deps(l2): upgrade SP1 from v5.0.8 (Turbo) to v6.0.1 (Hypercube) by avilagaston9 · Pull Request #6188 · lambdaclass/ethrex

avilagaston9 · 2026-02-11T20:50:32Z

Motivation

SP1 Hypercube is Succinct's new proof system built on multilinear polynomials, claiming up to 5x improvement for compute-heavy workloads and ~2x for precompile-heavy workloads like Ethereum proving. This PR upgrades ethrex from SP1 Turbo (v5.0.8) to SP1 Hypercube (v6.0.1) so we can benchmark the new prover against our existing Turbo baseline.

Description

Bump all SP1 dependencies to =6.0.1 and update all accelerated cryptography patches to v6 stable tags. Bump Rust toolchain from 1.90.0 to 1.91.0 (required by SP1 v6.0.1 transitive dependencies).

Rewrite the SP1 prover backend (crates/l2/prover/src/backend/sp1.rs) to use the v6 blocking API (sp1_sdk::blocking::EnvProver), which unifies CPU/CUDA/Mock behind the SP1_PROVER env var. Migrate to the new Elf type, builder pattern for prove (prove().mode().run()), and pk.verifying_key() for VK access.

Additional changes:

Remove the ecdsa patch (breaks p256's VerifyPrimitive in v6 — k256 brings its own via git)
Remove sp1-zkvm/embedded feature (removed in v6)
Update on-chain verifier contract to v6 (v6.0.0/SP1VerifierGroth16.sol from sp1-contracts main)
Temporarily disable aligned-sdk SP1 submission (sp1-sdk v5/v6 type mismatch)
Add docs/prover/sp1-hypercube-upgrade.md with full upgrade notes and next steps
Add missing .gitignore for crates/guest-program/ — when refactor(l1,l2): extract guest program to dedicated ethrex-guest crate #5995 extracted the guest program from crates/l2/, the out/ and target/ directories lost coverage from crates/l2/.gitignore which had a blanket out/ entry

How to Test

# Install SP1 v6 toolchain
sp1up --version 6.0.1

# Build guest program
cargo build --release -p ethrex-guest-program --features sp1,l2

# Check host compilation
cargo check -p ethrex --features sp1

Checklist

Updated STORE_SCHEMA_VERSION (crates/storage/lib.rs) if the PR includes breaking changes to the Store requiring a re-sync.

Switch the prover loop from prove() to prove_timed() so each batch logs a structured line with batch number and elapsed proving time (seconds and milliseconds). Add scripts/sp1_bench_metrics.sh that tails the prover log, collects results into a CSV, and prints a summary table on exit.

… of prover The guest program was moved from crates/l2/prover/src/ethrex_guest_program/ to crates/guest-program/, but the fallback VK paths in the deployer were not updated. This caused deploy-l1-sp1 to fail with "No such file or directory" when running outside Docker. CI was unaffected because it passes explicit VK paths via ETHREX_SP1_VERIFICATION_KEY_PATH in docker-compose.yaml.

…nce conflicts, and add --endless flag for continuous load generation. The load_test function now fetches the pending nonce (instead of latest) so re-runs pick up where the previous round left off. wait_until_all_included now tracks per-account target nonces instead of a flat tx_amount, which was incorrect for any run where the starting nonce was not zero.

…ndpoint. The script now fetches batch_gas_used, batch_tx_count, and batch_size from the L2 metrics endpoint (localhost:3702/metrics) for each proved batch, joining them with proving time into a single CSV and summary table.

…ing. It now parses all proving_time lines from the file, fetches batch metadata from Prometheus, prints the table, and exits immediately.

…last] contains last - first + 1 blocks, not last - first. A single-block batch was reported as size 0.

…ment. When enabled (default), the prover wraps each prove() call with timing and logs structured fields (proving_time_s, proving_time_ms). When disabled with --no-timed, the prover calls prove() directly without timing overhead. Also fixes a clippy as_conversions warning in the elapsed.as_millis() cast.

