SpecMamba: A Multi-Domain Mamba Framework

• SpecMamba is a term describing a range of Mamba-based methods that leverage speculative decoding and selective state space models for efficient long-context inference.
• It combines external drafter flexibility with linear-time sequence modeling, constant per-token memory, and tree drafting to achieve improved throughput and reduced memory usage.
• SpecMamba also spans spectral and spatial-spectral applications in hyperspectral imaging and point cloud processing, offering tailored adaptations for domain-specific challenges.

SpecMamba is a non-uniform term in the recent Mamba literature. In its most explicit usage, it denotes speculative decoding with Mamba-based drafters for LLM inference, where a selective state space model replaces a Transformer drafter in order to combine external-drafter flexibility with linear-time sequence modeling, constant per-token memory at inference, and lower memory usage than KV-cache-based alternatives (Choi et al., 1 Jun 2025). The same label is also used for an FPGA accelerator that makes speculative decoding practical for Mamba inference through system, algorithm, and hardware co-design (Zhong et al., 24 Sep 2025). In adjacent literatures, especially hyperspectral image classification and point cloud processing, “SpecMamba” is also used loosely as a shorthand for spectral, spectral-spatial, or spectral-informed Mamba variants such as SpectralMamba, S$^2$Mamba, SS-Mamba, and Spectral Informed Mamba (Yao et al., 2024).

1. Nomenclature and scope

The term has no single canonical meaning across all subfields represented in the current literature. Its most direct and formal use is in speculative decoding for autoregressive generation, but some summaries also use it as a convenient label for Mamba-style methods whose defining feature is spectral or spatial-spectral modeling rather than speculative decoding (Choi et al., 1 Jun 2025). This suggests that “SpecMamba” functions both as a proper method name and as a looser family label, depending on context.

Usage of “SpecMamba”	Formal paper title	Domain
Mamba drafter for speculative decoding	“Mamba Drafters for Speculative Decoding”	LLM decoding
FPGA speculative-decoding accelerator	“SpecMamba: Accelerating Mamba Inference on FPGA with Speculative Decoding”	Hardware acceleration
Loose shorthand for spectral Mamba variants	“SpectralMamba,” “S$^2$Mamba,” “SS-Mamba”	Hyperspectral imaging
Loose shorthand for spectral-informed Mamba	“Spectral Informed Mamba”	Point cloud processing

A common misconception is to treat every spectral or spatial Mamba variant as the same method. The hyperspectral literature actually uses distinct names—SpectralMamba, S$^2$Mamba, and SS-Mamba—each with different tokenization, scanning, or fusion mechanisms (Wang et al., 2024). A related boundary case appears in medical anomaly detection: SP-Mamba is a spatial-perception Mamba model, and the source summary explicitly notes that the paper does not mention a model called SpecMamba (Pan et al., 25 Jul 2025).

2. SpecMamba as speculative decoding with Mamba drafters

In autoregressive generation, speculative decoding speeds inference by combining a fast drafter model $M_q$ with a larger target model $M_p$. The drafter proposes a block of $\gamma$ draft tokens, the target model evaluates them in parallel, and verification proceeds token by token. The standard verification rule given in the paper is

$$\text{accept } \tilde{x}_t \text{ if } u < \min\left(1,\frac{p_t}{q_t}\right), \quad u\sim U[0,1].$$

If all $\gamma$ tokens are accepted, an extra token is sampled from the target model (Choi et al., 1 Jun 2025).

The motivation for Mamba-based drafters is a trade-off already present in speculative decoding. External drafters are plug-and-play and can be used with many target models, but may be slow and inherit quadratic Transformer costs. Self-speculation is faster and better aligned to the target model, but requires retraining for each target model. The Mamba drafter is proposed as a way to retain cross-model adaptability while avoiding Transformer attention’s quadratic complexity by using the linear structure of selective state space models (Choi et al., 1 Jun 2025).

