Class DeepSeekTransformerBuilder
java.lang.Object
io.github.kirstenali.deepj.transformer.DeepSeekTransformerBuilder
Builder for DeepSeek-style transformer stacks.
Assembles a TransformerStack of DeepSeekTransformerBlocks
(RMSNorm + Multi-Head Latent Attention + SwiGLU).
- See Also:
-
Constructor Summary
Constructors -
Method Summary
-
Constructor Details
-
DeepSeekTransformerBuilder
public DeepSeekTransformerBuilder()
-
-
Method Details
-
dModel
-
nHeads
-
dFF
-
nLayers
-
maxSeqLen
Maximum sequence length for the RoPE table. Required. -
qRank
Q latent dimension for MLA. Required. Typical value:dModel / 2. -
kvRank
KV latent dimension for MLA. Required. Typical value:dModel / 4. An incremental decoder may use this latent for a compressed cache. -
seed
-
random
-
build
-