feat: multiple instances of a circuit in single shard #1126

kunxian-xia · 2025-11-10T08:27:23Z

Summary

We want to prove multiple instances of a circuit in a single shard.

For example, in the case of ShardRam circuit, the 0-th shard often has ~2M records which is too large to fit in GPU memory.

Feat: bound ec sum in ecc quark prover #1130

kunxian-xia · 2025-11-12T14:41:06Z

.github/workflows/integration.yml

      - name: Install m4
        run: sudo apt-get install -y m4

+      - name: Run fibonacci (release) in 3 shards


Having 3 instances of ShardRam circuit in one shard.

Copilot

Pull Request Overview

This PR enables proving multiple instances of a circuit within a single shard, addressing memory constraints when handling large numbers of records (e.g., ~2M records in shard 0 of ShardRam circuit that exceed GPU memory capacity).

Key changes:

Introduced ChipInput structure to manage multiple circuit instances per chip
Refactored witness management to support chunking of circuit instances
Added configurable cross-shard access limits via CENO_CROSS_SHARD_LIMIT environment variable

Reviewed Changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
ceno_zkvm/src/tables/shard_ram.rs	Modified witness input type and EC sum extraction to return `SepticPoint` instead of vector
ceno_zkvm/src/structs.rs	Introduced `ChipInput` struct and refactored witness storage to support multiple circuit instances
ceno_zkvm/src/stats.rs	Updated to extract num_instances from new `ChipInput` structure
ceno_zkvm/src/scheme/verifier.rs	Updated chip proof iteration and commented out EC sum verification logic
ceno_zkvm/src/scheme/tests.rs	Adapted test to access witness via new `ChipInput` structure
ceno_zkvm/src/scheme/prover.rs	Refactored to use Vec instead of BTreeMap for chip proofs and iterate over chip inputs
ceno_zkvm/src/scheme/mock_prover.rs	Updated to iterate over witnesses using new `iter_sorted` method
ceno_zkvm/src/scheme.rs	Changed `chip_proofs` from BTreeMap to Vec to support multiple instances
ceno_zkvm/src/keygen.rs	Added circuit name to index mapping initialization
ceno_zkvm/src/instructions.rs	Added debug logging for record assignment
ceno_zkvm/src/e2e.rs	Added cross-shard access limit configuration and updated witness aggregation logic
.github/workflows/integration.yml	Added integration test with limited cross-shard accesses

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2025-11-13T03:12:51Z

ceno_zkvm/src/tables/shard_ram.rs

    type TableConfig = ShardRamConfig<E>;
    type FixedInput = ();
-    type WitnessInput = Vec<ShardRamInput<E>>;
+    type WitnessInput = [ShardRamInput<E>];


Changing WitnessInput from Vec<ShardRamInput<E>> to [ShardRamInput<E>] (unsized slice type) is a breaking API change that will prevent using this type directly in many contexts. Slices cannot be used as concrete types for struct fields or return values. Consider using &[ShardRamInput<E>] if you need a slice reference, or keep Vec<ShardRamInput<E>> if ownership is required.

Suggested change

type WitnessInput = [ShardRamInput<E>];

type WitnessInput = Vec<ShardRamInput<E>>;

Copilot · 2025-11-13T03:12:52Z

ceno_zkvm/src/scheme/verifier.rs

+            //     expected_septic_xy[0..SEPTIC_EXTENSION_DEGREE].into();
+            // let expected_septic_y: SepticExtension<E::BaseField> =
+            //     expected_septic_xy[SEPTIC_EXTENSION_DEGREE..].into();
+


Critical verification logic for EC sum validation has been commented out. This removes an important correctness check. Either remove the commented code if it's no longer needed, or if this is temporary for debugging/development, add a TODO comment explaining why it's disabled and when it should be re-enabled.

Suggested change

// TODO: EC sum verification is temporarily disabled for debugging/development.

// Re-enable the following assertions before production release to ensure proof correctness.

Copilot · 2025-11-13T03:12:52Z

ceno_zkvm/src/scheme/prover.rs

+                .circuit_name_to_index
+                .get(&circuit_name)
+                .cloned()
+                .expect("invalid circuit {} not exist in ceno zkvm");


The error message string doesn't include the actual circuit name variable. Change to .expect(&format!(\"invalid circuit {} not exist in ceno zkvm\", circuit_name)) to make the error message informative.

