Skip to content

v1.3.0 — INT8 Instruct Fix, v2 Models, Latent Control Nodes

Latest
Latest

Choose a tag to compare

View all tags
@EricRollei EricRollei released this 21 Feb 19:44
· 5 commits to main since this release
v1.3.0
ea06cf9

What's New

✅ INT8 Instruct Models Now Working (5 bugs fixed)

All Instruct INT8 variants (Distil and full) are fully operational. Fixes include:

  • blocks_to_swap no longer forced to 0 for INT8
  • Added _load_int8_block_swap() with CB/SCB guard hooks
  • Fixed INT8 model size estimate (40GB → 80GB) in memory budget
  • CB/SCB now correctly fixed after .to(device)
  • Quant-specific gb_per_block values (NF4: 0.72, INT8: 2.4, BF16: 4.7)

🐛 VAE Decode Crash Fix

Fixed super(): __class__ is not a type (NoneType) crash during VAE decode. The cache-clearing code was destroying Python __class__ closure cells needed by super() in Conv3d. Fixed with isinstance(val, type) guard.

📦 v2 Pre-Quantized Models on Hugging Face

All 6 INT8/NF4 models re-quantized with improved settings and uploaded:

🧪 Experimental Latent Control Nodes (Base Models)

New nodes for composition control, img2img, and custom noise injection:

  • Hunyuan Empty Latent — random noise at specific resolution/seed
  • Hunyuan Latent Noise Shaping — frequency filter, amplify, invert
  • Hunyuan Generate with Latent — all-in-one with optional image/latent inputs (composition, img2img, energy_map modes)

🛠️ Unified Generate V2 Node

Single node replacing all base-model generate variants — auto-detects NF4/INT8/BF16, handles block swap, memory budgets, and VRAM management.

Upgrade Notes

  • Existing workflows continue to work — all previous loader/generate nodes remain available.
  • v2 models recommended — download the v2 variants for improved block swap defaults. v1 models still work.
  • INT8 Instruct users: Update the node set and re-download models (v2 recommended). blocks_to_swap=28-31 for Instruct (full) INT8 on 96GB GPUs.
Assets 2
Loading