RocM Ninodes

ROCm Ninodes: ROCm-Optimized Nodes for ComfyUI (v2.2.9)

ROCm Ninodes provides ComfyUI nodes tuned for AMD GPUs with ROCm (e.g. gfx1151 / Strix Halo): VAE decode, KSampler, checkpoint/diffusion/GGUF/LoRA loaders, LTX2 prompt generation, SamplerCustomAdvanced drop-in, and performance/memory monitoring. Install via ComfyUI Manager, comfy node install rocm-ninodes, or clone into custom_nodes.

⬆️ Upgrade to v2 (Required for existing users)

If you were on v1.x, run the upgrade script to clean legacy files and ensure the new package layout is detected by ComfyUI.

Windows (PowerShell)

uv run python tools/upgrade_to_v2.py

Linux/Mac

uv run python tools/upgrade_to_v2.py

What it does:

• Backs up legacy rocm_nodes.py to backup/rocm_nodes.py.bak (if present)
• Removes any temporary temp_*.py files from earlier extractions
• Verifies rocm_nodes/ package structure is intact
• Prints next steps (restart ComfyUI)

After running:

1. Restart ComfyUI completely
2. Verify nodes appear under "ROCm Ninodes" categories
3. If nodes don’t appear, clear ComfyUI cache and restart again

ROCm Ninodes is a custom node collection tuned for AMD GPUs with ROCm (especially gfx1151). It includes optimized VAE decode, KSampler, checkpoint/diffusion/GGUF/LoRA loaders, LTX2 prompt generation, SamplerCustomAdvanced drop-in, and monitoring nodes to maximize performance on AMD hardware with mature ROCm drivers.

🚀 What's new in v2.2.9

• Fixed visible seams + extra frames in temporal VAE tiling (rocm_nodes/core/vae.py:673): The overlap-blend step was blending each new chunk against the truncated tail of the previous chunk (result_parts[-1]) instead of the cumulative result tensor. After chunk 1 was shortened by [1:] to drop the first decoded frame, the next blend used that 24/32-frame tail as its reference, so blend_frames alternated 32 → 24 → 32 → 24 across chunks. This left a visible seam every two chunks in the back half of long videos and appended 8 extra frames per 2 chunks (1393 frames instead of 1201 for a 50 s LTX job). Blending is now done against the cumulative result tensor with torch.cat (which returns a fresh tensor), so the count is exact and seams disappear.

🚀 What's new in v2.2.8

• Fixed Windows access violation on LTX video decode (rocm_nodes/core/vae.py): Removed the unconditional comfy_has_chunked_io = False override for LTX. Disabling chunked IO pushed ComfyUI's decode through out.to(device=cpu, dtype=fp32, copy=True), a combined op that crashes on Windows with a memmove access violation under the ZLUDA/ROCm backend. LTX now uses the stock chunked path, which streams writes into a pre-allocated CPU buffer — no combined cast, lower GPU peak memory, no crash.
• enable_temporal_tiling is now a 3-way combo (auto / enable / disable, default auto): auto enables temporal tiling when the estimated decode output exceeds 3 GB, and logs a one-line recommendation when the output is long but still under the threshold. Explicit enable / disable are honored verbatim. Boolean True / False from old workflows are accepted and mapped to enable / disable for backward compatibility.
• Visible previewer errors on Windows (rocm_nodes/core/sampler.py): the three ROCm sampler callbacks used to swallow the previewer's decode_latent_to_preview_image exception silently, which is why LTX video previews appeared to vanish on Windows (the taeltx_2 TAESD previewer hit the same ZLUDA crash internally). The exception type and message are now logged once per failed step.

🚀 What's new in v2.2.7

• Fixed Windows access violation in temporal VAE tiling (rocm_nodes/core/vae.py:700): The overlap-blend step was using in-place slice assignment (result_parts[-1][-blend_frames:] = ...), which triggers a fatal access violation on Windows under the ZLUDA/ROCm backend. The blend is now built with pure tensor arithmetic and torch.cat, then the list entry is replaced — no in-place write into a strided slice, so long LTX/WAN videos decode cleanly on Windows + AMD.

🚀 What's new in v2.2.6

• Fixed severe ROCm sampler slowdown on LTX 2.3 video (architecture.py, vae.py): Removed torch.backends.cuda.matmul.allow_tf32 = False which unconditionally disabled TF32-accelerated matrix multiplications on AMD GPUs. This forced all matmuls to fp32 precision, making DiT-based models like LTX Video ~2x slower. TF32 is now left at its ROCm default (True), restoring stock performance while keeping all other ROCm optimizations.

🚀 What's new in v2.2.5

• Fixed illegal memory access on LTX-2.3-22b (architecture.py): allow_fp16_accumulation=True is now skipped for RDNA 3.5 (gfx1151 / Strix Halo). This global matmul setting caused numerical drift and hipErrorIllegalAddress crashes in bf16 flow-matching models (LTX Video). Other AMD architectures (RDNA 2/3, CDNA) still enable it for performance.
• ROCm backend settings applied (sampler.py): ROCMSamplerCustomAdvanced now calls apply_rocm_backend_settings() and runs emergency_memory_cleanup() for high-memory models (factor > 5x) — consistent with the other ROCm samplers.
• VRAM safety check: Warns before sampling if < 2GB free or > 95% VRAM used, with actionable suggestions (disable CacheDiT, remove --reserve-vram, reduce frame count).
• Flash attention override: New advanced flash_attention parameter (auto / disable_flash / force_flash) allows toggling PyTorch SDP backends per-node. force_flash prefers flash but keeps math/efficient as fallback.
• Model architecture detection: detect_model_sampling_type() now returns model_architecture field ("ltx", "flux", "wan", "sd", "unknown") for better diagnostic logging.

🚀 What's new in v2.2.4

• Cancel responsiveness: ROCMSamplerCustomAdvanced callback and ROCM VAE Decode tiling loops now call model_management.throw_exception_if_processing_interrupted() each step/chunk — pressing Cancel stops generation immediately instead of waiting for the current operation to finish.

🚀 What's new in v2.2.3

• Fixed GPU crash on large models (sampler): Removed torch.backends.cuda.matmul.allow_tf32=False and all emergency_memory_cleanup() calls from ROCMSamplerCustomAdvanced. The TF32 flag is NVIDIA-specific and could trigger hipErrorLaunchFailure on ROCm 7.13 with 46GB+ models. Node now only adds logging and the enhanced callback — zero global torch state changes.

🚀 What's new in v2.2.2

• Fixed dark vertical bands (sampler): Removed allow_fp16_accumulation=True which caused numerical drift in bf16 flow-matching models (LTX), producing banding artifacts.
• Fixed first-frame oversaturation (VAE): Causal video VAEs (LTX) now keep model weights in native precision. Input tensors are still fp16 for memory efficiency, but the causal state no longer accumulates fp16 errors from frame 0.
• Clean interface: Removed optimize_for_video and precision_mode params. Video auto-detected from latent shape. Only compatibility_mode remains as an advanced debug toggle.

