H3 Chunked · PASS2 Relay Actual Calls (T8 EXP)

Prompt Relay in this pack is timeline routing: you declare local events at time positions, and the relay plan steers the model's prompt attention to the right content at the right moment instead of running one global prompt across the whole clip. In a chunked PASS2 run, that routing has to be bound to a specific segment's geometry - which is what the matching Relay Bind node does. This is the pair's other half.

Its job, in one line: after the full-clip PASS2 finishes, report the actual standalone Relay forwards and routed-attention calls that occurred. Not that they were wired. That they happened.

Inputs and outputs

segment_result is the T8_CHUNKED_PASS2_RESULT for the segment. segment_spec identifies it. runtime is the T8_CHUNKED_PASS2_RELAY_RUNTIME produced by the Relay Bind for that same segment - and, as with the EAV side, cross-wiring a runtime from a different chunk gets you a mismatch, not a merged report.

Outputs are cumulative_av_latent and report_json, and it's an output node, so the report prints on the node as well. The cumulative latent is the running full-video result you carry forward across segments; it's the thing you eventually decode.

The division of labour, because it catches people

Here's the part the node description is precise about and that you will otherwise get wrong: Relay+EAV combined calls belong to the EAV audit, not this one. So if a graph has both effects active on the same segment, MiniMaxH3ChunkedPass2EAVAuditEXPT8 counts the combined invocations and this node counts the standalone Relay ones. If you wire both and then compare against your expectation of "one node per effect", the numbers look short and you go hunting for a bug that isn't there.

The practical rule: run both audits when both effects are present, and read them as a division of one body of work, not two independent totals.

What a count means, and what it doesn't

A non-zero call count proves the routing reached the model rather than being swallowed by a wrapper or an attention backend. That's genuinely valuable - it's the difference between "my effect is in the graph" and "my effect executed".

It does not prove the routing helped, or that the events you declared are the right ones. Those are content questions. The pack's attitude throughout this family is that telemetry is a floor, not a certificate, and the Relay is not an exception: an empty event list is a legal, do-nothing configuration, and report_only deliberately installs no routing bias at all. Expect counts from a report-only run to describe routing attempts while the output stays numerically unchanged, and don't file that as a bug.

Also worth remembering: Relay doesn't reset your clock. The routing coordinates are absolute timeline positions, so the events you wrote for the whole clip stay meaningful when a window is sampled - which is exactly why the bind node validates geometry before trusting the pair.

Install

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

Manager: search MiniMax H3 Audio T8. Fully quit ComfyUI and relaunch - node classes are registered at startup, so a page refresh won't surface them. No packages to install; the requirements file is empty by design so the pack can never replace ComfyUI's Torch or CUDA stack. You need a recent ComfyUI with native H3 support. Weights are separate: H3 in models/diffusion_models, Qwen in models/text_encoders, video and audio VAEs in models/vae.

If you're loading one of the chunked example graphs, note the README's warning: several of them rely on third-party packs (KJNodes, Comfyroll, Easy-Use and so on) that this pack does not install, and the examples ship placeholder media you're expected to replace before queueing.

Snags

• Comparing this node's count against a both-effects total. Combined calls live in the EAV audit.
• A runtime from the wrong segment. Rejected, correctly.
• Assuming report-only runs count as zero. They route and report without applying bias; the counts are still counts.
• Old duplicate copies of the pack in custom_nodes, which have shadowed the live module before and produced errors from code you don't have. One copy.