…ess mode. The clap args now accept LOAD_TEST_RPC_URL, LOAD_TEST_TX_AMOUNT, and LOAD_TEST_ENDLESS environment variables as alternatives to --node, --tx-amount, and --endless flags respectively. CLI flags take precedence over env vars.

The guide (docs/l2/prover-benchmarking.md) explains how to use the localnet, prover timing, load test, and benchmark script together to measure proving performance. The agent workflow (docs/workflows/prover_benchmarking.md) provides step-by-step instructions for running a remote benchmark session, following the same format as the existing execution witness benchmarking workflow.

or in a loop depending on --endless, instead of a loop-with-break pattern.

…lient, tx_builder, and accounts instead of owned values, removing unnecessary clones at the call sites in run_round.

This removes the need for --no-timed; users pass --timed when they want proving time measurement, and omit it otherwise.

both the default and gpu-enabled cargo run commands.

--timed can be passed via make (e.g. make init-prover-sp1 PROVER_ARGS=--timed). Update docs to use the Makefile commands with GPU=true and PROVER_ARGS examples.

PROVER_ARGS, so timing is enabled with TIMED=true (e.g. make init-prover-sp1 GPU=true TIMED=true).

…alias. Clap already reads the env var, so no Makefile plumbing is needed. Reverts the TIMED/PROVER_ARGS additions to the Makefile targets.

…ummary instead of CSV. Add load-test Makefile target with env var configuration. Update docs to use Makefile commands throughout, add mempool limit warning, note that L2 must be running when collecting results, and reference PROVER_CLIENT_TIMED env var instead of --timed flag.

root Makefile target instead. Update docs to run make load-test from the repo root.

script. The report now includes a Server Specs section with hardware info detected from /proc/cpuinfo, /proc/meminfo, sysctl, and nvidia-smi.

user edits to prover benchmarking guide.

by default) in the prover benchmarking workflow.

batch assignments by whether the prover's type is needed. Previously the coordinator assigned a batch whenever ANY required proof was missing, causing provers to waste time re-proving batches they had already proved (e.g. an SP1 prover proving a batch that only needed an exec proof). The ProverBackend trait now has a prover_type() method, implemented for each backend (todo!() for ZisK and OpenVM which lack ProverType variants). The prover client stores its type at startup and sends it with every BatchRequest. The coordinator checks needed_proof_types.contains() before assigning. Also fixes the deploy-l1-sp1 Makefile target which used the removed --sp1.deploy-verifier flag — replaced with --sp1 and removed dummy verifier addresses for other backends. Updated benchmarking docs to reference deploy-l1-sp1 for SP1 benchmarks.

already exposes it via the prover_type() trait method.

…oad_test formatting. Commit 517e535 added prover_type to BatchRequest but didn't update the TDX quote-gen sender (outside the workspace), breaking the tdx lint/build CI jobs. The load_test formatting was also off since it's outside the main workspace.

…arted immediately after the L2 is up, not just before the prover. The prover proves batches faster than the sequencer fills them with transactions, so any delay in starting the load test results in empty batches being proved.

empty batches. The block producer and committer build and commit empty blocks/batches faster than the load test can start sending transactions (especially for erc20/fibonacci/io-heavy which have a contract deployment setup phase). By pre-compiling the binary in step 1b and running it directly in step 3, the load test starts immediately when the L2 is up.

…t recompilation. Previously each step (deploy, L2 start, prover start) triggered its own `cargo run` which recompiled the binary. Now step 1b builds everything once with all needed features, and all subsequent steps use `target/release/ethrex` directly. Also made the docs backend-agnostic (not SP1-specific) with placeholder flags and tables for SP1/RISC0/Exec, and made the branch/commit prompt mandatory (no default).

