MiniMax H3 LightX2V SLA Loader + Attention (Advanced EXP)

Every once in a while a pack ships a node that is basically a contract with the user: "do exactly these five things in this order, or it silently won't work." This is that node. The LightX2V Turbo-SLA path runs H3 at 4 steps with a learned router skipping about 85% of the attention blocks - a real speed play, the same distillation family as LCM/Turbo LoRAs but done in latent attention space. But it only works if the attention path is owned by this node. External SageAttention, Sol-Attn, or a stray LoRA on the wire will quietly break the assumption the whole thing is built on.

It's in T8/MiniMax H3/Performance/Experimental, and the "Advanced EXP" tag is doing honest work: the README's own conclusion for this line is that it passed mechanical verification at 256×256×22 INT8 - and explicitly does not claim quality, speed, audio, or general 16GB benefits. Treat it as a promising experiment you're helping validate, not a shipped speed hack.

How it works

The node loads the authenticated 4-step FL2VA Turbo-SLA LoRA (minimax_h3_fl2v_turbo_4step_v0.1_768p_sla_comfyui_bf16.safetensors - file size, SHA-256, metadata and all 208 patches are verified before it's trusted) and owns the LightX2V 85%-block-sparse Sage2 attention path for all 50 main H3 blocks.

The wiring contract, from the tooltip, is non-negotiable: connect the clean native H3 MODEL output from Dual-Clock, and do not place KJ Sage, Sol-Attn, FETA, BlockCache, STG or another LoRA before it. Then feed the same 4-step native_flow SIGMAS from Dual-Clock (video shift 6.0, audio shift 3.0). Those shifts aren't decoration - they're what the 4-step profile was trained against.

mode picks the behavior: apply_lightx2v_sla (default) runs the learned 85%-sparse router; dense_lora_control keeps the same LoRA but uses dense attention as a scientific A/B; disabled_identity changes nothing. base_policy defaults to auto_detect_exp - the official SLA base is BF16 FL2VA, and quantized bases are explicitly experimental, so official_bf16_only is the strict alternative. max_router_workspace_mib (512) is a fail-closed ceiling for the router's score/map workspace, not a whole-workflow VRAM limit.

Outputs

model (patched, goes to the sampler), runtime (an H3_T8_LIGHTX2V_SLA_RUNTIME - feed it to the SLA Audit node, that's its whole purpose), and report_json.

Installing it

Pack first (Manager → "MiniMax H3 Audio T8", or clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into ComfyUI/custom_nodes, restart). Then the real dependency: this node needs spas-sage-attn, matched to your Torch/CUDA build, and the pinned SLA LoRA in models/loras. The README is explicit that "installed latest version" proves nothing - Torch, CUDA, Python, GPU architecture and the Sage wheel must all line up.

Common issues

Runs but the report says a kernel fell back. Something else grabbed the attention path. Bypass every external attention node between Dual-Clock and this node. If you need KJ Sage in the same workflow, that's the other node (SLA + KJ Sage Composer), not this one.

Import error mentioning spas-sage-attn. Install the wheel for your exact Torch/CUDA and fully restart ComfyUI. A silent fallback isn't an option here - the node is designed to refuse rather than pretend.