perf: parallelize GGUF block dequantization with rayon

[jamesbrink]

Problem

GGUF quantized transformer loading is slow (~64s for Qwen-Image Q4 on i9-13900K). The bottleneck is sequential per-block dequantization in QuantizedQwenImageTransformer2DModel::new():

for i in 0..cfg.num_layers {  // 60 blocks, one at a time
    blocks.push(QwenImageTransformerBlock::new(cfg, vb_blocks.pp(i), dtype, cpu_device)?);
}

Each block calls dequant_tensor() → vb.get_no_shape(name)?.dequantize(device)?.to_dtype(dtype)? which expands Q4→F32→BF16 on CPU. With 60 independent blocks processed sequentially, only one CPU core is utilized.

Proposed Fix

Use rayon to parallelize block dequantization across CPU cores. The blocks are independent — no shared mutable state between them. On an i9-13900K (24 threads), this could reduce load time from ~64s to ~10-15s.

use rayon::prelude::*;

let blocks: Vec<QwenImageTransformerBlock> = (0..cfg.num_layers)
    .into_par_iter()
    .map(|i| QwenImageTransformerBlock::new(cfg, vb_blocks.pp(i), dtype, block_device))
    .collect::<Result<Vec<_>>>()?;

Scope

• crates/mold-inference/src/qwen_image/quantized_transformer.rs — parallelize block construction
• Same pattern applies to any other GGUF pipeline with block-level dequantization (Z-Image quantized transformer uses a similar loop)
• rayon is already available as a transitive dependency via candle

Environment

• CPU: i9-13900K (24 threads)
• Model: qwen-image:q4 (60 transformer blocks, ~10GB GGUF → ~40GB BF16 dequantized)
• Current load time: ~64s (release build)
• Expected: ~10-15s with parallel dequantization