attn' and not FLASH_ATTN_AVAILABLE:
+    # Flash Attention 3
+    if requested_mode == 'flash_attn_3':
+        if FLASH_ATTN_3_AVAILABLE:
+            return requested_mode
+        if FLASH_ATTN_2_AVAILABLE:
+            if debug:
+                debug.log(
+                    "Flash Attention 3 not available (requires Hopper+ GPU and flash-attn with FA3 support).\n"
+                    "Falling back to Flash Attention 2.",
+                    level="WARNING", category="setup", force=True
+                )
+            return 'flash_attn_2'
         error_msg = (
-            f"Cannot use 'flash_attn' attention mode: Flash Attention is not installed.\n"
-            f"\n"
-            f"Flash Attention provides speedup on some hardware through optimized CUDA kernels.\n"
-            f"Falling back to PyTorch SDPA (scaled dot-product attention).\n"
-            f"\n"
-            f"To fix this issue:\n"
-            f"  1. Install Flash Attention: pip install flash-attn\n"
-            f"  2. OR change attention_mode to 'sdpa' (default, always available)\n"
-            f"\n"
-            f"For more info: https://github.com/Dao-AILab/flash-attention"
+            "Cannot use 'flash_attn_3' attention mode: Flash Attention is not installed.\n"
+            "\n"
+            "Flash Attention 3 provides maximum speedup on Hopper+ GPUs through optimized CUDA kernels.\n"
+            "Falling back to PyTorch SDPA (scaled dot-product attention).\n"
+            "\n"
+            "To fix this issue:\n"
+            "  1. Install Flash Attention: pip install flash-attn\n"
+            "  2. OR change attention_mode to 'sdpa' (default, always available)\n"
+            "\n"
+            "For more info: https://github.com/Dao-AILab/flash-attention"
         )
         if debug:
             debug.log(error_msg, level="WARNING", category="setup", force=True)
-        
         return 'sdpa'
-
-    if requested_mode in ['sa2', 'sa3']:
-        if not SAGE_ATTN_AVAILABLE:
+    
+    # Flash Attention 2
+    if requested_mode == 'flash_attn_2':
+        if FLASH_ATTN_2_AVAILABLE:
+            return requested_mode
+        error_msg = (
+            "Cannot use 'flash_attn_2' attention mode: Flash Attention 2 is not installed.\n"
+            "\n"
+            "Flash Attention 2 provides speedup on Ampere+ GPUs through optimized CUDA kernels.\n"
+            "Falling back to PyTorch SDPA (scaled dot-product attention).\n"
+            "\n"
+            "To fix this issue:\n"
+            "  1. Install Flash Attention: pip install flash-attn\n"
+            "  2. OR change attention_mode to 'sdpa' (default, always available)\n"
+            "\n"
+            "For more info: https://github.com/Dao-AILab/flash-attention"
+        )
+        if debug:
+            debug.log(error_msg, level="WARNING", category="setup", force=True)
+        return 'sdpa'
+    
+    # SageAttention 3 (Blackwell)
+    if requested_mode == 'sageattn_3':
+        if SAGE_ATTN_3_AVAILABLE:
+            return requested_mode
+        if SAGE_ATTN_2_AVAILABLE:
             if debug:
-                 debug.log(f"SageAttention not installed. Falling back from '{requested_mode}' to Flash Attention 2...", level="WARNING", category="setup", force=True)
-            # Fallback to check FA2
-            return validate_flash_attention_availability('flash_attn', debug)
-
-        elif not SAGE_ATTN_VARLEN_AVAILABLE:
-             if debug:
-                 debug.log(f"SageAttention installed but 'sageattn_varlen' not found. Falling back from '{requested_mode}' to Flash Attention 2...", level="WARNING", category="setup", force=True)
-             # Fallback to check FA2
-             return validate_flash_attention_availability('flash_attn', debug)
-
-        # If the user explicitly requested sa3, we check for version compatibility.
-        # If version is unknown or insufficient, we fallback to sa2.
-        if requested_mode == 'sa3':
-             try:
-                 version = sageattention.__version__
-                 # Assuming sa3 requires at least a certain version or just presence of version string.
-                 # If we can read version, we assume it's compliant enough or user knows what they are doing.
-                 if debug:
-                     debug.log(f"SageAttention version {version} detected. Using installed kernel for 'sa3' mode.", category="setup", force=True)
-             except AttributeError:
-                 # Version unknown -> Assume it's an older version (sa2) and fallback
-                 if debug:
-                     debug.log("SageAttention version unknown (likely v2 or older). Falling back from 'sa3' to 'sa2'...", level="WARNING", category="setup", force=True)
-                 return validate_flash_attention_availability('sa2', debug)
-
-        pass
+                debug.log(
+                    "SageAttention 3 (Blackwell) not available (requires RTX 50xx GPU and sageattn3 package).\n"
+                    "Falling back to SageAttention 2.",
+                    level="WARNING", category="setup", force=True
+                )
+            return 'sageattn_2'
+        error_msg = (
+            "Cannot use 'sageattn_3' attention mode: SageAttention is not installed.\n"
+            "\n"
+            "SageAttention 3 provides maximum speedup on Blackwell (RTX 50xx) GPUs.\n"
+            "Falling back to PyTorch SDPA (scaled dot-product attention).\n"
+            "\n"
+            "To fix this issue:\n"
+            "  1. Install SageAttention: pip install sageattention\n"
+            "  2. For SA3 Blackwell support: pip install sageattn3\n"
+            "  3. OR change attention_mode to 'flash_attn_2' or 'sdpa'\n"
+            "\n"
+            "For more info: https://github.com/thu-ml/SageAttention"
+        )
+        if debug:
+            debug.log(error_msg, level="WARNING", category="setup", force=True)
+        return 'sdpa'
+    
+    # SageAttention 2
+    if requested_mode == 'sageattn_2':
+        if SAGE_ATTN_2_AVAILABLE:
+            return requested_mode
+        error_msg = (
+            "Cannot use 'sageattn_2' attention mode: SageAttention is not installed.\n"
+            "\n"
+            "SageAttention provides speedup on NVIDIA GPUs through optimized CUDA kernels.\n"
+            "Falling back to PyTorch SDPA (scaled dot-product attention).\n"
+            "\n"
+            "To fix this issue:\n"
+            "  1. Install SageAttention: pip install sageattention\n"
+            "  2. OR change attention_mode to 'flash_attn_2' or 'sdpa'\n"
+            "\n"
+            "For more info: https://github.com/thu-ml/SageAttention"
+        )
+        if debug:
+            debug.log(error_msg, level="WARNING", category="setup", force=True)
+        return 'sdpa'
     