🚀 What's new in v2.2.1

• Callback perf fix: Previewer created once (not every 5 steps), console timing at 10% milestones only — eliminates I/O overhead during sampling.
• Clean interface: Removed optimize_for_video and precision_mode noise parameters. Video is auto-detected from latent shape. Only compatibility_mode remains as an advanced toggle.

🚀 What's new in v2.2.0

• ROCm SamplerCustomAdvanced (rocm_nodes/core/sampler.py): Drop-in replacement for ComfyUI's V3 SamplerCustomAdvanced using the new define_schema() API. Preserves the exact same interface (noise, guider, sampler, sigmas, latent_image) while adding architecture-aware ROCm backend tuning, emergency memory defrag for high-memory models (LTX 128ch latents), enhanced per-step callback with ETA, and optional video-workflow mode. Category: ROCm Ninodes → Sampling.
• ROCm SamplerCustomAdvanced Benchmark: A/B comparison node that runs stock then ROCm-optimized on identical inputs, reporting timing, peak memory, speedup, and model info. Category: ROCm Ninodes → Sampling.
• Precision management: precision_mode (auto|fp32|bf16) and compatibility_mode toggle for the new custom sampler node.
• Full documentation: docs/SAMPLER_CUSTOM_ADVANCED.md covers implementation details, input/output reference, and usage for both agents and users.

🚀 What's new in v2.1.1

• ROCm KSampler auto-detection: Sampler now detects GPU architecture and model type automatically. No new toggles.
  • Flow-matching models (z-image, z-image-turbo, ernie, ideogram4) detected and logged
  • Pixel-space models (z-image-turbo) recognized — no VAE decode needed downstream
  • High-memory models (ideogram4, factor 11.6x) get automatic memory cleanup before/after sampling
  • Architecture-aware backend settings: allow_tf32=False, allow_fp16_accumulation=True
• Shared architecture detection: GPU architecture detection (detect_architecture()) extracted to shared module used by both VAE decode and KSampler.
• Updated Sampler Performance Monitor: Architecture-aware and model-type-aware recommendations.

🚀 What's new in v2.1.0

• LTX Video VAE support: Full support for Lightricks LTX VideoVAE (128 channels, 32× spatial compression). Automatic detection, fp16 optimized, full-video causal decode.
• z-image / z-image-turbo passthrough: Pixel-space VAEs (latent_channels=3, spatial_compression=1) detected automatically and processed as passthrough — no VAE decode needed.
• fp16 default on AMD: VAE decode defaults to fp16 on all AMD GPUs (Strix Halo, RDNA 2/3, CDNA). Doubles memory efficiency for large-channel VAEs.
• Per-architecture tuning: Automatic GPU architecture detection (gfx1151, gfx1100, gfx1030, CDNA) with per-arch tile sizes, precision, and batch caps.
• Temporal tiling for long videos (optional, disabled by default): Decode long LTX/WAN videos in temporal chunks with overlap to prevent OOM. See "ROCm VAE Decode" settings for enable_temporal_tiling.
• Chunked IO bypass for LTX: Disables giant pre-allocation for LTX VideoVAE, fixing crashes with PYTORCH_ALLOC_CONF=max_split_size_mb.
• Fixed memory estimation: Correctly accounts for all tensor dimensions (including temporal for 5D) and actual dtype size.
• Memory estimation for APU: Strix Halo unified memory now uses psutil for system-wide available RAM.

🚀 What's new in v2.0.11

• ROCm Text Generate LTX2 Prompt: Drop-in replacement for ComfyUI's TextGenerateLTX2Prompt — same inputs (clip, prompt, max_length, optional image, sampling params) and output. Tuned for ROCm/gfx1151 with device and memory handling; for best speed run ComfyUI with --use-pytorch-attention. Category: ROCm Ninodes → Generative AI.
• Generative AI category: LTX2 prompt node appears under ROCm Ninodes / Generative AI.

🚀 What's new in v2.0.10

• Checkpoint cache: ROCm Checkpoint Loader caches the last loaded checkpoint by name; repeat runs with the same checkpoint skip reload (saves time and avoids extra VRAM churn). Use use_cache (default on) and force_reload when you need a fresh load.
• FP8 / quantized checkpoint reuse: Fixed "Cannot set version_counter for inference tensor" when reusing a cached checkpoint that uses MixedPrecisionOps (e.g. LTX FP8 text encoder). Requires the corresponding fix in ComfyUI's comfy/ops.py (QuantizedTensor handling in _apply).

🚀 What's new in v2.0.1

• WAN VAE Video Fix: Fixed jitter/repeating frames issue for WAN models
  • Automatic WAN VAE detection with multiple fallback methods
  • Disabled chunking for WAN VAEs to preserve causal decoding chain
  • Matches native ComfyUI behavior exactly for WAN video processing

🚀 What's new in v2

• Stock-correct samplers with ROCm opt-in controls:
  • optimize_for_video (disable preview/progress on multi-frame latents)
  • precision_mode (auto|fp32|bf16) with ROCm bf16 safety guard
  • compatibility_mode to force pure stock behavior
• Advanced sampler is a 1:1 copy of ComfyUI's KSampler (Advanced), just categorized under ROCm Ninodes/Sampling.
• Native SDPA preference via PyTorch (no CUDA-only flags).
• Workflow update helper script to migrate node names.

🚀 What We Do

ROCm Ninodes transforms your AMD GPU experience in ComfyUI by providing:

• 🎯 ROCm-Optimized Nodes: Custom implementations of VAE decode, KSampler, and LoRA loading specifically tuned for AMD GPUs
• ⚡ Performance Boost: 15-78% faster generation times with better memory efficiency
• 🛡️ Memory Management: Gentle memory cleanup optimized for mature ROCm drivers
• 🔧 Easy Integration: Drop-in replacements for standard ComfyUI nodes
• 📊 Real-Time Monitoring: Built-in performance tracking and optimization recommendations

🔧 Quantized Model Support

ROCm Ninodes now includes comprehensive support for quantized models with automatic detection and optimization:

Supported Quantized Formats

• FP8 Models: Hardware-accelerated FP8 quantization (flux1-dev-fp8.safetensors)
• BFloat16: Native ROCm support with minimal overhead
• INT8/INT4: GGUF format support for WAN 2.2 models
• Automatic Detection: Detects quantized models from filename and dtype

Quantization-Specific Features

• 🛡️ Compatibility Mode: Automatically disables aggressive optimizations for quantized models
• 💾 Smart Memory Management: Quantization-aware memory allocation (FP8: 50% vs FP32, INT8: 25% vs FP32)
• 🔒 Dtype Preservation: Prevents forced dtype conversions that break quantized models
• 📹 Adaptive Video Processing: Smaller chunk sizes and tile sizes for quantized models
• ⚡ OOM Prevention: Lower default settings to prevent out-of-memory errors

User-Reported Issue Fixes

• Fixed OOM Errors: Lower default tile_size (512 vs 768) for better compatibility
• Fixed Quantized Model Breaking: Disabled batch optimization by default for quantized models
• Fixed Memory Management: Less aggressive cleanup for quantized models
• Fixed Video Processing: Adaptive chunk sizing based on frame count and available memory

