Merge pull request #441 from AInVFX/main

v2.5.24: Restores the MPS memory leak workaround that was accidentally removed during code cleanup in v2.5.23
fix(vae): restore MPS memory leak workaround removed in v2.5.23 cleanup
2025-12-24 09:52:32 +01:00 · 2025-12-24 09:50:19 +01:00 · 2025-12-23 21:09:01 -05:00 · 2025-12-24 03:02:34 +01:00 · 2025-12-24 02:49:42 +01:00 · 2025-12-24 02:22:46 +01:00
29 changed files with 1959 additions and 1010 deletions
@@ -16,15 +16,9 @@ We welcome code contributions from the community. Here's how to propose changes:
 1. Fork this repository to your own GitHub account.
 2. Create a new branch on your fork for your changes.
 3. Make your changes in this branch.
-4. When you are ready, submit a pull request to either:
-   - **`main` branch** - For stable features and bug fixes
-   - **`nightly` branch** - For experimental features or if you want faster integration (note: this branch changes more frequently and may be less stable)
+4. When you are ready, submit a pull request to the **`main`** branch.

-### Branch Guidelines
- **Main Branch**: Use this for production-ready changes, critical bug fixes, and well-tested features
- **Nightly Branch**: Use this for experimental features, cutting-edge improvements, or if you need your changes integrated quickly. Be aware that this branch is actively under development and may contain unstable features.
-
-We use the GitHub Flow workflow. Choose the appropriate target branch based on the nature of your contribution.
+We use the GitHub Flow workflow.

 Before submitting a pull request, please make sure your code adheres to the project's coding conventions and it has passed all tests. If you are adding features, please also add appropriate tests.

@@ -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:

@@ -34,7 +34,85 @@ 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
+
+**2025.12.24 - Version 2.5.24**
+
+- **🍎 Fix: MPS memory leak regression** - Restored MPS cache clearing after VAE encode/decode operations that was accidentally removed during code cleanup in v2.5.23
+
+**2025.12.24 - Version 2.5.23**
+
+- **🔒 Security: Prevent code execution in model loading** - Added protection against malicious .pth files by restricting deserialization to tensors only
+- **🎥 Fix: FFmpeg video writer reliability** - Resolved ffmpeg process hanging issues by redirecting stderr and adding buffer flush, with improved error messages for debugging *(thanks [@thehhmdb](https://github.com/thehhmdb))*
+- **⚡ Fix: GGUF VAE model support** - Enabled automatic weight dequantization for convolution operations, making GGUF-quantized VAE models fully functional *(thanks [@naxci1](https://github.com/naxci1))*
+- **🛡️ Fix: VAE slicing edge cases** - Protected against division by zero crashes when using small split sizes with high temporal downsampling *(thanks [@naxci1](https://github.com/naxci1))*
+- **🎨 Fix: LAB color transfer precision** - Resolved dtype mismatch errors during video upscaling by ensuring consistent float types before matrix operations
+- **🔧 Fix: PyTorch 2.9+ compatibility** - Extended Conv3d memory workaround to all PyTorch 2.9+ versions, fixing 3x VRAM usage on newer PyTorch releases
+- **📦 Fix: Bitsandbytes compatibility** - Added ValueError exception handling for Intel Gaudi version detection failures on non-Gaudi systems
+- **🍎 MPS: Memory optimization** - Reduced memory usage during encode/decode operations on Apple Silicon *(thanks [@s-cerevisiae](https://github.com/s-cerevisiae))*
+
+
+**2025.12.13 - Version 2.5.22**
+
+- **🎬 CLI: FFmpeg video backend with 10-bit support** - New `--video_backend ffmpeg` and `--10bit` flags enable x265 encoding with 10-bit color depth, reducing banding artifacts in gradients compared to 8-bit OpenCV output *(based on PR by [@thehhmdb](https://github.com/thehhmdb) - thank you!)*
+- **🍎 Fix: MPS bicubic upscaling compatibility** - Added CPU fallback for bicubic+antialias interpolation on PyTorch versions before 2.8.0, resolving RGBA alpha upscaling errors on Apple Silicon
+- **⚡ Fix: Cross-platform histogram matching** - Replaced scatter_ operation with argsort+index_select for improved reliability across CUDA, ROCm, and MPS backends
+- **🧹 MPS: Remove sync overhead** - Reverted unnecessary `torch.mps.synchronize()` calls introduced in v2.5.21 for consistent behavior with CUDA pipeline
+
+**2025.12.12 - Version 2.5.21**
+
+- **🛠️ Fix: GGUF dequantization error on MPS** - Resolved shape mismatch error introduced in 2.5.20 by skipping GGUF quantized buffers in precision conversion - these must remain in packed format for on-the-fly dequantization during inference
+- **🍎 MPS: Eliminate CPU sync overhead** - Skip unnecessary CPU tensor offload on Apple Silicon unified memory architecture, preventing sync stalls that caused slowdowns. Input images and output video now stay on MPS device throughout the pipeline
+- **⚡ MPS: Preload text embeddings** - Load text embeddings before Phase 1 encoding to avoid sync stall at Phase 2 start, improving timing accuracy and throughput
+- **🧹 MPS: Optimized model cleanup** - Skip redundant CPU movement before model deletion on unified memory
+
+**2025.12.12 - Version 2.5.20**
+
+- **⚡ 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**
+
+- **🔧 Fix: Older GPU compatibility (GTX 970, etc.)** - Runtime bf16 CUBLAS probe replaces compute capability heuristics, correctly detecting unsupported GPUs without affecting RTX 20XX
+
+**2025.12.05 - Version 2.5.16**
+
+- **🔧 Fix: Older GPU compatibility (GTX 970, etc.)** - Automatic fallback for GPUs without bfloat16 support
+- **🐛 Fix: Quality regression** - Reverted bfloat16 detection that was causing artifact issues
+- **📋 Debug: Environment info display** - Shows system info in debug mode to help with issue reporting
+- **📚 Docs: Simplified contribution workflow** - Streamlined to main branch only
+
+**2025.12.03 - Version 2.5.15**
+
+- **🍎 Fix: MPS compatibility** - Disable antialias for MPS tensors and fix bfloat16 arange issues
+- **⚡ Fix: Autocast device type** - Use proper device type attribute to prevent autocast errors
+- **📊 Memory: Accurate VRAM tracking** - Use max_memory_reserved for more precise peak reporting
+- **🔧 Fix: Triton compatibility** - Add shim for bitsandbytes 0.45+ / triton 3.0+ (fixes PyTorch 2.7 installation errors)
+
+**2025.12.01 - Version 2.5.14**
+
+- **🍎 Fix: MPS device comparison** - Normalize device strings to prevent unnecessary tensor movements
+- **📊 Memory: VRAM swap detection** - Peak stats now show GPU+swap breakdown when overflow occurs, with warning when swap detected
+- **🛡️ Memory: Enforce physical VRAM limit** - PyTorch now OOMs instead of silently swapping to shared memory (prevents extreme slowdowns on Windows)

