diff --git a/Cargo.lock b/Cargo.lock
index 7c9503073..292c0b4d9 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1414,7 +1414,7 @@ dependencies = [
 [[package]]
 name = "miden-gpu"
 version = "0.2.0"
-source = "git+https://github.com/0xPolygonMiden/miden-gpu?rev=9a55000ff8cc113c1f4acb42115c542c1bb8533c#9a55000ff8cc113c1f4acb42115c542c1bb8533c"
+source = "git+https://github.com/0xPolygonMiden/miden-gpu?rev=b8b3a4273dd76a3af2ba8cb7739c9e20823a8f70#b8b3a4273dd76a3af2ba8cb7739c9e20823a8f70"
 dependencies = [
  "bytemuck",
  "elsa",
diff --git a/prover/Cargo.toml b/prover/Cargo.toml
index 96e859ab4..dbb944ea9 100644
--- a/prover/Cargo.toml
+++ b/prover/Cargo.toml
@@ -18,7 +18,7 @@ crate-type = ["cdylib", "rlib"]
 
 [features]
 concurrent = ["processor/concurrent", "std", "winter-prover/concurrent"]
-default = ["webgpu", "async", "std"]
+default = ["std"]
 metal = ["dep:miden-gpu", "dep:pollster", "concurrent", "std"]
 webgpu = ["dep:miden-gpu", "miden-gpu/webgpu", "async"]
 async = ["maybe-async/async", "winter-prover/async", "dep:async-trait"]
@@ -34,11 +34,11 @@ async-trait = { version = "0.1", optional = true }
 elsa = { version = "1.9" }
 
 [target.'cfg(all(target_arch = "aarch64", target_os = "macos"))'.dependencies]
-miden-gpu = { git = "https://github.com/0xPolygonMiden/miden-gpu", rev = "9a55000ff8cc113c1f4acb42115c542c1bb8533c", default-features = false, optional = true }
+miden-gpu = { git = "https://github.com/0xPolygonMiden/miden-gpu", rev = "b8b3a4273dd76a3af2ba8cb7739c9e20823a8f70", default-features = false, optional = true }
 pollster = { version = "0.3", optional = true }
 
 [target.'cfg(target_family = "wasm")'.dependencies]
-miden-gpu = { git = "https://github.com/0xPolygonMiden/miden-gpu", rev = "9a55000ff8cc113c1f4acb42115c542c1bb8533c", default-features = false, optional = true }
+miden-gpu = { git = "https://github.com/0xPolygonMiden/miden-gpu", rev = "b8b3a4273dd76a3af2ba8cb7739c9e20823a8f70", default-features = false, optional = true }
 
 [target.'cfg(target_family = "wasm")'.dev-dependencies]
 wasm-bindgen-test = "0.3"
diff --git a/prover/src/gpu/webgpu/mod.rs b/prover/src/gpu/webgpu/mod.rs
index ae44f45c1..f1ef5d4b2 100644
--- a/prover/src/gpu/webgpu/mod.rs
+++ b/prover/src/gpu/webgpu/mod.rs
@@ -13,15 +13,13 @@ extern crate alloc;
 #[cfg(not(feature = "std"))]
 use alloc::boxed::Box;
 #[cfg(not(feature = "std"))]
-use alloc::vec;
-#[cfg(not(feature = "std"))]
-use alloc::vec::Vec;
+use alloc::{vec, vec::Vec};
 
 use air::{AuxRandElements, LagrangeKernelEvaluationFrame};
 use elsa::FrozenVec;
 use maybe_async::maybe_async;
 use miden_gpu::{
-    webgpu::{build_merkle_tree, get_wgpu_helper, RowHasher},
+    webgpu::{build_merkle_tree, get_or_init_wgpu_helper, init_wgpu_helper, RowHasher},
     HashFn,
 };
 use processor::{
@@ -29,6 +27,7 @@ use processor::{
     ONE,
 };
 use tracing::info_span;
+use tracing::{event, Level};
 use winter_prover::{
     crypto::{Digest, MerkleTree},
     matrix::{get_evaluation_offsets, ColMatrix, RowMatrix, Segment},
@@ -37,7 +36,6 @@ use winter_prover::{
     DefaultConstraintEvaluator, EvaluationFrame, Prover, StarkDomain, TraceInfo, TraceLde,
     TracePolyTable,
 };
-use tracing::{event, Level};
 
 use crate::{
     crypto::{RandomCoin, Rpo256},
@@ -49,6 +47,13 @@ use crate::{
 #[cfg(test)]
 mod tests;
 
+#[allow(clippy::uninit_vec)]
+pub unsafe fn uninit_vector_real<T>(length: usize) -> Vec<T> {
+    let mut vector = Vec::with_capacity(length);
+    vector.set_len(length);
+    vector
+}
+
 // CONSTANTS
 // ================================================================================================
 
@@ -102,15 +107,15 @@ where
         assert!(poly_offset < polys.num_base_cols());
 
         // allocate memory for the segment
-        // let data = if polys.num_base_cols() - poly_offset >= N {
-        //     // if we will fill the entire segment, we allocate uninitialized memory
-        //     unsafe { page_aligned_uninit_vector(domain_size) }
-        // } else {
-        // but if some columns in the segment will remain unfilled, we allocate memory
-        // initialized to zeros to make sure we don't end up with memory with
-        // undefined values
-        let data = alloc::vec![[E::BaseField::ZERO; N]; domain_size];
-        // };
+        let data = if polys.num_base_cols() - poly_offset >= N {
+            // if we will fill the entire segment, we allocate uninitialized memory
+            unsafe { uninit_vector_real(domain_size) }
+        } else {
+            // but if some columns in the segment will remain unfilled, we allocate memory
+            // initialized to zeros to make sure we don't end up with memory with
+            // undefined values
+            vec![[E::BaseField::ZERO; N]; domain_size]
+        };
 
