ci: update version string in docs

.github/workflows/pypi-gpu.yml

@@ -34,6 +34,7 @@ jobs:
         run: |
             mv pyproject.toml pyproject.toml.orig
             sed "s/ezkl/ezkl-gpu/" pyproject.toml.orig >pyproject.toml
+            sed "s/0\\.0\\.0/${RELEASE_TAG//v}/" pyproject.toml.orig >pyproject.toml
 
       - uses: actions-rs/toolchain@v1
         with:

.github/workflows/pypi.yml

@@ -233,6 +233,14 @@ jobs:
           python-version: 3.12
           architecture: x64
 
+      - name: Set pyproject.toml version to match github tag
+        shell: bash
+        env:
+          RELEASE_TAG: ${{ github.ref_name }}
+        run: |
+          mv pyproject.toml pyproject.toml.orig
+          sed "s/0\\.0\\.0/${RELEASE_TAG//v}/" pyproject.toml.orig >pyproject.toml
+
       - name: Set Cargo.toml version to match github tag
         shell: bash
         env:
@@ -242,7 +250,6 @@ jobs:
           sed "s/0\\.0\\.0/${RELEASE_TAG//v}/" Cargo.toml.orig >Cargo.toml
           mv Cargo.lock Cargo.lock.orig
           sed "s/0\\.0\\.0/${RELEASE_TAG//v}/" Cargo.lock.orig >Cargo.lock
-
       - name: Install required libraries
         shell: bash
         run: |

Cargo.toml

@@ -147,6 +147,10 @@ shellexpand = "3.1.0"
 runner = 'wasm-bindgen-test-runner'
 
 
+[[bench]]
+name = "zero_finder"
+harness = false
+
 [[bench]]
 name = "accum_dot"
 harness = false

benches/zero_finder.rs

@@ -0,0 +1,116 @@
+use std::thread;
+
+use criterion::{black_box, criterion_group, criterion_main, Criterion};
+use halo2curves::{bn256::Fr as F, ff::Field};
+use maybe_rayon::{
+    iter::{IndexedParallelIterator, IntoParallelRefIterator, ParallelIterator},
+    slice::ParallelSlice,
+};
+use rand::Rng;
+
+// Assuming these are your types
+#[derive(Clone)]
+enum ValType {
+    Constant(F),
+    AssignedConstant(usize, F),
+    Other,
+}
+
+// Helper to generate test data
+fn generate_test_data(size: usize, zero_probability: f64) -> Vec<ValType> {
+    let mut rng = rand::thread_rng();
+    (0..size)
+        .map(|_i| {
+            if rng.gen::<f64>() < zero_probability {
+                ValType::Constant(F::ZERO)
+            } else {
+                ValType::Constant(F::ONE) // Or some other non-zero value
+            }
+        })
+        .collect()
+}
+
+fn bench_zero_finding(c: &mut Criterion) {
+    let sizes = [
+        1_000,         // 1K
+        10_000,        // 10K
+        100_000,       // 100K
+        256 * 256 * 2, // Our specific case
+        1_000_000,     // 1M
+        10_000_000,    // 10M
+    ];
+
+    let zero_probability = 0.1; // 10% zeros
+
+    let mut group = c.benchmark_group("zero_finding");
+    group.sample_size(10); // Adjust based on your needs
+
+    for &size in &sizes {
+        let data = generate_test_data(size, zero_probability);
+
+        // Benchmark sequential version
+        group.bench_function(format!("sequential_{}", size), |b| {
+            b.iter(|| {
+                let result = data
+                    .iter()
+                    .enumerate()
+                    .filter_map(|(i, e)| match e {
+                        ValType::Constant(r) | ValType::AssignedConstant(_, r) => {
+                            (*r == F::ZERO).then_some(i)
+                        }
+                        _ => None,
+                    })
+                    .collect::<Vec<_>>();
+                black_box(result)
+            })
+        });
+
+        // Benchmark parallel version
+        group.bench_function(format!("parallel_{}", size), |b| {
+            b.iter(|| {
+                let result = data
+                    .par_iter()
+                    .enumerate()
+                    .filter_map(|(i, e)| match e {
+                        ValType::Constant(r) | ValType::AssignedConstant(_, r) => {
+                            (*r == F::ZERO).then_some(i)
+                        }
+                        _ => None,
+                    })
+                    .collect::<Vec<_>>();
+                black_box(result)
+            })
+        });
+
+        // Benchmark chunked parallel version
+        group.bench_function(format!("chunked_parallel_{}", size), |b| {
+            b.iter(|| {
+                let num_cores = thread::available_parallelism()
+                    .map(|n| n.get())
+                    .unwrap_or(1);
+                let chunk_size = (size / num_cores).max(100);
+
+                let result = data
+                    .par_chunks(chunk_size)
+                    .enumerate()
+                    .flat_map(|(chunk_idx, chunk)| {
+                        chunk
+                            .par_iter() // Make sure we use par_iter() here
+                            .enumerate()
+                            .filter_map(move |(i, e)| match e {
+                                ValType::Constant(r) | ValType::AssignedConstant(_, r) => {
+                                    (*r == F::ZERO).then_some(chunk_idx * chunk_size + i)
+                                }
+                                _ => None,
+                            })
+                    })
+                    .collect::<Vec<_>>();
+                black_box(result)
+            })
+        });
+    }
+    group.finish();
+}
+
+criterion_group!(benches, bench_zero_finding);
+criterion_main!(benches);

