Merge pull request #7 from spiceai/jeadie/25-04-15/upstream-spiceai

fix other things too

Author: Jeadie

candle-core/src/cpu_backend/mod.rs

@@ -3553,19 +3553,21 @@ impl BackendDevice for CpuDevice {
             }
             DType::BF16 => {
                 let mut data = Vec::with_capacity(elem_count);
-                let uniform = rand::distr::Uniform::new(bf16::from_f64(min), bf16::from_f64(max))
-                    .map_err(Error::wrap)?;
+                let normal: rand_distr::Uniform<f32> =
+                    rand_distr::Uniform::new(min as f32, max as f32).map_err(Error::wrap)?;
                 for _i in 0..elem_count {
-                    data.push(rng.sample::<bf16, _>(uniform))
+                    let sample: f32 = normal.sample(&mut rng);
+                    data.push(bf16::from_f32(sample));
                 }
                 Ok(CpuStorage::BF16(data))
             }
             DType::F16 => {
                 let mut data = Vec::with_capacity(elem_count);
-                let uniform = rand::distr::Uniform::new(f16::from_f64(min), f16::from_f64(max))
-                    .map_err(Error::wrap)?;
+                let normal: rand_distr::Uniform<f32> =
+                    rand_distr::Uniform::new(min as f32, max as f32).map_err(Error::wrap)?;
                 for _i in 0..elem_count {
-                    data.push(rng.sample::<f16, _>(uniform))
+                    let sample: f32 = normal.sample(&mut rng);
+                    data.push(f16::from_f32(sample));
                 }
                 Ok(CpuStorage::F16(data))
             }
@@ -3610,19 +3612,21 @@ impl BackendDevice for CpuDevice {
             }
             DType::BF16 => {
                 let mut data = Vec::with_capacity(elem_count);
-                let normal = rand_distr::Normal::new(bf16::from_f64(mean), bf16::from_f64(std))
-                    .map_err(Error::wrap)?;
+                let normal: rand_distr::Normal<f32> =
+                    rand_distr::Normal::new(mean as f32, std as f32).map_err(Error::wrap)?;
                 for _i in 0..elem_count {
-                    data.push(normal.sample(&mut rng))
+                    let sample: f32 = normal.sample(&mut rng);
+                    data.push(bf16::from_f32(sample));
                 }
                 Ok(CpuStorage::BF16(data))
             }
             DType::F16 => {
                 let mut data = Vec::with_capacity(elem_count);
-                let normal = rand_distr::Normal::new(f16::from_f64(mean), f16::from_f64(std))
-                    .map_err(Error::wrap)?;
+                let normal: rand_distr::Normal<f32> =
+                    rand_distr::Normal::new(mean as f32, std as f32).map_err(Error::wrap)?;
                 for _i in 0..elem_count {
-                    data.push(normal.sample(&mut rng))
+                    let sample: f32 = normal.sample(&mut rng);
+                    data.push(f16::from_f32(sample));
                 }
                 Ok(CpuStorage::F16(data))
             }

candle-core/src/pickle.rs

@@ -792,7 +792,7 @@ impl PthTensors {
 /// # Arguments
 /// * `path` - Path to the pth file.
 /// * `key` - Optional key to retrieve `state_dict` from the pth file. Sometimes the pth file
-///             contains multiple objects and the `state_dict` is the one we are interested in.
+///   contains multiple objects and the `state_dict` is the one we are interested in.
 pub fn read_all_with_key<P: AsRef<std::path::Path>>(
     path: P,
     key: Option<&str>,

candle-core/tests/quantized_tests.rs

@@ -6,7 +6,7 @@ use candle_core::{
     DType, Device, IndexOp, Module, Result, Tensor, Var,
 };
 use quantized::{k_quants, GgmlType};
-use rand::prelude::*;
+use rand::{prelude::*, random};
 
 const GGML_TEST_SIZE: usize = 32 * 128;
 
@@ -1110,13 +1110,11 @@ fn get_random_tensors(
     n: usize,
     device: &Device,
 ) -> Result<(Tensor, Tensor, Tensor)> {
-    let mut rng = StdRng::seed_from_u64(314159265358979);
-
     let lhs = (0..m * k)
-        .map(|_| rng.gen::<f32>() - 0.5)
+        .map(|_| random::<f32>() - 0.5)
         .collect::<Vec<_>>();
     let rhs = (0..n * k)
-        .map(|_| rng.gen::<f32>() - 0.5)
+        .map(|_| random::<f32>() - 0.5)
         .collect::<Vec<_>>();
 
     let lhs = Tensor::from_vec(lhs, (m, k), device)?;

candle-examples/examples/mamba-minimal/model.rs

@@ -21,7 +21,7 @@ impl Config {
     }
 
     fn dt_rank(&self) -> usize {
-        (self.d_model + 15) / 16
+        self.d_model.div_ceil(16)
     }
 
     fn d_conv(&self) -> usize {

candle-examples/examples/metavoice/main.rs

@@ -6,6 +6,7 @@ extern crate accelerate_src;
 
 use anyhow::Result;
 use clap::Parser;
+use rand::distr::Distribution;
 use std::io::Write;
 
 use candle_transformers::generation::LogitsProcessor;
@@ -16,7 +17,7 @@ use candle_transformers::models::quantized_metavoice::transformer as qtransforme
 use candle::{DType, IndexOp, Tensor};
 use candle_nn::VarBuilder;
 use hf_hub::api::sync::Api;
-use rand::{distributions::Distribution, SeedableRng};
+use rand::SeedableRng;
 
 pub const ENCODEC_NTOKENS: u32 = 1024;
 
@@ -250,7 +251,7 @@ fn main() -> Result<()> {
             let logits = logits.i(step)?.to_dtype(DType::F32)?;
             let logits = &(&logits / 1.0)?;
             let prs = candle_nn::ops::softmax_last_dim(logits)?.to_vec1::<f32>()?;
-            let distr = rand::distributions::WeightedIndex::new(prs.as_slice())?;
+            let distr = rand::distr::weighted::WeightedIndex::new(prs.as_slice())?;
             let sample = distr.sample(&mut rng) as u32;
             codes_.push(sample)
         }

candle-nn/src/loss.rs

@@ -7,7 +7,7 @@ use candle::{Result, Tensor};
 /// Arguments
 ///
 /// * [inp]: The input tensor of dimensions `N, C` where `N` is the batch size and `C` the number
-///          of categories. This is expected to contain log probabilities.
+///   of categories. This is expected to contain log probabilities.
 /// * [target]: The ground truth labels as a tensor of u32 of dimension `N`.
 ///
 /// The resulting tensor is a scalar containing the average value over the batch.
@@ -34,7 +34,7 @@ pub fn nll(inp: &Tensor, target: &Tensor) -> Result<Tensor> {
 /// Arguments
 ///
 /// * [inp]: The input tensor of dimensions `N, C` where `N` is the batch size and `C` the number
-///          of categories. This is expected to raw logits.
+///   of categories. This is expected to raw logits.
 /// * [target]: The ground truth labels as a tensor of u32 of dimension `N`.
 ///
 /// The resulting tensor is a scalar containing the average value over the batch.
@@ -56,9 +56,9 @@ pub fn mse(inp: &Tensor, target: &Tensor) -> Result<Tensor> {
 /// Arguments
 ///
 /// * [inp]: The input tensor of dimensions `N, C` where `N` is the batch size and `C` the number
-///          of categories. This is expected to raw logits.
+///   of categories. This is expected to raw logits.
 /// * [target]: The ground truth labels as a tensor of u32 of dimension `N, C` where `N` is the batch size and `C` the number
-///          of categories.
+///   of categories.
 ///
 /// The resulting tensor is a scalar containing the average value over the batch.
 pub fn binary_cross_entropy_with_logit(inp: &Tensor, target: &Tensor) -> Result<Tensor> {

candle-nn/src/var_builder.rs

@@ -423,7 +423,7 @@ impl SimpleBackend for candle::npy::NpzTensors {
     }
 
     fn contains_tensor(&self, name: &str) -> bool {
-        self.get(name).map_or(false, |v| v.is_some())
+        self.get(name).is_ok_and(|v| v.is_some())
     }
 }
 
@@ -461,7 +461,7 @@ impl SimpleBackend for candle::pickle::PthTensors {
     }
 
     fn contains_tensor(&self, name: &str) -> bool {
-        self.get(name).map_or(false, |v| v.is_some())
+        self.get(name).is_ok_and(|v| v.is_some())
     }
 }
 

candle-transformers/src/generation/mod.rs

@@ -1,5 +1,5 @@
 use candle::{DType, Error, Result, Tensor};
-use rand::{distributions::Distribution, SeedableRng};
+use rand::{distr::Distribution, SeedableRng};
 
 #[derive(Clone, PartialEq, Debug)]
 pub enum Sampling {
@@ -45,7 +45,7 @@ impl LogitsProcessor {
     }
 
     fn sample_multinomial(&mut self, prs: &Vec<f32>) -> Result<u32> {
-        let distr = rand::distributions::WeightedIndex::new(prs).map_err(Error::wrap)?;
+        let distr = rand::distr::weighted::WeightedIndex::new(prs).map_err(Error::wrap)?;
         let next_token = distr.sample(&mut self.rng) as u32;
         Ok(next_token)
     }

candle-transformers/src/models/dac.rs

@@ -99,7 +99,7 @@ impl EncoderBlock {
         let snake1 = Snake1d::new(dim / 2, vb.pp(3))?;
         let cfg1 = Conv1dConfig {
             stride,
-            padding: (stride + 1) / 2,
+            padding: stride.div_ceil(2),
             ..Default::default()
         };
         let conv1 = encodec::conv1d_weight_norm(dim / 2, dim, 2 * stride, cfg1, vb.pp(4))?;
@@ -191,7 +191,7 @@ impl DecoderBlock {
         let snake1 = Snake1d::new(in_dim, vb.pp(0))?;
         let cfg = ConvTranspose1dConfig {
             stride,
-            padding: (stride + 1) / 2,
+            padding: stride.div_ceil(2),
             ..Default::default()
         };
         let conv_tr1 = encodec::conv_transpose1d_weight_norm(

candle-transformers/src/models/flux/sampling.rs

@@ -6,8 +6,8 @@ pub fn get_noise(
     width: usize,
     device: &Device,
 ) -> Result<Tensor> {
-    let height = (height + 15) / 16 * 2;
-    let width = (width + 15) / 16 * 2;
+    let height = height.div_ceil(16) * 2;
+    let width = width.div_ceil(16) * 2;
     Tensor::randn(0f32, 1., (num_samples, 16, height, width), device)
 }
 
@@ -84,8 +84,8 @@ pub fn get_schedule(num_steps: usize, shift: Option<(usize, f64, f64)>) -> Vec<f
 
 pub fn unpack(xs: &Tensor, height: usize, width: usize) -> Result<Tensor> {
     let (b, _h_w, c_ph_pw) = xs.dims3()?;
-    let height = (height + 15) / 16;
-    let width = (width + 15) / 16;
+    let height = height.div_ceil(16);
+    let width = width.div_ceil(16);
     xs.reshape((b, height, width, c_ph_pw / 4, 2, 2))? // (b, h, w, c, ph, pw)
         .permute((0, 3, 1, 4, 2, 5))? // (b, c, h, ph, w, pw)
         .reshape((b, c_ph_pw / 4, height * 2, width * 2))