[PR #2562] [MERGED] Add HALO Debug Pack - AMD ROCm bf16 diagnostic tools #2598

New issue

Closed

opened 2026-10-28 21:11:04 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-28 21:11:04 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/ComfyUI-Manager/pull/2562
Author: @bkpaine1
Created: 1/30/2026
Status: ✅ Merged
Merged: 2/6/2026
Merged by: @ltdrdata

Base: main ← Head: add-halo-pack

📝 Commits (1)

3bb5f51 Add HALO Debug Pack - AMD ROCm bf16 diagnostic tools

📊 Changes

1 file changed (+11 additions, -0 deletions)

View changed files

📝 custom-node-list.json (+11 -0)

📄 Description

Summary

Adds HALO Debug Pack to the custom node list - diagnostic tools for debugging black image issues caused by numpy's lack of bfloat16 support.

Target Users

AMD ROCm users (especially Strix Halo APUs)
Older NVIDIA cards with bf16 issues
Apple Silicon / unified memory systems
Anyone experiencing black images from bf16 precision problems

Nodes Included

Node	Purpose
HALO VAE Decode (FP32)	Forces VAE decode to fp32, fixing numpy bf16 conversion
HALO VAE Encode (FP32)	Forces VAE encode to fp32 for img2img
HALO Latent Debug	Checks sampler output for NaN/dead latents
HALO Conditioning Debug	Checks text encoder output for NaN
HALO Model Debug	Inspects model dtype (bf16/fp16/fp32)

Repository

https://github.com/bkpaine1/halo_pack

Why This Matters

When PyTorch bf16 tensors get converted to numpy for image saving, numpy silently produces garbage because it doesn't support bfloat16. This causes the infamous "black image" problem. These nodes help users diagnose exactly where the issue occurs in their pipeline.

Thank you for maintaining ComfyUI-Manager!

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/Comfy-Org/ComfyUI-Manager/pull/2562 **Author:** [@bkpaine1](https://github.com/bkpaine1) **Created:** 1/30/2026 **Status:** ✅ Merged **Merged:** 2/6/2026 **Merged by:** [@ltdrdata](https://github.com/ltdrdata) **Base:** `main` ← **Head:** `add-halo-pack` --- ### 📝 Commits (1) - [`3bb5f51`](https://github.com/Comfy-Org/ComfyUI-Manager/commit/3bb5f51b8b57def491e3af38bc0e3cad385713d8) Add HALO Debug Pack - AMD ROCm bf16 diagnostic tools ### 📊 Changes **1 file changed** (+11 additions, -0 deletions) <details> <summary>View changed files</summary> 📝 `custom-node-list.json` (+11 -0) </details> ### 📄 Description ## Summary Adds **HALO Debug Pack** to the custom node list - diagnostic tools for debugging black image issues caused by numpy's lack of bfloat16 support. ## Target Users - AMD ROCm users (especially Strix Halo APUs) - Older NVIDIA cards with bf16 issues - Apple Silicon / unified memory systems - Anyone experiencing black images from bf16 precision problems ## Nodes Included | Node | Purpose | |------|---------| | HALO VAE Decode (FP32) | Forces VAE decode to fp32, fixing numpy bf16 conversion | | HALO VAE Encode (FP32) | Forces VAE encode to fp32 for img2img | | HALO Latent Debug | Checks sampler output for NaN/dead latents | | HALO Conditioning Debug | Checks text encoder output for NaN | | HALO Model Debug | Inspects model dtype (bf16/fp16/fp32) | ## Repository https://github.com/bkpaine1/halo_pack ## Why This Matters When PyTorch bf16 tensors get converted to numpy for image saving, numpy silently produces garbage because it doesn't support bfloat16. This causes the infamous "black image" problem. These nodes help users diagnose exactly where the issue occurs in their pipeline. Thank you for maintaining ComfyUI-Manager! --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-28 21:11:04 +00:00