lite-mamba 0.1.4

Pure-PyTorch lightweight Mamba with multi-dilated causal conv front-end

lite-mamba

A minimal, pure-PyTorch version of Mamba with a multi-dilated causal depthwise conv front-end. No CUDA/Triton build needed; works on CPU or GPU with standard PyTorch ops.

Install

pip install torch einops
pip install lite-mamba

Usage

from lite_mamba import Mamba
import torch

x = torch.randn(2, 128, 512)  # (batch, seq, d_model)
m = Mamba(d_model=512, d_conv=3, conv_dilations=(1,2,4,8))
y = m(x)
print(y.shape)  # (2, 128, 512)

API quick reference

Mamba(d_model, d_state=16, d_conv=4, conv_dilations=(1,), expand=2, dt_rank="auto", dt_min=0.001, dt_max=0.1, dt_init="random", dt_scale=1.0, dt_init_floor=1e-4, conv_bias=True, bias=False, use_fast_path=False, layer_idx=None, device=None, dtype=None)

• d_model (int, required): input/output embedding size.
• d_state (int, default 16): SSM state dimension per channel. Larger gives longer memory; increases compute.
• d_conv (int, default 4): depthwise conv kernel size for each branch.
• conv_dilations (tuple[int], default (1,)): dilation per branch. Multiple values create parallel dilated convs; effective receptive field is (d_conv-1)*dilation.
• expand (float, default 2): inner width multiplier; sets d_inner = expand * d_model.
• dt_rank (int or "auto", default "auto"): rank of delta projection. "auto" sets ceil(d_model/16).
• dt_min, dt_max (float, defaults 1e-3 / 1e-1): log-uniform range for delta initialization.
• dt_init ("random" | "constant", default "random") and dt_scale, dt_init_floor: control delta init magnitude/stability.
• conv_bias (bool, default True): include bias in depthwise convs.
• bias (bool, default False): include bias in input/output linear projections.
• use_fast_path (bool): ignored in this pure-PyTorch build; kept for API compatibility.
• layer_idx (int | None): identifier for streaming cache registration; required when using allocate_inference_cache + inference_params.
• device, dtype: standard module factory kwargs.

Inference / streaming helpers

• allocate_inference_cache(batch_size, max_seqlen, dtype=None): preallocates conv and SSM state buffers for step-wise decoding.
• step(hidden_states, conv_state, ssm_state): single-token forward (expects hidden_states with shape (B, 1, d_model)).
• forward(..., inference_params): if inference_params has cached states (with key_value_memory_dict and seqlen_offset), uses them for streaming.

Highlights

• Multi-branch causal dilated convs (weighted sum via learned gates).
• Pure Python: no custom C++/CUDA or Triton kernels.
• Streaming support via per-branch conv states and SSM state caching.

Practical setups

• Local modeling / small context: d_conv=3, conv_dilations=(1,2,4), d_state=8–16, expand=2.
• Longer context: widen conv_dilations (e.g., (1,2,4,8,16)) or increase d_state to 32; expect higher memory/compute.
• Streaming/AR decoding: call allocate_inference_cache once per layer, pass inference_params during forward; use step inside your generation loop.
• Stability first: keep dt_min >= 1e-4 and dt_init_floor small; leave defaults unless you observe drift or exploding activations.

Notes

• Set different conv_dilations to adjust receptive field; keep kernels small (e.g., 3–5) to avoid excessive padding.
• use_fast_path flag is ignored here (kept for API compatibility).
• Reference selective scan is implemented in PyTorch for portability; faster fused kernels are omitted intentionally.

License

Apache-2.0