docs/python/src/conf.py

@@ -1,7 +1,7 @@
 import ezkl
 
 project = 'ezkl'
-release = '16.2.4'
+release = '16.2.8'
 version = release
 
 

pyproject.toml

@@ -12,6 +12,7 @@ asyncio_mode = "auto"
 
 [project]
 name = "ezkl"
+version = "0.0.0"
 requires-python = ">=3.7"
 classifiers = [
     "Programming Language :: Rust",

src/circuit/ops/layouts.rs

@@ -30,6 +30,8 @@ use crate::{
 use super::*;
 use crate::circuit::ops::lookup::LookupOp;
 
+const ASCII_ALPHABET: &str = "abcdefghijklmnopqrstuvwxyz";
+
 /// Calculate the L1 distance between two tensors.
 /// ```
 /// use ezkl::tensor::Tensor;
@@ -2671,9 +2673,7 @@ pub fn greater<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
     rhs.expand(&broadcasted_shape)?;
 
     let diff = pairwise(config, region, &[lhs, rhs], BaseOp::Sub)?;
-
     let sign = sign(config, region, &[diff])?;
-
     equals(config, region, &[sign, create_unit_tensor(1)])
 }
 
@@ -5286,75 +5286,72 @@ pub(crate) fn decompose<F: PrimeField + TensorType + PartialOrd + std::hash::Has
     base: &usize,
     n: &usize,
 ) -> Result<ValTensor<F>, CircuitError> {
-    let input = values[0].clone();
+    let mut input = values[0].clone();
 
     let is_assigned = !input.all_prev_assigned();
 
-    let bases: ValTensor<F> = Tensor::from(
-        (0..*n)
-            .rev()
-            .map(|x| ValType::Constant(integer_rep_to_felt(base.pow(x as u32) as IntegerRep))),
+    if !is_assigned {
+        input = region.assign(&config.custom_gates.inputs[0], &input)?;
+    }
+
+    let mut bases: ValTensor<F> = Tensor::from(
+        // repeat it input.len() times
+        (0..input.len()).flat_map(|_| {
+            (0..*n)
+                .rev()
+                .map(|x| ValType::Constant(integer_rep_to_felt(base.pow(x as u32) as IntegerRep)))
+        }),
     )
     .into();
+    let mut bases_dims = input.dims().to_vec();
+    bases_dims.push(*n);
+    bases.reshape(&bases_dims)?;
 
-    let cartesian_coord = input
-        .dims()
-        .iter()
-        .map(|x| 0..*x)
-        .multi_cartesian_product()
-        .collect::<Vec<_>>();
+    let mut decomposed_dims = input.dims().to_vec();
+    decomposed_dims.push(*n + 1);
 