 **2025.11.30 - Version 2.5.13**

@@ -193,7 +271,7 @@ 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**

@@ -240,8 +318,8 @@ 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
@@ -270,7 +348,7 @@ 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

@@ -385,14 +463,21 @@ 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
@@ -747,9 +832,20 @@ 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) with 10-bit video output (requires FFMPEG)
+# 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 \
+    --video_backend ffmpeg \
+    --10bit
+
 # Multi-GPU processing with temporal overlap
 python inference_cli.py video.mp4 \
    --cuda_device 0,1 \
@@ -795,6 +891,8 @@ python inference_cli.py media_folder/ \
 - `<input>`: Input file (.mp4, .avi, .png, .jpg, etc.) or directory
 - `--output`: Output path (default: auto-generated in 'output/' directory)
 - `--output_format`: Output format: 'mp4' (video) or 'png' (image sequence). Default: auto-detect from input type
+- `--video_backend`: Video encoder backend: 'opencv' (default) or 'ffmpeg' (requires ffmpeg in PATH)
+- `--10bit`: Save 10-bit video with x265 codec and yuv420p10le pixel format (reduces banding in gradients). Without this flag, ffmpeg uses x264 (yuv420p) for maximum compatibility. Requires --video_backend ffmpeg
 - `--model_dir`: Model directory (default: ./models/SEEDVR2)

 **Model Selection:**
@@ -806,7 +904,8 @@ 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)

@@ -818,9 +917,8 @@ 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
@@ -832,7 +930,8 @@ 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)
@@ -843,8 +942,8 @@ 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
@@ -935,7 +1034,7 @@ For detailed contribution guidelines, see [CONTRIBUTING.md](CONTRIBUTING.md).
 2. Create your feature branch (`git checkout -b feature/AmazingFeature`)
 3. Commit your changes (`git commit -m 'Add some AmazingFeature'`)
 4. Push to the branch (`git push origin feature/AmazingFeature`)
-5. Open a Pull Request to **main** branch for stable features or **nightly** branch for experimental features
+5. Open a Pull Request to the **main** branch

 **Get Help:**
 - YouTube: [AInVFX Channel](https://www.youtube.com/@AInVFX)
@@ -947,7 +1046,7 @@ 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), [thehhmdb](https://github.com/thehhmdb), [s-cerevisiae](https://github.com/s-cerevisiae), [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

@@ -3,6 +3,7 @@ ComfyUI-SeedVR2_VideoUpscaler
 Official SeedVR2 integration for ComfyUI
 """

+from .src.optimization.compatibility import ensure_triton_compat  # noqa: F401
 from .src.interfaces import comfy_entrypoint, SeedVR2Extension

 __all__ = ["comfy_entrypoint", "SeedVR2Extension"]
@@ -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 = "2.5.13"
+version = "2.5.24"
 authors = [
    {name = "numz"},
    {name = "adrientoupet"}
 ]
 license = {file = "LICENSE"} 
 classifiers = [
-     "Operating System :: OS Independent",
-     "Environment :: GPU :: NVIDIA CUDA"
+     "Operating System :: OS Independent"
 ]
 dependencies = [
    "torch",
@@ -36,7 +36,7 @@ class UniformTrailingSamplingTimesteps(SamplingTimesteps):
        dtype: torch.dtype = torch.float32,
    ):
        # Create trailing timesteps with specified dtype
-        timesteps = torch.arange(1.0, 0.0, -1.0 / steps, device=device, dtype=dtype)
+        timesteps = torch.arange(1.0, 0.0, -1.0 / steps, device='cpu').to(device=device, dtype=dtype)

        # Shift timesteps.
        timesteps = shift * timesteps / (1 + (shift - 1) * timesteps)
@@ -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())

@@ -337,13 +337,23 @@ def edge_guided_alpha_upscale(
    rgb_edges = detect_edges_batch(images=rgb_normalized, method='sobel', debug=debug)
    
    # Step 1: Initial bicubic upscale provides smooth base before edge refinement
-    alpha_upscaled = F.interpolate(
-        input_alpha,
-        size=(H_out, W_out),
-        mode='bicubic',
-        align_corners=False,
-        antialias=True
-    ).clamp(0, 1)
+    # MPS on PyTorch < 2.8 doesn't support bicubic+antialias - use CPU fallback
+    try:
+        alpha_upscaled = F.interpolate(
+            input_alpha,
+            size=(H_out, W_out),
+            mode='bicubic',
+            align_corners=False,
+            antialias=True
+        ).clamp(0, 1)
+    except NotImplementedError:
+        alpha_upscaled = F.interpolate(
+            input_alpha.cpu(),
+            size=(H_out, W_out),
+            mode='bicubic',
+            align_corners=False,
+            antialias=True
+        ).to(device).clamp(0, 1)
    
    if is_binary_mask:
        if debug:
@@ -231,7 +231,11 @@ def encode_all_batches(
    if images is None:
        raise ValueError("Images to encode must be provided")
    else:
-        ctx['input_images'] = images
+        # MPS: keep on device to avoid sync overhead in Phase 4 color correction
+        if ctx['vae_device'].type == 'mps' and images.device.type != 'mps':
+            ctx['input_images'] = images.to(ctx['vae_device'])
+        else:
+            ctx['input_images'] = images
    
    # Get total frame count from context (set in video_upscaler before encoding)
    total_frames = ctx.get('total_frames', len(images))
@@ -700,7 +704,7 @@ 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
@@ -708,10 +712,11 @@ 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']:
-                    with torch.autocast(str(ctx['dit_device']), ctx['compute_dtype'], enabled=True):
+                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,
                            conditions=conditions,
@@ -859,7 +864,13 @@ def decode_all_batches(
    
