PART C · COMPLETE END-TO-END ARCHITECTURE
From training a model
to answering a question.
This page connects the two major phases of the LLM lifecycle. Training produces a model. Deployment makes it available. Inference uses it to process user input and generate output.
The full lifecycle
One model · Two operating phases
Training is model creation; inference is model usage.
PHASE 1 · PART A
Model Training Pipeline
How the model learns from data and how its parameters are updated.
Data collection
→Cleaning & transformation
→Training preparation
→Architecture & initialization
Pretraining
→Evaluation
→SFT & alignment
→Testing & optimization
→Deployment
Training output: A selected model version with learned weights and associated artifacts, including its tokenizer/configuration.
↓ packaged and made available by a serving system ↓
PHASE 2 · PART B
Model Inference Pipeline
How the deployed model processes a user request and generates an answer.
User input
→Tokenization
→Embeddings & positions
→Transformer computation
Attention & context
→Next-token probabilities
→Token generation
→Detokenize & deliver
Inference output: A generated response delivered to the user.
PART D · IMPORTANT DISTINCTION
Model training vs model inference
They use related model computations, but have different goals and execution behavior.
| Dimension | Model training | Model inference |
|---|---|---|
| Primary purpose | Learn patterns and capabilities from training data. | Use the learned model to respond to an input. |
| Data | Training examples, targets and validation data. | User prompts, instructions, context and multimodal inputs. |
| Weights | Updated by gradient-based optimization during training. | Normally fixed for the deployed model version. |
| Computation | Forward pass, loss, backpropagation, gradients and optimizer updates. | Forward computation through the network and output probability calculation. |
| Objective | Minimize a training objective and improve model behavior. | Select or sample output tokens according to a decoding strategy. |
| Typical mechanisms | Loss functions, optimizers, learning-rate schedules, batches and checkpoints. | Greedy decoding, temperature, top-k, top-p, caching and streaming. |
| When it happens | During pretraining and post-training / model development. | When an application sends an input to a deployed model. |
| Result | A trained model checkpoint / deployable model artifact. | A generated answer or other model output. |
Important technical note: Tokenization is used in both training and inference. The model learns its embedding and Transformer weights during training, then uses those learned parameters during inference. Inference normally does not update the deployed model's weights.
Follow the learning sequence
Continue into either pipeline
PART AModel Training
Explore the 14 stages from data preparation through deployment.
Open training pipeline →PART BModel Inference
Follow the interactive question-to-answer pipeline you already built.
Open inference lab →MODEL FOUNDATIONSArchitecture & representations
Revisit tokens, embeddings, attention and Transformer components.
Open detailed roadmap →