-    let mut output: Tensor<Tensor<ValType<F>>> = Tensor::new(None, input.dims())?;
+    let claimed_output = if region.witness_gen() {
+        input.decompose(*base, *n)?
+    } else {
+        let decomposed_len = decomposed_dims.iter().product();
+        let claimed_output = Tensor::new(
+            Some(&vec![ValType::Value(Value::unknown()); decomposed_len]),
+            &decomposed_dims,
+        )?;
 
-    let inner_loop_function =
-        |i: usize, region: &mut RegionCtx<F>| -> Result<Tensor<ValType<F>>, CircuitError> {
-            let coord = cartesian_coord[i].clone();
-            let slice = coord.iter().map(|x| *x..*x + 1).collect::<Vec<_>>();
-            let mut sliced_input = input.get_slice(&slice)?;
-            sliced_input.flatten();
+        claimed_output.into()
+    };
+    region.assign(&config.custom_gates.output, &claimed_output)?;
+    region.increment(claimed_output.len());
 
-            if !is_assigned {
-                sliced_input = region.assign(&config.custom_gates.inputs[0], &sliced_input)?;
-            }
+    let input_slice = input.dims().iter().map(|x| 0..*x).collect::<Vec<_>>();
+    let mut sign_slice = input_slice.clone();
+    sign_slice.push(0..1);
+    let mut rest_slice = input_slice.clone();
+    rest_slice.push(1..n + 1);
 
-            let mut claimed_output_slice = if region.witness_gen() {
-                sliced_input.decompose(*base, *n)?
-            } else {
-                Tensor::from(vec![ValType::Value(Value::unknown()); *n + 1].into_iter()).into()
-            };
+    let sign = claimed_output.get_slice(&sign_slice)?;
+    let rest = claimed_output.get_slice(&rest_slice)?;
 
-            claimed_output_slice =
-                region.assign(&config.custom_gates.inputs[1], &claimed_output_slice)?;
-            claimed_output_slice.flatten();
+    let sign = range_check(config, region, &[sign], &(-1, 1))?;
+    let rest = range_check(config, region, &[rest], &(0, (*base - 1) as i128))?;
 
-            region.increment(claimed_output_slice.len());
+    // equation needs to be constructed as ij,ij->i but for arbitrary n dims we need to construct this dynamically
+    // indices should map in order of the alphabet
+    // start with lhs
+    let lhs = ASCII_ALPHABET.chars().take(rest.dims().len()).join("");
+    let rhs = ASCII_ALPHABET.chars().take(rest.dims().len() - 1).join("");
+    let equation = format!("{},{}->{}", lhs, lhs, rhs);
 
-            // get the sign bit and make sure it is valid
-            let sign = claimed_output_slice.first()?;
-            let sign = range_check(config, region, &[sign], &(-1, 1))?;
+    // now add the rhs
 
-            // get the rest of the thing and make sure it is in the correct range
-            let rest = claimed_output_slice.get_slice(&[1..claimed_output_slice.len()])?;
+    let prod_decomp = einsum(config, region, &[rest.clone(), bases], &equation)?;
 
-            let rest = range_check(config, region, &[rest], &(0, (base - 1) as i128))?;
+    let signed_decomp = pairwise(config, region, &[prod_decomp, sign], BaseOp::Mult)?;
 
-            let prod_decomp = dot(config, region, &[rest, bases.clone()])?;
+    enforce_equality(config, region, &[input, signed_decomp])?;
 
-            let signed_decomp = pairwise(config, region, &[prod_decomp, sign], BaseOp::Mult)?;
-
-            enforce_equality(config, region, &[sliced_input, signed_decomp])?;
-
-            Ok(claimed_output_slice.get_inner_tensor()?.clone())
-        };
-
-    region.apply_in_loop(&mut output, inner_loop_function)?;
-
-    let mut combined_output = output.combine()?;
-    let mut output_dims = input.dims().to_vec();
-    output_dims.push(*n + 1);
-    combined_output.reshape(&output_dims)?;
-
-    Ok(combined_output.into())
+    Ok(claimed_output)
 }
 
 pub(crate) fn sign<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(

src/eth.rs

@@ -488,7 +488,8 @@ pub async fn deploy_da_verifier_via_solidity(
         }
     }
 