    # Pre-allocate final_video at the START of decode phase (before any batch processing)
    # This ensures we only need memory for final_video + 1 batch, not final_video + all batch_samples
-    target_device = ctx['tensor_offload_device'] if ctx['tensor_offload_device'] is not None else 'cpu'
+    # MPS: keep on device (unified memory, no benefit to CPU offload)
+    if ctx['tensor_offload_device'] is not None:
+        target_device = ctx['tensor_offload_device']
+    elif ctx['vae_device'].type == 'mps':
+        target_device = ctx['vae_device']
+    else:
+        target_device = 'cpu'
    channels_str = "RGBA" if C == 4 else "RGB"
    required_gb = (total_frames * true_h * true_w * C * 2) / (1024**3)
    debug.log(f"Pre-allocating output tensor: {total_frames} frames, {true_w}x{true_h}px, {channels_str} ({required_gb:.2f}GB)", 
@@ -1414,7 +1425,7 @@ 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'])
@@ -36,7 +36,7 @@ from .model_configuration import configure_runner
 from .infer import VideoDiffusionInfer
 from ..data.image.transforms.divisible_crop import DivisiblePad
 from ..data.image.transforms.na_resize import NaResize
-from ..optimization.compatibility import get_supported_compute_dtype
+from ..optimization.compatibility import COMPUTE_DTYPE, BFLOAT16_SUPPORTED
 from ..optimization.memory_manager import manage_tensor
 from ..utils.constants import get_script_directory

@@ -350,7 +350,12 @@ def setup_generation_context(
    vae_device = _normalize_device(vae_device)
    dit_offload_device = _normalize_device(dit_offload_device) if dit_offload_device is not None else None
    vae_offload_device = _normalize_device(vae_offload_device) if vae_offload_device is not None else None
-    tensor_offload_device = _normalize_device(tensor_offload_device) if tensor_offload_device is not None else None
+    # MPS unified memory: CPU offload causes sync overhead with no memory benefit
+    is_mps = dit_device.type == 'mps' or vae_device.type == 'mps'
+    if is_mps and tensor_offload_device is not None and str(tensor_offload_device) == 'cpu':
+        tensor_offload_device = None
+    else:
+        tensor_offload_device = _normalize_device(tensor_offload_device) if tensor_offload_device is not None else None
    
    # Set LOCAL_RANK to 0 for single-GPU inference mode
    # CLI multi-GPU uses CUDA_VISIBLE_DEVICES to restrict visibility per worker
@@ -372,7 +377,7 @@ def setup_generation_context(
        'dit_offload_device': dit_offload_device,
        'vae_offload_device': vae_offload_device,
        'tensor_offload_device': tensor_offload_device,
-        'compute_dtype': get_supported_compute_dtype(debug),  # Auto-detect best dtype for hardware
+        'compute_dtype': COMPUTE_DTYPE,
        'interrupt_fn': interrupt_fn,
        'video_transform': None,
        'text_embeds': None,
@@ -404,10 +409,10 @@ def setup_generation_context(
        )
        if ctx['compute_dtype'] == torch.float32:
            reason = "quality"
-        elif ctx['compute_dtype'] == torch.float16:
-            reason = "compatibility (GPU does not support bfloat16)"
+        elif not BFLOAT16_SUPPORTED:
+            reason = "compatibility (GPU lacks bfloat16 CUBLAS - 7B models unsupported, 3B may have artifacts)"
        else:
-            reason = "compatibility"
+            reason = "performance"
        debug.log(f"Unified compute dtype: {ctx['compute_dtype']} across entire pipeline for maximum {reason}", category="precision")
    
    return ctx
@@ -457,7 +462,7 @@ 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
        
@@ -529,8 +534,8 @@ 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,
@@ -819,4 +824,4 @@ 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)
@@ -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(str(device), 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(str(device), 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,
@@ -70,11 +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
+    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

@@ -171,7 +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')
+        attention_mode: Attention mode string ('sdpa', 'flash_attn_2', 'flash_attn_3', 'sageattn_2', or 'sageattn_3')
        
    Returns:
        Human-readable description string
@@ -181,7 +181,10 @@ def _describe_attention_mode(attention_mode: Optional[str]) -> str:
    
    mode_descriptions = {
        'sdpa': 'PyTorch SDPA',
-        'flash_attn': 'Flash Attention 2'
+        '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)
@@ -436,7 +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:
@@ -771,7 +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
        
@@ -792,6 +795,14 @@ def configure_runner(
    if debug is None:
        raise ValueError("Debug instance must be provided to configure_runner")
    
+    # 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, 
@@ -857,7 +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
@@ -1165,23 +1176,23 @@ 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
            compute_dtype = getattr(runner, '_compute_dtype', torch.bfloat16)            
@@ -1466,4 +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
@@ -146,7 +146,7 @@ def load_quantized_state_dict(checkpoint_path: str, device: torch.device = torch
                    handle_prefix="model.diffusion_model."
                )
    elif checkpoint_path.endswith('.pth'):
-        state = torch.load(checkpoint_path, map_location=device_str, mmap=True)
+        state = torch.load(checkpoint_path, map_location=device_str, mmap=True, weights_only=True)
    else:
        raise ValueError(f"Unsupported checkpoint format. Expected .safetensors or .pth, got: {checkpoint_path}")
    
@@ -393,6 +393,20 @@ class GGUFTensor(torch.Tensor):
                if debug:
                    debug.log(f"Error in {func.__name__} dequantization: {e}", level="WARNING", category="dit", force=True)
                raise
+
+        # Handle conv2d/conv3d operations (critical for GGUF VAE models)
+        # Conv3d layers (InflatedCausalConv3d) are not replaced by layer replacement
+        if func in {torch.nn.functional.conv2d, torch.nn.functional.conv3d}:
+            if len(args) >= 2 and isinstance(args[1], cls):  # weight is second arg
+                try:
+                    weight_tensor = args[1]
+                    dequantized_weight = weight_tensor.dequantize(device=args[0].device, dtype=args[0].dtype)
+                    new_args = (args[0], dequantized_weight) + args[2:]
+                    return func(*new_args, **kwargs)
+                except Exception as e:
+                    if debug:
+                        debug.log(f"Error in conv dequantization: {e}", level="WARNING", category="dit", force=True)
+                    raise
        
        # For ALL other operations, delegate to parent WITHOUT dequantization
        # This includes .cpu(), .to(), .device, .dtype, .shape, etc.
@@ -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]):
@@ -50,10 +51,12 @@ class AreaResize:

        resized_height, resized_width = round(height * scale), round(width * scale)

+        antialias = not (isinstance(image, torch.Tensor) and image.device.type == 'mps')
        return TVF.resize(
            image,
            size=(resized_height, resized_width),
            interpolation=self.interpolation,
+            antialias=antialias,
        )


@@ -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,
@@ -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]):
@@ -56,8 +57,9 @@ class SideResize:
        else:
            size = self.size

