ComfyUI Node
H3 Row Chunk
Splits the attention qkv projection and the MLP into row chunks so their intermediates never exceed CUDA's int32 element limit. Fixes 'CUDA error: an illegal memory access was encountered' on sequences past ~75k rows, and lowers peak activation memory. These projections are row-wise, so this computes the same function — it is not an approximation and costs no quality.
H3 Row Chunk
- model
- MODEL
- report
◄chunk_rows0►
CategoryNynxz/H3/Sampling
Inputs (2)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | A MiniMax H3 model. | |
| chunk_rows | INT | 00–1000000 | Rows per chunk. 0 picks a size from the sequence: small enough to stay under the int32 limit with margin AND to hold each widened intermediate near 1 GiB. Set a smaller number only if a run still faults — some kernels allocate workspace beyond their output. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |
| report | STRING | Where the int32 ceiling falls for this model, and what chunking does at a few clip lengths. |