MiniMax H3 RoPE Chunk Patch (Legacy) - Star7

Low-VRAM MiniMax H3 patch. Chunks QKV projection, fused split-half RoPE, and the large MLP expansion activation; preserves INT8/ConvRot weights and the upstream DiT block path for FP16/BF16, Sage, LoRA, and third-party compatibility. Can safely retry without AIMDO prefetch. Compatible with FP16 Exact Fix - Star7. The default Comfy Kitchen INT8 attention mode is approximate; select existing to preserve upstream attention math. The experimental hybrid_sm75_ck_sla_all_int8 mode is SM75-only and schedules whole sampling steps as CK/SLA/CK using ComfyUI's real sigma context; it uses the existing experimental SM75 All-INT8 SLA only in the middle region and never falls back from a selected SLA step. Strict SLA is dependency-free from Sage, includes a native target-audio guard, and errors instead of silently falling back.

By star7code·Created 11 days ago·Updated about 3 hours ago· 9

model

model

◄chunk_tokens8192►

◄auto_halve_on_oomtrue►

◄verbosetrue►

◄mlp_chunk_tokens8192►

◄disable_dynamic_prefetch实验功能已移除►

◄qkv_chunk_tokens8192►

◄reuse_mlp_weightstrue►

◄attention_backendcomfy_kitchen_int8►

CategoryStar7/MiniMax H3

Inputs (9)

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	STRING	实验功能已移除	提前加载下一层（实验功能已移除）。该字段仅用于兼容旧工作流，不再参与计算。
qkv_chunk_tokens	INT	81920–65536	H3 QKV 投影临时显存分块。SM75 为保护参考语音稳定性，0 或高于 4096 的设定会按 4096 运行；SM80+ 不设此质量上限。若自动降档开启，QKV OOM 时只降低 QKV 后重试。
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. Strict SLA targets 85% dynamic video-block sparsity, protects target-audio queries, and never falls back after failure. Choose hybrid_sm75_ck_sla_all_int8 for the faster but more approximate SM75 All-INT8 SLA middle region. Choose sla_sm75_qk_int8_pv_fp16 for the recommended SM75 path, sla_sm75_all_int8_experimental for the faster but lower-precision SM75 experiment, or sla_sm80+_qk_int8_pv_fp16 for SM80+.

Outputs (1)

Name	Type	Description
model	MODEL	—