Class QuantizedQwenFp8Loader

java.lang.Object
smile.llm.quant.QuantizedQwenFp8Loader

public final class QuantizedQwenFp8Loader extends Object
Installs native FP8 linears into a Qwen hybrid model from HuggingFace SafeTensors.

Official Qwen3.8-FP8 uses fine-grained weight_block_size=[128,128] with weight_scale_inv. Those weights are installed as Fp8BlockLinear (LibTorch _scaled_mm_v2). Tensor-scale FP8 checkpoints still use Fp8Linear.

DeltaNet in_proj_a/in_proj_b, norms, conv, A_log, dt_bias, embeddings, lm_head, and the vision tower stay on the dense residual load path.

  • Method Details

    • install

      public static void install(QwenModel model, Path dir, Device device, int tpSize, int tpRank, ScalarType outDtype, int modelLoaderThreads) throws IOException
      Installs FP8 (or block-dequant dense) projections for every hybrid block on one TP rank.
      Parameters:
      model - target rank model (shells already on device).
      dir - checkpoint directory.
      device - rank device.
      tpSize - tensor-parallel world size.
      tpRank - this rank.
      outDtype - compute dtype for dequant / Fp8Linear output.
      modelLoaderThreads - shard-read concurrency (0 = auto).
      Throws:
      IOException - if checkpoint IO fails.
    • isInstalledProjectionKey

      public static boolean isInstalledProjectionKey(String hfName)
      Returns whether hfName is owned by the FP8 installer (must not be force-fed into dense shells).
      Parameters:
      hfName - HuggingFace weight key.
      Returns:
      true when the FP8 installer owns this key.