Recommended Settings for Quantized Models

• Compatibility Mode: Enable for quantized models
• Tile Size: Use 512 (conservative) instead of 768
• Video Chunk Size: Use 2-4 frames instead of 8
• Batch Optimization: Disable for quantized models
• Memory Optimization: Disable aggressive cleanup

Our optimization approach focuses on three key areas:

1. ROCm-Specific Optimizations

• Environment Variables: Essential TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 for mature drivers
• Memory Allocation: Optimized settings (256MB chunks, 0.8 threshold) for better fragmentation control
• Precision Handling: Automatic selection of optimal precision for AMD hardware
• Attention Mechanisms: ROCm-tuned attention optimizations for better performance

2. Gentle Memory Management

• Single-Pass Cleanup: Efficient memory clearing without performance penalties
• Smart Monitoring: Real-time memory usage tracking and optimization
• Fragmentation Control: Proactive memory management to prevent OOM errors
• Mature Driver Support: Optimized for current ROCm drivers and libraries

3. Hardware-Specific Tuning

• gfx1151 Architecture: Specifically optimized for AMD Radeon 8060S and similar GPUs
• Unified Memory: Leverages AMD's unified memory architecture for better performance
• Conservative Batching: Smart batching strategies optimized for AMD GPU characteristics
• Tile Size Optimization: Optimal tile sizes (768-1024) for gfx1151 memory bandwidth

🎯 Real-World Performance Results

Tested on GMTek Evo-X2 Strix Halo (gfx1151) with 128GB Unified RAM:

🖼️ Image Generation (Flux)

• 1024x1024 generation: 500s → 110s (78% improvement!)

🎬 Image-to-Video Generation (WAN 2.2 i2v)

• 320x320px, 2s: 163s → 139s (15% improvement!)
• 320x320px, 17 frames: 98.33s → 92.78s (5.6% improvement!)
• 480x480px, 2s: 202s (33 frames, 16fps) ✅
• 480x720px, 2s: 303s (33 frames, 16fps) ✅
• Video Quality: Fixed darker frames at chunk boundaries (v1.0.29) ✅
• WAN VAE Jitter Fix: Fixed jitter/repeating frames issue for WAN models (v2.0.1) ✅

📊 Performance Metrics

• Memory efficiency: 50% reduction in attention memory requirements
• Stability: Significantly reduced OOM errors
• Scalability: Successfully handles up to 480x720px i2v generation
• Consistency: Stable performance across multiple runs (5.6% average improvement)

"Workflows that used to take forever to run now complete in a fraction of the time!" - Nino, GMTek Evo-X2 Owner

🔬 Detailed Benchmark Results (WAN 2.2 i2v, 320x320px, 17 frames)

Test Configuration:

• Model: WAN 2.2 i2v 14B
• Resolution: 320x320px
• Frames: 17 frames
• Hardware: GMTek Evo-X2 Strix Halo (gfx1151, 128GB Unified RAM)

With ROCm Ninodes optimizations:

• Run 1: ROCM Advanced KSampler: 20.77s | ROCM VAE Decode: 7.73s | Total: 92.78s
• Run 2: ROCM Advanced KSampler: 21.03s | ROCM VAE Decode: 7.41s | Total: 93.32s
• Average: 93.05s ⚡

Without ROCm Ninodes (standard ComfyUI):

• Run 1: Standard KSampler: 22.06s | Standard VAE Decode: 7.48s | Total: 98.33s
• Run 2: Standard KSampler: 22.71s | Standard VAE Decode: 7.20s | Total: 104.01s
• Average: 101.17s 🐌

Performance Improvement: 8.1% faster overall, 5.6% average improvement

🎯 Try It Now!

• Flux Image Generation - 78% performance improvement!
• WAN 2.2 Video Generation - 15% performance improvement!

🚀 Key Features

• ROCM-Specific Optimizations: Tuned specifically for AMD GPUs with ROCm 6.4+
• gfx1151 Architecture Support: Optimized for Strix Halo and similar architectures
• Performance Monitoring: Built-in performance analysis and optimization recommendations
• Memory Management: Advanced VRAM optimization for AMD GPUs
• Precision Optimization: Automatic precision selection for optimal ROCm performance

Features

ROCMOptimizedVAEDecode

• Optimized for gfx1151: Tuned tile sizes and memory management for your specific GPU
• ROCm-specific optimizations: Disables TF32, enables fp16 accumulation, optimizes for AMD GPUs
• Smart precision handling: Automatically selects optimal precision (fp32 for gfx1151)
• Memory management: Conservative batching strategy for AMD GPUs
• Performance monitoring: Built-in timing and logging

ROCMOptimizedVAEDecodeTiled

• Advanced tiling: More control over tile sizes and overlaps
• Temporal support: Optimized for video VAEs
• ROCm optimizations: Same optimizations as the main decode node

ROCMOptimizedKSampler

• Stock behavior with ROCm-safe toggles
• optimize_for_video: reduces host-device sync overhead on multi-frame latents
• precision_mode: auto|fp32|bf16 (guarded; no forced casts)
• compatibility_mode: revert to pure stock quickly

ROCMOptimizedKSamplerAdvanced

• 1:1 with ComfyUI KSampler (Advanced)
• Same ROCm toggles as the basic sampler (video, precision, compatibility)

ROCMSamplerCustomAdvanced

• Drop-in replacement for ComfyUI's V3 SamplerCustomAdvanced (same noise/guider/sampler/sigmas/latent_image interface)
• ROCm backend tuning: architecture detection, fp16 accumulation, TF32 disabled on AMD
• Emergency memory defrag for high-memory models (LTX Video 128ch, memory factor 5.5x)
• Enhanced per-step callback with ETA, timing, and video-workflow optimization
• precision_mode: auto|fp32|bf16
• compatibility_mode: pure stock fallback

ROCMSamplerCustomAdvancedBenchmark

• A/B benchmark: runs stock then ROCm-optimized on identical inputs
• Reports timing, peak memory, speedup %, model type, and GPU name
• Outputs both LATENT (from ROCm run) and a BENCHMARK_REPORT string

ROCMVAEPerformanceMonitor

• Device analysis: Shows your GPU information and current settings
• Performance tips: Provides specific recommendations for your hardware
• Optimal settings: Suggests best parameters for your setup

ROCMSamplerPerformanceMonitor

• Sampler analysis: Analyzes sampling performance and provides recommendations
• Optimal settings: Suggests best samplers and settings for your GPU
• Performance tips: Specific recommendations for sampling optimization

WindowsPaginationDiagnostic

• Error 1455 detection: Automatically detects Windows pagination errors
• Memory analysis: Checks system memory availability and usage
• Automatic fixes: Applies recommended environment variables and settings
• Step-by-step guidance: Provides detailed instructions for manual fixes
• Real-time monitoring: Shows current memory status and recommendations

🧪 Testing

Comprehensive Test Suite

The project includes a comprehensive test suite to ensure reliability and prevent regressions:

