Class QuantizedQwenFp8Loader
java.lang.Object
smile.llm.quant.QuantizedQwenFp8Loader
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 Summary
Modifier and TypeMethodDescriptionstatic voidinstall(QwenModel model, Path dir, Device device, int tpSize, int tpRank, ScalarType outDtype, int modelLoaderThreads) Installs FP8 (or block-dequant dense) projections for every hybrid block on one TP rank.static booleanisInstalledProjectionKey(String hfName) Returns whetherhfNameis owned by the FP8 installer (must not be force-fed into dense shells).
-
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 ondevice).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
Returns whetherhfNameis owned by the FP8 installer (must not be force-fed into dense shells).- Parameters:
hfName- HuggingFace weight key.- Returns:
truewhen the FP8 installer owns this key.
-