-    let contract = match call_to_account {
+    
+    match call_to_account {
         Some(call) => {
             deploy_single_da_contract(
                 client,
@@ -514,8 +515,7 @@ pub async fn deploy_da_verifier_via_solidity(
             )
             .await
         }
-    };
-    return contract;
+    }
 }
 
 async fn deploy_multi_da_contract(
@@ -630,7 +630,7 @@ async fn deploy_single_da_contract(
             // bytes memory _callData,
             PackedSeqToken(call_data.as_ref()),
             // uint256 _decimals,
-            WordToken(B256::from(decimals).into()),
+            WordToken(B256::from(decimals)),
             // uint[] memory _scales,
             DynSeqToken(
                 scales

src/graph/utilities.rs

@@ -142,8 +142,6 @@ use tract_onnx::prelude::SymbolValues;
 pub fn extract_tensor_value(
     input: Arc<tract_onnx::prelude::Tensor>,
 ) -> Result<Tensor<f32>, GraphError> {
-    use maybe_rayon::prelude::{IntoParallelRefIterator, ParallelIterator};
-
     let dt = input.datum_type();
     let dims = input.shape().to_vec();
 
@@ -156,7 +154,7 @@ pub fn extract_tensor_value(
     match dt {
         DatumType::F16 => {
             let vec = input.as_slice::<tract_onnx::prelude::f16>()?.to_vec();
-            let cast: Vec<f32> = vec.par_iter().map(|x| (*x).into()).collect();
+            let cast: Vec<f32> = vec.iter().map(|x| (*x).into()).collect();
             const_value = Tensor::<f32>::new(Some(&cast), &dims)?;
         }
         DatumType::F32 => {
@@ -165,61 +163,61 @@ pub fn extract_tensor_value(
         }
         DatumType::F64 => {
             let vec = input.as_slice::<f64>()?.to_vec();
-            let cast: Vec<f32> = vec.par_iter().map(|x| *x as f32).collect();
+            let cast: Vec<f32> = vec.iter().map(|x| *x as f32).collect();
             const_value = Tensor::<f32>::new(Some(&cast), &dims)?;
         }
         DatumType::I64 => {
             // Generally a shape or hyperparam
             let vec = input.as_slice::<i64>()?.to_vec();
-            let cast: Vec<f32> = vec.par_iter().map(|x| *x as f32).collect();
+            let cast: Vec<f32> = vec.iter().map(|x| *x as f32).collect();
             const_value = Tensor::<f32>::new(Some(&cast), &dims)?;
         }
         DatumType::I32 => {
             // Generally a shape or hyperparam
             let vec = input.as_slice::<i32>()?.to_vec();
-            let cast: Vec<f32> = vec.par_iter().map(|x| *x as f32).collect();
+            let cast: Vec<f32> = vec.iter().map(|x| *x as f32).collect();
             const_value = Tensor::<f32>::new(Some(&cast), &dims)?;
         }
         DatumType::I16 => {
             // Generally a shape or hyperparam
             let vec = input.as_slice::<i16>()?.to_vec();
-            let cast: Vec<f32> = vec.par_iter().map(|x| *x as f32).collect();
+            let cast: Vec<f32> = vec.iter().map(|x| *x as f32).collect();
             const_value = Tensor::<f32>::new(Some(&cast), &dims)?;
         }
         DatumType::I8 => {
             // Generally a shape or hyperparam
             let vec = input.as_slice::<i8>()?.to_vec();
-            let cast: Vec<f32> = vec.par_iter().map(|x| *x as f32).collect();
+            let cast: Vec<f32> = vec.iter().map(|x| *x as f32).collect();
             const_value = Tensor::<f32>::new(Some(&cast), &dims)?;
         }
         DatumType::U8 => {
             // Generally a shape or hyperparam
             let vec = input.as_slice::<u8>()?.to_vec();
-            let cast: Vec<f32> = vec.par_iter().map(|x| *x as f32).collect();
+            let cast: Vec<f32> = vec.iter().map(|x| *x as f32).collect();
             const_value = Tensor::<f32>::new(Some(&cast), &dims)?;
         }
         DatumType::U16 => {
             // Generally a shape or hyperparam
             let vec = input.as_slice::<u16>()?.to_vec();
-            let cast: Vec<f32> = vec.par_iter().map(|x| *x as f32).collect();
+            let cast: Vec<f32> = vec.iter().map(|x| *x as f32).collect();
             const_value = Tensor::<f32>::new(Some(&cast), &dims)?;
         }
         DatumType::U32 => {
             // Generally a shape or hyperparam
             let vec = input.as_slice::<u32>()?.to_vec();
-            let cast: Vec<f32> = vec.par_iter().map(|x| *x as f32).collect();
+            let cast: Vec<f32> = vec.iter().map(|x| *x as f32).collect();
             const_value = Tensor::<f32>::new(Some(&cast), &dims)?;
         }
         DatumType::U64 => {
             // Generally a shape or hyperparam
             let vec = input.as_slice::<u64>()?.to_vec();
-            let cast: Vec<f32> = vec.par_iter().map(|x| *x as f32).collect();
+            let cast: Vec<f32> = vec.iter().map(|x| *x as f32).collect();
             const_value = Tensor::<f32>::new(Some(&cast), &dims)?;
         }
         DatumType::Bool => {
             // Generally a shape or hyperparam
             let vec = input.as_slice::<bool>()?.to_vec();
-            let cast: Vec<f32> = vec.par_iter().map(|x| *x as usize as f32).collect();
+            let cast: Vec<f32> = vec.iter().map(|x| *x as usize as f32).collect();
             const_value = Tensor::<f32>::new(Some(&cast), &dims)?;
         }
         DatumType::TDim => {
@@ -227,7 +225,7 @@ pub fn extract_tensor_value(
             let vec = input.as_slice::<tract_onnx::prelude::TDim>()?.to_vec();
 