Error Prevention Tests

cd /path/to/ComfyUI/custom_nodes/rocm_ninodes
source /path/to/ComfyUI/.venv/bin/activate
python test_vae_error_scenarios.py

Test Coverage:

• ✅ AttributeError: 'dict' object has no attribute 'shape'
• ✅ IndexError: tuple index out of range
• ✅ ValueError: Expected numpy array with ndim 3 but got 4
• ✅ VAE Decode Input Formats: 5D vs 4D tensor handling
• ✅ Chunked Video Processing: Memory-safe chunking logic
• ✅ Tensor Shape Conversions: 5D→4D conversion validation
• ✅ Memory Calculation Edge Cases: Various tensor sizes
• ✅ Error Recovery Scenarios: Malformed input handling
• ✅ Performance Benchmarks: Decode timing tests

Test Results

Ran 9 tests in 0.032s
OK

Debug Data Collection

The nodes automatically collect debug data for optimization analysis:

• Location: test_data/debug/wan_vae_input_debug_{timestamp}.pkl
• Content: Tensor shapes, types, device info, and actual tensor data
• Usage: Run optimization tests and analyze performance

Performance Testing

# Run optimization tests
python test_vae_optimization.py

# Run error scenario tests  
python test_vae_error_scenarios.py

# Debug VAE decode issues
python debug_vae_decode.py

Test Data Structure

test_data/
├── debug/                    # Raw debug data from workflows
├── optimization/             # Optimization test results  
├── benchmarks/               # Performance benchmarks
└── README.md                # Test data documentation

🚀 ComfyUI Installation with uv

Complete Setup Guide

Tested on Manjaro Linux with GMTek Evo-X2 Strix Halo (gfx1151, 128GB Unified RAM)

🐧 Linux (Manjaro/Ubuntu/Arch/etc.)

1. Install uv (if not already installed):

# Install uv
curl -LsSf https://astral.sh/uv/install.sh | sh
source ~/.bashrc  # or ~/.zshrc

2. Clone and setup ComfyUI:

# Clone ComfyUI
git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI

# Create virtual environment with uv
uv venv
source .venv/bin/activate

# Install dependencies
uv pip install -r requirements.txt

# Install ROCm PyTorch nightly for gfx1151
uv pip uninstall torch torchaudio torchvision
uv pip install --index-url https://rocm.nightlies.amd.com/v2/gfx1151/ --pre torch torchaudio torchvision --upgrade

3. Start ComfyUI with optimized flags:

export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
uv run main.py --use-pytorch-cross-attention --highvram --cache-none

🪟 Windows (PowerShell)

1. Install uv (if not already installed):

# Install uv via pip
pip install uv

# Or download from: https://github.com/astral-sh/uv/releases

2. Clone and setup ComfyUI:

# Clone ComfyUI
git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI

# Create virtual environment with uv
uv venv
.venv\Scripts\Activate.ps1

# Install dependencies
uv pip install -r requirements.txt

# Install ROCm PyTorch nightly for gfx1151
uv pip uninstall torch torchaudio torchvision
uv pip install --index-url https://rocm.nightlies.amd.com/v2/gfx1151/ --pre torch torchaudio torchvision --upgrade

3. Start ComfyUI with optimized flags:

# Set environment variable
$env:TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL="1"

# Start ComfyUI
uv run main.py --use-pytorch-cross-attention --highvram --cache-none

Note for Windows users: ROCm support on Windows is limited. For best performance, consider using WSL2 with Ubuntu or dual-booting Linux.

📦 Plugin Installation

Method 1: ComfyUI CLI (Recommended)

Install ComfyUI CLI first:

pip install comfy-cli

Then install the plugin:

comfy node install rocm-ninodes

Method 2: Manual Installation

Prerequisites

For gfx1151 (Strix Halo) users, follow these setup steps:

🐧 Linux (Manjaro/Ubuntu/etc.)

1. Install ROCm PyTorch nightly build:

# Uninstall regular/CUDA PyTorch first
uv pip uninstall torch torchaudio torchvision

# Install ROCm nightly for gfx1151
uv pip install --index-url https://rocm.nightlies.amd.com/v2/gfx1151/ --pre torch torchaudio torchvision --upgrade

2. Start ComfyUI with optimized flags:

export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
uv run main.py --use-pytorch-cross-attention --highvram --cache-none

🪟 Windows (PowerShell)

1. Install ROCm PyTorch nightly build:

# Uninstall regular/CUDA PyTorch first
pip uninstall torch torchaudio torchvision

# Install ROCm nightly for gfx1151
pip install --index-url https://rocm.nightlies.amd.com/v2/gfx1151/ --pre torch torchaudio torchvision --upgrade

2. Start ComfyUI with optimized flags:

# Set environment variable
$env:TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL="1"

# Start ComfyUI
python main.py --use-pytorch-cross-attention --highvram --cache-none

Note for Windows users: ROCm support on Windows is limited. For best performance, consider using WSL2 with Ubuntu or dual-booting Linux.

Method 3: Git Clone

🐧 Linux/Mac:

cd ComfyUI/custom_nodes
git clone https://github.com/iGavroche/rocm-ninodes.git ComfyUI-ROCM-Optimized-VAE
cd ComfyUI-ROCM-Optimized-VAE
python install.py

🪟 Windows (PowerShell):

cd ComfyUI\custom_nodes
git clone https://github.com/iGavroche/rocm-ninodes.git ComfyUI-ROCM-Optimized-VAE
cd ComfyUI-ROCM-Optimized-VAE
python install.py

Method 4: Download ZIP

1. Download the latest release from GitHub
2. Extract to ComfyUI/custom_nodes/ComfyUI-ROCM-Optimized-VAE/
3. Run python install.py to verify installation

Windows users: Right-click the ZIP file → "Extract All" → Choose the ComfyUI/custom_nodes/ folder

Method 5: ComfyUI Manager (Future)

Coming soon - will be available through ComfyUI Manager

Post-Installation

1. Restart ComfyUI to load the new nodes
2. Verify Installation: Check that nodes appear in "ROCm Ninodes" in the node panel:
   • ROCm Ninodes/VAE: VAE Decode, VAE Decode Tiled, VAE Performance Monitor
   • ROCm Ninodes/Sampling: KSampler, KSampler Advanced, Sampler Performance Monitor
   • ROCm Ninodes/Generative AI: ROCm Text Generate LTX2 Prompt
   • ROCm Ninodes/Loaders: Checkpoint, Diffusion, GGUF, LoRA
3. Test Performance: Use the Performance Monitor nodes to verify optimizations

🔄 Workflow name migration (helper script)

We include a small helper to migrate workflow JSONs to the new node names.

Usage (Windows PowerShell):

uv run python scripts/update_workflows.py --paths "C:\path\to\ComfyUI\user\default\workflows" "comfyui_workflows"

What it does:

• Scans .json files and replaces legacy node names with the current ROCm naming
• Writes a .bak next to each changed file
• Shows a summary of changes

Node mappings applied:

• ROCMOptimizedUNetLoader → ROCmDiffusionLoader
• Ensures display names use "ROCm" capitalization in categories and titles

🔄 Plugin Updates

How to Update RocM Ninodes

🐧 Linux (Manjaro/Ubuntu/etc.)