-        # Resize to shortest edge
-        resized = TVF.resize(image, size, self.interpolation)
+        # Resize to shortest edge (disable antialias only for MPS tensors - not supported)
+        antialias = not (isinstance(image, torch.Tensor) and image.device.type == 'mps')
+        resized = TVF.resize(image, size, self.interpolation, antialias=antialias)
        
        # Apply max_size constraint if specified
        if self.max_size > 0:
@@ -69,6 +71,6 @@ class SideResize:
            if max(h, w) > self.max_size:
                scale = self.max_size / max(h, w)
                new_h, new_w = round(h * scale), round(w * scale)
-                resized = TVF.resize(resized, (new_h, new_w), self.interpolation)
+                resized = TVF.resize(resized, (new_h, new_w), self.interpolation, antialias=antialias)
        
        return resized
@@ -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",
@@ -100,16 +102,19 @@ class SeedVR2LoadDiTModel(io.ComfyNode):
                    )
                ),
                io.Combo.Input("attention_mode",
-                    options=["sdpa", "flash_attn"],
+                    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"
+                        "• 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 provides speedup through optimized CUDA kernels on compatible GPUs."
+                        "Flash Attention and SageAttention provide speedup through optimized CUDA kernels on compatible GPUs."
                    )
                ),
                io.Custom("TORCH_COMPILE_ARGS").Input("torch_compile_args",
@@ -142,23 +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(
@@ -19,7 +19,9 @@ from ..core.generation_utils import (
    setup_generation_context, 
    prepare_runner,
    compute_generation_info,
-    log_generation_start
+    log_generation_start,
+    load_text_embeddings,
+    script_directory
 )
 from ..optimization.memory_manager import (
    cleanup_text_embeddings,
@@ -349,13 +351,12 @@ 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")
@@ -438,6 +439,10 @@ class SeedVR2VideoUpscaler(io.ComfyNode):
            # Store cache context in ctx for use in generation phases
            ctx['cache_context'] = cache_context

+            # Preload text embeddings before Phase 1 to avoid sync stall in Phase 2
+            ctx['text_embeds'] = load_text_embeddings(script_directory, ctx['dit_device'], ctx['compute_dtype'], debug)
+            debug.log("Loaded text embeddings for DiT", category="dit")
+
            debug.log_memory_state("After model preparation", show_tensors=False, detailed_tensors=False)
            debug.end_timer("model_preparation", "Model preparation", force=True, show_breakdown=True)
            
@@ -15,8 +15,11 @@
 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
+# 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

@@ -61,37 +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
-    )
-
-
 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"
@@ -107,12 +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):
@@ -120,7 +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__()
@@ -144,11 +120,26 @@ 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 == '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, **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
            return pytorch_varlen_attention(
@@ -15,8 +15,11 @@
 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
+# 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

@@ -61,37 +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
-    )
-
-
 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"
@@ -107,12 +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):
@@ -120,7 +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__()
@@ -144,11 +120,26 @@ 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 == '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, **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
            return pytorch_varlen_attention(
@@ -17,6 +17,7 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from diffusers.models.attention_processor import Attention, SpatialNorm
+from diffusers.models.autoencoders.vae import DecoderOutput, DiagonalGaussianDistribution
 from diffusers.models.downsampling import Downsample2D
 from diffusers.models.lora import LoRACompatibleConv
 from diffusers.models.modeling_outputs import AutoencoderKLOutput
@@ -45,8 +46,6 @@ from .types import (
    CausalAutoencoderOutput,
    CausalDecoderOutput,
    CausalEncoderOutput,
-    DecoderOutput,
-    DiagonalGaussianDistribution,
    MemoryState,
    _inflation_mode_t,
    _memory_device_t,
@@ -1094,7 +1093,7 @@ class VideoAutoencoderKL(diffusers.AutoencoderKL):
    ):
        extra_cond_dim = kwargs.pop("extra_cond_dim") if "extra_cond_dim" in kwargs else None
        self.slicing_sample_min_size = slicing_sample_min_size
-        self.slicing_latent_min_size = slicing_sample_min_size // (2**temporal_scale_num)
+        self.slicing_latent_min_size = max(1, slicing_sample_min_size // (2**temporal_scale_num))

        super().__init__(
            in_channels=in_channels,
@@ -1225,6 +1224,10 @@ class VideoAutoencoderKL(diffusers.AutoencoderKL):
        
        output = causal_conv_gather_outputs(output)
        
+        # MPS memory leak workaround (pytorch/pytorch#155060)
+        if self.device.type == 'mps':
+            torch.mps.empty_cache()
+        
        # Only transfer back if needed
        return output if output.device == x.device else output.to(x.device)

@@ -1241,6 +1244,10 @@ class VideoAutoencoderKL(diffusers.AutoencoderKL):
        output = self.decoder(_z, memory_state=memory_state)
        output = causal_conv_gather_outputs(output)
        
+        # MPS memory leak workaround (pytorch/pytorch#155060)
+        if self.device.type == 'mps':
+            torch.mps.empty_cache()
+        
        # Only transfer back if needed
        return output if output.device == z.device else output.to(z.device)

@@ -1711,7 +1718,7 @@ class VideoAutoencoderKLWrapper(VideoAutoencoderKL):
        if split_size is not None:
            self.enable_slicing()
            self.slicing_sample_min_size = split_size
-            self.slicing_latent_min_size = split_size // self.temporal_downsample_factor
+            self.slicing_latent_min_size = max(1, split_size // self.temporal_downsample_factor)
        else:
            self.disable_slicing()
        for module in self.modules():
@@ -74,51 +74,3 @@ class CausalEncoderOutput(NamedTuple):

