⚙️ Local engine
Move from a design on paper to real inference in Portuguese.
Track map
⚙️ Installation and first inference
Run real weights on your machine.
⚙️ Router, context, and processing
Avoid routing an input to the wrong model.
Detailed content
Installation and first inference
Run real weights on your machine.
What it is
The clone includes the SDK and the practical application; the weights are downloaded separately from Hugging Face.
Why learn
Cloning the Git doesn’t automatically download the checkpoint.
Key concepts
versioned code
Python isolation
trained weights
downloaded files
What it is
The app uses CPU by default, letting you experiment without a compatible GPU.
Why learn
Don’t copy the published 33 ms to describe your machine.
Key concepts
general execution
NVIDIA acceleration
alternate device
real versions
What it is
The adaptation preloads only multilingual, using the corresponding subdirectory from the model repository.
Why learn
Preloading all models may be suitable for mixed traffic, but increases resident memory.
Key concepts
preloading
model in memory
bundle variant
weights and intermediates
What it is
The triage command accepts an optional message and subject, runs the four questions, and prints JSON.
Why learn
The command must exit with an error when the input is invalid or the model fails.
Key concepts
command interface
process result
predictions
operational decision
What it is
The fork starts from the upstream commit recorded in the documentation.
Why learn
If a result changes, compare these identities first.
Key concepts
exact commit
weights version
library
repeat conditions
What it is
Start by identifying the failure layer: import, download, loading, tokenization, or the forward pass.
Why learn
The lab exposes failures in the API and does not invent a substitute classification.
Key concepts
find the layer
local intervention
short run
recognizable failure
Router, context, and processing
Avoid routing an input to the wrong model.
What it is
Router.route can select a checkpoint without executing the neural network.
Why learn
The current code should take precedence over simplified diagrams.
Key concepts
pre-selection
explicit indication
writing system
specialized scheme
What it is
Portuguese uses the Latin alphabet, like English, and very short messages provide little evidence for automatic detection.
Why learn
A simple rule based on reliable product information can beat a language heuristic.
Key concepts
reliable metadata
little evidence
approximation
scope
What it is
The internal prompt reserves space for the question and options; the rest fits the state.
Why learn
The SDK may truncate the state.
Key concepts
tokenizer unit
question and options
ticket content
input loss
What it is
Each alternative needs enough bullets and text to remain distinguishable.
Why learn
One alternative is hierarchy: pick a family first, then a subcategory.
Key concepts
number of classes
selection stages
chained error
discriminative signal
What it is
The default Router keeps only a few hot models and can evict the least recently used one.
Why learn
Measurements must distinguish the hot path from the first load and the swaps.
Key concepts
eviction by use
loaded model
first load
serialized access
What it is
Laya groups questions from one state into a single call.
Why learn
Comparing milliseconds without declaring measurement boundaries leads to fragile conclusions.
Key concepts
batch of questions
median
latency tail
units per second