Method 1: Git Pull (Recommended)

cd ComfyUI/custom_nodes/ComfyUI-ROCM-Optimized-VAE
git pull origin main

Method 2: Fresh Install

# Remove old version
rm -rf ComfyUI/custom_nodes/ComfyUI-ROCM-Optimized-VAE

# Install latest version
cd ComfyUI/custom_nodes
git clone https://github.com/iGavroche/rocm-ninodes.git ComfyUI-ROCM-Optimized-VAE

🪟 Windows (PowerShell)

Method 1: Git Pull (For Existing Installations)

# Navigate to the plugin directory
cd ComfyUI\custom_nodes\ComfyUI-ROCM-Optimized-VAE

# Pull latest changes
git pull origin main

Method 2: Fresh Install (For New Installations)

# Navigate to custom_nodes directory
cd ComfyUI\custom_nodes

# Clone the repository
git clone https://github.com/iGavroche/rocm-ninodes.git ComfyUI-ROCM-Optimized-VAE

# Navigate into the plugin directory
cd ComfyUI-ROCM-Optimized-VAE

Method 3: Update Existing Installation (If git pull fails)

# Navigate to custom_nodes directory
cd ComfyUI\custom_nodes

# Remove old version
Remove-Item -Recurse -Force ComfyUI-ROCM-Optimized-VAE

# Clone fresh copy
git clone https://github.com/iGavroche/rocm-ninodes.git ComfyUI-ROCM-Optimized-VAE

# Navigate into the plugin directory
cd ComfyUI-ROCM-Optimized-VAE

After Updating

1. Restart ComfyUI to load the updated nodes
2. Check for new features in the node panel
3. Test workflows to ensure compatibility
4. Check the CHANGELOG for new features and fixes

Update Notifications

• GitHub Releases: Watch the repository for release notifications
• ComfyUI Manager: Future updates will be available through ComfyUI Manager
• Performance Updates: New optimizations are regularly added based on community feedback

🚀 Quick Start - Optimized Workflow

Ready to test the optimizations? Download the pre-configured workflow:

📥 Download Optimized Workflows

• Flux Image Generation - Complete Flux workflow with ROCM optimizations
• WAN 2.2 Video Generation - WAN 2.2 Image-to-Video workflow with ROCM optimizations

This workflow includes:

• ✅ ROCM VAE Decode (optimized for gfx1151)
• ✅ ROCM KSampler (with memory optimizations)
• ✅ Performance Monitors (to track improvements)
• ✅ Optimal Settings (tuned for Strix Halo)

How to use:

1. Download the workflow JSON file
2. Open in ComfyUI (drag & drop or File → Load)
3. Install missing nodes via ComfyUI Manager (if prompted)
4. Run and enjoy 78% faster generation! 🎉

📖 Node Usage Guide

Core Nodes

🖼️ ROCm VAE Decode

Location: ROCm Ninodes/VAE → ROCm VAE Decode

Purpose: Optimized VAE decoding for AMD GPUs with ROCm support. Detects VAE type automatically and applies optimal settings for standard, WAN, LTX Video, and pixel-space (z-image) VAEs.

How to Use:

1. Connect your LATENT output from a sampler to the samples input
2. Connect your VAE model to the vae input
3. Use default settings for most cases

Key Settings:

• tile_size: 768-2048 depending on architecture (default: 768)
• overlap: 96-128 for good quality (default: 96)
• precision_mode: "auto" selects optimal precision (fp16 default on AMD)
• Video (WAN, LTX): Decoded in one pass with causal preservation. fp16 used for memory efficiency.

LTX Video-specific:

• Latent channels: 128 (automatically detected)
• Spatial compression: 32× (tile sizes auto-scaled with 64px latent minimum)
• Enabled full-video causal decode; chunked IO disabled to avoid pre-allocation OOM
• Temporal tiling (optional, disabled by default): For very long videos (>200 frames). Decodes in overlapping temporal chunks with linear blending — no frame loss, no visible seams. Progress bar shown during decode.

Recommended temporal tiling settings for long LTX videos:

enable_temporal_tiling=True
temporal_chunk_size=16    # Latent frames per chunk (~121 output frames each)
temporal_overlap=2        # Latent frames overlap (blended for smooth transition)
last_frame_fix=True       # Prevents end-of-video artifacts

How temporal tiling works:

• Video is split into overlapping chunks along the temporal axis
• Each chunk is decoded independently (memory per chunk ~chunk_size, not full video)
• The first output frame of each chunk is dropped (incomplete temporal context)
• The next temporal_overlap × temporal_comp frames are linearly blended with the previous chunk's tail for a seamless transition
• Result: exact same frame count as a full decode — zero frame loss

Pixel-space (z-image / z-image-turbo):

• Detected automatically (latent_channels=3, compression=1)
• Passthrough decode — no VAE model loaded or dtype converted

Output: IMAGE - Decoded image tensor ready for saving or further processing

---

🎨 ROCm VAE Decode Tiled

Location: ROCm Ninodes/VAE → ROCm VAE Decode Tiled

Purpose: Advanced tiled VAE decode for very large images or video with temporal support.

How to Use:

1. Connect LATENT to samples input
2. Connect VAE model to vae input
3. Use for images >1024x1024 or when you need temporal control for video
4. Adjust temporal_size and temporal_overlap for video workflows

Key Settings:

• tile_size: 768-1024 (default: 768)
• temporal_size: 64 frames for video (default: 64)
• temporal_overlap: 8 frames overlap for video (default: 8)

Output: IMAGE - Decoded image/video tensor

---

🎲 ROCm KSampler

Location: ROCm Ninodes/Sampling → ROCm KSampler

Purpose: Optimized KSampler with ROCm-specific optimizations and progress reporting. Auto-detects GPU architecture and model type — no manual configuration needed.

How to Use:

1. Connect your MODEL to model input
2. Connect CONDITIONING (positive and negative) from your CLIP text encode nodes
3. Connect LATENT image (usually from Empty Latent Image node)
4. Set your desired steps, cfg, sampler_name, and scheduler
5. Connect the output LATENT to your VAE Decode node

Key Settings:

• steps: 20-30 for most cases
• cfg: 7.0-8.0 for gfx1151 (default: 8.0)
• sampler_name: Euler, Heun, or dpmpp_2m work well with ROCm
• optimize_for_video: Enable for multi-frame latents (disables previews)
• precision_mode: "auto" selects optimal precision (fp16 on AMD, fp32 on CPU)

Auto-detection (no toggles needed):

• Architecture: Detects gfx1151/gfx1100 (fp16), gfx942 (bf16) — applies backend settings automatically
• Flow-matching (z-image, ernie, ideogram4): Logged and handled optimally
• Pixel-space (z-image-turbo): Recognized — no VAE decode needed downstream
• High-memory (ideogram4): Automatic emergency_memory_cleanup before sampling

Output: LATENT - Sampled latent tensor (connect to VAE Decode)

Progress: Shows real-time progress in both UI and terminal with step-by-step information

---

🎯 ROCm KSampler Advanced

Location: ROCm Ninodes/Sampling → ROCm KSampler Advanced

Purpose: Advanced KSampler with step control and more options.