     return requested_mode
 
 
+@torch._dynamo.disable
+def call_flash_attn_2_varlen(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, **kwargs):
+    """
+    Wrapper for Flash Attention 2 flash_attn_varlen_func that handles tensor-to-scalar conversion.
+    
+    Flash Attention 2 supports dropout_p and window_size parameters.
+    Works on Ampere+ GPUs (RTX 30xx, 40xx, A100, etc.).
+    
+    This function is excluded from torch.compile because:
+    1. flash_attn is a C++ extension that can't be compiled anyway
+    2. It requires Python int scalars for max_seqlen parameters
+    3. Disabling compilation here keeps the rest of the model compilable
+    
+    Args:
+        q: Query tensor (total_seq, heads, head_dim)
+        k: Key tensor (total_seq, heads, head_dim)
+        v: Value tensor (total_seq, heads, head_dim)
+        cu_seqlens_q: Cumulative sequence lengths for queries
+        cu_seqlens_k: Cumulative sequence lengths for keys
+        max_seqlen_q: Maximum query sequence length (can be tensor or int)
+        max_seqlen_k: Maximum key sequence length (can be tensor or int)
+        **kwargs: Additional arguments (dropout_p, softmax_scale, causal, window_size, deterministic)
+        
+    Returns:
+        Attention output tensor (total_seq, heads, head_dim)
+    """
+    if not FLASH_ATTN_2_AVAILABLE:
+        raise ImportError("Flash Attention 2 is not available")
+    
+    # Convert tensor max_seqlen to Python int if needed
+    if torch.is_tensor(max_seqlen_q):
+        max_seqlen_q = int(max_seqlen_q.item())
+    if torch.is_tensor(max_seqlen_k):
+        max_seqlen_k = int(max_seqlen_k.item())
+    
+    return flash_attn_2_varlen_func(
+        q=q,
+        k=k,
+        v=v,
+        cu_seqlens_q=cu_seqlens_q,
+        cu_seqlens_k=cu_seqlens_k,
+        max_seqlen_q=max_seqlen_q,
+        max_seqlen_k=max_seqlen_k,
+        **kwargs
+    )
+
+
+@torch._dynamo.disable
+def call_flash_attn_3_varlen(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, **kwargs):
+    """
+    Wrapper for Flash Attention 3 flash_attn_varlen_func that handles tensor-to-scalar conversion.
+    
+    Flash Attention 3 is faster than FA2 but does NOT support dropout_p and window_size.
+    Works on Hopper+ GPUs (H100, etc.) - requires flash_attn_interface package.
+    
+    This function is excluded from torch.compile because:
+    1. flash_attn is a C++ extension that can't be compiled anyway
+    2. It requires Python int scalars for max_seqlen parameters
+    3. Disabling compilation here keeps the rest of the model compilable
+    
+    Args:
+        q: Query tensor (total_seq, heads, head_dim)
+        k: Key tensor (total_seq, heads, head_dim)
+        v: Value tensor (total_seq, heads, head_dim)
+        cu_seqlens_q: Cumulative sequence lengths for queries
+        cu_seqlens_k: Cumulative sequence lengths for keys
+        max_seqlen_q: Maximum query sequence length (can be tensor or int)
+        max_seqlen_k: Maximum key sequence length (can be tensor or int)
+        **kwargs: Additional arguments (softmax_scale, causal, deterministic)
+                  Note: dropout_p and window_size are ignored (not supported by FA3)
+        
+    Returns:
+        Attention output tensor (total_seq, heads, head_dim)
+    """
+    if not FLASH_ATTN_3_AVAILABLE:
+        raise ImportError("Flash Attention 3 is not available")
+    
+    # Convert tensor max_seqlen to Python int if needed
+    if torch.is_tensor(max_seqlen_q):
+        max_seqlen_q = int(max_seqlen_q.item())
+    if torch.is_tensor(max_seqlen_k):
+        max_seqlen_k = int(max_seqlen_k.item())
+    
+    # FA3 doesn't support dropout_p and window_size - filter them out
+    fa3_kwargs = {key: val for key, val in kwargs.items() if key not in ('dropout_p', 'window_size')}
+    
+    # FA3 returns a tuple (output, softmax_lse), we only need output
+    return flash_attn_3_varlen_func(
+        q=q,
+        k=k,
+        v=v,
+        cu_seqlens_q=cu_seqlens_q,
+        cu_seqlens_k=cu_seqlens_k,
+        max_seqlen_q=max_seqlen_q,
+        max_seqlen_k=max_seqlen_k,
+        seqused_q=None,
+        seqused_k=None,
+        **fa3_kwargs
+    )[0]
+
+
+@torch._dynamo.disable
+def call_sage_attn_2_varlen(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, **kwargs):
+    """
+    Wrapper for SageAttention 2 sageattn_varlen that handles tensor-to-scalar conversion.
+    
+    SageAttention 2 provides optimized attention for NVIDIA GPUs with native varlen support.
+    Works on most modern NVIDIA GPUs.
+    
+    This function is excluded from torch.compile because:
+    1. SageAttention is a C++ extension that can't be compiled anyway
+    2. It requires Python int scalars for max_seqlen parameters
+    3. Disabling compilation here keeps the rest of the model compilable
+    
+    Args:
+        q: Query tensor (total_seq, heads, head_dim)
+        k: Key tensor (total_seq, heads, head_dim)
+        v: Value tensor (total_seq, heads, head_dim)
+        cu_seqlens_q: Cumulative sequence lengths for queries
+        cu_seqlens_k: Cumulative sequence lengths for keys
+        max_seqlen_q: Maximum query sequence length (can be tensor or int)
+        max_seqlen_k: Maximum key sequence length (can be tensor or int)
+        **kwargs: Additional arguments (causal supported, others ignored)
+        
+    Returns:
+        Attention output tensor (total_seq, heads, head_dim)
+    """
+    if not SAGE_ATTN_2_AVAILABLE:
+        raise ImportError("SageAttention 2 is not available")
+    
+    # Convert tensor max_seqlen to Python int if needed
+    if torch.is_tensor(max_seqlen_q):
+        max_seqlen_q = int(max_seqlen_q.item())
+    if torch.is_tensor(max_seqlen_k):
+        max_seqlen_k = int(max_seqlen_k.item())
+    
+    # SageAttention requires half precision (fp16/bf16)
+    out_dtype = q.dtype
+    half_dtypes = (torch.float16, torch.bfloat16)
+    
+    if not (q.dtype == k.dtype == v.dtype):
+        k = k.to(q.dtype)
+        v = v.to(q.dtype)
+    
+    if q.dtype not in half_dtypes:
+        q = q.to(torch.bfloat16)
+        k = k.to(torch.bfloat16)
+        v = v.to(torch.bfloat16)
+    
+    is_causal = kwargs.get('causal', False)
+    sm_scale = 1.0 / (q.shape[-1] ** 0.5)
+    
+    out = sageattn_varlen(
+        q, k, v,
+        cu_seqlens_q, cu_seqlens_k,
+        max_seqlen_q, max_seqlen_k,
+        is_causal, sm_scale
+    )
+    
+    return out.to(out_dtype) if out.dtype != out_dtype else out
+
+
+@torch._dynamo.disable
+def call_sage_attn_3_varlen(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, **kwargs):
+    """
+    Wrapper for SageAttention 3 (Blackwell) that converts varlen format to batched format.
+    
+    SageAttention 3 / Blackwell provides maximum performance on RTX 50xx series GPUs.
+    However, it only supports batched attention (uniform sequence lengths), not varlen.
+    
+    This wrapper detects uniform-length batches and reshapes accordingly.
+    For variable-length sequences, it automatically falls back to SageAttention 2.
+    
+    This function is excluded from torch.compile because:
+    1. SageAttention is a C++ extension that can't be compiled anyway
+    2. It requires Python int scalars for max_seqlen parameters
+    3. The varlen-to-batched conversion involves dynamic shapes
+    4. Disabling compilation here keeps the rest of the model compilable
+    
+    Args:
+        q: Query tensor (total_seq, heads, head_dim)
+        k: Key tensor (total_seq, heads, head_dim)
+        v: Value tensor (total_seq, heads, head_dim)
+        cu_seqlens_q: Cumulative sequence lengths for queries
+        cu_seqlens_k: Cumulative sequence lengths for keys
+        max_seqlen_q: Maximum query sequence length (can be tensor or int)
+        max_seqlen_k: Maximum key sequence length (can be tensor or int)
+        **kwargs: Additional arguments (passed to SA2 fallback if needed)
+        
+    Returns:
+        Attention output tensor (total_seq, heads, head_dim)
+    """
+    if not SAGE_ATTN_3_AVAILABLE:
+        raise ImportError("SageAttention 3 (Blackwell) is not available")
+    
+    # Convert tensor max_seqlen to Python int if needed
+    if torch.is_tensor(max_seqlen_q):
+        max_seqlen_q = int(max_seqlen_q.item())
+    if torch.is_tensor(max_seqlen_k):
+        max_seqlen_k = int(max_seqlen_k.item())
+    
+    # Check if all sequences have uniform length (required for SA3 batched API)
+    # SA3/Blackwell uses batched attention, not varlen, so we need uniform lengths
+    seq_lens_q = cu_seqlens_q[1:] - cu_seqlens_q[:-1]
+    seq_lens_k = cu_seqlens_k[1:] - cu_seqlens_k[:-1]
+    
+    uniform_q = (seq_lens_q == seq_lens_q[0]).all()
+    uniform_k = (seq_lens_k == seq_lens_k[0]).all()
+    
+    if not (uniform_q and uniform_k):
+        # Fall back to SA2 for variable-length sequences
+        # This is expected behavior - SA3 Blackwell doesn't support varlen natively
+        if SAGE_ATTN_2_AVAILABLE:
+            return call_sage_attn_2_varlen(
+                q, k, v, cu_seqlens_q, cu_seqlens_k,
+                max_seqlen_q, max_seqlen_k, **kwargs
+            )
+        raise RuntimeError(
+            "SageAttention 3 (Blackwell) requires uniform sequence lengths, "
+            "and SageAttention 2 is not available as fallback. "
+            "Please install sageattention package or use flash_attn/sdpa instead."
+        )
+    
+    # Extract batch dimensions
+    batch_size = len(cu_seqlens_q) - 1
+    seq_len_q = int(seq_lens_q[0].item())
+    seq_len_k = int(seq_lens_k[0].item())
+    heads = q.shape[1]
+    dim = q.shape[2]
+    
+    # SageAttention requires half precision (fp16/bf16)
+    out_dtype = q.dtype
+    half_dtypes = (torch.float16, torch.bfloat16)
+    
+    if not (q.dtype == k.dtype == v.dtype):
+        k = k.to(q.dtype)
+        v = v.to(q.dtype)
+    
+    if q.dtype not in half_dtypes:
+        q = q.to(torch.bfloat16)
+        k = k.to(torch.bfloat16)
+        v = v.to(torch.bfloat16)
+    
+    # Reshape varlen (total_seq, heads, dim) -> batched (batch, seq, heads, dim)
+    q_batched = q.view(batch_size, seq_len_q, heads, dim)
+    k_batched = k.view(batch_size, seq_len_k, heads, dim)
+    v_batched = v.view(batch_size, seq_len_k, heads, dim)
+    
+    # SA3/Blackwell expects (batch, heads, seq, dim) layout
+    q_batched = q_batched.transpose(1, 2)  # (batch, heads, seq, dim)
+    k_batched = k_batched.transpose(1, 2)
+    v_batched = v_batched.transpose(1, 2)
+    
+    # Call SA3 Blackwell
+    out = sageattn_blackwell(q_batched, k_batched, v_batched, per_block_mean=False)
+    
+    # Reshape back to varlen format (total_seq, heads, dim)
+    out = out.transpose(1, 2).reshape(-1, heads, dim).contiguous()
+    
+    return out.to(out_dtype) if out.dtype != out_dtype else out
+
+
 # 2. Triton - Required for torch.compile with inductor backend
 try:
     import triton
@@ -226,21 +643,34 @@ NVIDIA_CONV3D_MEMORY_BUG_WORKAROUND = _check_conv3d_memory_bug()
 if not os.environ.get("SEEDVR2_OPTIMIZATIONS_LOGGED"):
     os.environ["SEEDVR2_OPTIMIZATIONS_LOGGED"] = "1"
     
-    # Flash Attention & Triton status
-    has_both = FLASH_ATTN_AVAILABLE and TRITON_AVAILABLE
-    has_neither = not FLASH_ATTN_AVAILABLE and not TRITON_AVAILABLE
+    # Build status strings
+    sage_status = "✅" if SAGE_ATTN_AVAILABLE else "❌"
+    flash_status = "✅" if FLASH_ATTN_AVAILABLE else "❌"
+    triton_status = "✅" if TRITON_AVAILABLE else "❌"
     
-    if has_both:
-        print("⚡ SeedVR2 optimizations check: Flash Attention ✅ | Triton ✅")
-    elif has_neither:
-        print("⚠️  SeedVR2 optimizations check: Flash Attention ❌ | Triton ❌")
-        print("💡 For best performance: pip install flash-attn triton")
-    elif FLASH_ATTN_AVAILABLE:
-        print("⚡ SeedVR2 optimizations check: Flash Attention ✅ | Triton ❌")
-        print("💡 Install Triton for torch.compile: pip install triton")
-    else:  # TRITON_AVAILABLE only
-        print("⚠️  SeedVR2 optimizations check: Flash Attention ❌ | Triton ✅")
-        print("💡 Install Flash Attention for faster inference: pip install flash-attn")
+    # Count available optimizations
+    available = [SAGE_ATTN_AVAILABLE, FLASH_ATTN_AVAILABLE, TRITON_AVAILABLE]
+    num_available = sum(available)
+    
+    if num_available == 3:
+        print(f"⚡ SeedVR2 optimizations check: SageAttention {sage_status} | Flash Attention {flash_status} | Triton {triton_status}")
+    elif num_available == 0:
+        print(f"⚠️  SeedVR2 optimizations check: SageAttention {sage_status} | Flash Attention {flash_status} | Triton {triton_status}")
+        print("💡 For best performance: pip install sageattention flash-attn triton")
+    else:
+        icon = "⚡" if num_available >= 2 else "⚠️ "
+        print(f"{icon} SeedVR2 optimizations check: SageAttention {sage_status} | Flash Attention {flash_status} | Triton {triton_status}")
+        
+        # Build install suggestions for missing packages
+        missing = []
+        if not SAGE_ATTN_AVAILABLE:
+            missing.append("sageattention")
+        if not FLASH_ATTN_AVAILABLE:
+            missing.append("flash-attn")
+        if not TRITON_AVAILABLE:
+            missing.append("triton")
+        if missing:
+            print(f"💡 Optional: pip install {' '.join(missing)}")
     
     # Conv3d workaround status (if applicable)
     if NVIDIA_CONV3D_MEMORY_BUG_WORKAROUND:
@@ -266,83 +696,38 @@ def _probe_bfloat16_support() -> bool:
 BFLOAT16_SUPPORTED = _probe_bfloat16_support()
 COMPUTE_DTYPE = torch.bfloat16 if BFLOAT16_SUPPORTED else torch.float16
 
-def log_system_capabilities(debug=None):
-    """Log installed attention backends and versions at startup."""
-    if not debug:
-        return
-
-    # SageAttention
-    sa_status = "Available" if SAGE_ATTN_AVAILABLE else "Not Installed"
-    if SAGE_ATTN_AVAILABLE:
-        try:
-            sa_version = sageattention.__version__
-            sa_status += f" (v{sa_version})"
-        except AttributeError:
-            sa_status += " (Version Unknown)"
-
-    # FlashAttention
-    fa_status = "Available" if FLASH_ATTN_AVAILABLE else "Not Installed"
-
-    # Triton
-    triton_status = "Available" if TRITON_AVAILABLE else "Not Installed"
-
-    debug.log(f"Attention Backends: SageAttention={sa_status} | FlashAttention={fa_status} | Triton={triton_status}", category="info", force=True)
-
-def detect_high_end_system() -> Dict[str, Any]:
-    """
-    Detect high-end systems (16GB+ VRAM, etc.) and return optimized defaults.
-
-    Returns:
-        Dict with recommended settings or empty if no specific optimizations found.
-    """
-    optimizations = {}
-    try:
-        # Basic VRAM check
-        if torch.cuda.is_available():
-            device = torch.device("cuda:0")
-            props = torch.cuda.get_device_properties(device)
-            total_vram_gb = props.total_memory / (1024**3)
-
-            # High-end GPU check (e.g., 5070ti/4080/4090/etc with >15GB VRAM)
-            if total_vram_gb >= 15.5:
-                optimizations['high_vram'] = True
-                optimizations['recommended_dtype'] = 'bf16' if BFLOAT16_SUPPORTED else 'fp16'
-                # For 16GB cards, BlockSwap might still be useful for 7B models but maybe less aggressive
-                optimizations['block_swap_recommendation'] = 'moderate'
-
-    except Exception:
-        pass
-
-    return optimizations
-
 
 def call_rope_with_stability(method, *args, **kwargs):
     """
     Call RoPE method with stability fixes:
     1. Clear cache if available
-    2. Disable autocast to prevent numerical issues
+    2. Disable autocast to prevent numerical issues (CUDA only)
     This prevents artifacts in FP8/mixed precision models.
     """
     if hasattr(method, 'cache_clear'):
         method.cache_clear()
     
-    with torch.cuda.amp.autocast(enabled=False):
+    # Only use CUDA autocast context on CUDA devices
+    # MPS has no CUDA autocast to disable
+    if torch.cuda.is_available():
+        with torch.cuda.amp.autocast(enabled=False):
+            return method(*args, **kwargs)
+    else:
         return method(*args, **kwargs)
     
     
-class FP8CompatibleDiT(torch.nn.Module):
+class CompatibleDiT(torch.nn.Module):
     """
     Wrapper for DiT models with automatic compatibility management + advanced optimizations
     
