Features#
Getting
Access values
Setting
Intervene on values
Scan and Validate
Debug tensor shapes
Operations
Edit values
Modules
Apply modules
Custom Functions
Add thing to the Intervention Graph
Gradients
Intervene on gradients
Early Stopping
Save computation time
Conditional Interventions
Use If Needed
Cross Prompts
Edit in one pass
Generation
Generate multiple tokens
Model Editing
Add persistent interventions
Remote Execution
Use our servers
Sessions
Do many traces in one request
Streaming
Send remote values to local
Iterative Interventions
Make loops
LORA
Train one
vLLM Support
Fast inference
Report Issues#
NNsight and NDIF are open-source and you can report issues, read, and clone the full source at ndif-team/nnsight. Also check out https://discuss.ndif.us/ to ask questions about our features or suggest new ones.
- Conditional Interventions
- Cross-Prompt Intervention
- Custom Functions
- Early Stopping
- Editing
- Getting Values
- Gradients
- Iterative Interventions
- LoRA
- Model Editing
- Modules
- Multiple Token Generation
- Operations
- Remote Execution
- Remote Model Considerations & System Limits
- Scan and Validate
- Sessions
- Setting Values
- Streaming
- vLLM Support