             let cast: Result<Vec<f32>, GraphError> = vec
-                .par_iter()
+                .iter()
                 .map(|x| match x.to_i64() {
                     Ok(v) => Ok(v as f32),
                     Err(_) => match x.to_i64() {
@@ -1136,23 +1134,21 @@ pub fn new_op_from_onnx(
                         a: crate::circuit::utils::F32(exponent),
                     })
                 }
-            } else {
-                if let Some(c) = inputs[0].opkind().get_mutable_constant() {
-                    inputs[0].decrement_use();
-                    deleted_indices.push(0);
-                    if c.raw_values.len() > 1 {
-                        unimplemented!("only support scalar base")
-                    }
-
-                    let base = c.raw_values[0];
-
-                    SupportedOp::Nonlinear(LookupOp::Exp {
-                        scale: scale_to_multiplier(input_scales[1]).into(),
-                        base: base.into(),
-                    })
-                } else {
-                    unimplemented!("only support constant base or pow for now")
+            } else if let Some(c) = inputs[0].opkind().get_mutable_constant() {
+                inputs[0].decrement_use();
+                deleted_indices.push(0);
+                if c.raw_values.len() > 1 {
+                    unimplemented!("only support scalar base")
                 }
+
+                let base = c.raw_values[0];
+
+                SupportedOp::Nonlinear(LookupOp::Exp {
+                    scale: scale_to_multiplier(input_scales[1]).into(),
+                    base: base.into(),
+                })
+            } else {
+                unimplemented!("only support constant base or pow for now")
             }
         }
         "Div" => {

src/tensor/mod.rs

@@ -638,42 +638,44 @@ impl<T: Clone + TensorType> Tensor<T> {
     where
         T: Send + Sync,
     {
-        if indices.is_empty() {
+        // Fast path: empty indices or full tensor slice
+        if indices.is_empty()
+            || indices.iter().map(|x| x.end - x.start).collect::<Vec<_>>() == self.dims
+        {
             return Ok(self.clone());
         }
+
+        // Validate dimensions
         if self.dims.len() < indices.len() {
             return Err(TensorError::DimError(format!(
                 "The dimensionality of the slice {:?} is greater than the tensor's {:?}",
                 indices, self.dims
             )));
-        } else if indices.iter().map(|x| x.end - x.start).collect::<Vec<_>>() == self.dims {
-            // else if slice is the same as dims, return self
-            return Ok(self.clone());
         }
 