     Precision Handling:
     - FP8: Keeps native FP8 parameters (memory efficient), converts inputs/outputs to compute_dtype for arithmetic
-    - FP16: Uses native FP16 precision throughout
-    - BFloat16: Uses native BFloat16 precision throughout
-    - Float32: Uses full precision for maximum quality
+    - FP16/BFloat16/Float32: Uses native precision throughout
+    - GGUF: On-the-fly dequantization to compute_dtype
+    - MPS: Forces all parameters to compute_dtype (unified memory requires dtype consistency)
     - RoPE: Converted from FP8 to compute_dtype for numerical consistency
     
     Optimizations:
-    - Flash Attention: Automatic optimization of attention layers
     - RoPE Stabilization: Error handling for numerical stability in mixed precision
     - MPS Compatibility: Unified dtype conversion for Apple Silicon backends
     """
@@ -365,9 +750,12 @@ class FP8CompatibleDiT(torch.nn.Module):
             self.debug.start_timer("_convert_rope_freqs")
             self._convert_rope_freqs(target_dtype=self.compute_dtype)
             self.debug.end_timer("_convert_rope_freqs", "RoPE freqs conversion")
-            
-            if hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
-                self.debug.log(f"Also converting NaDiT parameters/buffers for MPS backend", category="setup", force=True)
+        
+        # MPS requires unified dtype for all parameters/buffers (no autocast fallback)
+        # Apply to ALL model types (FP8, FP16, GGUF) when dtype differs from compute_dtype
+        if not skip_conversion and hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
+            if self.model_dtype != self.compute_dtype:
+                self.debug.log(f"Converting NaDiT parameters/buffers to {self.compute_dtype} for MPS backend", category="setup", force=True)
                 self.debug.start_timer("_force_nadit_precision")
                 self._force_nadit_precision(target_dtype=self.compute_dtype)
                 self.debug.end_timer("_force_nadit_precision", "NaDiT parameters/buffers conversion")
@@ -377,11 +765,6 @@ class FP8CompatibleDiT(torch.nn.Module):
         self.debug.start_timer("_stabilize_rope_computations")
         self._stabilize_rope_computations()
         self.debug.end_timer("_stabilize_rope_computations", "RoPE stabilization")
-
-        # 🚀 FLASH ATTENTION OPTIMIZATION (Phase 2)
-        self.debug.start_timer("_apply_flash_attention_optimization")
-        self._apply_flash_attention_optimization()
-        self.debug.end_timer("_apply_flash_attention_optimization", "Flash Attention application")
     
     def _detect_model_dtype(self) -> torch.dtype:
         """Detect main model dtype"""
@@ -504,216 +887,6 @@ class FP8CompatibleDiT(torch.nn.Module):
         
         if rope_count > 0:
             self.debug.log(f"Stabilized {rope_count} RoPE modules", category="success")
-
-    def _apply_flash_attention_optimization(self) -> None:
-        """🚀 FLASH ATTENTION OPTIMIZATION - 30-50% speedup of attention layers"""
-        attention_layers_optimized = 0
-        flash_attention_available = self._check_flash_attention_support()
-        
-        for name, module in self.dit_model.named_modules():
-            # Identify all attention layers
-            if self._is_attention_layer(name, module):
-                # Apply optimization based on availability
-                if self._optimize_attention_layer(name, module, flash_attention_available):
-                    attention_layers_optimized += 1
-        
-        if not flash_attention_available:
-            self.debug.log("Flash Attention not available, using PyTorch SDPA as fallback", category="info", force=True)
-    
-    def _check_flash_attention_support(self) -> bool:
-        """Check if Flash Attention is available"""
-        # Check PyTorch SDPA (includes Flash Attention on H100/A100)
-        if hasattr(torch.nn.functional, 'scaled_dot_product_attention'):
-            return True
-        
-        # Check flash-attn package (uses module-level check from top of file)
-        return FLASH_ATTN_AVAILABLE
-    
-    def _is_attention_layer(self, name: str, module: torch.nn.Module) -> bool:
-        """Identify if a module is an attention layer"""
-        module_type_name = type(module).__name__
-
-        # Skip FlashAttentionVarlen modules - they manage their own attention backends
-        if module_type_name == 'FlashAttentionVarlen':
-            return False
-
-        attention_keywords = [
-            'attention', 'attn', 'self_attn', 'cross_attn', 'mhattn', 'multihead',
-            'transformer_block', 'dit_block'
-        ]
-        
-        # Check by name
-        if any(keyword in name.lower() for keyword in attention_keywords):
-            return True
-        
-        # Check by module type
-        module_type = module_type_name.lower()
-        if any(keyword in module_type for keyword in attention_keywords):
-            return True
-        
-        # Check by attributes (modules with q, k, v projections)
-        if hasattr(module, 'q_proj') or hasattr(module, 'qkv') or hasattr(module, 'to_q'):
-            return True
-        
-        return False
-    
-    def _optimize_attention_layer(self, name: str, module: torch.nn.Module, flash_attention_available: bool) -> bool:
-        """Optimize a specific attention layer"""
-        try:
-            # Save original forward method
-            if not hasattr(module, '_original_forward'):
-                module._original_forward = module.forward
-            
-            # Create new optimized forward method
-            if flash_attention_available:
-                optimized_forward = self._create_flash_attention_forward(module, name)
-            else:
-                optimized_forward = self._create_sdpa_forward(module, name)
-            
-            # Replace forward method
-            module.forward = optimized_forward
-            return True
-            
-        except Exception as e:
-            self.debug.log(f"Failed to optimize attention layer '{name}': {e}", level="WARNING", category="dit", force=True)
-            return False
-    
-    def _create_flash_attention_forward(self, module: torch.nn.Module, layer_name: str):
-        """Create optimized forward with Flash Attention"""
-        original_forward = module._original_forward
-        
-        def flash_attention_forward(*args, **kwargs):
-            try:
-                # Try to use Flash Attention via SDPA
-                return self._sdpa_attention_forward(original_forward, module, *args, **kwargs)
-            except Exception as e:
-                # Fallback to original implementation
-                self.debug.log(f"Flash Attention failed for {layer_name}, using original: {e}", level="WARNING", category="dit", force=True)
-                return original_forward(*args, **kwargs)
-        
-        return flash_attention_forward
-    
-    def _create_sdpa_forward(self, module: torch.nn.Module, layer_name: str):
-        """Create optimized forward with PyTorch SDPA"""
-        original_forward = module._original_forward
-        
-        def sdpa_forward(*args, **kwargs):
-            try:
-                return self._sdpa_attention_forward(original_forward, module, *args, **kwargs)
-            except Exception as e:
-                # Fallback to original implementation
-                return original_forward(*args, **kwargs)
-        
-        return sdpa_forward
-    
-    def _sdpa_attention_forward(self, original_forward, module: torch.nn.Module, *args, **kwargs):
-        """Optimized forward pass using SDPA (Scaled Dot Product Attention)"""
-        # Detect if we can intercept and optimize this layer
-        if len(args) >= 1 and isinstance(args[0], torch.Tensor):
-            input_tensor = args[0]
-            
-            # Check dimensions to ensure it's standard attention
-            if len(input_tensor.shape) >= 3:  # [batch, seq_len, hidden_dim] or similar
-                try:
-                    return self._optimized_attention_computation(module, input_tensor, *args[1:], **kwargs)
-                except:
-                    pass
-        
-        # Fallback to original implementation
-        return original_forward(*args, **kwargs)
-    
-    def _optimized_attention_computation(self, module: torch.nn.Module, input_tensor: torch.Tensor, *args, **kwargs):
-        """Optimized attention computation with SDPA"""
-        # Try to detect standard attention format
-        batch_size, seq_len = input_tensor.shape[:2]
-        
-        # Check if module has standard Q, K, V projections
-        if hasattr(module, 'qkv') or (hasattr(module, 'q_proj') and hasattr(module, 'k_proj') and hasattr(module, 'v_proj')):
-            return self._compute_sdpa_attention(module, input_tensor, *args, **kwargs)
-        
-        # If no standard format detected, use original
-        return module._original_forward(input_tensor, *args, **kwargs)
-    
-    def _compute_sdpa_attention(self, module: torch.nn.Module, x: torch.Tensor, *args, **kwargs):
-        """Optimized SDPA computation for standard attention modules"""
-        try:
-            # Case 1: Module with combined QKV projection
-            if hasattr(module, 'qkv'):
-                qkv = module.qkv(x)
-                # Reshape to separate Q, K, V
-                batch_size, seq_len, _ = qkv.shape
-                qkv = qkv.reshape(batch_size, seq_len, 3, -1)
-                q, k, v = qkv.unbind(dim=2)
-                
-            # Case 2: Separate Q, K, V projections
-            elif hasattr(module, 'q_proj') and hasattr(module, 'k_proj') and hasattr(module, 'v_proj'):
-                q = module.q_proj(x)
-                k = module.k_proj(x)
-                v = module.v_proj(x)
-            else:
-                # Unsupported format, use original
-                return module._original_forward(x, *args, **kwargs)
-            
-            # Detect number of heads
-            head_dim = getattr(module, 'head_dim', None)
-            num_heads = getattr(module, 'num_heads', None)
-            
-            if head_dim is None or num_heads is None:
-                # Try to guess from dimensions
-                hidden_dim = q.shape[-1]
-                if hasattr(module, 'num_heads'):
-                    num_heads = module.num_heads
-                    head_dim = hidden_dim // num_heads
-                else:
-                    # Reasonable defaults
-                    head_dim = 64
-                    num_heads = hidden_dim // head_dim
-            
-            # Reshape for multi-head attention
-            batch_size, seq_len = q.shape[:2]
-            q = q.view(batch_size, seq_len, num_heads, head_dim).transpose(1, 2)
-            k = k.view(batch_size, seq_len, num_heads, head_dim).transpose(1, 2)
-            v = v.view(batch_size, seq_len, num_heads, head_dim).transpose(1, 2)
-            
-            if hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
-                attn_output = torch.nn.functional.scaled_dot_product_attention(
-                    q, k, v,
-                    dropout_p=0.0,
-                    is_causal=False
-                )
-            else:
-                # Use optimized SDPA - PyTorch 2.3+ API with CUDNN support, fallback for older versions
-                if hasattr(torch.nn.attention, 'sdpa_kernel'):
-                    ctx = torch.nn.attention.sdpa_kernel([
-                        torch.nn.attention.SDPBackend.FLASH_ATTENTION,
-                        torch.nn.attention.SDPBackend.EFFICIENT_ATTENTION,
-                        torch.nn.attention.SDPBackend.CUDNN_ATTENTION,
-                        torch.nn.attention.SDPBackend.MATH])
-                else:
-                    ctx = torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=True, enable_mem_efficient=True)
-                
-                with ctx:
-                    attn_output = torch.nn.functional.scaled_dot_product_attention(
-                        q, k, v,
-                        dropout_p=0.0,
-                        is_causal=False
-                    )
-            
-            # Reshape back
-            attn_output = attn_output.transpose(1, 2).contiguous().view(
-                batch_size, seq_len, num_heads * head_dim
-            )
-            
-            # Output projection if it exists
-            if hasattr(module, 'out_proj') or hasattr(module, 'o_proj'):
-                proj = getattr(module, 'out_proj', None) or getattr(module, 'o_proj', None)
-                attn_output = proj(attn_output)
-            
-            return attn_output
-            
-        except Exception as e:
-            # In case of error, use original implementation
-            return module._original_forward(x, *args, **kwargs)
     