…name sp1_bench_metrics.sh to bench_metrics.sh since the script is backend-agnostic, fix the load test targets table in the workflow doc to show -t flag values instead of Makefile targets, and add a note clarifying that the hardcoded keys in the deploy command are default test keys from fixtures/.

fix the load test progress percentage to be relative to the current round (not the absolute nonce), make run_round return Result so endless mode survives individual round failures, remove a duplicate debug log in the proof coordinator, add bash 4+ note to bench_metrics.sh, and clarify that LOAD_TEST_RPC_URL must point to the L2 port for benchmarking.

backend that is not in the deployment's required proof types, instead of silently sending an empty batch response. The prover now logs an error and surfaces the rejection clearly, rather than endlessly polling for batches it will never be assigned.

github-actions · 2026-02-11T20:53:25Z

Lines of code report

Total lines added: 50
Total lines removed: 39
Total lines changed: 89

Detailed view

+-----------------------------------------------+-------+------+
| File                                          | Lines | Diff |
+-----------------------------------------------+-------+------+
| ethrex/crates/guest-program/build.rs          | 199   | +6   |
+-----------------------------------------------+-------+------+
| ethrex/crates/l2/prover/src/backend/error.rs  | 43    | +5   |
+-----------------------------------------------+-------+------+
| ethrex/crates/l2/prover/src/backend/sp1.rs    | 202   | +20  |
+-----------------------------------------------+-------+------+
| ethrex/crates/l2/prover/src/prover.rs         | 246   | +19  |
+-----------------------------------------------+-------+------+
| ethrex/crates/l2/sequencer/l1_proof_sender.rs | 553   | -39  |
+-----------------------------------------------+-------+------+

…dirs). When the guest program was extracted from crates/l2/ to crates/guest-program/ in PR #5995, the out/ and target/ directories lost coverage from crates/l2/.gitignore which had a blanket out/ entry. Without this, built ELF binaries and verification keys show up as untracked files.

…rmatting in sp1.rs and l1_proof_sender.rs, and regenerate Cargo.lock. SP1 Hypercube (v6) introduced a dependency on protobuf-compiler via sp1-prover-types, which was missing from the GitHub Actions runners.

requiring Rust >1.90.0), update tooling/Cargo.lock to match the base branch, and automate the substrate-bn sp1-lib version workaround in the Makefile's check-cargo-lock target so CI can validate the SP1 guest lockfile without manual cargo cache patching.

(CpuProver / CudaProver) to match the approach used on main. The gpu feature flag selects the CUDA prover at compile time instead of relying on the SP1_PROVER environment variable at runtime.

.verifying_key() resolves on the CpuProver's proving key type.

"Cannot start a runtime from within a runtime" panic. The CudaProver builder internally calls block_on, which conflicts with the prover's tokio runtime. CpuProver is unaffected. The fix runs once via OnceLock.

tokio runtime nesting panics. The SP1 v6 blocking SDK uses block_on() internally for proving, execution, and verification — not just initialization. Since the prover runs inside a tokio runtime, all these calls must run on threads without an active runtime.

…kwell) attempt. The pre-built sp1-gpu-server v6.0.0-rc.1 binary crashes with an AllocError on compute capability 12.0 GPUs. Report covers three issues encountered (CUDA 12 library mismatch, CUDA_VISIBLE_DEVICES propagation, and the fatal AllocError), workarounds applied, and recommendations to use Ampere/Hopper GPUs until SP1 adds Blackwell support.

…nces. The standalone native rollups integration guide is removed since the implementation details now live in the open PRs (#6186, #6248). The L2 roadmap is updated to reflect current progress: distributed proving is done (#6158), SP1 Hypercube upgrade is in progress (#6188), native rollups EXECUTE precompile and L2 PoC are in progress (#6186, #6248), and custom contract errors are in progress (#6206).

… patch tags, Docker tags, CI workflows, and the on-chain verifier contract path. Remove the three RC-specific workarounds that are no longer needed: the sp1-lib patch entry (stable semver now resolves correctly), the Makefile cargo-cache patching for substrate-bn's sp1-lib constraint, and the temporary sp1-contracts PR branch (v6.0.0 contracts are merged to main). Regenerate all three lockfiles.

# Conflicts: # crates/l2/prover/src/prover.rs # crates/l2/sequencer/proof_coordinator.rs

introduced by the merge.