-        // if indices weren't specified we fill them in as required
-        let mut full_indices = indices.to_vec();
+        // Pre-allocate the full indices vector with capacity
+        let mut full_indices = Vec::with_capacity(self.dims.len());
+        full_indices.extend_from_slice(indices);
 
-        for i in 0..(self.dims.len() - indices.len()) {
-            full_indices.push(0..self.dims()[indices.len() + i])
-        }
+        // Fill remaining dimensions
+        full_indices.extend((indices.len()..self.dims.len()).map(|i| 0..self.dims[i]));
 
-        let cartesian_coord: Vec<Vec<usize>> = full_indices
+        // Pre-calculate total size and allocate result vector
+        let total_size: usize = full_indices
             .iter()
-            .cloned()
-            .multi_cartesian_product()
-            .collect();
-
-        let res: Vec<T> = cartesian_coord
-            .par_iter()
-            .map(|e| {
-                let index = self.get_index(e);
-                self[index].clone()
-            })
-            .collect();
+            .map(|range| range.end - range.start)
+            .product();
+        let mut res = Vec::with_capacity(total_size);
 
+        // Calculate new dimensions once
         let dims: Vec<usize> = full_indices.iter().map(|e| e.end - e.start).collect();
 
+        // Use iterator directly without collecting into intermediate Vec
+        for coord in full_indices.iter().cloned().multi_cartesian_product() {
+            let index = self.get_index(&coord);
+            res.push(self[index].clone());
+        }
+
         Tensor::new(Some(&res), &dims)
     }
 

src/tensor/val.rs

@@ -1,12 +1,12 @@
 use crate::{circuit::region::ConstantsMap, fieldutils::felt_to_integer_rep};
-use maybe_rayon::slice::Iter;
+use maybe_rayon::slice::{Iter, ParallelSlice};
 
 use super::{
     ops::{intercalate_values, pad, resize},
     *,
 };
 use halo2_proofs::{arithmetic::Field, circuit::Cell, plonk::Instance};
-use maybe_rayon::iter::{FilterMap, IntoParallelIterator, ParallelIterator};
+use maybe_rayon::iter::{FilterMap, ParallelIterator};
 
 pub(crate) fn create_constant_tensor<
     F: PrimeField + TensorType + std::marker::Send + std::marker::Sync + PartialOrd,
@@ -455,7 +455,7 @@ impl<F: PrimeField + TensorType + PartialOrd + std::hash::Hash> ValTensor<F> {
         }
     }
 
-    /// Returns the number of constants in the [ValTensor].
+    /// Returns an iterator over the [ValTensor]'s constants.
     pub fn create_constants_map_iterator(
         &self,
     ) -> FilterMap<Iter<'_, ValType<F>>, fn(&ValType<F>) -> Option<(F, ValType<F>)>> {
@@ -473,20 +473,48 @@ impl<F: PrimeField + TensorType + PartialOrd + std::hash::Hash> ValTensor<F> {
         }
     }
 