     def forward(self, *args, **kwargs):
         """
diff --git a/src/optimization/memory_manager.py b/src/optimization/memory_manager.py
index 592e37e..f5ca8af 100644
--- a/src/optimization/memory_manager.py
+++ b/src/optimization/memory_manager.py
@@ -10,8 +10,9 @@ import gc
 import sys
 import time
 import psutil
+import platform
 from typing import Tuple, Dict, Any, Optional, List, Union
-    
+
 
 def _device_str(device: Union[torch.device, str]) -> str:
     """Normalized uppercase device string for comparison and logging. MPS variants → 'MPS'."""
@@ -19,6 +20,31 @@ def _device_str(device: Union[torch.device, str]) -> str:
     return 'MPS' if s.startswith('MPS') else s
 
 
+def is_mps_available() -> bool:
+    """Check if MPS (Apple Metal) backend is available."""
+    return hasattr(torch.backends, 'mps') and torch.backends.mps.is_available()
+
+
+def is_cuda_available() -> bool:
+    """Check if CUDA backend is available."""
+    return torch.cuda.is_available()
+
+
+def get_gpu_backend() -> str:
+    """Get the active GPU backend type.
+    
+    Returns:
+        'cuda': NVIDIA CUDA
+        'mps': Apple Metal Performance Shaders
+        'cpu': No GPU backend available
+    """
+    if is_cuda_available():
+        return 'cuda'
+    if is_mps_available():
+        return 'mps'
+    return 'cpu'
+
+
 def get_device_list(include_none: bool = False, include_cpu: bool = False) -> List[str]:
     """
     Get list of available compute devices for SeedVR2
@@ -37,14 +63,14 @@ def get_device_list(include_none: bool = False, include_cpu: bool = False) -> Li
     has_mps = False
     
     try:
-        if hasattr(torch, "cuda") and hasattr(torch.cuda, "is_available") and torch.cuda.is_available():
+        if is_cuda_available():
             devs += [f"cuda:{i}" for i in range(torch.cuda.device_count())]
             has_cuda = True
     except Exception:
         pass
     
     try:
-        if hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+        if is_mps_available():
             devs.append("mps")  # MPS doesn't use device indices
             has_mps = True
     except Exception:
@@ -66,7 +92,7 @@ def get_device_list(include_none: bool = False, include_cpu: bool = False) -> Li
     result.extend(devs)
     
     return result if result else []
-    
+
 
 def get_basic_vram_info(device: Optional[torch.device] = None) -> Dict[str, Any]:
     """
@@ -80,13 +106,13 @@ def get_basic_vram_info(device: Optional[torch.device] = None) -> Dict[str, Any]
         dict: {"free_gb": float, "total_gb": float} or {"error": str}
     """
     try:
-        if torch.cuda.is_available():
+        if is_cuda_available():
             if device is None:
                 device = torch.device("cuda:0")
             elif not isinstance(device, torch.device):
                 device = torch.device(device)
             free_memory, total_memory = torch.cuda.mem_get_info(device)
-        elif hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
+        elif is_mps_available():
             # MPS doesn't support per-device queries or mem_get_info
             # Use system memory as proxy
             mem = psutil.virtual_memory()
@@ -106,29 +132,13 @@ def get_basic_vram_info(device: Optional[torch.device] = None) -> Dict[str, Any]
 # Initial VRAM check at module load
 vram_info = get_basic_vram_info(device=None)
 if "error" not in vram_info:
-    backend = "MPS" if (hasattr(torch.backends, 'mps') and torch.backends.mps.is_available()) else "CUDA"
+    backend = "MPS" if is_mps_available() else "CUDA"
     print(f"📊 Initial {backend} memory: {vram_info['free_gb']:.2f}GB free / {vram_info['total_gb']:.2f}GB total")
 else:
     print(f"⚠️ Memory check failed: {vram_info['error']} - No available backend!")
 
 
-def _enforce_vram_limit() -> None:
-    """
-    Enforce VRAM limit to physical capacity to prevent silent swap to system RAM.
-    Called once at module load. No-op on MPS or unsupported platforms.
-    """
-    if not torch.cuda.is_available():
-        return
-    try:
-        for i in range(torch.cuda.device_count()):
-            torch.cuda.set_per_process_memory_fraction(1.0, i)
-    except Exception:
-        pass
-
-_enforce_vram_limit()
-
-
-def get_vram_usage(device: Optional[torch.device] = None, debug: Optional['Debug'] = None) -> Tuple[float, float, float]:
+def get_vram_usage(device: Optional[torch.device] = None, debug: Optional['Debug'] = None) -> Tuple[float, float, float, float]:
     """
     Get current VRAM usage metrics for monitoring.
     Used for tracking memory consumption during processing.
@@ -138,29 +148,30 @@ def get_vram_usage(device: Optional[torch.device] = None, debug: Optional['Debug
         debug: Optional debug instance for logging
     
