chore: bump rand 0.10 and rand_distr 0.6 with API migration

CHANGELOG.md

@@ -7,6 +7,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Changed — dependencies
+- Bumped `rand` 0.9 → 0.10.2 and `rand_distr` 0.5 → 0.6.0. `candle-core` still
+  pulls `rand` 0.9 / `rand_distr` 0.5 for itself; the crate now depends on both
+  major lines.
+- Migrated call sites for rand 0.10: `Rng` → `RngExt`, `StdRng::from_os_rng()` →
+  `rand::make_rng::<StdRng>()` (typed), and `&mut impl Rng` bounds on sampling /
+  training helpers to `RngExt`.
+
 ### Fixed — checkpointing
 - **Resume was implicit and could silently do nothing.** `train()` loaded any checkpoint it
   found with no flag and no way to opt out, restoring the step counter with it — so

Cargo.toml

@@ -28,8 +28,8 @@ ureq = { version = "3", features = ["json"] }
 gemm = { version = "0.19", features = ["f16"] }
 serde = { version = "1.0", features = ["derive"] }
 serde_yaml = "0.9"
-rand = "0.9"
-rand_distr = "0.5"
+rand = "0.10"
+rand_distr = "0.6"
 bytemuck = "1.25.2"
 half = "2.7"
 clap = { version = "4.6.4", features = ["derive"] }

src/generation/codegen_generate.rs

@@ -184,7 +184,7 @@ impl CodeGenGenerator {
         // same value at every step.
         let mut rng = match self.seed {
             Some(seed) => StdRng::seed_from_u64(seed),
-            None => StdRng::from_os_rng(),
+            None => rand::make_rng::<StdRng>(),
         };
 
         let mut cache: Option<Vec<KVCache>> = None;

src/generation/glm_generate.rs

@@ -4,7 +4,6 @@ use crate::generation::sampling::sample;
 use crate::glm::attention_mask::build_glm_mask;
 use crate::glm::model::GLMModel;
 use rand::rngs::StdRng;
-use rand::SeedableRng;
 
 pub struct GLMGenerator {
     model: GLMModel,
@@ -48,7 +47,7 @@ impl GLMGenerator {
     /// No blanks — generates new tokens after the prompt.
     pub fn generate(&self, prompt_token_ids: &[u32]) -> Result<Vec<u32>> {
         let device = self.model.embedding.weight.device();
-        let mut rng = StdRng::from_os_rng();
+        let mut rng: StdRng = rand::make_rng();
         let mut generated = prompt_token_ids.to_vec();
 
         for _step in 0..self.max_new_tokens {
@@ -89,7 +88,7 @@ impl GLMGenerator {
     /// - earlier positions within the same blank (causal within-span)
     pub fn fill_blanks(&self, context: &[u32], blank_lens: &[usize]) -> Result<Vec<u32>> {
         let device = self.model.embedding.weight.device();
-        let mut rng = StdRng::from_os_rng();
+        let mut rng: StdRng = rand::make_rng();
         let total_blank_tokens: usize = blank_lens.iter().sum();
 
         let mut all_tokens = context.to_vec();

src/generation/sampling.rs

@@ -1,5 +1,5 @@
 use candle_core::{DType, Result, Tensor};
-use rand::Rng;
+use rand::RngExt;
 
 /// Sample one token from `logits`.
 ///
@@ -12,7 +12,7 @@ pub fn sample(
     top_p: f64,
     repetition_penalty: f64,
     seen_tokens: &[u32],
-    rng: &mut impl Rng,
+    rng: &mut impl RngExt,
 ) -> Result<u32> {
     let logits = if logits.dtype() != DType::F32 {
         logits.to_dtype(DType::F32)?

src/training/data.rs

@@ -2,7 +2,7 @@ use crate::tokenizer::CodeGenTokenizer;
 use candle_core::Result;
 use rand::rngs::StdRng;
 use rand::seq::SliceRandom;
-use rand::Rng;
+use rand::RngExt;
 use rand::SeedableRng;
 use std::io::Read;
 use std::path::Path;

src/training/train.rs

@@ -7,7 +7,7 @@ use candle_core::{DType, Device, Result, Tensor, Var};
 use candle_nn::{AdamW, Optimizer, ParamsAdamW};
 use half;
 use rand::rngs::StdRng;
-use rand::{Rng, SeedableRng};
+use rand::{RngExt, SeedableRng};
 use safetensors::tensor::TensorView;
 use safetensors::SafeTensors;
 
@@ -40,7 +40,7 @@ pub struct GLMTrainer {
 /// Corrupt a sequence for masked denoising: some positions are replaced by the
 /// mask token or a random token, and their originals become the labels.
 /// Positions left alone get `-1`, which [`cross_entropy_loss`] ignores.
-fn corrupt(config: &GLMConfig, tokens: &[u32], rng: &mut impl Rng) -> (Vec<u32>, Vec<i64>) {
+fn corrupt(config: &GLMConfig, tokens: &[u32], rng: &mut impl RngExt) -> (Vec<u32>, Vec<i64>) {
     let mask_token_id = config.vocab_size as u32 - 1;
     let mut inputs = tokens.to_vec();
     let mut labels = vec![-1i64; tokens.len()];