PDMD 2-NFE LoRA for MiniMax-H3 — ComfyUI format

ComfyUI-format conversion of the PDMD 2-NFE LoRA — the 2-step (2 NFE) student distilled from MiniMax-H3-33B with Projected Distribution Matching Distillation, rank 128, covering attention projections and both feed-forward layers of all 50 transformer blocks + 2 token-refiner blocks. (Companion repo for the 4-NFE variant: Iwannapose/minimax_h3_pdmd_4nfe_comfyui.)

Files

The two are numerically identical (the source fp32 file is an exact cast of the bf16 one); the bf16 file is the one to use. Both are larger than the source because the fused qkv_proj stores the block-diagonal zero padding explicitly.

What the conversion does

Source keys are Diffusers PEFT names (transformer.transformer_blocks.N.attn.to_q.lora_A.weight …); target keys follow the ComfyUI H3 layout (same scheme as the proven minimax_h3_fl2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors):

1. q/k/v → fused qkv_proj: A = [A_q; A_k; A_v] (rows), B = block_diag(B_q, B_k, B_v), row order [q; k; v] (A [384, 5376], B [21504, 384]).
2. SwiGLU remap: diffusers SwiGLU outputs [value; gate], ComfyUI's H3 _swiglu_eager expects [gate; up] — the two 14336-row halves of every mlp.fc1.lora_B are swapped.
3. alpha entries: qkv_proj = 384 (3 × 128), out_proj/fc1/fc2 = 128, so ComfyUI's alpha/rank scale = 1.0 = the PDMD fusion scale (W += (B @ A), alpha/rank = 128/128).
4. to_out.0 → attn.out_proj, ff.net.0.proj → mlp.fc1, ff.net.2 → mlp.fc2; transformer_blocks.N → blocks.N, token_refiner.refiner_blocks.N → token_refiner.blocks.N.

Verified: all 468 numeric checks pass against each original (exact in fp32), the key set is identical to the known-good turbo LoRA, and all 208 target keys exist with matching dimensions in minimax_h3_ref2va_int8_convrot.safetensors. The conversion output was also diffed tensor-by-tensor against an independent community converter — identical.

Usage (ComfyUI)

• Node: LoraLoaderModelOnly (model part only — the H3 text encoder is loaded separately).
• Strength: 1.0 (any other value scales the distillation delta and changes the 2-step behavior — do not use it as a partial-style LoRA).
• Sample at 2 denoising steps (the student's operating point); H3 scheduler config (shift 12 video / 3 audio), no CFG (H3 is guidance-distilled).