     Returns:
-        tuple: (allocated_gb, reserved_gb, max_reserved_gb)
-               Returns (0, 0, 0) if no GPU available
+        tuple: (allocated_gb, reserved_gb, peak_allocated_gb, peak_reserved_gb)
+               Returns (0, 0, 0, 0) if no GPU available
     """
     try:
-        if torch.cuda.is_available():
+        if is_cuda_available():
             if device is None:
                 device = torch.device("cuda:0")
             elif not isinstance(device, torch.device):
                 device = torch.device(device)
             allocated = torch.cuda.memory_allocated(device) / (1024**3)
             reserved = torch.cuda.memory_reserved(device) / (1024**3)
-            max_reserved = torch.cuda.max_memory_reserved(device) / (1024**3)
-            return allocated, reserved, max_reserved
-        elif hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
+            peak_allocated = torch.cuda.max_memory_allocated(device) / (1024**3)
+            peak_reserved = torch.cuda.max_memory_reserved(device) / (1024**3)
+            return allocated, reserved, peak_allocated, peak_reserved
+        elif is_mps_available():
             # MPS doesn't support per-device queries - uses global memory tracking
             allocated = torch.mps.current_allocated_memory() / (1024**3)
             reserved = torch.mps.driver_allocated_memory() / (1024**3)
-            max_allocated = allocated  # MPS doesn't track peak separately
-            return allocated, reserved, max_allocated
+            # MPS doesn't track peak separately
+            return allocated, reserved, allocated, reserved
     except Exception as e:
         if debug:
             debug.log(f"Failed to get VRAM usage: {e}", level="WARNING", category="memory", force=True)
-    return 0.0, 0.0, 0.0
+    return 0.0, 0.0, 0.0, 0.0
 
 
 def get_ram_usage(debug: Optional['Debug'] = None) -> Tuple[float, float, float, float]:
@@ -251,17 +262,17 @@ def clear_memory(debug: Optional['Debug'] = None, deep: bool = False, force: boo
         # Use existing function for memory info
         mem_info = get_basic_vram_info(device=None)
         
-        if "error" not in mem_info:
+        if "error" not in mem_info and mem_info["total_gb"] > 0:
             # Check VRAM/MPS memory pressure (5% free threshold)
             free_ratio = mem_info["free_gb"] / mem_info["total_gb"]
             if free_ratio < 0.05:
                 should_clear = True
                 if debug:
-                    backend = "MPS" if (hasattr(torch.backends, 'mps') and torch.backends.mps.is_available()) else "VRAM"
+                    backend = "Unified Memory" if is_mps_available() else "VRAM"
                     debug.log(f"{backend} pressure: {mem_info['free_gb']:.2f}GB free of {mem_info['total_gb']:.2f}GB", category="memory")
         
         # For non-MPS systems, also check system RAM separately
-        if not should_clear and not (hasattr(torch.backends, 'mps') and torch.backends.mps.is_available()):
+        if not should_clear and not is_mps_available():
             mem = psutil.virtual_memory()
             if mem.available < mem.total * 0.05:
                 should_clear = True
@@ -284,10 +295,10 @@ def clear_memory(debug: Optional['Debug'] = None, deep: bool = False, force: boo
     if debug:
         debug.start_timer(gpu_timer)
     
-    if torch.cuda.is_available():
+    if is_cuda_available():
         torch.cuda.empty_cache()
         torch.cuda.ipc_collect()
-    elif hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
+    elif is_mps_available():
         torch.mps.empty_cache()
     
     if debug:
@@ -324,7 +335,7 @@ def clear_memory(debug: Optional['Debug'] = None, deep: bool = False, force: boo
                 handle = _os_memory_lib.GetCurrentProcess()
                 _os_memory_lib.SetProcessWorkingSetSize(handle, -1, -1)
                 
-            elif hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
+            elif is_mps_available():
                 # macOS with MPS
                 import ctypes  # Import only when needed
                 import ctypes.util
@@ -401,7 +412,7 @@ def reset_vram_peak(device: Optional[torch.device] = None, debug: Optional['Debu
     if debug and debug.enabled:
         debug.log("Resetting VRAM peak memory statistics", category="memory")
     try:
-        if torch.cuda.is_available():
+        if is_cuda_available():
             if device is None:
                 device = torch.device("cuda:0")
             elif not isinstance(device, torch.device):
@@ -691,7 +702,7 @@ def manage_model_device(model: torch.nn.Module, target_device: torch.device, mod
         
         if has_blockswap_config:
             is_blockswap_model = True
-            # Get the actual model (handle FP8CompatibleDiT wrapper)
+            # Get the actual model (handle CompatibleDiT wrapper)
             if hasattr(model, "dit_model"):
                 actual_model = model.dit_model
 
diff --git a/src/utils/constants.py b/src/utils/constants.py
index 2b202c8..d077e71 100644
--- a/src/utils/constants.py
+++ b/src/utils/constants.py
@@ -4,7 +4,7 @@ Only includes constants actually used in the codebase
 """
 
 # Version information
-__version__ = "2.5.17"
+__version__ = "2.5.20"
 
 import os
 import warnings
diff --git a/src/utils/debug.py b/src/utils/debug.py
index e819120..716c3da 100644
--- a/src/utils/debug.py
+++ b/src/utils/debug.py
@@ -10,16 +10,33 @@ import torch
 import gc
 from typing import Optional, List, Dict, Any, Union
 from datetime import datetime
-from ..optimization.memory_manager import get_vram_usage, get_basic_vram_info, get_ram_usage, reset_vram_peak
+import platform
+from ..optimization.memory_manager import (
+    get_vram_usage, 
+    get_basic_vram_info, 
+    get_ram_usage, 
+    reset_vram_peak,
+    is_mps_available,
+    is_cuda_available
+)
 from ..utils.constants import __version__
 
 
-def _format_peak_with_swap(peak_gb: float, total_vram_gb: float) -> str:
-    """Format peak memory, showing swap breakdown if overflow occurred."""
-    if total_vram_gb > 0 and peak_gb > total_vram_gb:
-        swap_gb = peak_gb - total_vram_gb
-        return f"{peak_gb:.2f}GB ({total_vram_gb:.0f}GB GPU + {swap_gb:.2f}GB swap)"
-    return f"{peak_gb:.2f}GB"
+def _format_peak_with_overflow(peak_gb: float, total_vram_gb: float) -> str:
+    """Format peak reserved memory, showing overflow breakdown on Windows.
+    
+    Args:
+        peak_gb: Peak reserved memory from PyTorch
+        total_vram_gb: Physical GPU VRAM capacity
+    """
+    if total_vram_gb <= 0:
+        return f"{peak_gb:.2f}GB reserved"
+    
+    overflow_gb = peak_gb - total_vram_gb
+    if overflow_gb <= 0 or platform.system() != 'Windows':
+        return f"{peak_gb:.2f}GB reserved"
+    
+    return f"{peak_gb:.2f}GB reserved ({total_vram_gb:.0f}GB GPU + {overflow_gb:.2f}GB overflow)"
 
 
 class Debug:
@@ -80,7 +97,8 @@ class Debug:
         self.vram_history: List[float] = []
         self.active_timer_stack: List[str] = [] 
         self.timer_namespace: str = ""
-        self.phase_vram_peaks: Dict[str, float] = {}
+        self.phase_vram_peaks_alloc: Dict[str, float] = {}
+        self.phase_vram_peaks_rsv: Dict[str, float] = {}
         self.phase_ram_peaks: Dict[str, float] = {}
         
     @torch._dynamo.disable  # Skip tracing to avoid datetime.now() warnings
@@ -125,26 +143,33 @@ class Debug:
 
     def print_header(self, cli: bool = False) -> None:
         """Print the header with banner - always displayed"""
-        # Intro logo
-        self.log("", category="none", force=True)
-        self.log(" ╔══════════════════════════════════════════════════════════╗", category="none", force=True)
-        self.log(" ║ ███████ ███████ ███████ ██████  ██    ██ ██████  ███████ ║", category="none", force=True)
-        self.log(" ║ ██      ██      ██      ██   ██ ██    ██ ██   ██      ██ ║", category="none", force=True)
-        self.log(" ║ ███████ █████   █████   ██   ██ ██    ██ ██████  █████   ║", category="none", force=True)
-        self.log(" ║      ██ ██      ██      ██   ██  ██  ██  ██   ██ ██      ║", category="none", force=True)
-        self.log(" ║ ███████ ███████ ███████ ██████    ████   ██   ██ ███████ ║", category="none", force=True)
+        # Temporarily disable timestamps for clean header display
+        original_timestamps = self.show_timestamps
+        self.show_timestamps = False
         
