BOLT
PYTHON
Bolt(
d_model,
num_heads,
*,
latent_dim=32,
bias=False,
dropout=0.0,
causal=True,
backend="auto",
autotune_kernels=True,
use_sdpa=True,
position=None,
)BoltAttentionis an alias ofBolt.
MAIN METHODS
bolt(x) | Full-sequence forward. |
prefill(x, start_pos=0) | Return output and generation cache. |
decode_step(x, cache) | Advance the recurrent cache. |
project_cache_state(...) | Build cache state from hidden inputs. |
prefill_with_cache(...) | Prefill helper with explicit cache output. |
set_backend(...) | Setauto,nativeorpytorch. |
STANDARD ATTENTION
PYTHON
Attention(
d_model,
num_heads,
*,
bias=False,
dropout=0.0,
causal=True,
backend="auto",
position=None,
)BOLT MODEL
PYTHON
BoltModel(
vocab_size=50257,
context=2048,
layers=6,
dim=384,
heads=6,
latent_dim=32,
position="rope",
backend="auto",
)The complete model configuration also supports FFN, residual, normalization, dropout, embedding tying and MLBricks FFNBrick/ResController options.