-    /// Returns the number of constants in the [ValTensor].
+    /// Returns a map of the constants in the [ValTensor].
     pub fn create_constants_map(&self) -> ConstantsMap<F> {
-        match self {
-            ValTensor::Value { inner, .. } => inner
-                .par_iter()
-                .filter_map(|x| {
-                    if let ValType::Constant(v) = x {
-                        Some((*v, x.clone()))
-                    } else {
-                        None
-                    }
-                })
-                .collect(),
-            ValTensor::Instance { .. } => ConstantsMap::new(),
+        let threshold = 1_000_000; // Tuned using the benchmarks
+
+        if self.len() < threshold {
+            match self {
+                ValTensor::Value { inner, .. } => inner
+                    .par_iter()
+                    .filter_map(|x| {
+                        if let ValType::Constant(v) = x {
+                            Some((*v, x.clone()))
+                        } else {
+                            None
+                        }
+                    })
+                    .collect(),
+                ValTensor::Instance { .. } => ConstantsMap::new(),
+            }
+        } else {
+            // Use parallel for larger arrays
+            let num_cores = std::thread::available_parallelism()
+                .map(|n| n.get())
+                .unwrap_or(1);
+            let chunk_size = (self.len() / num_cores).max(100_000);
+
+            match self {
+                ValTensor::Value { inner, .. } => inner
+                    .par_chunks(chunk_size)
+                    .flat_map(|chunk| {
+                        chunk
+                            .par_iter() // Make sure we use par_iter() here
+                            .filter_map(|x| {
+                                if let ValType::Constant(v) = x {
+                                    Some((*v, x.clone()))
+                                } else {
+                                    None
+                                }
+                            })
+                    })
+                    .collect(),
+                ValTensor::Instance { .. } => ConstantsMap::new(),
+            }
         }
     }
 
@@ -878,70 +906,161 @@ impl<F: PrimeField + TensorType + PartialOrd + std::hash::Hash> ValTensor<F> {
 
     /// remove constant zero values constants
     pub fn remove_const_zero_values(&mut self) {
-        match self {
-            ValTensor::Value { inner: v, dims, .. } => {
-                *v = v
-                    .clone()
-                    .into_par_iter()
-                    .filter_map(|e| {
-                        if let ValType::Constant(r) = e {
-                            if r == F::ZERO {
-                                return None;
+        let size_threshold = 1_000_000; // Tuned using the benchmarks
+
+        if self.len() < size_threshold {
+            match self {
+                ValTensor::Value { inner: v, dims, .. } => {
+                    *v = v
+                        .clone()
+                        .into_iter()
+                        .filter_map(|e| {
+                            if let ValType::Constant(r) = e {
+                                if r == F::ZERO {
+                                    return None;
+                                }
+                            } else if let ValType::AssignedConstant(_, r) = e {
+                                if r == F::ZERO {
+                                    return None;
+                                }
                             }
-                        } else if let ValType::AssignedConstant(_, r) = e {
-                            if r == F::ZERO {
-                                return None;
-                            }
-                        }
-                        Some(e)
-                    })
-                    .collect();
-                *dims = v.dims().to_vec();
+                            Some(e)
+                        })
+                        .collect();
+                    *dims = v.dims().to_vec();
+                }
+                ValTensor::Instance { .. } => {}
+            }
+        } else {
+            // Use parallel for larger arrays
+            let num_cores = std::thread::available_parallelism()
+                .map(|n| n.get())
+                .unwrap_or(1);
+            let chunk_size = (self.len() / num_cores).max(100_000);
+
+            match self {
+                ValTensor::Value { inner: v, dims, .. } => {
+                    *v = v
+                        .par_chunks_mut(chunk_size)
+                        .flat_map(|chunk| {
+                            chunk
+                                .par_iter_mut() // Make sure we use par_iter() here
+                                .filter_map(|e| {
+                                    if let ValType::Constant(r) = e {
+                                        if *r == F::ZERO {
+                                            return None;
+                                        }
+                                    } else if let ValType::AssignedConstant(_, r) = e {
+                                        if *r == F::ZERO {
+                                            return None;
+                                        }
+                                    }
+                                    Some(e.clone())
+                                })
+                        })
+                        .collect();
+                    *dims = v.dims().to_vec();
+                }
+                ValTensor::Instance { .. } => {}
             }
-            ValTensor::Instance { .. } => {}
         }
     }
 
