Module init 

Weight initialization utilities (C-INIT-001).

Provides rand_normal_seeded for proper random normal initialization, matching HuggingFace LLaMA’s normal(0, initializer_range).

Replaces the legacy sinusoidal sin(i * const) * scale placeholder that caused a 16x convergence gap vs PyTorch (entrenar#309).

References:

• Touvron et al. (2023) LLaMA: arxiv 2302.13971
• He et al. (2015) Kaiming init: arxiv 1502.01852
• HuggingFace LlamaPreTrainedModel._init_weights

Constants

INITIALIZER_RANGE

Default initializer range matching HuggingFace LLaMA config.

Functions

get_init_seed

Get the current initialization seed.

lock_init_seed

Lock the init-seed critical section and set the seed atomically.

rand_normal_seeded

Generate n random normal values with mean=0 and std=INITIALIZER_RANGE.

set_init_seed

Set the global initialization seed (called from training config).