 class CausalDecoderOutput(NamedTuple):
    sample: torch.Tensor
-
-
-class DecoderOutput:
-    """Output of decoding method - matches diffusers.models.autoencoders.vae.DecoderOutput"""
-    def __init__(self, sample: torch.Tensor, commit_loss: Optional[torch.Tensor] = None):
-        self.sample = sample
-        self.commit_loss = commit_loss
-
-
-class DiagonalGaussianDistribution:
-    """Matches diffusers.models.autoencoders.vae.DiagonalGaussianDistribution exactly."""
-    def __init__(self, parameters: torch.Tensor, deterministic: bool = False):
-        self.parameters = parameters
-        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
-        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
-        self.deterministic = deterministic
-        self.std = torch.exp(0.5 * self.logvar)
-        self.var = torch.exp(self.logvar)
-        if self.deterministic:
-            self.var = self.std = torch.zeros_like(
-                self.mean, device=self.parameters.device, dtype=self.parameters.dtype
-            )
-
-    def sample(self, generator: Optional[torch.Generator] = None) -> torch.Tensor:
-        if self.deterministic:
-            return self.mode()
-        sample = torch.randn(
-            self.mean.shape,
-            generator=generator,
-            device=self.parameters.device,
-            dtype=self.parameters.dtype,
-        )
-        return self.mean + self.std * sample
-
-    def mode(self) -> torch.Tensor:
-        return self.mean
-
-    def kl(self, other: Optional["DiagonalGaussianDistribution"] = None) -> torch.Tensor:
-        if other is None:
-            return 0.5 * torch.sum(
-                self.mean.pow(2) + self.var - 1.0 - self.logvar,
-                dim=[1, 2, 3],
-            )
-        return 0.5 * torch.sum(
-            (self.mean - other.mean).pow(2) / other.var
-            + self.var / other.var - 1.0 - self.logvar + other.logvar,
-            dim=[1, 2, 3],
-        )
@@ -15,6 +15,7 @@ from typing import Optional, Tuple, Literal, Callable, Union
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+from diffusers.models.autoencoders.vae import DiagonalGaussianDistribution
 from einops import rearrange
 from ....common.half_precision_fixes import safe_pad_operation

@@ -35,7 +36,6 @@ from .types import (
    CausalAutoencoderOutput,
    CausalDecoderOutput,
    CausalEncoderOutput,
-    DiagonalGaussianDistribution,
    MemoryState,
    _inflation_mode_t,
    _memory_device_t,
@@ -733,7 +733,7 @@ class VideoAutoencoderKL(nn.Module):
        if slicing_sample_min_size is None:
            slicing_sample_min_size = temporal_downsample_factor
        self.slicing_sample_min_size = slicing_sample_min_size
-        self.slicing_latent_min_size = slicing_sample_min_size // (2**temporal_scale_num)
+        self.slicing_latent_min_size = max(1, slicing_sample_min_size // (2**temporal_scale_num))

        # pass init params to Encoder
        self.encoder = Encoder3D(
@@ -886,7 +886,7 @@ class VideoAutoencoderKL(nn.Module):
        if split_size is not None:
            self.enable_slicing()
            self.slicing_sample_min_size = split_size
-            self.slicing_latent_min_size = split_size // self.temporal_downsample_factor
+            self.slicing_latent_min_size = max(1, split_size // self.temporal_downsample_factor)
        else:
            self.disable_slicing()
        for module in self.modules():
@@ -950,7 +950,7 @@ class VideoAutoencoderKLWrapper(VideoAutoencoderKL):
            self.disable_slicing()
        self.slicing_sample_min_size = split_size
        if split_size is not None:
-            self.slicing_latent_min_size = split_size // self.temporal_downsample_factor
+            self.slicing_latent_min_size = max(1, split_size // self.temporal_downsample_factor)
        for module in self.modules():
            if isinstance(module, InflatedCausalConv3d):
                module.set_memory_device(memory_device)
@@ -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
@@ -5,51 +5,546 @@ Contains FP8/FP16 compatibility layers and wrappers for different model architec
 Extracted from: seedvr2.py (lines 1045-1630)
 """

-import torch
+# 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."""
+    if 'triton.ops.matmul_perf_model' in sys.modules:
+        return
+    
+    try:
+        from triton.ops.matmul_perf_model import early_config_prune  # noqa: F401
+        return
+    except (ImportError, ModuleNotFoundError, AttributeError):
+        pass
+    
+    if 'triton.ops' not in sys.modules:
+        sys.modules['triton.ops'] = types.ModuleType('triton.ops')
+    
+    matmul_perf = types.ModuleType('triton.ops.matmul_perf_model')
+    matmul_perf.early_config_prune = lambda configs, *a, **kw: configs
+    matmul_perf.estimate_matmul_time = lambda *a, **kw: 0.0
+    
+    sys.modules['triton.ops'].matmul_perf_model = matmul_perf
+    sys.modules['triton.ops.matmul_perf_model'] = matmul_perf
+
+
+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, ValueError):
+        # Installation broken, not present, or version detection failed - 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 os

-# Flash Attention & Triton Compatibility Layer
-# 1. Flash Attention - speedup for attention operations
-try:
-    from flash_attn import flash_attn_varlen_func
-    FLASH_ATTN_AVAILABLE = True
-except ImportError:
-    flash_attn_varlen_func = None
-    FLASH_ATTN_AVAILABLE = False

-def validate_flash_attention_availability(requested_mode: str, debug=None) -> str:
+# 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:
+    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
+
+# 2. Flash Attention 2 (wider compatibility, supports dropout/window)
+flash_attn_2_varlen_func = None
+FLASH_ATTN_2_AVAILABLE = False
+try:
+    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_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_attention_mode(requested_mode: str, debug=None) -> str:
    """
-    Validate Flash Attention availability and warn if fallback needed.
+    Validate attention mode availability with automatic fallback.
    
    Args:
-        requested_mode: Either 'flash_attn' or 'sdpa'
+        requested_mode: 'sdpa', 'flash_attn_2', 'flash_attn_3', 'sageattn_2', or 'sageattn_3'
        debug: Optional debug instance for logging
        
    Returns:
-        Validated mode ('flash_attn' or 'sdpa')
+        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'
+    
+    # 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(
+                    "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
@@ -97,11 +592,11 @@ def validate_gguf_availability(operation: str = "load GGUF model", debug=None) -
        raise RuntimeError(f"GGUF library required to {operation}")


-# 4. NVIDIA Conv3d Memory Bug - Workaround for PyTorch 2.9-2.10 + cuDNN >= 91002
+# 4. NVIDIA Conv3d Memory Bug - Workaround for PyTorch >= 2.9 + cuDNN >= 91002
 def _check_conv3d_memory_bug():
    """
    Check if Conv3d memory bug workaround needed.
-    Bug: PyTorch 2.9-2.10 with cuDNN >= 91002 uses 3x memory for Conv3d 
+    Bug: PyTorch 2.9+ with cuDNN >= 91002 uses 3x memory for Conv3d 
    with fp16/bfloat16 due to buggy dispatch layer.
    """
    try:
@@ -127,7 +622,8 @@ def _check_conv3d_memory_bug():
        parts = version_str.split('.')
        torch_version = tuple(int(p) for p in parts[:2])
        
-        if not ((2, 9) <= torch_version <= (2, 10)):
+        # Bug affects PyTorch 2.9 and later versions
+        if torch_version < (2, 9):
            return False
        
        if not hasattr(torch.backends.cudnn, 'version'):
@@ -144,60 +640,38 @@ def _check_conv3d_memory_bug():
 NVIDIA_CONV3D_MEMORY_BUG_WORKAROUND = _check_conv3d_memory_bug()