         Segment::new_with_buffer(data, polys, poly_offset, offsets, twiddles)
     }
@@ -234,7 +239,7 @@ where
             offsets.len().ilog2(),
             now.elapsed().as_millis()
         );
-        let helper = get_wgpu_helper().unwrap();
+        let helper = get_or_init_wgpu_helper().await;
 
         // build constraint evaluation commitment
         #[cfg(feature = "std")]
@@ -246,7 +251,7 @@ where
         let rpo_padded_segment_idx = rpo_requires_padding.then_some(num_base_columns / RATE);
         let mut row_hasher =
             RowHasher::new(helper, lde_domain_size, rpo_requires_padding, self.webgpu_hash_fn);
-        let mut rpo_padded_segment: Vec<[Felt; RATE]>;
+        let rpo_padded_segment: Vec<[Felt; RATE]>;
         for (segment_idx, segment) in segments.iter().enumerate() {
             // check if the segment requires padding
             if rpo_padded_segment_idx.map_or(false, |pad_idx| pad_idx == segment_idx) {
@@ -263,11 +268,6 @@ where
                         s
                     })
                     .collect();
-                // For rpx, skip this step
-                if self.webgpu_hash_fn == HashFn::Rpo256 {
-                    let rpo_pad_column = num_base_columns % RATE;
-                    rpo_padded_segment.iter_mut().for_each(|row| row[rpo_pad_column] = ONE);
-                }
                 row_hasher.update(helper, &rpo_padded_segment);
                 assert_eq!(segments.len() - 1, segment_idx, "padded segment should be the last");
                 break;
@@ -337,6 +337,7 @@ impl<
         domain: &StarkDomain<Felt>,
         webgpu_hash_fn: HashFn,
     ) -> (Self, TracePolyTable<E>) {
+        init_wgpu_helper().await;
         // extend the main execution trace and build a Merkle tree from the extended trace
         let (main_segment_lde, main_segment_tree, main_segment_polys) =
             build_trace_commitment_sync::<E, Felt, H>(main_trace, domain);
@@ -602,6 +603,7 @@ async fn build_trace_commitment<
         fft::interpolate_poly(&mut poly, &inv_twiddles);
         poly
     });
+    let helper = get_or_init_wgpu_helper().await;
 
     // extend the execution trace and generate hashes on the gpu
     let lde_segments = FrozenVec::new();
@@ -609,9 +611,8 @@ async fn build_trace_commitment<
     let num_base_columns = trace.num_base_cols();
     let rpo_requires_padding = num_base_columns % RATE != 0;
     let rpo_padded_segment_idx = rpo_requires_padding.then_some(num_base_columns / RATE);
-    let mut row_hasher =
-        RowHasher::new(get_wgpu_helper().unwrap(), lde_domain_size, rpo_requires_padding, hash_fn);
-    let mut rpo_padded_segment: Vec<[Felt; RATE]>;
+    let mut row_hasher = RowHasher::new(&helper, lde_domain_size, rpo_requires_padding, hash_fn);
+    let rpo_padded_segment: Vec<[Felt; RATE]>;
     let mut lde_segment_generator = SegmentGenerator::new(trace_polys, domain);
     let mut lde_segment_iter = lde_segment_generator.gen_segment_iter().enumerate();
     for (segment_idx, segment) in &mut lde_segment_iter {
@@ -630,21 +631,14 @@ async fn build_trace_commitment<
                     s
                 })
                 .collect();
-            // skip this in case of Rpx
-            if hash_fn == HashFn::Rpo256 {
-                let rpo_pad_column = num_base_columns % RATE;
-                rpo_padded_segment.iter_mut().for_each(|row| row[rpo_pad_column] = ONE);
-            }
-            row_hasher.update(get_wgpu_helper().unwrap(), &rpo_padded_segment);
+            row_hasher.update(&helper, &rpo_padded_segment);
             assert!(lde_segment_iter.next().is_none(), "padded segment should be the last");
             break;
         }
-        row_hasher.update(get_wgpu_helper().unwrap(), segment);
+        row_hasher.update(&helper, segment);
     }
-    let row_hashes = row_hasher.finish(get_wgpu_helper().unwrap()).await.unwrap();
-    let tree_nodes = build_merkle_tree(get_wgpu_helper().unwrap(), &row_hashes, hash_fn)
-        .await
-        .unwrap();
+    let row_hashes = row_hasher.finish(&helper).await.unwrap();
+    let tree_nodes = build_merkle_tree(&helper, &row_hashes, hash_fn).await.unwrap();
     // aggregate segments at the same time as the GPU generates the merkle tree nodes
     let lde_segments = lde_segments.into_vec().into_iter().map(|p| *p).collect();
     let trace_lde = RowMatrix::from_segments(lde_segments, num_base_columns);
diff --git a/prover/src/lib.rs b/prover/src/lib.rs
index ee0077dd7..69ccc8809 100644
--- a/prover/src/lib.rs
+++ b/prover/src/lib.rs
@@ -9,6 +9,7 @@ use core::marker::PhantomData;
 use maybe_async::maybe_async;
 
 use air::{AuxRandElements, ProcessorAir, PublicInputs};
+#[cfg(any(feature = "metal", feature = "webgpu"))]
 use miden_gpu::HashFn;
 use processor::{
     crypto::{