SP1 v6.0.1 transitive dependencies require Rust 1.91. Also fix duplicate ProverTypeNotNeeded enum variant and match arm introduced by the merge with main (both branches added the same variant independently).

propagation to satisfy the ethrex-prover crate's strict clippy lints (expect_used, unwrap_used, panic are all denied). Add BackendError::Initialization variant for setup failures. Use process::exit(1) for the one-time prover startup path where recovery is not possible. Pin vergen to 9.0.6 to avoid a trait mismatch with vergen-git2 1.0.7 caused by vergen-lib version split.

…revious hash corresponded to Rust 1.90.0 and caused all Nix-based CI jobs (Test, TDX, Reorg Tests) to fail with a fixed-output derivation hash mismatch.

…ss lines to satisfy the formatter (sp1.rs:55).

…event the resolver from upgrading to 9.1.0 which uses vergen-lib 9.1.0, breaking compatibility with vergen-git2 1.0.7 (which requires vergen-lib 0.1.6). Without this pin, CI merge refs that re-resolve dependencies hit a trait mismatch: vergen::Rustc does not implement vergen_lib::entries::Add.

avilagaston9 · 2026-05-12T19:19:53Z

Closed for inactivity.

avilagaston9 added 30 commits February 9, 2026 11:02

Make benchmark script read the log file once and exit instead of tail…

a236c08

…ing. It now parses all proving_time lines from the file, fetches batch metadata from Prometheus, prints the table, and exits immediately.

Fix off-by-one in batch_size metric: a batch spanning blocks [first, …

a56a257

…last] contains last - first + 1 blocks, not last - first. A single-block batch was reported as size 0.

Extract load test round logic into run_round function and use it once

a786cd1

or in a loop depending on --endless, instead of a loop-with-break pattern.

Change load_test and wait_until_all_included to take references for c…

48ec1c1

…lient, tx_builder, and accounts instead of owned values, removing unnecessary clones at the call sites in run_round.

Change --timed flag default to false so it must be explicitly opted in.

9fd1d07

This removes the need for --no-timed; users pass --timed when they want proving time measurement, and omit it otherwise.

Clarify GPU acceleration option in prover benchmarking guide by showing

131c650

both the default and gpu-enabled cargo run commands.

Add PROVER_ARGS variable to Makefile prover targets so extra flags like

8375cb2

--timed can be passed via make (e.g. make init-prover-sp1 PROVER_ARGS=--timed). Update docs to use the Makefile commands with GPU=true and PROVER_ARGS examples.

Use PROVER_CLIENT_TIMED env var in Makefile prover targets instead of

d0fbcaa

PROVER_ARGS, so timing is enabled with TIMED=true (e.g. make init-prover-sp1 GPU=true TIMED=true).

Use PROVER_CLIENT_TIMED env var directly instead of a TIMED Makefile …

baae118

…alias. Clap already reads the env var, so no Makefile plumbing is needed. Reverts the TIMED/PROVER_ARGS additions to the Makefile targets.

Remove load-test target from crates/l2/Makefile and use the existing

ff9e680

root Makefile target instead. Update docs to run make load-test from the repo root.

Add automatic server specs detection (CPU, RAM, GPU) to the benchmark

f87ed2a

script. The report now includes a Server Specs section with hardware info detected from /proc/cpuinfo, /proc/meminfo, sysctl, and nvidia-smi.

Add batches-to-prove prompt to agent setup instructions and include

db185aa

user edits to prover benchmarking guide.

Clarify that LOAD_TEST_RPC_URL must point to the L2 node RPC (port 1729

7e49164

by default) in the prover benchmarking workflow.

Remove redundant prover_type field from Prover struct — the backend

8bd6f49

already exposes it via the prover_type() trait method.

avilagaston9 changed the base branch from main to bench/prover-tooling February 12, 2026 13:10

avilagaston9 added 4 commits February 12, 2026 10:27

Update CI workflows to install protoc and use SP1 v6.0.0-rc.1, fix fo…

6fa474a

…rmatting in sp1.rs and l1_proof_sender.rs, and regenerate Cargo.lock. SP1 Hypercube (v6) introduced a dependency on protobuf-compiler via sp1-prover-types, which was missing from the GitHub Actions runners.

