MiniMax H3 RoPE Chunk Patch (Legacy) - Star7
The 'Legacy' H3 chunk node is just the new one under an old name.
- model
- model
The display name says it: MiniMax H3 RoPE Chunk Patch (Legacy) - Star7. This is not a separate, older, worse implementation of chunking. It is the same node as MiniMax H3 Activation Chunk - Star7, still registered under its old class ID so that workflows saved before the package was renamed keep loading without errors. In the source it's literally a subclass:
class MiniMaxH3RoPEChunkPatch(MiniMaxH3ActivationChunkStar7):
"""Legacy class ID retained only so existing workflows keep loading."""
DEPRECATED = True
So everything that's true of the main node is true here: it chunks the three temporary activation peaks that make H3's long packed sequences OOM a consumer card - the QKV projection, the fused split-half RoPE, and the MLP SwiGLU expansion - without touching sampling, the VAE, or output resolution. Same inputs, same output, same behavior.
The inputs and output
Identical to the main node: model (MODEL) in, model (MODEL) out. Then chunk_tokens (RoPE, default 8192), mlp_chunk_tokens (default 8192 - the one to lower first when VRAM is tight), qkv_chunk_tokens (quality-capped at 4096 on SM75/RTX 20-series, uncapped on SM80+), auto_halve_on_oom (default on), reuse_mlp_weights (default on), verbose, and disable_dynamic_prefetch (a compatibility placeholder that does nothing). The attention_backend enum is the same six-way choice: existing, comfy_kitchen_int8 (default - approximate INT8), and the four SLA modes.
When would you ever see this node?
When you download an H3 workflow from a few versions back. "Legacy" here doesn't mean "broken" - it means the author froze the old ID in place so your saved graphs don't suddenly show a missing-node error. If it runs, it runs the exact same patching code as the new node. If you want to tidy up, swap it for the Activation Chunk node in your graph; if you leave it, nothing is lost.
The one genuinely useful troubleshooting note: if ComfyUI tells you MiniMaxH3RoPEChunkPatch is missing, that's the pack not being installed, not this node being retired. Install star7code/minimax-h3-chunk-star7 - ComfyUI Manager, search MiniMax H3 Activation Chunk - Star7, or:
cd ComfyUI/custom_nodes
git clone https://github.com/star7code/minimax-h3-chunk-star7.git
Then restart. The legacy ID resolves as soon as the pack loads, because it's bundled in the same __init__.py. Zero extra dependencies, no model downloads - this pack is self-contained.
Bottom line: you'll almost never add this node on purpose. You'll encounter it inside someone else's workflow, and the correct response is a shrug. It works, it's the modern code, and the only reason it has a separate page is that its class name survived a rename.
Inputs (10)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| chunk_tokens | INT | 81920–65536 | H3 RoPE sequence tokens per chunk. 2080 Ti 22GB: use 8192 after a safe 4096 validation run. |
| auto_halve_on_oom | BOOLEAN | true | If a reducible RoPE, MLP, or QKV temporary chunk OOMs, halve only that stage and retry down to 256 tokens. Zero still tries one full sequence first, then may auto-reduce. |
| verbose | BOOLEAN | true | Print compact one-time configuration and shape summaries to the ComfyUI console. |
| mlp_chunk_tokens | INT | 81920–65536 | H3 MLP tokens per chunk. Keeps the upstream block path while streaming the large expansion activation. The default 8192 is validated on the 22GB reference workflow. |
| disable_dynamic_prefetch | COMBO | off | Off by default and zero-overhead. Auto proactively uses bounded out_proj chunks when current free VRAM cannot safely hold the estimated full contraction; this may reduce speed. |
| qkv_chunk_tokens | INT | 81920–65536 | H3 QKV projection workspace chunk. Zero tries the full sequence first; automatic reduction lowers only QKV after a QKV projection OOM. |
| out_proj_chunk_tokens | INT | 40960–65536 | Internal compatibility value retained for old workflows. Attention output memory protection selects its own safe tile automatically. |
| reuse_mlp_weights | BOOLEAN | true | Reuse prepared QKV/MLP weight snapshots across token chunks when safe. Falls back to streamed preparation on VRAM pressure. |
| attention_backend | COMBO | comfy_kitchen_int8 | existing keeps the incoming attention patch (for example KJ Sage). comfy_kitchen_int8 selects ComfyUI's native INT8 attention and overrides an earlier MiniMax Sage patch. SLA uses fixed Top-K Q128/K64 routing. Sol uses Q64/K64 threshold routing; the SM80+ recommended Sol mode directly calls NVIDIA official BF16 exact+approx Sol-Attn. SM75 exposes the native All-INT8 Sol path; it keeps exact selected blocks and centroid contributions for unselected blocks while quantizing PV. Hybrid modes schedule CK/SLA/CK or CK/Sol/CK by complete sampling step. Strict sparse modes never silently fall back. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |