Server-Side WebAssembly by Danilo Chiarlone

Author:Danilo Chiarlone [Danilo Chiarlone]
Language: eng
Format: epub
ISBN: 9781633436206
Publisher: Manning Publications Co.
Published: 2025-08-05T00:00:00+00:00

Overall, the full code can be seen in listing 5.9.

Listing 5.9 lib.rs file from smartcms_ml_guest

#[allow(warnings)] mod bindings; use tokenizers::Tokenizer; use rand::prelude::*; use ndarray::Array; use std::fs; use bindings::{Guest, wasi::nn::{ graph::{Graph, load, ExecutionTarget, GraphEncoding}, tensor::{Tensor, TensorType}, }}; struct Component; impl Guest for Component { fn storygen() -> String { let model_data = fs::read("fixture [CA]/modified_model.onnx").unwrap(); #A let graph = load(&[model_data], [CA]GraphEncoding::Onnx, ExecutionTarget [CA]::Cpu).unwrap(); #A let exec_context = Graph::init_execution_ [CA]context(&graph).unwrap(); #A let tokenizer = [CA]Tokenizer::from_file("fixture/ [CA]tokenizer.json").unwrap(); #B let prompt = "Once upon a time"; #C let encoding = tokenizer. [CA]encode(prompt, true).unwrap(); #C let mut input_ids = [CA]encoding.get_ids().to_vec(); #C let mut rng = thread_rng(); #D let eos_token_id = tokenizer.token_ [CA]to_id("<EOS>").unwrap_or(0); #D for _ in 0..100 { #E let sequence_length = input_ids.len(); #F let dimensions = vec! [CA][1, sequence_length as u32]; #F let input_f32: Vec<f32> = input_ [CA]ids.iter().map(|&id| id as f32).collect(); #F let mut input_data = Vec:: [CA]with_capacity(input_f32.len() * 4); #F for &val in &input_f32 { #F input_data.extend_from_slice [CA](&val.to_ne_bytes()); #F } #F let tensor = Tensor::new [CA](&dimensions, TensorType::Fp32, &input_data); #F exec_context.set_input [CA]("input_ids", tensor).unwrap(); #F exec_context.compute().unwrap(); #G let output_data = exec_context. [CA]get_output("logits").unwrap().data(); #H let output_f32 = bytes_to_f32_vec [CA](&output_data); #H let vocab_size = output_f32. [CA]len() / sequence_length; #I let start = (sequence_length - 1) [CA]* vocab_size; #I let end = sequence_length * vocab_size; #I let last_token_logits = [CA]&output_f32[start..end]; #I let probabilities = softmax [CA](last_token_logits); #J let dist = rand::distributions:: [CA]WeightedIndex::new(&probabilities).unwrap(); #K let next_token = dist.sample(&mut rng); #K input_ids.push(next_token as u32); #K let generated_text = [CA]tokenizer.decode(&input_ids, true).unwrap(); #L if generated_text.ends_with("

") { #M break; #M } #M if next_token as u32 == eos_token_id { #M break; #M } #M } tokenizer.decode(&input_ids, true) [CA].unwrap().trim().to_string() #N } } pub fn bytes_to_f32_vec(data: &[u8]) -> Vec<f32> { #O data.chunks_exact(4) .map(|c| f32::from_le_bytes(c.try_into().unwrap())) .collect() } fn softmax(logits: &[f32]) -> Vec<f32> { #P let logits_array = Array::from_vec(logits.to_vec()); let exp_logits = logits_array.mapv(f32::exp); let sum_exp = exp_logits.sum(); (exp_logits / sum_exp).to_vec() } bindings::export!(Component with_types_in bindings);

#A Loading the ONNX model and setting up the graph and execution context

#B Loading the tokenizer from the provided file

#C Defining the prompt and encoding it into token IDs

#D Initializing the random number generator and setting the EOS token

#E Looping up to 100 iterations to generate tokens

#F Preparing and setting the input tensor

#G Running the model inference

#H Retrieving and converting the model’s output

#I Extracting the logits for the last token in the sequence

#J Applying softmax to calculate token probabilities

#K Sampling the next token based on probabilities and appending it to the sequence

#L Decoding the generated text

#M Breaking the loop if a natural end to the story is detected

#N Returning the final generated story

#O The function to convert bytes to f32 values

#P The softmax function to compute the probability distribution

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.