Add support for mamba, alternative to transformers

[geoman2]

Describe the feature
Is it possible to add support to Mamba, a deep learning architecture focused on long sequence modeling, for details please see https://en.wikipedia.org/wiki/Deep_learning

Your proposal
Just asking

[JoanFM]

Have u tried to build an Executor using mamba? I do not think this would be a problem at all

[jigarpatel1007]

I’d like to formally propose expanding Jina Serve’s architecture support to include Mamba-based models for long-sequence processing workloads.

---

Mamba has demonstrated promising results as a linear-time alternative to transformers, especially in scenarios requiring efficient processing of long sequences (e.g., time series, logs, streaming text). Adding Mamba support aligns well with Jina’s modularity and future-facing AI service goals.

---

🧠 Proposed Design Consideration

Rather than integrating Mamba as a core dependency, I propose:

• Developing a MambaExecutor template (similar to other model executors)
• Exposing the Mamba config (e.g., sequence length, hidden dimensions) via Executor.args or environment variables
• Making it deployable through Jina Hub or Flow YAML for on-demand use

This would keep Jina Serve agnostic of any one architecture while giving users the flexibility to adopt Mamba where it makes sense.

---

🔧 Optional Integration Points

• Add to Executor examples with a new module like MambaEncoder
• Provide a starter implementation based on state-spaces/mamba
• Extend monitoring/tracing hooks to track long-sequence processing performance

---

Let me know if this aligns with the roadmap. Thanks.