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: |

.github/workflows/swift-pm.yml

@@ -0,0 +1,129 @@
+name: Build and Publish EZKL iOS SPM package
+
+on:
+  push:
+    tags:
+      # Only support SemVer versioning tags
+      - 'v[0-9]+.[0-9]+.[0-9]+'
+      - '[0-9]+.[0-9]+.[0-9]+'
+
+jobs:
+  build-and-update:
+    runs-on: macos-latest
+    env:
+      EZKL_SWIFT_PACKAGE_REPO: github.com/zkonduit/ezkl-swift-package.git
+
+    steps:
+      - name: Checkout EZKL
+        uses: actions/checkout@v3
+
+      - name: Extract TAG from github.ref_name
+        run: |
+          # github.ref_name is provided by GitHub Actions and contains the tag name directly.
+          TAG="${{ github.ref_name }}"
+          echo "Original TAG: $TAG"
+          # Remove leading 'v' if present to match the Swift Package Manager version format.
+          NEW_TAG=${TAG#v}
+          echo "Stripped TAG: $NEW_TAG"
+          echo "TAG=$NEW_TAG" >> $GITHUB_ENV
+
+      - name: Install Rust (nightly)
+        uses: actions-rs/toolchain@v1
+        with:
+          toolchain: nightly
+          override: true
+
+      - name: Build EzklCoreBindings
+        run: CONFIGURATION=release cargo run --bin ios_gen_bindings --features "ios-bindings uuid camino uniffi_bindgen" --no-default-features
+
+      - name: Clone ezkl-swift-package repository
+        run: |
+          git clone https://${{ env.EZKL_SWIFT_PACKAGE_REPO }}
+
+      - name: Copy EzklCoreBindings
+        run: |
+          rm -rf ezkl-swift-package/Sources/EzklCoreBindings
+          cp -r build/EzklCoreBindings ezkl-swift-package/Sources/
+
+      - name: Copy Test Files
+        run: |
+          rm -rf ezkl-swift-package/Tests/EzklAssets/*
+          cp tests/assets/kzg ezkl-swift-package/Tests/EzklAssets/kzg.srs
+          cp tests/assets/input.json ezkl-swift-package/Tests/EzklAssets/input.json
+          cp tests/assets/model.compiled ezkl-swift-package/Tests/EzklAssets/network.ezkl
+          cp tests/assets/settings.json ezkl-swift-package/Tests/EzklAssets/settings.json
+
+      - name: Check for changes
+        id: check_changes
+        run: |
+          cd ezkl-swift-package
+          if git diff --quiet Sources/EzklCoreBindings Tests/EzklAssets; then
+            echo "no_changes=true" >> $GITHUB_OUTPUT
+          else
+            echo "no_changes=false" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Set up Xcode environment
+        if: steps.check_changes.outputs.no_changes == 'false'
+        run: |
+          sudo xcode-select -s /Applications/Xcode.app/Contents/Developer
+          sudo xcodebuild -license accept
+
+      - name: Run Package Tests
+        if: steps.check_changes.outputs.no_changes == 'false'
+        run: |
+          cd ezkl-swift-package
+          xcodebuild test \
+            -scheme EzklPackage \
+            -destination 'platform=iOS Simulator,name=iPhone 15 Pro,OS=17.5' \
+            -resultBundlePath ../testResults
+
+      - name: Run Example App Tests
+        if: steps.check_changes.outputs.no_changes == 'false'
+        run: |
+          cd ezkl-swift-package/Example
+          xcodebuild test \
+            -project Example.xcodeproj \
+            -scheme EzklApp \
+            -destination 'platform=iOS Simulator,name=iPhone 15 Pro,OS=17.5' \
+            -parallel-testing-enabled NO \
+            -resultBundlePath ../../exampleTestResults \
+            -skip-testing:EzklAppUITests/EzklAppUITests/testButtonClicksInOrder
+
+      - name: Setup Git
+        run: |
+          cd ezkl-swift-package
+          git config user.name "GitHub Action"
+          git config user.email "action@github.com"
+          git remote set-url origin https://zkonduit:${EZKL_SWIFT_PACKAGE_REPO_TOKEN}@${{ env.EZKL_SWIFT_PACKAGE_REPO }}
+        env:
+          EZKL_SWIFT_PACKAGE_REPO_TOKEN: ${{ secrets.EZKL_PORTER_TOKEN }}
+
+      - name: Commit and Push Changes
+        if: steps.check_changes.outputs.no_changes == 'false'
+        run: |
+          cd ezkl-swift-package
+          git add Sources/EzklCoreBindings Tests/EzklAssets
+          git commit -m "Automatically updated EzklCoreBindings for EZKL"
+          
+          if ! git push origin; then
+            echo "::error::Failed to push changes to ${{ env.EZKL_SWIFT_PACKAGE_REPO }}. Please ensure that EZKL_PORTER_TOKEN has the correct permissions."
+            exit 1
+          fi
+
+      - name: Tag the latest commit
+        run: |
+          cd ezkl-swift-package
+          source $GITHUB_ENV
+          
+          # Tag the latest commit on the current branch
+          if git rev-parse "$TAG" >/dev/null 2>&1; then
+            echo "Tag $TAG already exists locally. Skipping tag creation."
+          else
+            git tag "$TAG"
+          fi
+
+          if ! git push origin "$TAG"; then
+            echo "::error::Failed to push tag '$TAG' to ${{ env.EZKL_SWIFT_PACKAGE_REPO }}. Please ensure EZKL_PORTER_TOKEN has correct permissions."
+            exit 1
+          fi

