MiniMax H3 Enhance-A-Video + Strict Sage (Advanced EXP)

This is the node you reach for when you want both Enhance-A-Video's temporal boost and SageAttention's speed on H3, and you don't want to gamble on whether a third-party Sage patch just silently bypassed your FETA math. It's a composer: one node owns the H3 FETA route (the same cross-frame-interaction trick as the standalone MiniMaxH3EnhanceAVideoT8Advanced) and a strict SageAttention backend. It refuses external Sage object patches and it never quietly falls back to PyTorch attention - if Sage can't run, you get an error, not a slower-but-silently-wrong run.

Why does that matter? SageAttention is one of the most fiddly pieces of the modern ComfyUI stack. It needs a wheel matched to your exact Torch/CUDA/Python/GPU-architecture combo - people lose hours to sageattention is not new enough ... or could not determine CUDA architecture because the wheel's smXX doesn't match their card. The common failure mode in a DIY setup is that your attention patch silently doesn't engage and you keep generating at full PyTorch speed without noticing. This composer is designed to make that failure loud.

The mechanism: it wires FETA into H3's attention and routes the actual attention calls to a strict SageAttention HND backend, so a third-party whole-block patch can't wrap around and skip the FETA route. On the author's own hardware, a 1152×640×124 Stock20 run logged 1000 FETA measurements and 1000 Sage calls with zero failures or fallbacks - a mechanical verification, not a quality claim.

Inputs worth knowing:

• model - native H3 with no attention/object/block patch. For turbo8_alpha8, only the corrected 208-module Alpha8 bypass LoRA goes in front.
• sigmas - exact SIGMAS from your sampler; strict contract check.
• task_scope - visual (T2VA/I2VA/FL2VA/L2VA) or reference (Ref2VA/Hybrid, Stock20 only). It's explicit so task routing can't silently drift.
• mode - disabled runs neither Sage nor FETA and returns the original model; report_only still runs Sage but doesn't apply the FETA gain; apply_exp runs both.
• tau (default 4) - experimental FETA weight; a candidate, not an H3 optimum.
• max_workspace_mib and g_hard_limit - the score-buffer budget and the fail-closed gain guard, same as the standalone EAV node.

Outputs: model, runtime (H3_T8_EAV_RUNTIME for the audit node), and report_json.

The one real trap: this node is an exclusive owner. The tooltip is blunt - don't stack standalone STG, BlockCache, Prompt Relay, Sage, Long Video, or another guidance hook on top. If you need a combination, the pack has dedicated composers (there's an STG one, a BlockCache one, a Prompt Relay one). Mixing your own attention stack over the top is exactly the bypass this node exists to prevent.

Install is the standard pack route - ComfyUI Manager, search "MiniMax H3 Audio T8", or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8

Restart, and don't expect the pack to hand you SageAttention: no forced pip deps, no model downloads. The Sage wheel itself is on you to match to your environment. If you're getting Sage errors, run the pack's Environment Audit node with attention_backend set to sage_attention - it'll tell you whether the wheel, the CUDA arch, or the KJ-required symbols are the mismatch.