-        # Version number with dynamic padding to maintain visual alignment with any version length
-        version_text = f"v{__version__}"
-        prefix = " 💻 CLI mode · " if cli else " "
-        suffix = "© ByteDance Seed · NumZ · AInVFX "
-        emoji_compensation = 1 if cli else 0
-        padding_width = 59 - len(prefix) - len(version_text) - len(suffix) - 2 - emoji_compensation
-        padding = " " * max(1, padding_width)
-        self.log(f" ║{prefix}{version_text}{padding} {suffix}║", category="none", force=True)
-
-        self.log(" ╚══════════════════════════════════════════════════════════╝", category="none", force=True)
+        # ASCII art logo
         self.log("", category="none", force=True)
+        self.log("", category="none", force=True)
+        self.log("███████╗███████╗███████╗██████╗ ██╗   ██╗██████╗     ██████╗       ███████╗", category="none", force=True, indent_level=1)
+        self.log("██╔════╝██╔════╝██╔════╝██╔══██╗██║   ██║██╔══██╗    ╚════██╗      ██╔════╝", category="none", force=True, indent_level=1)
+        self.log("███████╗█████╗  █████╗  ██║  ██║██║   ██║██████╔╝     █████╔╝      ███████╗", category="none", force=True, indent_level=1)
+        self.log("╚════██║██╔══╝  ██╔══╝  ██║  ██║╚██╗ ██╔╝██╔══██╗    ██╔═══╝       ╚════██║", category="none", force=True, indent_level=1)
+        self.log("███████║███████╗███████╗██████╔╝ ╚████╔╝ ██║  ██║    ███████╗  ██╗ ███████║", category="none", force=True, indent_level=1)
+        self.log("╚══════╝╚══════╝╚══════╝╚═════╝   ╚═══╝  ╚═╝  ╚═╝    ╚══════╝  ╚═╝ ╚══════╝", category="none", force=True, indent_level=1)
+        # Version and credits - left/right aligned to logo width
+        version_text = f"v{__version__}"
+        cli_indicator = "💻 CLI · " if cli else ""
+        left_part = f"{cli_indicator}{version_text}"
+        right_part = "© ByteDance Seed · NumZ · AInVFX"
+        logo_width = 75
+        emoji_compensation = 1 if cli else 0
+        padding = logo_width - len(left_part) - len(right_part) - emoji_compensation
+        self.log(f"{left_part}{' ' * max(1, padding)}{right_part}", category="none", force=True, indent_level=1)
+        self.log("━" * logo_width, category="none", force=True, indent_level=1)
+        self.log("", category="none", force=True)
+        
+        # Restore timestamps setting
+        self.show_timestamps = original_timestamps
         
         # Environment info - only in debug mode
         if self.enabled:
@@ -174,7 +199,7 @@ class Debug:
         cuda_ver = getattr(torch.version, 'cuda', None) or "N/A"
         
         # GPU
-        if torch.cuda.is_available():
+        if is_cuda_available():
             try:
                 props = torch.cuda.get_device_properties(0)
                 gpu_str = f"{props.name} ({round(props.total_memory / (1024**3))}GB)"
@@ -182,19 +207,37 @@ class Debug:
             except Exception:
                 gpu_str = "CUDA"
                 cudnn_ver = "N/A"
-        elif getattr(getattr(torch, 'mps', None), 'is_available', lambda: False)():
+        elif is_mps_available():
             gpu_str = "Apple Silicon (MPS)"
             cudnn_ver = "N/A"
         else:
             gpu_str = "CPU"
             cudnn_ver = "N/A"
         
-        # Flash Attn & Triton - reuse existing module constants
+        # Flash Attn, SageAttn & Triton - reuse existing module constants
         try:
-            from ..optimization.compatibility import FLASH_ATTN_AVAILABLE, TRITON_AVAILABLE
-            flash_str, triton_str = ("✓" if FLASH_ATTN_AVAILABLE else "✗"), ("✓" if TRITON_AVAILABLE else "✗")
+            from ..optimization.compatibility import (
+                FLASH_ATTN_2_AVAILABLE, FLASH_ATTN_3_AVAILABLE,
+                SAGE_ATTN_2_AVAILABLE, SAGE_ATTN_3_AVAILABLE,
+                TRITON_AVAILABLE
+            )
+            fa_parts = []
+            if FLASH_ATTN_3_AVAILABLE:
+                fa_parts.append("3")
+            if FLASH_ATTN_2_AVAILABLE:
+                fa_parts.append("2")
+            flash_str = f"v{','.join(fa_parts)} ✓" if fa_parts else "✗"
+            
+            sa_parts = []
+            if SAGE_ATTN_3_AVAILABLE:
+                sa_parts.append("3")
+            if SAGE_ATTN_2_AVAILABLE:
+                sa_parts.append("2")
+            sage_str = f"v{','.join(sa_parts)} ✓" if sa_parts else "✗"
+            
+            triton_str = "✓" if TRITON_AVAILABLE else "✗"
         except ImportError:
-            flash_str = triton_str = "?"
+            flash_str = sage_str = triton_str = "?"
         
         # ComfyUI version
         comfy_str = None
@@ -207,7 +250,7 @@ class Debug:
         
         # Print
         self.log(f"OS: {os_str} | GPU: {gpu_str}", category="info")
-        self.log(f"Python: {py_ver} | PyTorch: {torch_ver} | Flash Attn: {flash_str} | Triton: {triton_str}", category="info")
+        self.log(f"Python: {py_ver} | PyTorch: {torch_ver} | FlashAttn: {flash_str} | SageAttn: {sage_str} | Triton: {triton_str}", category="info")
         cuda_line = f"CUDA: {cuda_ver} | cuDNN: {cudnn_ver}"
         self.log(f"{cuda_line} | ComfyUI: {comfy_str}" if comfy_str else cuda_line, category="info")
         self.log("", category="none")
@@ -381,9 +424,12 @@ class Debug:
         if show_diff and self.memory_checkpoints:
             self._log_memory_diff(current_metrics=memory_info, force=force)
 
-        # Warn if swap detected (peak > physical VRAM)
-        if memory_info['vram_total'] > 0 and memory_info['vram_peak_since_last'] > memory_info['vram_total']:
-            self.log("VRAM swap detected - severe slowdown expected. Consider optimizing (e.g., reduce resolution, batch_size, enable BlockSwap, VAE tiling...).", 
+        # Overflow warning (Windows only - WDDM can page to system RAM)
+        overflow = memory_info.get('vram_overflow', 0.0)
+        
+        if overflow > 0 and platform.system() == 'Windows':
+            self.log(f"VRAM overflow: {overflow:.2f}GB paged to system RAM - severe slowdown expected. "
+                     "Consider optimizing (e.g., reduce resolution, batch size, enable BlockSwap, VAE tiling...).",
                      level="WARNING", category="memory", force=True)
 
         # Log detailed analysis if requested
@@ -395,10 +441,15 @@ class Debug:
 
         # Update phase peaks if we're in an active phase
         if self.current_phase:
-            if memory_info['vram_peak_since_last'] > 0:
-                self.phase_vram_peaks[self.current_phase] = max(
-                    self.phase_vram_peaks.get(self.current_phase, 0),
-                    memory_info['vram_peak_since_last']
+            if memory_info['vram_peak_alloc'] > 0:
+                self.phase_vram_peaks_alloc[self.current_phase] = max(
+                    self.phase_vram_peaks_alloc.get(self.current_phase, 0),
+                    memory_info['vram_peak_alloc']
+                )
+            if memory_info['vram_peak_rsv'] > 0:
+                self.phase_vram_peaks_rsv[self.current_phase] = max(
+                    self.phase_vram_peaks_rsv.get(self.current_phase, 0),
+                    memory_info['vram_peak_rsv']
                 )
             if memory_info['ram_process'] > 0:
                 self.phase_ram_peaks[self.current_phase] = max(
@@ -410,13 +461,18 @@ class Debug:
         reset_vram_peak(device=None, debug=self)
     
     def _collect_memory_metrics(self) -> Dict[str, Any]:
-        """Collect current memory metrics efficiently."""
+        """Collect current memory metrics."""
+        is_mps = is_mps_available()
+        has_gpu = is_mps or is_cuda_available()
+        
         metrics = {
             'vram_allocated': 0.0,
             'vram_reserved': 0.0,
             'vram_free': 0.0,
             'vram_total': 0.0,
-            'vram_peak_since_last': 0.0,
+            'vram_peak_alloc': 0.0,
+            'vram_peak_rsv': 0.0,
+            'vram_overflow': 0.0,
             'ram_process': 0.0,
             'ram_available': 0.0,
             'ram_total': 0.0,
@@ -425,46 +481,36 @@ class Debug:
             'summary_ram': ""
         }
         
