mamba2¶
Model type |
|
Class |
|
Task |
|
Source |
|
Description¶
Mamba2/SSD causal language model with multi-head SSM.
Pure Mamba2 model — all layers are Mamba2Block (no attention). Carries conv_state + ssm_state per layer (no KV cache).
State shapes (per layer): conv_state: (batch, conv_dim, conv_kernel - 1) ssm_state: (batch, num_heads, head_dim, state_size)
HuggingFace reference: Mamba2ForCausalLM.
Usage¶
mobius build --model <MODEL_ID> output_dir/
from mobius import build
model = build("<MODEL_ID>")