The architectural intuition is direct. A Transformer drafter must manage a KV cache whose size grows with context length. A Mamba drafter keeps only a fixed-size state, so memory remains essentially constant during decoding, and the paper emphasizes constant per-token memory at inference, linear-time sequence modeling, and the absence of KV-cache growth. This is presented as especially advantageous in long-context settings, where drafting speed and memory usage remain low even as context grows (Choi et al., 1 Jun 2025).

Empirically, the paper reports that Mamba-based drafters outperform existing external drafting methods and are comparable to state-of-the-art self-speculation approaches while using less memory and maintaining cross-model adaptability. On GSM-8K, Mamba achieves nearly $2\times$ throughput of Pythia-410M in sampling setup, and in long-context settings it reaches similar throughput to EAGLE while using up to $20$ GB less memory (Choi et al., 1 Jun 2025).

3. Tree drafting, batch generation, and adaptive search

SpecMamba in the speculative-decoding sense is not limited to replacing a Transformer drafter with an SSM. A central methodological contribution is tree-structured drafting. A tree configuration is written as

$^2$0

where $^2$1 is the number of new nodes sampled at the $^2$2-th generation step, and the total batch size at depth $^2$3 is

$^2$4

Wider trees increase candidate diversity and can improve acceptance length, but they also increase drafting cost (Choi et al., 1 Jun 2025).

The paper argues that Mamba is particularly well suited to tree drafting because each next-token prediction requires only the current state. Expanding multiple children from a node therefore reduces to copying the state and sampling, whereas Transformer branching requires copying and managing a KV cache for each branch. To reduce overhead further, the implementation pre-allocates state caches for each batch size and uses CUDA graphs to reuse computation graphs for fixed batch shapes such as $^2$5 (Choi et al., 1 Jun 2025).

Tree selection is made adaptive through a multi-armed bandit formulation. The candidate tree set is

$^2$6

and UCB is used with

$^2$7

The appendix relates reward to speculative-decoding speedup via

$^2$8

and

$^2$9

The reward is derived from the inverse of this expression (Choi et al., 1 Jun 2025).

The reported ablations indicate that tree drafting improves acceptance length and throughput. With tree configuration $^2$0, the paper gives MT-bench examples in which JSRed reaches acceptance length $^2$1, latency $^2$2, throughput $^2$3, while JSViolet reaches acceptance length $^2$4, latency $^2$5, throughput $^2$6. The paper also evaluates $^2$7 and $^2$8, noting that Mamba remains relatively stable across configurations while Transformer drafters degrade more as tree depth grows (Choi et al., 1 Jun 2025).

4. FPGA SpecMamba and hardware co-design

The FPGA paper uses SpecMamba as a proper method name for an accelerator that targets Mamba inference with speculative decoding. Its starting point is that Mamba is attractive for long-sequence modeling because of linear-time sequence modeling, but autoregressive decoding remains memory-bound. The paper attributes this to repeated loading of weights and intermediate data through linear projection layers, a 1D convolution, an SSM layer, and residual and normalization operations, with throughput limited by memory bandwidth rather than raw compute (Zhong et al., 24 Sep 2025).

Three challenges are identified. First, hidden state backtracking is difficult because Mamba compresses the past into a fixed-size hidden state rather than retaining a Transformer-style KV cache. Second, tree-based parallel verification is incompatible with naive SSM execution because Mamba is a sequential state machine with causal hidden-state dependencies. Third, the draft and target models impose a hardware workload mismatch: draft generation is memory-bound, while target verification over many candidates is more compute-heavy and parallel (Zhong et al., 24 Sep 2025).

SpecMamba addresses these issues through system, algorithm, and hardware co-design. At the system level, it uses memory-aware hybrid backtracking: off-chip hidden-state storage for the draft model and on-chip activation caching for the target model. At the algorithm level, it introduces FIFO-based tree verification with tiling. From the SSM update

$^2$9

the paper observes inter-token dependency but no intra-token dependency, so the hidden state tensor can be split into tiles of size $M_q$0, processed over $M_q$1 iterations, and scheduled with breadth-first traversal. The resulting FIFO storage requirement is at most

$M_q$2

for a tree with $M_q$3 nodes (Zhong et al., 24 Sep 2025).