Suggested change

.expect("invalid circuit {} not exist in ceno zkvm");

.expect(&format!("invalid circuit {} not exist in ceno zkvm", circuit_name));

hero78119

first pass, overall lgtm, some stuff still pending clarify

hero78119 · 2025-11-13T02:26:19Z

ceno_zkvm/src/instructions.rs

        // TODO: selector is the only structural witness
        // this is workaround, as call `construct_circuit` will not initialized selector
        // we can remove this one all opcode unittest migrate to call `build_gkr_iop_circuit`
+        tracing::debug!("assign {} records for {}", steps.len(), Self::name());


This log a bit duplicate with https://github.com/scroll-tech/ceno/blob/master/ceno_zkvm/src/instructions/riscv/rv32im.rs#L461-L465 and suggest just need keep one

clean up this log by push new commit

ceno_zkvm/src/e2e.rs

hero78119 · 2025-11-13T03:10:50Z

ceno_zkvm/src/e2e.rs

+        let shard_ram_witnesses = zkvm_witness.get_circuit_witness(&ShardRamCircuit::<E>::name());
+
+        if let Some(shard_ram_witnesses) = shard_ram_witnesses {
+            let shard_ram_digest: SepticPoint<E::BaseField> = shard_ram_witnesses


nitpick: better name shard_ram_ec_point or ec_point

renamed to shard_ram_ec_sum

hero78119 · 2025-11-13T03:28:23Z

ceno_zkvm/src/scheme/prover.rs

-                    }
-                })
-                .or_else(|| {
-                    vk.get_cs().is_static_circuit().then(|| {


i am checking how new logic work after removing this. Because there are some circuit which only has fixed commitment and they will not present in witness rmm

push new commits to remove is_static_circuit because now EVERY circuit got 1 structural_witness as selector, thus this mechanism does not needed

hero78119 · 2025-11-13T03:29:33Z

ceno_zkvm/src/structs.rs

-            self.num_instances
-                .insert(
+        let circuit_inputs = global_input
+            .chunks(shard_ctx.max_num_cross_shard_accesses)


par_chunks better

fixed in new commit

hero78119 · 2025-11-13T03:39:12Z

ceno_zkvm/src/scheme.rs

    pub pi_evals: Vec<E>,
-    pub chip_proofs: BTreeMap<usize, ZKVMChipProof<E>>,
+    // each circuit may have multiple proof instances
+    pub chip_proofs: Vec<(usize, ZKVMChipProof<E>)>,


soundness: i think we need to have some mechanism to check, w.g. maximal proofs allow for a chip. E.g. what if prover fill same proof multiple times

I made this part to be BTreeMap<usize, Vec<ZKVMChipProof<E>>> then at least constrain OMC init proof can only be ONE proof

what if prover fill same proof multiple times

Then the offline memory checking prod_r == prod_w must fail.

hero78119 · 2025-11-13T07:47:59Z

ceno_zkvm/src/scheme/prover.rs

+            }
+            if num_instances.is_empty() {
+                // we need to drain respective fixed when num_instances is 0
+                if cs.num_fixed() > 0 {


FYI: here we drain num_fixed from mles otherwise their order will be mismatched

hero78119 · 2025-11-13T07:51:01Z

ceno_zkvm/src/structs.rs

-                .insert(OC::name(), vec![witness[0].num_instances()])
-                .is_none()
+        let num_instances = vec![witness[0].num_instances()];
+        let input = ChipInput::new(


we need to assign empty witness as placeholder to witnesses to skip respective fixed mles, see
https://github.com/scroll-tech/ceno/pull/1126/files#r2522083303

hero78119 · 2025-11-13T07:51:13Z

ceno_zkvm/src/structs.rs

+        let input = ChipInput::new(
+            TC::name(),
+            witness,
+            if num_instances > 0 {


same reasoning https://github.com/scroll-tech/ceno/pull/1126/files#r2522096832

hero78119 · 2025-11-13T07:52:03Z

ceno_zkvm/src/scheme/verifier.rs

                        .into(),
                ));
            }
+            if shard_id == 0 && circuit_vk.get_cs().with_omc_init_only() && proofs.len() != 1 {


