feat(prover): add CLI option to run prover with max allocation

prover/Cargo.toml

@@ -61,8 +61,8 @@ circuit_sequencer_api = "=0.150.4"
 zkevm_test_harness = "=0.150.4"
 
 # GPU proving dependencies
-wrapper_prover = { package = "zksync-wrapper-prover", version = "=0.150.4" }
-shivini = "=0.150.4"
+wrapper_prover = { package = "zksync-wrapper-prover", version = "=0.150.6" }
+shivini = "=0.150.6"
 
 # Core workspace dependencies
 zksync_multivm = { path = "../core/lib/multivm", version = "0.1.0" }

prover/crates/bin/prover_fri/src/gpu_prover_job_processor.rs

@@ -5,6 +5,7 @@ pub mod gpu_prover {
     use anyhow::Context as _;
     use shivini::{
         gpu_proof_config::GpuProofConfig, gpu_prove_from_external_witness_data, ProverContext,
+        ProverContextConfig,
     };
     use tokio::task::JoinHandle;
     use zksync_config::configs::{fri_prover_group::FriProverGroupConfig, FriProverConfig};
@@ -82,7 +83,15 @@ pub mod gpu_prover {
             address: SocketAddress,
             zone: Zone,
             protocol_version: ProtocolSemanticVersion,
+            max_allocation: Option<usize>,
         ) -> Self {
+            let prover_context = match max_allocation {
+                Some(max_allocation) => ProverContext::create_with_config(
+                    ProverContextConfig::default().with_maximum_device_allocation(max_allocation),
+                )
+                .expect("failed initializing gpu prover context"),
+                None => ProverContext::create().expect("failed initializing gpu prover context"),
+            };
             Prover {
                 blob_store,
                 public_blob_store,
@@ -91,8 +100,7 @@ pub mod gpu_prover {
                 setup_load_mode,
                 circuit_ids_for_round_to_be_proven,
                 witness_vector_queue,
-                prover_context: ProverContext::create()
-                    .expect("failed initializing gpu prover context"),
+                prover_context,
                 address,
                 zone,
                 protocol_version,

prover/crates/bin/prover_fri/src/main.rs

@@ -139,6 +139,7 @@ async fn main() -> anyhow::Result<()> {
         public_blob_store,
         pool,
         circuit_ids_for_round_to_be_proven,
+        opt.max_allocation,
         notify,
     )
     .await
@@ -178,6 +179,7 @@ async fn get_prover_tasks(
     public_blob_store: Option<Arc<dyn ObjectStore>>,
     pool: ConnectionPool<Prover>,
     circuit_ids_for_round_to_be_proven: Vec<CircuitIdRoundTuple>,
+    _max_allocation: Option<usize>,
     _init_notifier: Arc<Notify>,
 ) -> anyhow::Result<Vec<JoinHandle<anyhow::Result<()>>>> {
     use crate::prover_job_processor::{load_setup_data_cache, Prover};
@@ -213,6 +215,7 @@ async fn get_prover_tasks(
     public_blob_store: Option<Arc<dyn ObjectStore>>,
     pool: ConnectionPool<Prover>,
     circuit_ids_for_round_to_be_proven: Vec<CircuitIdRoundTuple>,
+    max_allocation: Option<usize>,
     init_notifier: Arc<Notify>,
 ) -> anyhow::Result<Vec<JoinHandle<anyhow::Result<()>>>> {
     use gpu_prover_job_processor::gpu_prover;
@@ -245,6 +248,7 @@ async fn get_prover_tasks(
         address.clone(),
         zone.clone(),
         protocol_version,
+        max_allocation,
     );
     let producer = shared_witness_vector_queue.clone();
 
@@ -295,4 +299,6 @@ pub(crate) struct Cli {
     pub(crate) config_path: Option<std::path::PathBuf>,
     #[arg(long)]
     pub(crate) secrets_path: Option<std::path::PathBuf>,
+    #[arg(long)]
+    pub(crate) max_allocation: Option<usize>,
 }