-        # VRAM metrics
-        if torch.cuda.is_available() or (hasattr(torch.backends, 'mps') and torch.backends.mps.is_available()):
-            metrics['vram_allocated'], metrics['vram_reserved'], current_global_peak = get_vram_usage(device=None, debug=self)
-
-            # Calculate peak since last log_memory_state
-            # This captures the actual peak that occurred between calls
-            metrics['vram_peak_since_last'] = current_global_peak
-            
+        if has_gpu:
+            metrics['vram_allocated'], metrics['vram_reserved'], metrics['vram_peak_alloc'], metrics['vram_peak_rsv'] = get_vram_usage(device=None, debug=self)
             vram_info = get_basic_vram_info(device=None)
             
-            if "error" not in vram_info:
+            if "error" not in vram_info and vram_info["total_gb"] > 0:
                 metrics['vram_free'] = vram_info["free_gb"]
                 metrics['vram_total'] = vram_info["total_gb"]
+                metrics['vram_overflow'] = max(0.0, metrics['vram_peak_rsv'] - metrics['vram_total'])
                 
-                backend = "MPS" if (hasattr(torch.backends, 'mps') and torch.backends.mps.is_available()) else "VRAM"
-                peak_str = _format_peak_with_swap(metrics['vram_peak_since_last'], metrics['vram_total'])
-                metrics['summary_vram'] = (f"  [{backend}] {metrics['vram_allocated']:.2f}GB allocated / "
-                        f"{metrics['vram_reserved']:.2f}GB reserved / "
-                        f"Peak: {peak_str} / "
-                        f"{metrics['vram_free']:.2f}GB free / "
-                        f"{metrics['vram_total']:.2f}GB total")
-            else:
-                metrics['summary_vram'] = ""
-        else:
-            metrics['summary_vram'] = ""
+                backend = "Unified Memory" if is_mps else "VRAM"
+                metrics['summary_vram'] = (
+                    f"  [{backend}] {metrics['vram_allocated']:.2f}GB allocated / "
+                    f"{metrics['vram_reserved']:.2f}GB reserved / "
+                    f"Peak: {metrics['vram_peak_alloc']:.2f}GB / "
+                    f"{metrics['vram_free']:.2f}GB free / "
+                    f"{metrics['vram_total']:.2f}GB total"
+                )
+            
+            self.vram_history.append(metrics['vram_reserved'])
         
-        # RAM metrics using new function
+        # RAM metrics
         metrics['ram_process'], metrics['ram_available'], metrics['ram_total'], metrics['ram_others'] = get_ram_usage(debug=self)
         
         if metrics['ram_total'] > 0:
-            metrics['summary_ram'] = (f"  [RAM] {metrics['ram_process']:.2f}GB process / "
-                      f"{metrics['ram_others']:.2f}GB others / "
-                      f"{metrics['ram_available']:.2f}GB free / "
-                      f"{metrics['ram_total']:.2f}GB total")
-        else:
-            metrics['summary_ram'] = ""
-        
-        # Update VRAM history for tracking
-        if torch.cuda.is_available() or (hasattr(torch.backends, 'mps') and torch.backends.mps.is_available()):
-            self.vram_history.append(metrics['vram_allocated'])
+            metrics['summary_ram'] = (
+                f"  [RAM] {metrics['ram_process']:.2f}GB process / "
+                f"{metrics['ram_others']:.2f}GB others / "
+                f"{metrics['ram_available']:.2f}GB free / "
+                f"{metrics['ram_total']:.2f}GB total"
+            )
         
         return metrics
     
@@ -592,8 +638,8 @@ class Debug:
             self.log(f"Memory changes: {', '.join(diffs)}", category="memory", force=force, indent_level=1)
     
     def log_peak_memory_summary(self, force: bool = True) -> None:
-        """Display peak memory usage across all phases (VRAM and RAM combined)"""
-        if not self.phase_vram_peaks and not self.phase_ram_peaks:
+        """Display peak memory usage across all phases."""
+        if not self.phase_vram_peaks_alloc and not self.phase_ram_peaks:
             return
         
         phase_names = {
@@ -603,9 +649,9 @@ class Debug:
             'phase4': 'Post-processing'
         }
         
-        is_mps = hasattr(torch.backends, 'mps') and torch.backends.mps.is_available() and not torch.cuda.is_available()
+        is_mps = is_mps_available()
         
-        # Get total VRAM for swap detection (reuse existing function)
+        # Get total VRAM for overflow formatting (Windows only)
         total_vram_gb = 0.0
         if not is_mps:
             vram_info = get_basic_vram_info(device=None)
@@ -616,25 +662,29 @@ class Debug:
         self.log("────────────────────────", category="none", force=force)
         self.log("Peak memory by phase:", category="memory", force=force)
         
-        all_phases = sorted(set(self.phase_vram_peaks.keys()) | set(self.phase_ram_peaks.keys()))
+        all_phases = sorted(set(self.phase_vram_peaks_alloc.keys()) | set(self.phase_ram_peaks.keys()))
         for phase_key in all_phases:
             phase_num = phase_key[-1]
             phase_name = phase_names.get(phase_key, phase_key)
-            vram = self.phase_vram_peaks.get(phase_key, 0)
+            alloc = self.phase_vram_peaks_alloc.get(phase_key, 0)
+            rsv = self.phase_vram_peaks_rsv.get(phase_key, 0)
             ram = self.phase_ram_peaks.get(phase_key, 0)
             
             if is_mps:
-                self.log(f"  Phase {phase_num} ({phase_name}): {vram:.2f}GB", category="memory", force=force)
+                self.log(f"{phase_num}. {phase_name}: {alloc:.2f}GB", category="memory", indent_level=1, force=force)
             else:
-                self.log(f"  Phase {phase_num} ({phase_name}): {_format_peak_with_swap(vram, total_vram_gb)} | RAM {ram:.2f}GB", category="memory", force=force)
+                rsv_str = _format_peak_with_overflow(rsv, total_vram_gb)
+                self.log(f"{phase_num}. {phase_name}: VRAM {alloc:.2f}GB allocated, {rsv_str} | RAM {ram:.2f}GB", category="memory", indent_level=1, force=force)
+        
+        overall_alloc = max(self.phase_vram_peaks_alloc.values()) if self.phase_vram_peaks_alloc else 0
+        overall_rsv = max(self.phase_vram_peaks_rsv.values()) if self.phase_vram_peaks_rsv else 0
+        overall_ram = max(self.phase_ram_peaks.values()) if self.phase_ram_peaks else 0
         
         if is_mps:
-            overall = max(self.phase_vram_peaks.values()) if self.phase_vram_peaks else 0
-            self.log(f"Overall Peak: {overall:.2f}GB", category="memory", force=force)
+            self.log(f"Overall peak: {overall_alloc:.2f}GB", category="memory", force=force)
         else:
-            overall_vram = max(self.phase_vram_peaks.values()) if self.phase_vram_peaks else 0
-            overall_ram = max(self.phase_ram_peaks.values()) if self.phase_ram_peaks else 0
-            self.log(f"Overall peak: {_format_peak_with_swap(overall_vram, total_vram_gb)} | RAM {overall_ram:.2f}GB", category="memory", force=force)
+            overall_rsv_str = _format_peak_with_overflow(overall_rsv, total_vram_gb)
+            self.log(f"Overall peak: VRAM {overall_alloc:.2f}GB allocated, {overall_rsv_str} | RAM {overall_ram:.2f}GB", category="memory", force=force)
     
     @torch._dynamo.disable  # Skip tracing to avoid time.time() warnings
     def _store_checkpoint(self, label: str, metrics: Dict[str, Any]) -> None:
@@ -744,6 +794,7 @@ class Debug:
         self.timer_durations.clear()
         self.timer_messages.clear()
         self.active_timer_stack.clear()
-        self.phase_vram_peaks.clear()
+        self.phase_vram_peaks_alloc.clear()
+        self.phase_vram_peaks_rsv.clear()
         self.phase_ram_peaks.clear()
         self.current_phase = None
\ No newline at end of file