Nodes/ComfyUI_JR_MiniMaxH3Node/JR H3 Unified Acceleration v2 (Experimental)
ComfyUI Node

JR H3 Unified Acceleration v2 (Experimental)

Experimental H3 sparse backend selection. Start from a MODEL without attention patches. Legacy is the conservative default; select core to evaluate the native chunked producer. Auto may choose legacy before sampling, never after a CUDA/runtime failure. Core currently excludes TST, tau_profile, custom legacy quantization and allow_compile.

By Goldlionren·Created 2 months ago·Updated 4 days ago· 58
JR H3 Unified Acceleration v2 (Experimental)
  • model
  • model
◄enabletrue►
◄sage_attentionsageattn_qk_int8_pv_fp8_cuda++►
◄allow_compilefalse►
◄enable_low_vram_attentiontrue►
◄head_chunks4►
◄enable_low_vram_ffntrue►
◄ffn_chunks4►
◄ffn_seq_threshold4096►
◄enable_sol_attntrue►
◄tau1.30►
◄start_percent0.20►
◄end_percent0.90►
◄min_tokens12288►
◄int8_qktrue►
◄int8_pvtrue►
◄sink_conditioningexact_kv_and_rows►
◄mortonfalse►
◄morton_curve2d_frame►
◄verbosefalse►
◄use_tmafalse►
◄dense_blocks►
◄sparse_backendlegacy_kijai►
◄extra_tokens256►
◄tau_profile—►
◄enable_tstfalse►
◄tst_strength0.10►
CategoryJR MiniMax H3/Optimization

Inputs (27)

NameTypeDefaultDescription
modelMODEL—
enableBOOLEANtrue—
sage_attentionCOMBOsageattn_qk_int8_pv_fp8_cuda++8 options: disabled, auto, sageattn_qk_int8_pv_fp16_cuda, sageattn_qk_int8_pv_fp16_triton, sageattn_qk_int8_pv_fp8_cuda, sageattn_qk_int8_pv_fp8_cuda++, +2
allow_compileBOOLEANfalse—
enable_low_vram_attentionBOOLEANtrue—
head_chunksINT41–56—
enable_low_vram_ffnBOOLEANtrue—
ffn_chunksINT41–64—
ffn_seq_thresholdINT4096256–262144—
enable_sol_attnBOOLEANtrue—
tauFLOAT1.300–4—
start_percentFLOAT0.200–1—
end_percentFLOAT0.900–1—
min_tokensINT122880–1048576—
int8_qkBOOLEANtrue—
int8_pvBOOLEANtrue—
sink_conditioningCOMBOexact_kv_and_rows3 options: exact_kv, exact_kv_and_rows, off
mortonBOOLEANfalse—
morton_curveCOMBO2d_frame2 options: 3d, 2d_frame
verboseBOOLEANfalse—
use_tmaBOOLEANfalse—
dense_blocksSTRING—
sparse_backendCOMBOlegacy_kijai4 options: legacy_kijai, core, auto, disabled
extra_tokensINT2560–256—
tau_profileoptSTRING—
enable_tstoptBOOLEANfalseExperimental temporal Q correction. Keep Morton and compile off.
tst_strengthoptFLOAT0.100–1Independent of Sol tau. Start with 0.1; compare motion and prompt adherence.

Outputs (1)

NameTypeDescription
modelMODEL—