-def get_supported_compute_dtype(debug=None) -> torch.dtype:
-    """
-    Detect the best supported compute dtype for the current hardware.
-    
-    bfloat16 requires CUDA compute capability 8.0+ (Ampere and newer).
-    Older CUDA GPUs fall back to float16 to avoid CUBLAS_STATUS_NOT_SUPPORTED errors.
-    
-    Args:
-        debug: Optional debug instance for logging warnings
-    
-    Returns:
-        torch.bfloat16 if supported, torch.float16 for older CUDA GPUs
-    """
-    try:
-        # CUDA: Check compute capability
-        if torch.cuda.is_available():
-            major, minor = torch.cuda.get_device_capability()
-            # bfloat16 requires compute capability 8.0+ (Ampere: A100, RTX 30xx)
-            if major < 8:
-                if debug:
-                    gpu_name = torch.cuda.get_device_name()
-                    debug.log(
-                        f"GPU '{gpu_name}' (compute capability {major}.{minor}) does not support bfloat16. "
-                        f"Using float16 instead - 7B models are not supported and will render black, 3B models may have artifacts.",
-                        level="WARNING", category="precision", force=True
-                    )
-                return torch.float16
-        
-        # Default: bfloat16 (MPS, CPU, or CUDA 8.0+)
-        return torch.bfloat16
-        
-    except Exception:
-        return torch.bfloat16
-
-
 # Log all optimization status once globally (cross-process) using environment variable
 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:
@@ -206,33 +680,55 @@ if not os.environ.get("SEEDVR2_OPTIMIZATIONS_LOGGED"):
        print(f"🔧 Conv3d workaround active: PyTorch {torch_ver}, cuDNN {cudnn_ver} (fixing VAE 3x memory bug)")


+# Bfloat16 CUBLAS support
+def _probe_bfloat16_support() -> bool:
+    if not torch.cuda.is_available():
+        return True
+    try:
+        a = torch.randn(8, 8, dtype=torch.bfloat16, device='cuda:0')
+        _ = torch.matmul(a, a)
+        del a
+        return True
+    except RuntimeError as e:
+        if "CUBLAS_STATUS_NOT_SUPPORTED" in str(e):
+            return False
+        raise
+
+BFLOAT16_SUPPORTED = _probe_bfloat16_support()
+COMPUTE_DTYPE = torch.bfloat16 if BFLOAT16_SUPPORTED else torch.float16
+
+
 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
    """
@@ -255,9 +751,12 @@ 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")
@@ -267,11 +766,6 @@ 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"""
@@ -333,8 +827,11 @@ class FP8CompatibleDiT(torch.nn.Module):
                    param.data = param.data.to(target_dtype)
                converted_count += 1
                
-        # Also convert buffers
+        # Also convert buffers (skip GGUF quantized buffers - they have tensor_type attribute)
        for name, buffer in self.dit_model.named_buffers():
+            # Skip GGUF quantized buffers - these must stay in packed format for on-the-fly dequantization
+            if hasattr(buffer, 'tensor_type'):
+                continue
            if buffer.dtype != target_dtype:
                if buffer.device.type == "mps":
                    temp_cpu = buffer.data.to("cpu")
@@ -394,210 +891,6 @@ 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"""
-        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 = type(module).__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):
        """
@@ -10,8 +10,40 @@ 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'."""
+    s = str(device).upper()
+    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]:
    """
@@ -31,14 +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:
@@ -60,7 +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]:
    """
@@ -74,13 +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()
@@ -100,13 +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 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.
@@ -116,29 +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_allocated_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_allocated = torch.cuda.max_memory_allocated(device) / (1024**3)
-            return allocated, reserved, max_allocated
-        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]:
@@ -229,17 +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
@@ -262,10 +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:
@@ -302,7 +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
@@ -379,7 +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):
@@ -591,7 +624,7 @@ def manage_tensor(
    target_dtype = dtype if dtype is not None else current_dtype
    
    # Check if movement is actually needed
-    needs_device_move = current_device != target_device
+    needs_device_move = _device_str(current_device) != _device_str(target_device)
    needs_dtype_change = dtype is not None and current_dtype != target_dtype
    
    if not needs_device_move and not needs_dtype_change:
@@ -609,8 +642,8 @@ def manage_tensor(
    
    # Log the movement
    if debug:
-        current_device_str = str(current_device).upper()
-        target_device_str = str(target_device).upper()
+        current_device_str = _device_str(current_device)
+        target_device_str = _device_str(target_device)
        
        dtype_info = ""
        if needs_dtype_change:
@@ -669,7 +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

@@ -681,8 +714,8 @@ def manage_model_device(model: torch.nn.Module, target_device: torch.device, mod
    
    # Extract device type for comparison (both are torch.device objects)
    target_type = target_device.type
-    current_device_upper = str(current_device).upper()
-    target_device_upper = str(target_device).upper()
+    current_device_upper = _device_str(current_device)
+    target_device_upper = _device_str(target_device)

    # Compare normalized device types
    if current_device_upper == target_device_upper and not is_blockswap_model:
@@ -737,10 +770,10 @@ def _handle_blockswap_model_movement(runner: Any, model: torch.nn.Module,
                actual_source_device = param.device
                break
        
-        source_device_desc = str(actual_source_device).upper() if actual_source_device else str(target_device).upper()
+        source_device_desc = _device_str(actual_source_device) if actual_source_device else _device_str(target_device)
        
        if debug:
-            debug.log(f"Moving {model_name} from {source_device_desc} to {str(target_device).upper()} ({reason or 'model caching'})", category="general")
+            debug.log(f"Moving {model_name} from {source_device_desc} to {_device_str(target_device)} ({reason or 'model caching'})", category="general")
        
        # Enable bypass to allow movement
        set_blockswap_bypass(runner=runner, bypass=True, debug=debug)
@@ -755,7 +788,7 @@ def _handle_blockswap_model_movement(runner: Any, model: torch.nn.Module,
        model.zero_grad(set_to_none=True)
        
        if debug:
-            debug.end_timer(timer_name, f"BlockSwap model offloaded to {str(target_device).upper()}")
+            debug.end_timer(timer_name, f"BlockSwap model offloaded to {_device_str(target_device)}")
        
        return True
        
@@ -775,10 +808,10 @@ def _handle_blockswap_model_movement(runner: Any, model: torch.nn.Module,
                actual_current_device = param.device
                break
        