More robust soundness: omc table can only present one proof

hero78119 · 2025-11-13T07:52:56Z

ceno_zkvm/src/scheme/verifier.rs

            && shard_id > 0
        {
            PCS::write_commitment(fixed_commit, &mut transcript).map_err(ZKVMError::PCSError)?;
        }


Hi @darth-cy please refer new verifier logic here to support
one shard + one chip + multiple proofs + respective soundness check

hero78119

Code diff for my change during code review

05c7504...adadd5b

Pending issue

need refactor mock_proving to support multi-proofs of a chip in a shard
remove environment loading and move to command line argument
retain ecc check https://github.com/scroll-tech/ceno/pull/1126/files#r2523080132

hero78119 · 2025-11-13T11:21:33Z

ceno_zkvm/src/scheme/verifier.rs

-
-            assert_eq!(&ecc_proof.sum.x, &expected_septic_x);
-            assert_eq!(&ecc_proof.sum.y, &expected_septic_y);
+            // let expected_septic_xy = cs


this will be retain in #1130

wip

df10648

kunxian-xia mentioned this pull request Nov 11, 2025

[milestones] ethproofs & benchmark #1095

Open

23 tasks

kunxian-xia added 4 commits November 11, 2025 23:58

wip 2

a014be0

wip3

598166e

fix

e6f9380

clippy

05c7504

kunxian-xia commented Nov 12, 2025

View reviewed changes

kunxian-xia requested review from Copilot and hero78119 and removed request for hero78119 November 12, 2025 15:14

Copilot AI reviewed Nov 13, 2025

View reviewed changes

hero78119 reviewed Nov 13, 2025

View reviewed changes

hero78119 added 5 commits November 13, 2025 14:30

remove is_static_circuit

1ff3c01

chores: renaming and par_chunks

af7d466

retain appending empty rmm to skip fixed mles correctly

46c75cc

soundness: check proof length in verifier and reject > 1 omc init proof

02ef251

clippy fix and rename

d0d4547

hero78119 reviewed Nov 13, 2025

View reviewed changes

hero78119 added 3 commits November 13, 2025 16:13

format && parallel

051132a

fix num_instance transcript order and mock_prover

f28d67a

shift rotation var of instances

adadd5b

hero78119 approved these changes Nov 13, 2025

View reviewed changes

hero78119 reviewed Nov 13, 2025

View reviewed changes

Merge branch 'master' into feat/multiple_instances

3c5adc0

hero78119 enabled auto-merge November 13, 2025 11:22

hero78119 added this pull request to the merge queue Nov 13, 2025

Merged via the queue into master with commit 07fc421 Nov 13, 2025
4 checks passed

hero78119 deleted the feat/multiple_instances branch November 13, 2025 11:59

hero78119 mentioned this pull request Nov 14, 2025

[recursion] ecc verifier + single chunk verify #1131

Open

3 tasks

	type WitnessInput = [ShardRamInput<E>];
	type WitnessInput = Vec<ShardRamInput<E>>;



	// TODO: EC sum verification is temporarily disabled for debugging/development.
	// Re-enable the following assertions before production release to ensure proof correctness.

	.expect("invalid circuit {} not exist in ceno zkvm");
	.expect(&format!("invalid circuit {} not exist in ceno zkvm", circuit_name));

feat: multiple instances of a circuit in single shard #1126

feat: multiple instances of a circuit in single shard #1126

Uh oh!

Conversation

kunxian-xia commented Nov 10, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull Request Overview

Reviewed Changes

Uh oh!

Copilot AI Nov 13, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Nov 13, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Nov 13, 2025

Choose a reason for hiding this comment

Uh oh!

hero78119 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

hero78119 Nov 13, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

hero78119 Nov 13, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

hero78119 left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

kunxian-xia commented Nov 10, 2025 •

edited

Loading

hero78119 Nov 13, 2025 •

edited

Loading

hero78119 Nov 13, 2025 •

edited

Loading

hero78119 left a comment •

edited

Loading