optimization,profiling,performance
Before optimizing, I help you profile; then we focus on the real bottlenecks with concrete code or architectural changes.
9,967 technical terms and definitions
Before optimizing, I help you profile; then we focus on the real bottlenecks with concrete code or architectural changes.
To optimize latency, profile the stack: network, tokenizer, model, quantization, batching. Fix the slowest layer first (profiling before tuning).
Optimizers update weights using gradients. Adam adapts learning rate per parameter. SGD is simpler. Learning rate controls step size.
Options framework enables hierarchical RL through temporally extended actions with initiation termination and policy.
Temporal abstractions in RL.
Optuna is hyperparameter optimization framework. Efficient search.
Orca Mini applies Orca techniques to small models. Punch above weight.
Orca learns from GPT-4 reasoning traces. Microsoft research.
An orchestrator routes each request to the best model or tool (cheap vs. expensive, code vs. chat) and can chain multiple steps into a workflow.
Organic residues on surfaces.
Lower-cost organic substrate.
Organic interposers use polymer materials providing lower-cost alternative to silicon.
Organic-inorganic hybrid low-k.
General term for orientation mapping.
Constrain convolutions to be orthogonal.
Initialize with orthogonal matrices.
Third-party packaging and testing services.
Outsourced Semiconductor Assembly and Test companies provide backend services for packaged chips.
Large precipitates grow at expense of small.
Multi-modal instruction tuning.
Process exceeding control limits.
OOD detection identifies inputs unlike training data. Model should abstain or flag. Safety critical.
Indicators of special cause variation.
Out-of-distribution inputs differ from training data potentially causing failures.
Tool not meeting specifications.
Words not in the model's vocabulary (less common with subword tokenization).
Outbound logistics handles product distribution from factories to customers.
Release of trapped gases.
Find data points far from normal.
Outlier detection identifies devices with abnormal parametric values indicating process excursions or latent reliability risks.
Detect and handle outliers. Remove, clip, or transform.
Library for structured text generation and prompting.
Outlines guarantees structured LLM output. JSON schema, regex. Reliable parsing.
Extend image beyond boundaries.
Outpainting extends images beyond original boundaries generating consistent extrapolations.
Extend image beyond its original boundaries.
Output constraints limit response characteristics like length style or content.
Output filters block unsafe responses before delivery to users.
Filter model outputs with content classifiers. Block harmful content before returning to user.
Post-process model outputs to remove harmful sensitive or off-policy content.
Check generations before returning.
Extra etch time to ensure complete removal accounting for variations.
Over-processing waste performs unnecessary operations or exceeds requirements.
Decline benign requests unnecessarily.
Duplicate rare examples.
Over-travel is additional probe displacement beyond touchdown ensuring stable contact and compensating for non-planarity.
Composite yield from wafer to ship.
Overconfidence occurs when models express higher certainty than warranted by accuracy.
Additional etch to clear residues ensuring completeness.
Overfitting: model memorizes training data, fails on new data. Underfitting: model too simple. Goal is good generalization.