Replace SP1 ProverClient::from_env() with compile-time feature flags

e46b2dd

(CpuProver / CudaProver) to match the approach used on main. The gpu feature flag selects the CUDA prover at compile time instead of relying on the SP1_PROVER environment variable at runtime.

Add missing ProvingKey trait import in l1_proof_sender so that

33449de

.verifying_key() resolves on the CpuProver's proving key type.

avilagaston9 force-pushed the bench/sp1-hypercube branch from 72e054e to 1c43180 Compare February 12, 2026 17:14

Spawn SP1 CudaProver initialization on a separate OS thread to avoid

216a422

"Cannot start a runtime from within a runtime" panic. The CudaProver builder internally calls block_on, which conflicts with the prover's tokio runtime. CpuProver is unaffected. The fix runs once via OnceLock.

avilagaston9 force-pushed the bench/sp1-hypercube branch from 1c43180 to 216a422 Compare February 12, 2026 17:46

avilagaston9 added 2 commits February 12, 2026 15:29

Base automatically changed from bench/prover-tooling to main February 20, 2026 13:48

avilagaston9 mentioned this pull request Feb 24, 2026

docs(l2): add L2 roadmap #6114

Closed

1 task

avilagaston9 changed the title ~~Upgrade SP1 from v5.0.8 (Turbo) to v6.0.0-rc.1 (Hypercube)~~ deps(l2): upgrade SP1 from v5.0.8 (Turbo) to v6.0.0 (Hypercube) Feb 25, 2026

avilagaston9 added 3 commits February 25, 2026 12:23

Merge remote-tracking branch 'origin/main' into bench/sp1-hypercube

9cde048

# Conflicts: # crates/l2/prover/src/prover.rs # crates/l2/sequencer/proof_coordinator.rs

Regenerate Cargo.lock after merging main to resolve dependency versions

6b2a6f8

introduced by the merge.

Bump SP1 from v6.0.0 to v6.0.1 and Rust toolchain from 1.90.0 to 1.91.0.

12223d0

SP1 v6.0.1 transitive dependencies require Rust 1.91. Also fix duplicate ProverTypeNotNeeded enum variant and match arm introduced by the merge with main (both branches added the same variant independently).

avilagaston9 changed the title ~~deps(l2): upgrade SP1 from v5.0.8 (Turbo) to v6.0.0 (Hypercube)~~ deps(l2): upgrade SP1 from v5.0.8 (Turbo) to v6.0.1 (Hypercube) Feb 25, 2026

github-actions Bot added the L2 Rollup client label Feb 25, 2026

github-project-automation Bot added this to ethrex_l2 Feb 25, 2026

avilagaston9 added 4 commits February 25, 2026 15:23

Update Nix fenix toolchain hash for Rust 1.91.0 in service.nix. The p…

6084984

…revious hash corresponded to Rust 1.90.0 and caused all Nix-based CI jobs (Test, TDX, Reorg Tests) to fail with a fixed-output derivation hash mismatch.

Apply cargo fmt to init_prover_setup — break chained method call acro…

d77f04c

…ss lines to satisfy the formatter (sp1.rs:55).

avilagaston9 closed this May 12, 2026

github-project-automation Bot moved this to Done in ethrex_l2 May 12, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

deps(l2): upgrade SP1 from v5.0.8 (Turbo) to v6.0.1 (Hypercube)#6188

deps(l2): upgrade SP1 from v5.0.8 (Turbo) to v6.0.1 (Hypercube)#6188
avilagaston9 wants to merge 53 commits into
mainfrom
bench/sp1-hypercube

avilagaston9 commented Feb 11, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Feb 11, 2026 •

edited

Loading

Uh oh!

avilagaston9 commented May 12, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

avilagaston9 commented Feb 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Description

How to Test

Checklist

Uh oh!

github-actions Bot commented Feb 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Lines of code report

Uh oh!

avilagaston9 commented May 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

avilagaston9 commented Feb 11, 2026 •

edited

Loading

github-actions Bot commented Feb 11, 2026 •

edited

Loading

avilagaston9 commented May 12, 2026 •

edited

Loading