At the hardware level, the design adopts parallel linear computation and sequential SSM computation with overlap. Linear layers are parallelized across $M_q$4 tokens, with output-channel blocks $M_q$5 and input-channel tiles $M_q$6, so the total linear-layer computation takes $M_q$7 cycles. The SSM stage, using fully unrolled element-wise multiplication units, processes one tile of one token per cycle and takes $M_q$8 cycles. The scheduling choice is to make $M_q$9 approximately match $M_p$0, so linear-layer work and SSM work overlap (Zhong et al., 24 Sep 2025).

Implementation is reported on AMD VHK158 and VCK190 platforms using Vitis HLS and Vivado Design Flow, with throughput measured using PYNQ and power using Xilinx BEAM. Models are quantized to INT4; the target model is Mamba2-2.7B and draft models are Mamba2-130M, 370M, and 780M. The abstract reports a $M_p$1 speedup over GPU baselines and a $M_p$2 improvement over prior FPGA solutions, together with $M_p$3 and $M_p$4 higher energy efficiency, respectively. The detailed comparisons include 313.4 tokens/s on VHK158 versus 93 tokens/s for LightMamba on U280 in the HBM scenario, and 20.6 tokens/s versus 7.21 tokens/s on VCK190 in the DDR scenario (Zhong et al., 24 Sep 2025).

5. Spectral and spatial-spectral uses in hyperspectral imaging

In hyperspectral image classification, “SpecMamba” often serves as a loose umbrella for Mamba-style methods that replace attention with selective structured state space models in order to model long-range dependencies with linear complexity rather than Transformer self-attention with $M_p$5 cost. This motivation is explicit in SpectralMamba, S$M_p$6Mamba, and SS-Mamba, all of which treat the spectrum as an ordered or near-continuous sequence while incorporating spatial context through distinct mechanisms (Yao et al., 2024).

SpectralMamba models hyperspectral data at two levels. In spatial-spectral space it uses a gated spatial-spectral merging module,

$M_p$7

to learn a dynamical mask by efficient convolutions. In hidden state space it applies a Mamba/S6 block with input-dependent $M_p$8. Between these stages it introduces piece-wise sequential scanning,

$M_p$9

which converts approximately continuous spectra into shorter sequences while retaining short- and long-term context. On Houston2013, the paper reports $\gamma$0, $\gamma$1, $\gamma$2, Params $\gamma$3, and MACs $\gamma$4 (Yao et al., 2024).

S$\gamma$5Mamba is explicitly framed as a spatial-spectral state space model for hyperspectral image classification. It operates on an input patch $\gamma$6, extracts features with two selective SSM branches, and fuses them with a Spatial-spectral Mixture Gate. The spatial branch uses Patch Cross Scanning over four routes to capture local-to-long-range spatial context, while the spectral branch uses Bi-directional Spectral Scanning over

$\gamma$7

to capture inter-band semantic progression in both directions. Fusion is handled by learnable matrices $\gamma$8 and a thresholded competitive combination

$\gamma$9

The reported results are OA $$\text{accept } \tilde{x}_t \text{ if } u < \min\left(1,\frac{p_t}{q_t}\right), \quad u\sim U[0,1].$$0 on Indian Pines, OA $$\text{accept } \tilde{x}_t \text{ if } u < \min\left(1,\frac{p_t}{q_t}\right), \quad u\sim U[0,1].$$1 on Pavia University, and OA $$\text{accept } \tilde{x}_t \text{ if } u < \min\left(1,\frac{p_t}{q_t}\right), \quad u\sim U[0,1].$$2 on Houston 2013, with only about $$\text{accept } \tilde{x}_t \text{ if } u < \min\left(1,\frac{p_t}{q_t}\right), \quad u\sim U[0,1].$$3M parameters (Wang et al., 2024).