How to Use:

1. Same as basic KSampler, but with additional controls:
   • start_at_step: Start sampling from a specific step (default: 0)
   • end_at_step: End sampling at a specific step (default: 10000)
   • add_noise: Enable/disable noise addition
   • return_with_leftover_noise: Enable to return with leftover noise

Use Cases:

• Img2Img workflows: Use start_at_step to control denoising strength
• Inpainting: Use step ranges for precise control
• Video workflows: Enable optimize_for_video for better performance

Output: LATENT - Sampled latent tensor

---

📦 ROCm Checkpoint Loader

Location: ROCm Ninodes/Loaders → ROCm Checkpoint Loader

Purpose: Optimized checkpoint loading with memory management.

How to Use:

1. Select your checkpoint file from the dropdown
2. Connect outputs to your workflow:
   • MODEL → Connect to KSampler
   • CLIP → Connect to CLIP Text Encode nodes
   • VAE → Connect to VAE Decode node

Key Settings:

• use_rocm_optimizations: Enable for best performance (default: True)
• memory_optimization: Enable for better VRAM usage (default: True)

Outputs: MODEL, CLIP, VAE

---

🔄 ROCm Diffusion Loader

Location: ROCm Ninodes/Loaders → ROCm Diffusion Loader

Purpose: Load UNet/Diffusion models separately (useful for Flux and other models).

How to Use:

1. Select your UNet model file
2. Connect MODEL output to your KSampler
3. Use with separate VAE and CLIP loaders

Output: MODEL

---

🎨 ROCm LoRA Loader

Location: ROCm Ninodes/Loaders → ROCm LoRA Loader

Purpose: Load and apply LoRA weights to models with memory optimization.

How to Use:

1. Connect your MODEL to model input
2. Connect your CLIP to clip input (optional)
3. Select LoRA file from dropdown
4. Set strength_model and strength_clip (0.0-2.0, default: 1.0)
5. Connect outputs back to your workflow

Key Settings:

• strength_model: LoRA strength for model (default: 1.0)
• strength_clip: LoRA strength for CLIP (default: 1.0)

Outputs: MODEL, CLIP (if provided)

---

📊 Performance Monitoring Nodes

🔍 ROCm VAE Performance Monitor

Location: ROCm Ninodes/VAE → ROCm VAE Performance Monitor

Purpose: Analyze VAE performance and get optimization recommendations.

How to Use:

1. Connect your VAE model to the vae input
2. Set test_resolution (default: 1024) - this is just for recommendations, not actual testing
3. Connect outputs to Show Text nodes to display results:
   • DEVICE_INFO → Shows GPU information and VAE settings
   • PERFORMANCE_TIPS → Provides specific optimization tips
   • OPTIMAL_SETTINGS → Shows recommended settings for your GPU

Example Workflow Connection:

VAE → ROCm VAE Performance Monitor
     ├─ DEVICE_INFO → Show Text (to display GPU info)
     ├─ PERFORMANCE_TIPS → Show Text (to display tips)
     └─ OPTIMAL_SETTINGS → Show Text (to display settings)

Outputs:

• DEVICE_INFO: GPU name, VAE dtype, output device
• PERFORMANCE_TIPS: Specific recommendations for your hardware
• OPTIMAL_SETTINGS: Recommended tile_size, overlap, precision, etc.

Note: This node doesn't run actual benchmarks - it analyzes your VAE and provides recommendations based on your GPU.

---

📈 ROCm Sampler Performance Monitor

Location: ROCm Ninodes/Sampling → ROCm Sampler Performance Monitor

Purpose: Analyze sampler performance and get optimization recommendations. Auto-detects model type (flow-matching, pixel-space, standard) and GPU architecture.

How to Use:

1. Connect your MODEL to the model input
2. Set test_steps (default: 20) — just for recommendations, not actual testing
3. Connect outputs to Show Text nodes to display results:
   • DEVICE_INFO → Shows GPU, model architecture, sampling type, latent channels, memory factor
   • PERFORMANCE_TIPS → Model-type-specific tips (flow matching, large latent format, etc.)
   • OPTIMAL_SETTINGS → Architecture-aware recommendations

Example Workflow Connection:

MODEL → ROCm Sampler Performance Monitor
       ├─ DEVICE_INFO → Show Text
       ├─ PERFORMANCE_TIPS → Show Text
       └─ OPTIMAL_SETTINGS → Show Text

Outputs:

• DEVICE_INFO: Architecture, APU mode, model device/dtype, sampling type, latent channels, memory factor
• PERFORMANCE_TIPS: Model-type-specific tips (e.g. "Flow-matching model — use euler/dpmpp_sde, sgm_uniform scheduler")
• OPTIMAL_SETTINGS: Architecture-aware recommended samplers, schedulers, precision

Note: This node provides recommendations based on your GPU and model, not actual performance measurements.

---

⚡ ROCm Flux Benchmark

Location: ROCm Ninodes/Benchmark → ROCm Flux Benchmark

Purpose: Run comprehensive benchmark tests on your Flux workflow.

How to Use:

1. Connect your MODEL, VAE, and CLIP to the inputs
2. Configure test parameters:
   • test_resolutions: Comma-separated resolutions like "256x320,512x512,1024x1024"
   • test_steps: Number of sampling steps (default: 20)
   • test_cfg_values: Comma-separated CFG values like "1.0,3.5,8.0"
3. Connect all 4 outputs to Show Text nodes to display results:
   • BENCHMARK_RESULTS → Summary of benchmark results
   • PERFORMANCE_CHART → Detailed timing for each resolution
   • OPTIMIZATION_RECOMMENDATIONS → Specific recommendations
   • MEMORY_ANALYSIS → Memory usage information

Example Workflow Connection:

MODEL ─┐
VAE   ├─→ ROCm Flux Benchmark
CLIP  ─┘
       ├─ BENCHMARK_RESULTS → Show Text
       ├─ PERFORMANCE_CHART → Show Text
       ├─ OPTIMIZATION_RECOMMENDATIONS → Show Text
       └─ MEMORY_ANALYSIS → Show Text

Outputs:

• BENCHMARK_RESULTS: Average decode times and peak memory for each resolution
• PERFORMANCE_CHART: Detailed timing breakdown
• OPTIMIZATION_RECOMMENDATIONS: Specific tips for your GPU
• MEMORY_ANALYSIS: Total VRAM, current usage, device information

Note: This node actually runs benchmarks - it will decode test latents and measure performance. This may take several minutes.

---

💾 ROCm Memory Optimizer

Location: ROCm Ninodes/Memory → ROCm Memory Optimizer

Purpose: Monitor and optimize GPU memory usage.

