Tessera Hypernetwork

Generate per-session LoRA adapters for inference tasks using hypernetwork synthesis.

Version: 1.4.0

Features

• Metadata-to-LoRA: Generate adapters from structured user metadata (JSON)
• Text-to-LoRA: Generate adapters from natural language descriptions
• Doc-to-LoRA with SHINE: Generate adapters from document content using SHINE (ICML 2026) for long-context internalization
• Base Model Management: Download, cache, and serve base models with vLLM integration
• FastAPI: Modern async Python web framework
• OpenAI-compatible API: Easy integration with existing tooling

Installation

pip install tessera-hypernetwork

Quick Start

Generate LoRA Adapters

From metadata (JSON string or file):

tessera generate \
  --from-metadata '{"task": "classification", "domain": "medical"}' \
  --base-model mistralai/Mistral-7B-Instruct-v0.2 \
  --rank 16 \
  --save ./adapter.safetensors

From text description:

tessera generate \
  --from-text "Medical diagnosis assistant" \
  --base-model mistralai/Mistral-7B-Instruct-v0.2 \
  --rank 16 \
  --save ./adapter.safetensors

From document:

tessera generate \
  --from-doc ./document.txt \
  --base-model mistralai/Mistral-7B-Instruct-v0.2 \
  --rank 16 \
  --save ./adapter.safetensors

Base Model Management

Download a base model from HuggingFace Hub:

tessera model pull mistralai/Mistral-7B-Instruct-v0.2
tessera model pull meta-llama/Llama-3.1-8B-Instruct
tessera model pull deepseek-ai/DeepSeek-R1-Distill-Qwen-7B

Start vLLM with a base model:

tessera model serve-model mistralai/Mistral-7B-Instruct-v0.2 --port 8000
tessera model serve-model mistralai/Mistral-7B-Instruct-v0.2 --gpu-memory-utilization 0.9
tessera model serve-model mistralai/Mistral-7B-Instruct-v0.2 --quantization awq

List cached base models:

tessera model list-models

Remove a cached model:

tessera model remove mistralai/Mistral-7B-Instruct-v0.2

Start Tessera Server

Start the hypernetwork server (with auto vLLM):

tessera serve --port 8080 --base-model mistralai/Mistral-7B-Instruct-v0.2

Start the hypernetwork server (standalone):

tessera serve --port 8080 --host 0.0.0.0

Check Server Health

tessera health --url http://localhost:8080

List Available Models

tessera list

Commands

Generate

Generate LoRA adapters from metadata, text, or documents.

Options:

• --from-metadata: JSON metadata string or file path
• --from-text: Natural language description
• --from-doc: Document content or file path
• --base-model: Base model identifier (default: mistralai/Mistral-7B-Instruct-v0.2)
• --rank: LoRA rank (default: 16)
• --save: Output path for safetensors file (required)
• --mode: Generation mode: doc, metadata, or text (auto-inferred if not specified)

Model Management

Manage base models for vLLM serving.

tessera model pull <model_id> Download a base model from HuggingFace Hub and cache locally.

tessera model serve-model <model_id> Start vLLM with a specified base model.

Options:

• --port: Port to serve on (default: 8000)
• --gpu-memory-utilization: GPU memory utilization fraction (e.g., 0.9)
• --tensor-parallel-size: Tensor parallel size (default: 1)
• --quantization: Quantization method (e.g., awq, gptq, bitsandbytes)
• --max-model-len: Maximum model length (default: 8192)

tessera model list-models List all locally cached base models.

tessera model remove <model_id> Remove a cached base model to free disk space.

Serve

Start the Tessera hypernetwork server.

Options:

• --port: Port to serve on (default: 8080)
• --host: Host to bind to (default: 0.0.0.0)
• --qdrant-url: Qdrant vector database URL (optional)
• --workers: Number of worker processes (default: 1)
• --base-model: Base model to auto-start vLLM with (e.g., mistralai/Mistral-7B-Instruct-v0.2)
• --vllm-port: Port for vLLM server (default: 8000)

Health

Check server health status.

Options:

• --url: Server URL (default: http://localhost:8080)

List

List available base models and their dimensions, plus cached models.

LoRAX Adapter Management

Import, list, and unload adapters:

Import an adapter:

tessera lorax import-adapter \
  --path ./adapter.safetensors \
  --name my-adapter \
  --base-model mistralai/Mistral-7B-Instruct-v0.2 \
  --server-url http://localhost:8080

List loaded adapters:

tessera lorax list-adapters --server-url http://localhost:8080

Unload an adapter:

tessera lorax unload --name my-adapter --server-url http://localhost:8080

API Endpoints

The hypernetwork service provides a FastAPI server with the following endpoints:

• POST /v1/generate - Generate a LoRA adapter for a given prompt
• GET /health - Health check endpoint
• POST /v1/adapters - Import adapter safetensors
• GET /v1/adapters - List loaded adapters
• DELETE /v1/adapters/{name} - Unload adapter

License

Apache-2.0