Class QwenWeightShard

java.lang.Object

smile.llm.model.qwen.QwenWeightShard

public final class QwenWeightShard
extends Object

Applies TP weight slices for Qwen module parameter names.

Method Summary

Modifier and Type	Method	Description
static Tensor	shard(String smileName, Tensor src, QwenModelArgs args, TensorShardSpec shard)	Returns a tensor suitable for loadStateDict / FP8 install on one TP rank.
static Tensor	shardScaleInv(String smileWeightName, Tensor src, QwenModelArgs args, TensorShardSpec shard)	Shards a block-scale inverse tensor [ceil(N/128), ceil(K/128)] using the same TP rules as shard(String, Tensor, QwenModelArgs, TensorShardSpec) for the matching weight name.

Methods inherited from class Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Method Details

shard

public static Tensor shard(String smileName,
 Tensor src,
 QwenModelArgs args,
 TensorShardSpec shard)

Returns a tensor suitable for loadStateDict / FP8 install on one TP rank. Caller owns any newly created view; may return src unchanged.

Parameters:

smileName - Smile weight name.

src - full-rank source tensor.

args - model hyperparameters.

shard - TP shard spec, or null for no sharding.

Returns:

sharded or original tensor.

shardScaleInv

public static Tensor shardScaleInv(String smileWeightName,
 Tensor src,
 QwenModelArgs args,
 TensorShardSpec shard)

Shards a block-scale inverse tensor [ceil(N/128), ceil(K/128)] using the same TP rules as shard(String, Tensor, QwenModelArgs, TensorShardSpec) for the matching weight name.

Parameters:

smileWeightName - Smile weight name for the matching linear.

src - full-rank scale-inv tensor.

args - model hyperparameters.

shard - TP shard spec, or null for no sharding.

Returns:

sharded or original tensor.