How to Use:

1. Place this node anywhere in your workflow (it doesn't need inputs)
2. Configure optimization settings:
   • optimization_level: conservative, balanced, or aggressive
   • enable_gc: Enable Python garbage collection (default: True)
   • clear_cache: Clear CUDA cache (default: True)
   • cleanup_frequency: How often to run cleanup (default: 10 operations)
3. Connect outputs to Show Text nodes:
   • MEMORY_STATUS → Current memory usage
   • OPTIMIZATION_LOG → What optimizations were performed
   • RECOMMENDATIONS → Memory optimization suggestions

Example Workflow Connection:

ROCm Memory Optimizer (no inputs needed)
├─ MEMORY_STATUS → Show Text
├─ OPTIMIZATION_LOG → Show Text
└─ RECOMMENDATIONS → Show Text

Outputs:

• MEMORY_STATUS: Allocated, reserved, free, and total VRAM
• OPTIMIZATION_LOG: What cleanup operations were performed
• RECOMMENDATIONS: Suggestions based on current memory usage

Use Cases:

• Place after large operations to monitor memory
• Use in loops to prevent memory buildup
• Check memory before starting large workflows

---

🔗 Connecting Performance Nodes to Outputs

All performance monitoring nodes output STRING values that need to be connected to Show Text nodes to be visible in ComfyUI.

Step-by-Step: Adding Performance Monitoring to Your Workflow

1. Add the performance node (e.g., ROCm VAE Performance Monitor)
2. Connect the required inputs (e.g., VAE model)
3. Add Show Text nodes for each output:
   • Right-click → Add Node → Show Text (or search for "Show Text")
   • You'll need one Show Text node per output
4. Connect the outputs:
   • Drag from the performance node's output to the Show Text node's text input
5. Run your workflow - the text will appear in the UI showing the results

Example: Complete VAE Performance Monitor Setup

[VAE Model] 
    ↓
[ROCm VAE Performance Monitor]
    ├─ DEVICE_INFO ──→ [Show Text] ──→ (displays in UI)
    ├─ PERFORMANCE_TIPS ──→ [Show Text] ──→ (displays in UI)
    └─ OPTIMAL_SETTINGS ──→ [Show Text] ──→ (displays in UI)

Example: Complete Flux Benchmark Setup

[MODEL] ─┐
[VAE]   ├─→ [ROCm Flux Benchmark]
[CLIP]  ─┘
         ├─ BENCHMARK_RESULTS ──→ [Show Text]
         ├─ PERFORMANCE_CHART ──→ [Show Text]
         ├─ OPTIMIZATION_RECOMMENDATIONS ──→ [Show Text]
         └─ MEMORY_ANALYSIS ──→ [Show Text]

Tip: You can connect multiple Show Text nodes to the same output if you want to display the same information in multiple places, or use different Show Text nodes for each output to organize your workflow better.

---

Basic Usage

1. Replace your standard VAE Decode node with "ROCm VAE Decode"
2. Replace your standard KSampler with "ROCm KSampler"
3. Use the default settings (optimized for gfx1151)
4. Enable "use_rocm_optimizations" for best performance

Advanced Usage

• VAE Settings:

  • Tile Size: 768-1024 works well for gfx1151 (default: 768)
  • Overlap: 96-128 provides good quality (default: 96)
  • Precision: "auto" selects optimal for your GPU
  • Batch Optimization: Keep enabled for better memory usage

• Sampler Settings:

  • Precision: "auto" selects fp32 for gfx1151
  • Memory Optimization: Keep enabled for better VRAM usage
  • Attention Optimization: Keep enabled for faster sampling
  • Samplers: Euler, Heun, dpmpp_2m work well with ROCm
  • CFG: 7.0-8.0 is optimal for gfx1151

Performance Tips for gfx1151

• Use fp32 precision (automatically selected)
• Tile size 768-1024 for 1024x1024 images
• Enable all ROCm optimizations
• Use tiled decode for images larger than 1024x1024

Expected Performance Improvements

Based on gfx1151 architecture optimizations:

• VAE Decode: 15-25% faster, 20-30% better VRAM usage
• Sampling: 10-20% faster sampling with better memory management
• Overall Workflow: 20-40% faster end-to-end generation
• Memory efficiency: 25-35% better VRAM usage overall
• Stability: Reduced OOM errors with better memory management
• Quality: Maintained or improved output quality

Troubleshooting

🚨 Quick Fix for Common Windows Errors

If you see these errors:

• fatal: couldn't find remote ref ComfyUI-ROCM-Optimized-VAE
• does not appear to be a git repository
• Le module « .venv » n'a pas pu être chargé

Quick Solution:

# 1. Navigate to ComfyUI directory
cd C:\ComfyUI

# 2. Activate virtual environment
.venv\Scripts\Activate.ps1

# 3. Navigate to custom_nodes
cd custom_nodes

# 4. Clone the plugin (if not already installed)
git clone https://github.com/iGavroche/rocm-ninodes.git ComfyUI-ROCM-Optimized-VAE

# 5. Navigate into the plugin directory
cd ComfyUI-ROCM-Optimized-VAE

# 6. Run the installer
python install.py

🪟 Windows Pagination Error Fixes (Error 1455)

If you encounter the error "Le fichier de pagination est insuffisant pour terminer cette opération" (os error 1455):

🚨 Quick Fix (Recommended)

Use the new Windows Pagination Diagnostic node in ComfyUI:

1. Add "Windows Pagination Diagnostic" node from "ROCm Ninodes/Diagnostics"
2. Connect it to your workflow
3. Run it to automatically diagnose and fix the issue

Method 1: Environment Variable (Immediate Fix)

# Set environment variable before starting ComfyUI
$env:PYTORCH_CUDA_ALLOC_CONF = "expandable_segments:True,max_split_size_mb:512"
$env:PYTORCH_HIP_ALLOC_CONF = "expandable_segments:True"
python main.py

Method 2: Batch File Solution

Create a start_comfyui.bat file in your ComfyUI directory:

@echo off
set PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,max_split_size_mb:512
set PYTORCH_HIP_ALLOC_CONF=expandable_segments:True
python main.py
pause

Method 3: PowerShell Profile (Permanent)

Add to your PowerShell profile:

# Open PowerShell profile
notepad $PROFILE

# Add these lines:
$env:PYTORCH_CUDA_ALLOC_CONF = "expandable_segments:True,max_split_size_mb:512"
$env:PYTORCH_HIP_ALLOC_CONF = "expandable_segments:True"

Method 4: System Environment Variable (Permanent)

1. Press Win + R, type sysdm.cpl, press Enter
2. Click "Environment Variables"
3. Under "User variables", click "New"
4. Variable name: PYTORCH_CUDA_ALLOC_CONF
5. Variable value: expandable_segments:True,max_split_size_mb:512
6. Click "New" again
7. Variable name: PYTORCH_HIP_ALLOC_CONF
8. Variable value: expandable_segments:True
9. Click OK and restart ComfyUI

Method 5: Increase Windows Paging File (Most Effective)