-        current_device_desc = str(actual_current_device).upper() if actual_current_device else "OFFLOAD"
+        current_device_desc = _device_str(actual_current_device) if actual_current_device else "OFFLOAD"
        
        if debug:
-            debug.log(f"Moving {model_name} from {current_device_desc} to {str(target_device).upper()} ({reason or 'inference requirement'})", category="general")
+            debug.log(f"Moving {model_name} from {current_device_desc} to {_device_str(target_device)} ({reason or 'inference requirement'})", category="general")
        
        timer_name = f"{model_name.lower()}_to_gpu"
        if debug:
@@ -818,7 +851,7 @@ def _handle_blockswap_model_movement(runner: Any, model: torch.nn.Module,
                    blocks_on_gpu = model._block_swap_config.get('total_blocks', 32) - model._block_swap_config.get('blocks_swapped', 16)
                    total_blocks = model._block_swap_config.get('total_blocks', 32)
                    main_device = model._block_swap_config.get('main_device', 'GPU')
-                    debug.log(f"BlockSwap blocks restored to configured devices ({blocks_on_gpu}/{total_blocks} blocks on {str(main_device).upper()})", category="success")
+                    debug.log(f"BlockSwap blocks restored to configured devices ({blocks_on_gpu}/{total_blocks} blocks on {_device_str(main_device)})", category="success")
                else:
                    debug.log("BlockSwap blocks restored to configured devices", category="success")

@@ -865,8 +898,8 @@ def _standard_model_movement(model: torch.nn.Module, current_device: torch.devic
    
    # Log the movement with full device strings
    if debug:
-        current_device_str = str(current_device).upper()
-        target_device_str = str(target_device).upper()
+        current_device_str = _device_str(current_device)
+        target_device_str = _device_str(target_device)
        debug.log(f"Moving {model_name} from {current_device_str} to {target_device_str} ({reason})", category="general")

    # Start timer based on direction
@@ -891,7 +924,7 @@ def _standard_model_movement(model: torch.nn.Module, current_device: torch.devic
    
    # End timer
    if debug:
-        debug.end_timer(timer_name, f"{model_name} moved to {str(target_device).upper()}")
+        debug.end_timer(timer_name, f"{model_name} moved to {_device_str(target_device)}")
    
    return True

@@ -1017,15 +1050,17 @@ def cleanup_dit(runner: Any, debug: Optional['Debug'] = None, cache_model: bool
        
        # Move model off GPU if needed
        if param_device.type not in ['meta', 'cpu']:
-            # Get offload target - default to 'cpu' if not configured or set to 'none'
-            offload_target = getattr(runner, '_dit_offload_device', None)
-            if offload_target is None or offload_target == 'none':
-                offload_target = torch.device('cpu')
-            
-            # Move model off GPU (either for caching or before deletion)
-            reason = "model caching" if cache_model else "releasing GPU memory"
-            manage_model_device(model=runner.dit, target_device=offload_target, model_name="DiT", 
-                               debug=debug, reason=reason, runner=runner)
+            # MPS: skip CPU movement before deletion (unified memory, just causes sync)
+            if param_device.type == 'mps' and not cache_model:
+                if debug:
+                    debug.log("DiT on MPS - skipping CPU movement before deletion", category="cleanup")
+            else:
+                offload_target = getattr(runner, '_dit_offload_device', None)
+                if offload_target is None or offload_target == 'none':
+                    offload_target = torch.device('cpu')
+                reason = "model caching" if cache_model else "releasing GPU memory"
+                manage_model_device(model=runner.dit, target_device=offload_target, model_name="DiT", 
+                                   debug=debug, reason=reason, runner=runner)
        elif param_device.type == 'meta' and debug:
            debug.log("DiT on meta device - keeping structure for cache", category="cleanup")
    except StopIteration:
@@ -1093,15 +1128,17 @@ def cleanup_vae(runner: Any, debug: Optional['Debug'] = None, cache_model: bool
        
        # Move model off GPU if needed
        if param_device.type not in ['meta', 'cpu']:
-            # Get offload target - default to 'cpu' if not configured or set to 'none'
-            offload_target = getattr(runner, '_vae_offload_device', None)
-            if offload_target is None or offload_target == 'none':
-                offload_target = torch.device('cpu')
-            
-            # Move model off GPU (either for caching or before deletion)
-            reason = "model caching" if cache_model else "releasing GPU memory"
-            manage_model_device(model=runner.vae, target_device=offload_target, model_name="VAE", 
-                               debug=debug, reason=reason, runner=runner)
+            # MPS: skip CPU movement before deletion (unified memory, just causes sync)
+            if param_device.type == 'mps' and not cache_model:
+                if debug:
+                    debug.log("VAE on MPS - skipping CPU movement before deletion", category="cleanup")
+            else:
+                offload_target = getattr(runner, '_vae_offload_device', None)
+                if offload_target is None or offload_target == 'none':
+                    offload_target = torch.device('cpu')
+                reason = "model caching" if cache_model else "releasing GPU memory"
+                manage_model_device(model=runner.vae, target_device=offload_target, model_name="VAE", 
+                                   debug=debug, reason=reason, runner=runner)
        elif param_device.type == 'meta' and debug:
            debug.log("VAE on meta device - keeping structure for cache", category="cleanup")
    except StopIteration:
@@ -381,6 +381,8 @@ def _rgb_to_lab_batch(rgb: Tensor, device: torch.device, matrix: Tensor, epsilon
    rgb_flat = rgb_linear.permute(0, 2, 3, 1).reshape(-1, 3)
    del rgb_linear
    
+    # Ensure dtype consistency for matrix multiplication
+    rgb_flat = rgb_flat.to(dtype=matrix.dtype)
    xyz_flat = torch.matmul(rgb_flat, matrix.T)
    del rgb_flat
    
@@ -452,6 +454,8 @@ def _lab_to_rgb_batch(lab: Tensor, device: torch.device, matrix_inv: Tensor, eps
    xyz_flat = xyz.permute(0, 2, 3, 1).reshape(-1, 3)
    del xyz
    
+    # Ensure dtype consistency for matrix multiplication
+    xyz_flat = xyz_flat.to(dtype=matrix_inv.dtype)
    rgb_linear_flat = torch.matmul(xyz_flat, matrix_inv.T)
    del xyz_flat
    
@@ -490,6 +494,7 @@ def _histogram_matching_channel(source: Tensor, reference: Tensor, device: torch
    # Sort both arrays
    source_sorted, source_indices = torch.sort(source_flat)
    reference_sorted, _ = torch.sort(reference_flat)
+    del reference_flat
    