SS-Mamba takes a dual-token approach. A Spectral-Spatial Token Generation Module converts an HSI cube into a spatial token sequence and a spectral token sequence; stacked Spectral-Spatial Mamba Blocks then process the two streams with separate basic Mamba blocks and a feature-enhancement module guided by center-region information. The enhancement mechanism computes a fused center feature

$$\text{accept } \tilde{x}_t \text{ if } u < \min\left(1,\frac{p_t}{q_t}\right), \quad u\sim U[0,1].$$4

produces gating weights with $$\text{accept } \tilde{x}_t \text{ if } u < \min\left(1,\frac{p_t}{q_t}\right), \quad u\sim U[0,1].$$5, broadcasts them to token dimensions, and modulates both streams element-wise. With default settings $$\text{accept } \tilde{x}_t \text{ if } u < \min\left(1,\frac{p_t}{q_t}\right), \quad u\sim U[0,1].$$6, $$\text{accept } \tilde{x}_t \text{ if } u < \min\left(1,\frac{p_t}{q_t}\right), \quad u\sim U[0,1].$$7, $$\text{accept } \tilde{x}_t \text{ if } u < \min\left(1,\frac{p_t}{q_t}\right), \quad u\sim U[0,1].$$8, $$\text{accept } \tilde{x}_t \text{ if } u < \min\left(1,\frac{p_t}{q_t}\right), \quad u\sim U[0,1].$$9, and $\gamma$0, the reported main results are OA $\gamma$1 on Indian Pines, OA $\gamma$2 on Pavia University, and OA $\gamma$3 on Houston (Huang et al., 2024).

Taken together, these hyperspectral methods support a narrower, domain-specific reading of “SpecMamba” as shorthand for spectral or spatial-spectral Mamba frameworks rather than speculative decoding. A plausible implication is that the naming overlap comes from the prominence of spectral-sequence modeling in hyperspectral data and the suitability of selective SSMs for long ordered band structures.

6. Related spectral-informed usages and boundary cases

A further extension of the naming pattern appears in point cloud processing, where Spectral Informed Mamba is explicitly summarized as “SpecMamba / SI-Mamba.” The method addresses a problem specific to Mamba on point clouds: traversal order matters, and naive 2D or 3D grid-based traversals are view dependent and may not respect surface adjacency. The proposed solution builds a patch-connectivity graph, forms the random-walk Laplacian

$\gamma$4

computes low-frequency eigenvectors, and uses them to define isometry-invariant traversals (Bahri et al., 6 Mar 2025).

Three mechanisms structure the method. Spectral-Aware Spectral Traversing uses forward and reverse orderings from eigenvectors for classification. Hierarchical Local Traversal recursively partitions tokens with binary codes derived from spectral components for segmentation. Traverse-Aware Repositioning restores masked tokens to their original positions in Mamba-based masked autoencoding, and reconstruction is trained with Chamfer distance

$\gamma$5

The results reported in the summary include ScanObjectNN pretrained accuracies of $\gamma$6, $\gamma$7, and $\gamma$8 on OBJ-BG, OBJ-ONLY, and PB-T50-RS, respectively, as well as ShapeNetPart pretrained segmentation of $\gamma$9 for the HLT variant (Bahri et al., 6 Mar 2025).

By contrast, SP-Mamba in medical anomaly detection is a spatial-perception Mamba framework rather than a SpecMamba method. It combines a ResNet34 encoder, window-sliding prototype learning, a Circular-Hilbert scanning-based Mamba decoder, and anomaly scoring based on reconstruction, prototype distance, concentration, and contrast. The source summary explicitly treats any association with “SpecMamba” as interpretive rather than formal. This boundary matters because it shows that the Mamba literature contains multiple similarly abbreviated names whose relation is conceptual rather than nominal (Pan et al., 25 Jul 2025).

Across these usages, the stable element is not a single architecture but a recurring design pattern: Mamba is adopted when linear-time sequence modeling, fixed-size state, and structured scanning or traversal offer a more suitable inductive bias than attention or classic recurrence. Where the term is used strictly, it refers to speculative decoding with Mamba drafters or to its FPGA realization; where it is used loosely, it denotes spectral or spectral-informed Mamba variants in data modalities whose structure is naturally sequential, traversable, or banded.