MCPcopy Create free account
hub / github.com/LargeWorldModel/LWM / setup

Method setup

lwm/llama.py:629–656  ·  view source on GitHub ↗
(self)

Source from the content-addressed store, hash-verified

627 precision: Optional[Union[jax.lax.Precision, str]]=None
628
629 def setup(self) -> None:
630 config = self.config
631
632 self.w1 = nn.Dense(
633 config.intermediate_size,
634 dtype=self.dtype,
635 param_dtype=self.param_dtype,
636 use_bias=False,
637 kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
638 precision=self.precision,
639 )
640 self.w2 = nn.Dense(
641 config.hidden_size,
642 dtype=self.dtype,
643 param_dtype=self.param_dtype,
644 use_bias=False,
645 kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
646 precision=self.precision,
647 )
648 self.w3 = nn.Dense(
649 config.intermediate_size,
650 dtype=self.dtype,
651 param_dtype=self.param_dtype,
652 use_bias=False,
653 kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
654 precision=self.precision,
655 )
656 self.dropout = nn.Dropout(rate=self.config.resid_pdrop)
657
658 def __call__(self, x: jnp.ndarray, deterministic: bool = True) -> jnp.ndarray:
659 x = self.w2(nn.silu(self.w1(x)) * self.w3(x))

Callers

nothing calls this directly

Calls

no outgoing calls

Tested by

no test coverage detected