How paradigma-inc/limite-violetto works
Limite Violetto is a vLLM plugin that serves Limite mathematical-language-model checkpoints. The principal workflow is a client request to vLLM, plugin registration of the checkpoint configuration and model architecture, validated model construction and weight loading, then autoregressive inference and a response. The graph preserves the model’s attention, feed-forward, and residual-mixing stages. Source excerpts are partial; relationships not shown in excerpts or documented in the README are omitted.
Main entry points: modeling.py Causal model · register.py Plugin registration · configuration.py Checkpoint configuration · validation.py Runtime validation
Components 16
Plugin integration
- Plugin registration
src/limite_vllm/register.py
Checkpoint contract
- Model identities
src/limite_vllm/contract.py - Checkpoint configuration
src/limite_vllm/configuration.py - Runtime validation
src/limite_vllm/validation.py
Model execution
- Causal model
src/limite_vllm/modeling.py - Decoder layers
src/limite_vllm/modeling.py - Token embeddings
src/limite_vllm/modeling.py - Logit head
src/limite_vllm/modeling.py - Weight loading
src/limite_vllm/modeling.py
Model features
- Attention
src/limite_vllm/_components.py - Feed-forward MLP
src/limite_vllm/_components.py - Residual mixing
src/limite_vllm/_components.py - Rotary embeddings
src/limite_vllm/_components.py
Other components
- Serving client
- vLLM serving
- HF checkpoint
Connections 19
- Serving client → vLLM serving: sends request
README.md - vLLM serving → Plugin registration: loads plugin
README.md - Plugin registration → Model identities: uses identities
src/limite_vllm/register.py - Plugin registration → Checkpoint configuration: registers config
src/limite_vllm/register.py - Plugin registration → Causal model: registers model
src/limite_vllm/register.py - HF checkpoint → vLLM serving: provides model
README.md - vLLM serving → Checkpoint configuration: parses config
README.md - Causal model → Runtime validation: validates config
src/limite_vllm/modeling.py - Runtime validation → Checkpoint configuration: checks architecture
src/limite_vllm/validation.py - Causal model → Token embeddings: embeds tokens
src/limite_vllm/modeling.py - Causal model → Decoder layers: runs layers
src/limite_vllm/modeling.py - Decoder layers → Attention: uses attention
src/limite_vllm/modeling.py - Decoder layers → Feed-forward MLP: uses MLP
src/limite_vllm/modeling.py - Decoder layers → Residual mixing: mixes residuals
src/limite_vllm/modeling.py - Decoder layers → Rotary embeddings: applies rotary
src/limite_vllm/modeling.py - Causal model → Logit head: computes logits
src/limite_vllm/modeling.py - HF checkpoint → Weight loading: supplies weights
src/limite_vllm/modeling.py - Weight loading → Causal model: loads parameters
src/limite_vllm/modeling.py - vLLM serving → Serving client: returns generation
README.md