1. Press Win + R, type sysdm.cpl, press Enter
2. Click "Advanced" tab → "Performance Settings" → "Advanced" tab
3. Under "Virtual memory", click "Change"
4. Uncheck "Automatically manage paging file size for all drives"
5. Select your system drive (usually C:)
6. Select "Custom size"
7. Set Initial size: 16384 MB (16 GB)
8. Set Maximum size: 32768 MB (32 GB)
9. Click "Set", then "OK", then restart ComfyUI

Method 6: PowerShell Script (Advanced)

Create fix_pagination.ps1:

# Fix Windows pagination error 1455
Write-Host "Applying Windows pagination fixes..." -ForegroundColor Green

# Set environment variables
$env:PYTORCH_CUDA_ALLOC_CONF = "expandable_segments:True,max_split_size_mb:512"
$env:PYTORCH_HIP_ALLOC_CONF = "expandable_segments:True"
$env:PYTORCH_CUDA_MEMORY_POOL_TYPE = "expandable_segments"

# Check memory
$memory = Get-WmiObject -Class Win32_PhysicalMemory | Measure-Object -Property Capacity -Sum
$totalGB = [math]::Round($memory.Sum / 1GB, 2)
Write-Host "Total RAM: $totalGB GB" -ForegroundColor Yellow

if ($totalGB -lt 16) {
    Write-Host "WARNING: Less than 16GB RAM detected. Consider increasing paging file." -ForegroundColor Red
}

Write-Host "Environment variables set. Starting ComfyUI..." -ForegroundColor Green
python main.py

Run with: powershell -ExecutionPolicy Bypass -File fix_pagination.ps1

🔧 Advanced Windows Troubleshooting

Memory Issues on Windows:

# Check available memory
Get-WmiObject -Class Win32_PhysicalMemory | Measure-Object -Property Capacity -Sum

# Set additional memory management
$env:PYTORCH_CUDA_ALLOC_CONF = "expandable_segments:True,max_split_size_mb:512"

ROCm Installation Issues:

# Verify ROCm installation
rocm-smi

# Check PyTorch ROCm support
python -c "import torch; print(torch.cuda.is_available()); print(torch.version.hip)"

Git Issues on Windows:

# Fix line ending issues
git config --global core.autocrlf true

# Reset repository if corrupted
cd custom_nodes
rmdir /s ComfyUI-ROCM-Optimized-VAE
git clone https://github.com/iGavroche/rocm-ninodes.git ComfyUI-ROCM-Optimized-VAE

If you experience issues:

1. Check the Performance Monitor node for recommendations
2. Try reducing tile size if you get OOM errors
3. Ensure you're using ROCm-compatible PyTorch
4. Check that "use_rocm_optimizations" is enabled

ROCm Requirements:

• PyTorch with ROCm support (nightly build recommended)
• ROCm 6.4+ (you're using 6.4)

uv-Specific Issues

🐧 Linux (Manjaro/Ubuntu/etc.)

1. uv not found after installation:

   # Add to your shell profile (~/.bashrc or ~/.zshrc)
   echo 'export PATH="$HOME/.cargo/bin:$PATH"' >> ~/.bashrc
   source ~/.bashrc
   

2. Virtual environment not activating:

   # Make sure you're in the ComfyUI directory
   cd ComfyUI
   source .venv/bin/activate
   

3. PyTorch ROCm installation fails:

   # Clear uv cache and retry
   uv cache clean
   uv pip install --index-url https://rocm.nightlies.amd.com/v2/gfx1151/ --pre torch torchaudio torchvision --upgrade
   

4. Permission issues with uv:

   # Install uv for current user only
   curl -LsSf https://astral.sh/uv/install.sh | sh
   

🪟 Windows (PowerShell)

1. uv not found after installation:

   # Add uv to PATH or use full path
   $env:PATH += ";C:\Users\$env:USERNAME\.cargo\bin"
   # Or restart PowerShell after installation
   

2. Virtual environment not activating:

   # Make sure you're in the ComfyUI directory
   cd ComfyUI
   .venv\Scripts\Activate.ps1
   

3. PowerShell execution policy:

   # If you get execution policy errors
   Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser
   

4. Git not found:

   # Install Git for Windows from: https://git-scm.com/download/win
   # Or use GitHub Desktop
   

5. "fatal: couldn't find remote ref" error:

   # This happens when trying to git pull before cloning
   # Solution: Clone the repository first
   cd ComfyUI\custom_nodes
   git clone https://github.com/iGavroche/rocm-ninodes.git ComfyUI-ROCM-Optimized-VAE
   

6. "does not appear to be a git repository" error:

   # This happens when the directory isn't a git repository
   # Solution: Clone the repository first
   cd ComfyUI\custom_nodes
   git clone https://github.com/iGavroche/rocm-ninodes.git ComfyUI-ROCM-Optimized-VAE
   

7. Virtual environment activation fails:

   # Make sure you're in the ComfyUI directory (not custom_nodes)
   cd C:\ComfyUI
   
   # Try activating the virtual environment
   .venv\Scripts\Activate.ps1
   
   # If that fails, try this alternative
   & ".venv\Scripts\Activate.ps1"
   

Windows-Specific Issues

1. ROCm not working on Windows:

   • ROCm has limited Windows support
   • Consider using WSL2 with Ubuntu for better compatibility
   • Or dual-boot Linux for optimal performance

2. PowerShell execution policy:

   # If you get execution policy errors
   Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser
   

3. Python path issues:

   # Make sure Python is in your PATH
   python --version
   # If not found, add Python to PATH or use full path
   

4. Git not found:

   • Install Git for Windows from https://git-scm.com/download/win
   • Or use GitHub Desktop for GUI-based cloning

Technical Details

Optimizations Applied:

1. Memory Management: Conservative batching for AMD GPUs
2. Precision: fp32 preferred over bf16 for gfx1151
3. Tile Sizing: Optimized for gfx1151 memory bandwidth
4. ROCm Settings: Disabled TF32, enabled fp16 accumulation
5. Batch Processing: Improved batch size calculation

Architecture-Specific Tuning:

• gfx1151: Optimized tile sizes (768-1024)
• Memory: Conservative memory allocation
• Precision: fp32 for best ROCm performance
• Batching: AMD-optimized batch sizes

Testing

ROCM Ninodes includes a comprehensive test suite to ensure performance and correctness.

Quick Start

cd tests
./run_tests.sh

Test Categories

• Performance Tests: Validate timing targets (78% Flux improvement, 5.6% WAN improvement)
• Correctness Tests: Verify tensor shapes and data formats
• Integration Tests: Full ComfyUI workflow testing
• Mock Data Tests: Tests using synthetic data when real data unavailable

Data Capture

Enable debug mode to capture real workflow data for testing:

export ROCM_NINODES_DEBUG=1
# Run your ComfyUI workflows
# Data will be saved to test_data/captured/

Performance Targets

• Flux Checkpoint Load: <30s
• Flux VAE Decode: <10s
• WAN Sampling: <100s
• WAN VAE Decode: <10s

Documentation

• Testing Guide: Comprehensive testing documentation
• Architecture: System architecture and constraints
• Rules: Development rules and best practices

Contributing

Feel free to submit issues or pull requests to improve the optimizations for your specific use case.