H3 Unified Acceleration

There are a handful of well-known speed tricks for MiniMax H3 on a consumer GPU - SageAttention for fast attention, low-VRAM attention and chunked feed-forward patches, and Sol-Attn as a fallback dense backend. The JR H3 Unified Acceleration node is a single switchboard that composes all of them in a fixed order: KJ Sage → Low VRAM Attention → Chunk FeedForward → Sol-Attn. You feed it a MODEL, it returns a patched MODEL ready for the sampler.

Here's the honest part, up front: this node does not install or vendor the actual accelerators. They're external dependencies that this pack deliberately refuses to copy - you bring kijai/ComfyUI-KJNodes (for the Sage/Low-VRAM/FFN patches), the SageAttention variant your chosen mode needs (sageattention or sageattn3), and kijai/ComfyUI-SolAttn_triton plus its Triton runtime. They're only resolved at execution time, so a missing one won't stop the other JR nodes from loading - but this node will error when you actually run it. The README is also explicit about install order: Sol must be installed last, so Sage remains the previous dense backend for scenarios where Sol isn't applicable.

The switches that matter

• model in → patched model out. The rest is configuration.
• sage_attention - pick the backend: disabled, auto, int8 qk/pv variants over CUDA or Triton, sageattn3, or sageattn3_per_block_mean. The default is the cuda++ int8 variant.
• enable_low_vram_attention + head_chunks (default 4) - split attention heads to cap peak memory.
• enable_low_vram_ffn + ffn_chunks (default 4) + ffn_seq_threshold - chunk the feed-forward over sequence length. The threshold is the sequence length below which chunking kicks in.
• enable_sol_attn + tau (default 1.3) + start_percent/end_percent (0.2→0.9) - Sol-Attn's soft-attention parameters and the sampling window where it applies.
• int8_qk / int8_pv - quantization of the QK and PV attention paths; both default on.
• sink_conditioning (exact_kv, exact_kv_and_rows, off), morton / morton_curve, use_tma, dense_blocks, allow_compile (default off - compilation is slow to build, fast at runtime), verbose.

Every enable is a true bypass - flipping one off removes the patch rather than simulating it with a chunk value.

Recommended chain

The README's acceleration chain puts this node in the middle of the model-prep pipeline:

Load Diffusion Model → Turbo LoRA (external) → Reserved VRAM Setter (optional)
  → H3 Unified Acceleration → JR H3 Adaptive Cache (optional/experimental)
  → Sigma Shift → Basic Guider / Scheduler

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Goldlionren/ComfyUI_JR_MiniMaxH3Node
<your-comfyui-python> -m pip install -r ComfyUI_JR_MiniMaxH3Node/requirements.txt

or ComfyUI Manager → search ComfyUI_JR_MiniMaxH3Node. Then install KJNodes, SageAttention, and SolAttn_triton separately (Sol last), per their own docs.

Troubleshooting

• Runs fine until this node, then errors - a missing external accelerator. The error message tells you which one; install it (SageAttention variants are finicky about CUDA/Triton versions).
• No speedup / slower - allow_compile off means no kernel fusion; some backends only pay off at longer sequences. Check ffn_seq_threshold and min_tokens against your actual clip length.
• Sol-Attn-specific artifacts - tau, start_percent, and end_percent are the tuning knobs; the window defaults (20–90%) are a reasonable starting point, not gospel.