iGavroche-rocm-ninodes/docs/SAMPLER_CUSTOM_ADVANCED.md

ROCM SamplerCustomAdvanced

Overview

A drop-in replacement for ComfyUI's built-in SamplerCustomAdvanced (V3 API, model/sampling/custom_sampling) with ROCm-aware optimizations.

Menu location: ROCm Ninodes/Sampling > ROCM SamplerCustomAdvanced

Why This Exists

The stock SamplerCustomAdvanced in comfy_extras/nodes_custom_sampler.py accepts five pluggable sub-components:

noise (RandomNoise)      ─┐
guider (CFGGuider)       ─┤
sampler (KSamplerSelect) ─┤── SamplerCustomAdvanced → (output, denoised_output)
sigmas (BasicScheduler)  ─┤
latent_image (...)       ─┘

This is the standard pattern for LTX Video, LTX 2.3 AV, and many other advanced ComfyUI workflows. The ROCm version preserves the exact same interface while adding:

Optimization	What It Does	Impact
Architecture detection	Detects AMD GPU family (gfx110x, gfx1151, gfx90a, etc.)	Tunes backend settings per GPU
ROCm backend tuning	Disables TF32, enables fp16 accumulation on AMD	Small throughput gain
Emergency memory defrag	Frees fragmented VRAM before sampling	Prevents OOM on large models
Precision management	Auto-selects fp16 vs bf16 based on GPU capability	Matches hardware capability
Enhanced callback	Per-step timing, ETA, preview generation	Real-time visibility
Video workflow mode	Disables previews for multi-frame latents	Faster long-video generation
Post-sample cleanup	Gentle memory defrag after generation	Reduces fragmentation

Inputs

Input	Type	Required	Notes
noise	NOISE	Yes	e.g. RandomNoise
guider	GUIDER	Yes	e.g. CFGGuider, BasicGuider
sampler	SAMPLER	Yes	e.g. KSamplerSelect
sigmas	SIGMAS	Yes	e.g. BasicScheduler
latent_image	LATENT	Yes	e.g. LTXVConcatAVLatent
optimize_for_video	BOOLEAN	No (default: False)	Disables previews when latent is 5D with T > 1
precision_mode	COMBO	No (default: auto)	auto, fp32, bf16 — ROCm precision hint
compatibility_mode	BOOLEAN	No (default: False)	When True, skips all ROCm optimizations (pure stock behavior)

Outputs

Output	Type	Description
output	LATENT	The sampled latent
denoised_output	LATENT	The denoised (x0) prediction, if available

Usage

Replace SamplerCustomAdvanced in an Existing Workflow

1. Delete the stock SamplerCustomAdvanced node
2. Add ROCM SamplerCustomAdvanced from ROCm Ninodes/Sampling
3. Wire the same connections (noise, guider, sampler, sigmas, latent_image)
4. Optionally configure optimize_for_video, precision_mode, or compatibility_mode
5. Run — the console will show per-step progress, model detection, and completion

LTX 2.3 / LTX Video Workflows

For LTX-based workflows (like LTX2.3-Director-App-Mariah.json), the latent_image input typically comes from LTXVConcatAVLatent. ROCMSamplerCustomAdvanced detects the flow-matching model type (128ch latent, memory factor 5.5x) and applies appropriate memory management before and after sampling.

Benchmark Node

ROCMSamplerCustomAdvancedBenchmark (ROCm Ninodes/Sampling > ROCm SamplerCustomAdvanced Benchmark) runs both the stock and ROCm-optimized sampler in sequence on identical inputs and outputs a comparison string.

Inputs: Same as SamplerCustomAdvanced.

Outputs:

• LATENT — result from the ROCm-optimized run
• BENCHMARK_REPORT — multiline string with timing and memory comparison

Example output:

Benchmark Results
──────────────────────────────────────────────────
Stock:  42.35s, 5840MB peak
ROCm:   41.12s, 5720MB peak
Speed:  +3.0%
Memory: +120MB delta
Steps:  40
Model:  flow (128ch, mem factor 5.5x)
GPU:    AMD Radeon Graphics (gfx1151)
──────────────────────────────────────────────────

Implementation Details

See rocm_nodes/core/sampler.py:

ROCMSamplerCustomAdvanced      (V3 io.ComfyNode, define_schema)
ROCMSamplerCustomAdvancedBenchmark  (V1 legacy API, INPUT_TYPES)

The node uses the ComfyUI V3 (ComfyNode) API via comfy_api.latest.io, which ComfyUI's execution system detects automatically from the class inheritance.