    # Quantile mapping
    n_source = len(source_sorted)
@@ -503,12 +508,15 @@ def _histogram_matching_channel(source: Tensor, reference: Tensor, device: torch
        ref_indices = (source_quantiles * (n_reference - 1)).long()
        ref_indices.clamp_(0, n_reference - 1)
        matched_sorted = reference_sorted[ref_indices]
-        del source_quantiles, ref_indices
+        del source_quantiles, ref_indices, reference_sorted
    
-    # Reconstruct with matched values
-    matched_flat = torch.empty_like(source_flat)
-    matched_flat.scatter_(0, source_indices, matched_sorted)
-    del source_flat, reference_flat, source_sorted, source_indices, reference_sorted, matched_sorted
+    del source_sorted, source_flat
+    
+    # Reconstruct using argsort (portable across CUDA/ROCm/MPS)
+    inverse_indices = torch.argsort(source_indices)
+    del source_indices
+    matched_flat = matched_sorted[inverse_indices]
+    del matched_sorted, inverse_indices
    
    return matched_flat.reshape(original_shape)

@@ -748,11 +756,15 @@ def _histogram_match_1d(source: Tensor, reference: Tensor, device: torch.device)
        ref_indices = (source_quantiles * (n_reference - 1)).long()
        ref_indices.clamp_(0, n_reference - 1)
        matched_sorted = reference_sorted[ref_indices]
-        del source_quantiles, ref_indices
+        del source_quantiles, ref_indices, reference_sorted
    
-    matched = torch.empty_like(source)
-    matched.scatter_(0, source_indices, matched_sorted)
-    del source_sorted, source_indices, reference_sorted, matched_sorted
+    del source_sorted
+    
+    # Reconstruct using argsort (portable across CUDA/ROCm/MPS)
+    inverse_indices = torch.argsort(source_indices)
+    del source_indices
+    matched = matched_sorted[inverse_indices]
+    del matched_sorted, inverse_indices
    
    return matched

@@ -4,7 +4,7 @@ Only includes constants actually used in the codebase
 """

 # Version information
-__version__ = "2.5.13"
+__version__ = "2.5.24"

 import os
 import warnings
@@ -10,10 +10,35 @@ 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_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:
    """
    Unified debug logging for generation pipeline and BlockSwap monitoring
@@ -53,7 +78,7 @@ class Debug:
        "device": "🖥️",       # Device info
        "file": "📂",         # File operations
        "alpha": "👻",        # Alpha operations
-        "star": "⭐",         # Star
+        "starlove": "⭐💝",   # Star + love
        "dialogue": "💬",     # Dialogue
        "none" : "",
    }
@@ -72,7 +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
@@ -117,34 +143,125 @@ 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:
+            self._print_environment_info(cli)
+
+    def _print_environment_info(self, cli: bool = False) -> None:
+        """Print concise environment info for bug reports - zero cost when debug disabled"""
+        import platform
+        import sys
+        
+        # OS
+        os_name = platform.system()
+        if os_name == "Windows":
+            os_str = f"Windows ({platform.version()})"
+        elif os_name == "Darwin":
+            os_str = f"macOS {platform.mac_ver()[0]}"
+        else:
+            try:
+                distro = platform.freedesktop_os_release()
+                os_str = f"{distro.get('NAME', 'Linux')} {distro.get('VERSION_ID', '')}"
+            except (OSError, AttributeError):
+                os_str = f"Linux {platform.release()}"
+        
+        # Python & PyTorch & CUDA
+        py_ver = f"{sys.version_info.major}.{sys.version_info.minor}.{sys.version_info.micro}"
+        torch_ver = torch.__version__
+        cuda_ver = getattr(torch.version, 'cuda', None) or "N/A"
+        
+        # GPU
+        if is_cuda_available():
+            try:
+                props = torch.cuda.get_device_properties(0)
+                gpu_str = f"{props.name} ({round(props.total_memory / (1024**3))}GB)"
+                cudnn_ver = str(torch.backends.cudnn.version()) if torch.backends.cudnn.is_available() else "N/A"
+            except Exception:
+                gpu_str = "CUDA"
+                cudnn_ver = "N/A"
+        elif is_mps_available():
+            gpu_str = "Apple Silicon (MPS)"
+            cudnn_ver = "N/A"
+        else:
+            gpu_str = "CPU"
+            cudnn_ver = "N/A"
+        
+        # Flash Attn, SageAttn & Triton - reuse existing module constants
+        try:
+            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 = sage_str = triton_str = "?"
+        
+        # ComfyUI version
+        comfy_str = None
+        if not cli:
+            try:
+                from comfyui_version import __version__ as comfy_ver
+                comfy_str = comfy_ver
+            except ImportError:
+                pass
+        
+        # Print
+        self.log(f"OS: {os_str} | GPU: {gpu_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")

    def print_footer(self) -> None:
        """Print the footer with links - always displayed"""
        self.log("", category="none", force=True)
        self.log("────────────────────────", category="none", force=True)
-        self.log("Questions? Updates? Watch the videos, star the repo & join us!", category="dialogue", force=True)
+        self.log("Questions? Updates? Watch, star & sponsor if you can!", category="dialogue", force=True)
        self.log("https://www.youtube.com/@AInVFX", category="generation", force=True)
-        self.log("https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler", category="star", force=True)
+        self.log("https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler", category="starlove", force=True)
    
    @torch._dynamo.disable  # Skip tracing to avoid time.time() warnings
    def start_timer(self, name: str, force: bool = False) -> None:
@@ -307,7 +424,15 @@ class Debug:
        if show_diff and self.memory_checkpoints:
            self._log_memory_diff(current_metrics=memory_info, force=force)

-       # Log detailed analysis if requested
+        # 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
        if detailed_tensors and tensor_stats.get('details'):
            self._log_detailed_tensor_analysis(details=tensor_stats['details'], force=force)
                
@@ -316,10 +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(
@@ -331,13 +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,
@@ -346,45 +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"
-                metrics['summary_vram'] = (f"  [{backend}] {metrics['vram_allocated']:.2f}GB allocated / "
-                        f"{metrics['vram_reserved']:.2f}GB reserved / "
-                        f"Peak: {metrics['vram_peak_since_last']:.2f}GB / "
-                        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
    
@@ -512,8 +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 = {
@@ -523,31 +649,42 @@ 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 overflow formatting (Windows only)
+        total_vram_gb = 0.0
+        if not is_mps:
+            vram_info = get_basic_vram_info(device=None)
+            if "error" not in vram_info:
+                total_vram_gb = vram_info["total_gb"]
        
        self.log("", category="none", force=force)
        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}): VRAM {vram:.2f}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: VRAM {overall_vram:.2f}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:
@@ -657,6 +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