.github/workflows/update-ios-package.yml

@@ -1,85 +0,0 @@
-name: Build and Publish EZKL iOS SPM package
-
-on:
-  workflow_dispatch:
-    inputs:
-      tag:
-        description: "The tag to release"
-        required: true
-  push:
-    tags:
-      - "*"
-
-jobs:
-  build-and-update:
-    runs-on: macos-latest
-
-    steps:
-      - name: Checkout EZKL
-        uses: actions/checkout@v3
-
-      - name: Install Rust
-        uses: actions-rs/toolchain@v1
-        with:
-          toolchain: nightly
-          override: true
-
-      - name: Build EzklCoreBindings
-        run: CONFIGURATION=release cargo run --bin ios_gen_bindings --features "ios-bindings uuid camino uniffi_bindgen" --no-default-features
-
-      - name: Clone ezkl-swift-package repository
-        run: |
-          git clone https://github.com/zkonduit/ezkl-swift-package.git
-
-      - name: Copy EzklCoreBindings
-        run: |
-          rm -rf ezkl-swift-package/Sources/EzklCoreBindings
-          cp -r build/EzklCoreBindings ezkl-swift-package/Sources/
-
-      - name: Copy Test Files
-        run: |
-          rm -rf ezkl-swift-package/Tests/EzklAssets/*
-          
-          cp tests/assets/kzg ezkl-swift-package/Tests/EzklAssets/kzg.srs
-          cp tests/assets/input.json ezkl-swift-package/Tests/EzklAssets/input.json
-          cp tests/assets/model.compiled ezkl-swift-package/Tests/EzklAssets/network.ezkl
-          cp tests/assets/settings.json ezkl-swift-package/Tests/EzklAssets/settings.json
-
-      - name: Set up Xcode environment
-        run: |
-          sudo xcode-select -s /Applications/Xcode.app/Contents/Developer
-          sudo xcodebuild -license accept
-
-      - name: Run Package Tests
-        run: |
-          cd ezkl-swift-package
-          xcodebuild test \
-            -scheme EzklPackage \
-            -destination 'platform=iOS Simulator,name=iPhone 15 Pro,OS=17.5' \
-            -resultBundlePath ../testResults
-
-      - name: Run Example App Tests
-        run: |
-          cd ezkl-swift-package/Example
-          xcodebuild test \
-            -project Example.xcodeproj \
-            -scheme EzklApp \
-            -destination 'platform=iOS Simulator,name=iPhone 15 Pro,OS=17.5' \
-            -parallel-testing-enabled NO \
-            -resultBundlePath ../../exampleTestResults \
-            -skip-testing:EzklAppUITests/EzklAppUITests/testButtonClicksInOrder
-
-      - name: Commit and Push Changes to feat/ezkl-direct-integration
-        run: |
-          cd ezkl-swift-package
-          git config user.name "GitHub Action"
-          git config user.email "action@github.com"
-          git add Sources/EzklCoreBindings
-          git add Tests/EzklAssets
-          git commit -m "Automatically updated EzklCoreBindings for EZKL"
-          git tag ${{ github.event.inputs.tag }}
-          git remote set-url origin https://zkonduit:${EZKL_PORTER_TOKEN}@github.com/zkonduit/ezkl-swift-package.git
-          git push origin
-          git push origin tag ${{ github.event.inputs.tag }}
-        env:
-          EZKL_PORTER_TOKEN: ${{ secrets.EZKL_PORTER_TOKEN }}

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
@@ -286,3 +290,11 @@ rustflags = ["-C", "relocation-model=pic"]
 lto = "fat"
 codegen-units = 1
 # panic = "abort"
+
+
+[package.metadata.wasm-pack.profile.release]
+wasm-opt = [
+    "-O4",
+    "--flexible-inline-max-function-size",
+    "4294967295",
+]

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 = '16.2.1'
+release = '16.2.9'
 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/bindings/universal.rs

@@ -141,10 +141,11 @@ pub(crate) fn gen_vk(
     .map_err(|e| EZKLError::InternalError(format!("Failed to create verifying key: {}", e)))?;
 
     let mut serialized_vk = Vec::new();
-    vk.write(&mut serialized_vk, halo2_proofs::SerdeFormat::RawBytes)
-        .map_err(|e| {
-            EZKLError::InternalError(format!("Failed to serialize verifying key: {}", e))
-        })?;
+    vk.write(
+        &mut serialized_vk,
+        halo2_proofs::SerdeFormat::RawBytesUnchecked,
+    )
+    .map_err(|e| EZKLError::InternalError(format!("Failed to serialize verifying key: {}", e)))?;
 
     Ok(serialized_vk)
 }
@@ -165,7 +166,7 @@ pub(crate) fn gen_pk(
     let mut reader = BufReader::new(&vk[..]);
     let vk = VerifyingKey::<G1Affine>::read::<_, GraphCircuit>(
         &mut reader,
-        halo2_proofs::SerdeFormat::RawBytes,
+        halo2_proofs::SerdeFormat::RawBytesUnchecked,
         circuit.settings().clone(),
     )
     .map_err(|e| EZKLError::InternalError(format!("Failed to deserialize verifying key: {}", e)))?;
@@ -197,7 +198,7 @@ pub(crate) fn verify(
     let mut reader = BufReader::new(&vk[..]);
     let vk = VerifyingKey::<G1Affine>::read::<_, GraphCircuit>(
         &mut reader,
-        halo2_proofs::SerdeFormat::RawBytes,
+        halo2_proofs::SerdeFormat::RawBytesUnchecked,
         circuit_settings.clone(),
     )
     .map_err(|e| EZKLError::InternalError(format!("Failed to deserialize vk: {}", e)))?;
@@ -277,7 +278,7 @@ pub(crate) fn verify_aggr(
     let mut reader = BufReader::new(&vk[..]);
     let vk = VerifyingKey::<G1Affine>::read::<_, AggregationCircuit>(
         &mut reader,
-        halo2_proofs::SerdeFormat::RawBytes,
+        halo2_proofs::SerdeFormat::RawBytesUnchecked,
         (),
     )
     .map_err(|e| EZKLError::InternalError(format!("Failed to deserialize vk: {}", e)))?;
@@ -365,7 +366,7 @@ pub(crate) fn prove(
     let mut reader = BufReader::new(&pk[..]);
     let pk = ProvingKey::<G1Affine>::read::<_, GraphCircuit>(
         &mut reader,
-        halo2_proofs::SerdeFormat::RawBytes,
+        halo2_proofs::SerdeFormat::RawBytesUnchecked,
         circuit.settings().clone(),
     )
     .map_err(|e| EZKLError::InternalError(format!("Failed to deserialize proving key: {}", e)))?;
@@ -487,7 +488,7 @@ pub(crate) fn vk_validation(vk: Vec<u8>, settings: Vec<u8>) -> Result<bool, EZKL
     let mut reader = BufReader::new(&vk[..]);
     let _ = VerifyingKey::<G1Affine>::read::<_, GraphCircuit>(
         &mut reader,
-        halo2_proofs::SerdeFormat::RawBytes,
+        halo2_proofs::SerdeFormat::RawBytesUnchecked,
         circuit_settings,
     )
     .map_err(|e| EZKLError::InternalError(format!("Failed to deserialize verifying key: {}", e)))?;
@@ -504,7 +505,7 @@ pub(crate) fn pk_validation(pk: Vec<u8>, settings: Vec<u8>) -> Result<bool, EZKL
     let mut reader = BufReader::new(&pk[..]);
     let _ = ProvingKey::<G1Affine>::read::<_, GraphCircuit>(
         &mut reader,
-        halo2_proofs::SerdeFormat::RawBytes,
+        halo2_proofs::SerdeFormat::RawBytesUnchecked,
         circuit_settings,
     )
     .map_err(|e| EZKLError::InternalError(format!("Failed to deserialize proving key: {}", e)))?;

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;
@@ -418,10 +420,6 @@ pub fn dot<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
     values[0].remove_indices(&mut removal_indices, true)?;
     values[1].remove_indices(&mut removal_indices, true)?;
 
-    let elapsed = global_start.elapsed();
-    trace!("filtering const zero indices took: {:?}", elapsed);
-
-    let start = instant::Instant::now();
     let mut inputs = vec![];
     let block_width = config.custom_gates.output.num_inner_cols();
 
@@ -429,37 +427,22 @@ pub fn dot<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
     for (i, input) in values.iter_mut().enumerate() {
         input.pad_to_zero_rem(block_width, ValType::Constant(F::ZERO))?;
         let inp = {
-            let (res, len) = region.assign_with_duplication(
-                &config.custom_gates.inputs[i],
-                input,
-                &config.check_mode,
-                false,
-            )?;
+            let (res, len) = region
+                .assign_with_duplication_unconstrained(&config.custom_gates.inputs[i], input)?;
             assigned_len = len;
             res.get_inner()?
         };
         inputs.push(inp);
     }
 
-    let elapsed = start.elapsed();
-    trace!("assigning inputs took: {:?}", elapsed);
-
     // Now we can assign the dot product
     // time this step
-    let start = instant::Instant::now();
     let accumulated_dot = accumulated::dot(&[inputs[0].clone(), inputs[1].clone()], block_width)?;
-    let elapsed = start.elapsed();
-    trace!("calculating accumulated dot took: {:?}", elapsed);
-
-    let start = instant::Instant::now();
-    let (output, output_assigned_len) = region.assign_with_duplication(
+    let (output, output_assigned_len) = region.assign_with_duplication_constrained(
         &config.custom_gates.output,
         &accumulated_dot.into(),
         &config.check_mode,
-        true,
     )?;
-    let elapsed = start.elapsed();
-    trace!("assigning output took: {:?}", elapsed);
 
     // enable the selectors
     if !region.is_dummy() {
@@ -1000,7 +983,6 @@ fn select<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
     region: &mut RegionCtx<F>,
     values: &[ValTensor<F>; 2],
 ) -> Result<ValTensor<F>, CircuitError> {
-    let start = instant::Instant::now();
     let (mut input, index) = (values[0].clone(), values[1].clone());
     input.flatten();
 
@@ -1028,9 +1010,6 @@ fn select<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
     let (_, assigned_output) =
         dynamic_lookup(config, region, &[index, output], &[dim_indices, input])?;
 
-    let end = start.elapsed();
-    trace!("select took: {:?}", end);
-
     Ok(assigned_output)
 }
 
@@ -1092,7 +1071,6 @@ pub(crate) fn dynamic_lookup<F: PrimeField + TensorType + PartialOrd + std::hash
     lookups: &[ValTensor<F>; 2],
     tables: &[ValTensor<F>; 2],
 ) -> Result<(ValTensor<F>, ValTensor<F>), CircuitError> {
-    let start = instant::Instant::now();
     // if not all lookups same length err
     if lookups[0].len() != lookups[1].len() {
         return Err(CircuitError::MismatchedLookupLength(
@@ -1126,28 +1104,20 @@ pub(crate) fn dynamic_lookup<F: PrimeField + TensorType + PartialOrd + std::hash
     }
     let table_len = table_0.len();
 
-    trace!("assigning tables took: {:?}", start.elapsed());
-
     // now create a vartensor of constants for the dynamic lookup index
     let table_index = create_constant_tensor(F::from(dynamic_lookup_index as u64), table_len);
     let _table_index =
         region.assign_dynamic_lookup(&config.dynamic_lookups.tables[2], &table_index)?;
 
-    trace!("assigning table index took: {:?}", start.elapsed());
-
     let lookup_0 = region.assign(&config.dynamic_lookups.inputs[0], &lookup_0)?;
     let lookup_1 = region.assign(&config.dynamic_lookups.inputs[1], &lookup_1)?;
     let lookup_len = lookup_0.len();
 
-    trace!("assigning lookups took: {:?}", start.elapsed());
-
     // now set the lookup index
     let lookup_index = create_constant_tensor(F::from(dynamic_lookup_index as u64), lookup_len);
 
     let _lookup_index = region.assign(&config.dynamic_lookups.inputs[2], &lookup_index)?;
 
-    trace!("assigning lookup index took: {:?}", start.elapsed());
-
     let mut lookup_block = 0;
 
     if !region.is_dummy() {
@@ -1194,9 +1164,6 @@ pub(crate) fn dynamic_lookup<F: PrimeField + TensorType + PartialOrd + std::hash
     region.increment_dynamic_lookup_index(1);
     region.increment(lookup_len);
 
-    let end = start.elapsed();
-    trace!("dynamic lookup took: {:?}", end);
-
     Ok((lookup_0, lookup_1))
 }
 
@@ -1441,7 +1408,6 @@ pub(crate) fn linearize_element_index<F: PrimeField + TensorType + PartialOrd +
     dim: usize,
     is_flat_index: bool,
 ) -> Result<ValTensor<F>, CircuitError> {
-    let start_time = instant::Instant::now();
     let index = values[0].clone();
     if !is_flat_index {
         assert_eq!(index.dims().len(), dims.len());
@@ -1515,9 +1481,6 @@ pub(crate) fn linearize_element_index<F: PrimeField + TensorType + PartialOrd +
 
     region.apply_in_loop(&mut output, inner_loop_function)?;
 
-    let elapsed = start_time.elapsed();
-    trace!("linearize_element_index took: {:?}", elapsed);
-
     Ok(output.into())
 }
 
@@ -1949,16 +1912,11 @@ pub fn sum<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
 
     region.flush()?;
     // time this entire function run
-    let global_start = instant::Instant::now();
-
     let mut values = values.clone();
 
     // this section has been optimized to death, don't mess with it
     values[0].remove_const_zero_values();
 
-    let elapsed = global_start.elapsed();
-    trace!("filtering const zero indices took: {:?}", elapsed);
-
     // if empty return a const
     if values[0].is_empty() {
         return Ok(create_zero_tensor(1));
@@ -1970,12 +1928,8 @@ pub fn sum<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
     let input = {
         let mut input = values[0].clone();
         input.pad_to_zero_rem(block_width, ValType::Constant(F::ZERO))?;
-        let (res, len) = region.assign_with_duplication(
-            &config.custom_gates.inputs[1],
-            &input,
-            &config.check_mode,
-            false,
-        )?;
+        let (res, len) =
+            region.assign_with_duplication_unconstrained(&config.custom_gates.inputs[1], &input)?;
         assigned_len = len;
         res.get_inner()?
     };
@@ -1983,11 +1937,10 @@ pub fn sum<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
     // Now we can assign the dot product
     let accumulated_sum = accumulated::sum(&input, block_width)?;
 
-    let (output, output_assigned_len) = region.assign_with_duplication(
+    let (output, output_assigned_len) = region.assign_with_duplication_constrained(
         &config.custom_gates.output,
         &accumulated_sum.into(),
         &config.check_mode,
-        true,
     )?;
 
     // enable the selectors
@@ -2053,13 +2006,10 @@ pub fn prod<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
 ) -> Result<ValTensor<F>, CircuitError> {
     region.flush()?;
     // time this entire function run
-    let global_start = instant::Instant::now();
 
     // this section has been optimized to death, don't mess with it
     let removal_indices = values[0].get_const_zero_indices();
 
-    let elapsed = global_start.elapsed();
-    trace!("finding const zero indices took: {:?}", elapsed);
     // if empty return a const
     if !removal_indices.is_empty() {
         return Ok(create_zero_tensor(1));
@@ -2070,12 +2020,8 @@ pub fn prod<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
     let input = {
         let mut input = values[0].clone();
         input.pad_to_zero_rem(block_width, ValType::Constant(F::ONE))?;
-        let (res, len) = region.assign_with_duplication(
-            &config.custom_gates.inputs[1],
-            &input,
-            &config.check_mode,
-            false,
-        )?;
+        let (res, len) =
+            region.assign_with_duplication_unconstrained(&config.custom_gates.inputs[1], &input)?;
         assigned_len = len;
         res.get_inner()?
     };
@@ -2083,11 +2029,10 @@ pub fn prod<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
     // Now we can assign the dot product
     let accumulated_prod = accumulated::prod(&input, block_width)?;
 
-    let (output, output_assigned_len) = region.assign_with_duplication(
+    let (output, output_assigned_len) = region.assign_with_duplication_constrained(
         &config.custom_gates.output,
         &accumulated_prod.into(),
         &config.check_mode,
-        true,
     )?;
 
     // enable the selectors
@@ -2440,7 +2385,6 @@ pub(crate) fn pairwise<F: PrimeField + TensorType + PartialOrd + std::hash::Hash
     let orig_lhs = lhs.clone();
     let orig_rhs = rhs.clone();
 
-    let start = instant::Instant::now();
     let first_zero_indices = HashSet::from_iter(lhs.get_const_zero_indices());
     let second_zero_indices = HashSet::from_iter(rhs.get_const_zero_indices());
 
@@ -2455,7 +2399,6 @@ pub(crate) fn pairwise<F: PrimeField + TensorType + PartialOrd + std::hash::Hash
         BaseOp::Sub => second_zero_indices.clone(),
         _ => return Err(CircuitError::UnsupportedOp),
     };
-    trace!("setting up indices took {:?}", start.elapsed());
 
     if lhs.len() != rhs.len() {
         return Err(CircuitError::DimMismatch(format!(
@@ -2480,7 +2423,6 @@ pub(crate) fn pairwise<F: PrimeField + TensorType + PartialOrd + std::hash::Hash
 
     // Now we can assign the dot product
     // time the calc
-    let start = instant::Instant::now();
     let op_result = match op {
         BaseOp::Add => add(&inputs),
         BaseOp::Sub => sub(&inputs),
@@ -2491,20 +2433,13 @@ pub(crate) fn pairwise<F: PrimeField + TensorType + PartialOrd + std::hash::Hash
         error!("{}", e);
         halo2_proofs::plonk::Error::Synthesis
     })?;
-    trace!("pairwise {} calc took {:?}", op.as_str(), start.elapsed());
 
-    let start = instant::Instant::now();
     let assigned_len = op_result.len() - removal_indices.len();
     let mut output = region.assign_with_omissions(
         &config.custom_gates.output,
         &op_result.into(),
         &removal_indices,
     )?;
-    trace!(
-        "pairwise {} input assign took {:?}",
-        op.as_str(),
-        start.elapsed()
-    );
 
     // Enable the selectors
     if !region.is_dummy() {
@@ -2671,9 +2606,7 @@ 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 +5219,72 @@ 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/circuit/ops/region.rs

@@ -671,22 +671,17 @@ impl<'a, F: PrimeField + TensorType + PartialOrd + std::hash::Hash> RegionCtx<'a
     }
 
     /// Assign a valtensor to a vartensor with duplication
-    pub fn assign_with_duplication(
+    pub fn assign_with_duplication_unconstrained(
         &mut self,
         var: &VarTensor,
         values: &ValTensor<F>,
-        check_mode: &crate::circuit::CheckMode,
-        single_inner_col: bool,
     ) -> Result<(ValTensor<F>, usize), Error> {
         if let Some(region) = &self.region {
             // duplicates every nth element to adjust for column overflow
-            let (res, len) = var.assign_with_duplication(
+            let (res, len) = var.assign_with_duplication_unconstrained(
                 &mut region.borrow_mut(),
-                self.row,
                 self.linear_coord,
                 values,
-                check_mode,
-                single_inner_col,
                 &mut self.assigned_constants,
             )?;
             Ok((res, len))
@@ -695,7 +690,37 @@ impl<'a, F: PrimeField + TensorType + PartialOrd + std::hash::Hash> RegionCtx<'a
                 self.row,
                 self.linear_coord,
                 values,
-                single_inner_col,
+                false,
+                &mut self.assigned_constants,
+            )?;
+            Ok((values.clone(), len))
+        }
+    }
+
+    /// Assign a valtensor to a vartensor with duplication
+    pub fn assign_with_duplication_constrained(
+        &mut self,
+        var: &VarTensor,
+        values: &ValTensor<F>,
+        check_mode: &crate::circuit::CheckMode,
+    ) -> Result<(ValTensor<F>, usize), Error> {
+        if let Some(region) = &self.region {
+            // duplicates every nth element to adjust for column overflow
+            let (res, len) = var.assign_with_duplication_constrained(
+                &mut region.borrow_mut(),
+                self.row,
+                self.linear_coord,
+                values,
+                check_mode,
+                &mut self.assigned_constants,
+            )?;
+            Ok((res, len))
+        } else {
+            let (_, len) = var.dummy_assign_with_duplication(
+                self.row,
+                self.linear_coord,
+                values,
+                true,
                 &mut self.assigned_constants,
             )?;
             Ok((values.clone(), len))