-    /// gets constants
+    /// filter constant zero values constants
     pub fn get_const_zero_indices(&self) -> Vec<usize> {
-        match self {
-            ValTensor::Value { inner: v, .. } => v
-                .par_iter()
-                .enumerate()
-                .filter_map(|(i, e)| {
-                    if let ValType::Constant(r) = e {
-                        if *r == F::ZERO {
-                            return Some(i);
+        let size_threshold = 1_000_000; // Tuned using the benchmarks
+
+        if self.len() < size_threshold {
+            // Use single-threaded for smaller arrays
+            match &self {
+                ValTensor::Value { inner: v, .. } => v
+                    .iter()
+                    .enumerate()
+                    .filter_map(|(i, e)| {
+                        match e {
+                            // Combine both match arms to reduce branching
+                            ValType::Constant(r) | ValType::AssignedConstant(_, r) => {
+                                (*r == F::ZERO).then_some(i)
+                            }
+                            _ => None,
                         }
-                    } else if let ValType::AssignedConstant(_, r) = e {
-                        if *r == F::ZERO {
-                            return Some(i);
-                        }
-                    }
-                    None
-                })
-                .collect(),
-            ValTensor::Instance { .. } => vec![],
+                    })
+                    .collect(),
+                ValTensor::Instance { .. } => vec![],
+            }
+        } else {
+            // Use parallel for larger arrays
+            let num_cores = std::thread::available_parallelism()
+                .map(|n| n.get())
+                .unwrap_or(1);
+            let chunk_size = (self.len() / num_cores).max(100_000);
+
+            match &self {
+                ValTensor::Value { inner: v, .. } => v
+                    .par_chunks(chunk_size)
+                    .enumerate()
+                    .flat_map(|(chunk_idx, chunk)| {
+                        chunk
+                            .par_iter() // Make sure we use par_iter() here
+                            .enumerate()
+                            .filter_map(move |(i, e)| match e {
+                                ValType::Constant(r) | ValType::AssignedConstant(_, r) => {
+                                    (*r == F::ZERO).then_some(chunk_idx * chunk_size + i)
+                                }
+                                _ => None,
+                            })
+                    })
+                    .collect::<Vec<_>>(),
+                ValTensor::Instance { .. } => vec![],
+            }
         }
     }
 
-    /// gets constants
+    /// gets constant indices
     pub fn get_const_indices(&self) -> Vec<usize> {
-        match self {
-            ValTensor::Value { inner: v, .. } => v
-                .par_iter()
-                .enumerate()
-                .filter_map(|(i, e)| {
-                    if let ValType::Constant(_) = e {
-                        Some(i)
-                    } else if let ValType::AssignedConstant(_, _) = e {
-                        Some(i)
-                    } else {
-                        None
-                    }
-                })
-                .collect(),
-            ValTensor::Instance { .. } => vec![],
+        let size_threshold = 1_000_000; // Tuned using the benchmarks
+
+        if self.len() < size_threshold {
+            // Use single-threaded for smaller arrays
+            match &self {
+                ValTensor::Value { inner: v, .. } => v
+                    .iter()
+                    .enumerate()
+                    .filter_map(|(i, e)| {
+                        match e {
+                            // Combine both match arms to reduce branching
+                            ValType::Constant(_) | ValType::AssignedConstant(_, _) => Some(i),
+                            _ => None,
+                        }
+                    })
+                    .collect(),
+                ValTensor::Instance { .. } => vec![],
+            }
+        } else {
+            // Use parallel for larger arrays
+            let num_cores = std::thread::available_parallelism()
+                .map(|n| n.get())
+                .unwrap_or(1);
+            let chunk_size = (self.len() / num_cores).max(100_000);
+
+            match &self {
+                ValTensor::Value { inner: v, .. } => v
+                    .par_chunks(chunk_size)
+                    .enumerate()
+                    .flat_map(|(chunk_idx, chunk)| {
+                        chunk
+                            .par_iter() // Make sure we use par_iter() here
+                            .enumerate()
+                            .filter_map(move |(i, e)| match e {
+                                ValType::Constant(_) | ValType::AssignedConstant(_, _) => {
+                                    Some(chunk_idx * chunk_size + i)
+                                }
+                                _ => None,
+                            })
+                    })
+                    .collect::<Vec<_>>(),
+                ValTensor::Instance { .. } => vec![],
+            }
         }
     }