minimax-h3-conditioner-fp8

MiniMax-H3 conditioner, e4m3 (fine-tuned)

The Qwen3-VL conditioner MiniMax-H3 reads hidden_states[50] from, narrowed to FP8 e4m3 with one float32 scale per 128 elements of a row.

Which weights. The stock MiniMaxAI/MiniMax-H3 text_encoder/ with the 9 tensors of abenliao/minimax-h3-conditioner-patch spliced in -- layers.31..39.self_attn.o_proj.weight, the only tensors the fine-tune changed. Applied byte-for-byte and read back to verify before quantising.

gate_proj and up_proj are already joined into mlp.gate_up_proj.weight. That is not a convenience: the loader keeps streamed weights in pinned write-combined host memory, which cannot be read back cheaply to join anything, so the join has to happen here or not at all.

66.7 GB -> 36.5 GB, of which 31.2 GB is e4m3 and the rest is what stays bf16.

Load with H3_TEXT_ENCODER=<this dir>. Add H3_STREAM_WEIGHTS=1 and it is 2.3 GB resident instead of 25.9.