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/model.rs

@@ -1226,6 +1226,7 @@ impl Model {
                 values.iter().map(|v| v.dims()).collect_vec()
             );
 
+            let start = instant::Instant::now();
             match &node {
                 NodeType::Node(n) => {
                     let res = if node.is_constant() && node.num_uses() == 1 {
@@ -1363,6 +1364,7 @@ impl Model {
                     results.insert(*idx, full_results);
                 }
             }
+            debug!("------------ layout of {} took {:?}", idx, start.elapsed());
         }
 
         // we do this so we can support multiple passes of the same model and have deterministic results (Non-assigned inputs etc... etc...)

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)
     }
 
@@ -831,7 +833,7 @@ impl<T: Clone + TensorType> Tensor<T> {
         num_repeats: usize,
         initial_offset: usize,
     ) -> Result<Tensor<T>, TensorError> {
-        let mut inner: Vec<T> = vec![];
+        let mut inner: Vec<T> = Vec::with_capacity(self.inner.len());
         let mut offset = initial_offset;
         for (i, elem) in self.inner.clone().into_iter().enumerate() {
             if (i + offset + 1) % n == 0 {
@@ -860,20 +862,22 @@ impl<T: Clone + TensorType> Tensor<T> {
         num_repeats: usize,
         initial_offset: usize,
     ) -> Result<Tensor<T>, TensorError> {
-        let mut inner: Vec<T> = vec![];
-        let mut indices_to_remove = std::collections::HashSet::new();
-        for i in 0..self.inner.len() {
-            if (i + initial_offset + 1) % n == 0 {
-                for j in 1..(1 + num_repeats) {
-                    indices_to_remove.insert(i + j);
-                }
-            }
-        }
+        // Pre-calculate capacity to avoid reallocations
+        let estimated_size = self.inner.len() - (self.inner.len() / n) * num_repeats;
+        let mut inner = Vec::with_capacity(estimated_size);
 
-        let old_inner = self.inner.clone();
-        for (i, elem) in old_inner.into_iter().enumerate() {
-            if !indices_to_remove.contains(&i) {
-                inner.push(elem.clone());
+        // Use iterator directly instead of creating intermediate collections
+        let mut i = 0;
+        while i < self.inner.len() {
+            // Add the current element
+            inner.push(self.inner[i].clone());
+
+            // If this is an nth position (accounting for offset)
+            if (i + initial_offset + 1) % n == 0 {
+                // Skip the next num_repeats elements
+                i += num_repeats + 1;
+            } else {
+                i += 1;
             }
         }
 

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![],
+            }
         }
     }
 

src/tensor/var.rs

@@ -494,16 +494,56 @@ impl VarTensor {
         }
     }
 
+    /// Assigns specific values (`ValTensor`) to the columns of the inner tensor but allows for column wrapping for accumulated operations.
+    pub fn assign_with_duplication_unconstrained<
+        F: PrimeField + TensorType + PartialOrd + std::hash::Hash,
+    >(
+        &self,
+        region: &mut Region<F>,
+        offset: usize,
+        values: &ValTensor<F>,
+        constants: &mut ConstantsMap<F>,
+    ) -> Result<(ValTensor<F>, usize), halo2_proofs::plonk::Error> {
+        match values {
+            ValTensor::Instance { .. } => unimplemented!("duplication is not supported on instance columns. increase K if you require more rows."),
+            ValTensor::Value { inner: v, dims , ..} => {
+
+                let duplication_freq = self.block_size();
+
+                let num_repeats = self.num_inner_cols();
+
+                let duplication_offset = offset;
+
+                // duplicates every nth element to adjust for column overflow
+                let v = v.duplicate_every_n(duplication_freq, num_repeats, duplication_offset).unwrap();
+                let mut res: ValTensor<F> = {
+                    v.enum_map(|coord, k| {
+                    let cell = self.assign_value(region, offset, k.clone(), coord, constants)?;
+                    Ok::<_, halo2_proofs::plonk::Error>(cell)
+
+                })?.into()};
+                let total_used_len = res.len();
+                res.remove_every_n(duplication_freq, num_repeats, duplication_offset).unwrap();
+
+                res.reshape(dims).unwrap();
+                res.set_scale(values.scale());
+
+                Ok((res, total_used_len))
+            }
+        }
+    }
+
     /// Assigns specific values (`ValTensor`) to the columns of the inner tensor but allows for column wrapping for accumulated operations.
     /// Duplication occurs by copying the last cell of the column to the first cell next column and creating a copy constraint between the two.
-    pub fn assign_with_duplication<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
+    pub fn assign_with_duplication_constrained<
+        F: PrimeField + TensorType + PartialOrd + std::hash::Hash,
+    >(
         &self,
         region: &mut Region<F>,
         row: usize,
         offset: usize,
         values: &ValTensor<F>,
         check_mode: &CheckMode,
-        single_inner_col: bool,
         constants: &mut ConstantsMap<F>,
     ) -> Result<(ValTensor<F>, usize), halo2_proofs::plonk::Error> {
         let mut prev_cell = None;
@@ -512,34 +552,16 @@ impl VarTensor {
             ValTensor::Instance { .. } => unimplemented!("duplication is not supported on instance columns. increase K if you require more rows."),
             ValTensor::Value { inner: v, dims , ..} => {
 
-                let duplication_freq = if single_inner_col {
-                    self.col_size()
-                } else {
-                    self.block_size()
-                };
-
-                let num_repeats = if single_inner_col {
-                    1
-                } else {
-                    self.num_inner_cols()
-                };
-
-                let duplication_offset = if single_inner_col {
-                    row
-                } else {
-                    offset
-                };
+                let duplication_freq = self.col_size();
+                let num_repeats = 1;
+                let duplication_offset = row;
 
                 // duplicates every nth element to adjust for column overflow
                 let v = v.duplicate_every_n(duplication_freq, num_repeats, duplication_offset).unwrap();
                 let mut res: ValTensor<F> = {
                     v.enum_map(|coord, k| {
 
-                    let step = if !single_inner_col {
-                        1
-                    } else {
-                        self.num_inner_cols()
-                    };
+                    let step = self.num_inner_cols();
 
                     let (x, y, z) = self.cartesian_coord(offset + coord * step);
                     if matches!(check_mode, CheckMode::SAFE) && coord > 0 && z == 0 && y == 0 {
@@ -549,11 +571,13 @@ impl VarTensor {
 
                     let cell = self.assign_value(region, offset, k.clone(), coord * step, constants)?;
 
-                    if single_inner_col {
-                    if z == 0 {
+                    let at_end_of_column = z == duplication_freq - 1;
+                    let at_beginning_of_column = z == 0;
+
+                    if at_end_of_column {
                         // if we are at the end of the column, we need to copy the cell to the next column
                         prev_cell = Some(cell.clone());
-                    } else if coord > 0 && z == 0 && single_inner_col {
+                    } else if coord > 0 && at_beginning_of_column  {
                         if let Some(prev_cell) = prev_cell.as_ref() {
                             let cell = cell.cell().ok_or({
                                 error!("Error getting cell: {:?}", (x,y));
@@ -563,10 +587,10 @@ impl VarTensor {
                                 halo2_proofs::plonk::Error::Synthesis})?;
                             region.constrain_equal(prev_cell,cell)?;
                         } else {
-                            error!("Error copy-constraining previous value: {:?}", (x,y));
+                            error!("Previous cell was not set");
                             return Err(halo2_proofs::plonk::Error::Synthesis);
                         }
-                    }}
+                    }
 
                     Ok(cell)
 
@@ -577,20 +601,6 @@ impl VarTensor {
                 res.reshape(dims).unwrap();
                 res.set_scale(values.scale());
 
-                if matches!(check_mode, CheckMode::SAFE) {
-                     // during key generation this will be 0 so we use this as a flag to check
-                     // TODO: this isn't very safe and would be better to get the phase directly
-                    let res_evals = res.int_evals().unwrap();
-                    let is_assigned = res_evals
-                    .iter()
-                    .all(|&x| x == 0);
-                    if !is_assigned {
-                        assert_eq!(
-                           values.int_evals().unwrap(),
-                           res_evals
-                    )};
-                }
-
                 Ok((res, total_used_len))
             }
         }