Tolerance parameter setting

[ExcellentHH]

Hi, I am trying to verify some models using version 9edd47bbf6d15d685b3494f1019289412f255abc but I'm having trouble setting the tolerance parameter. The details are as follows:

1. Execute ezkl gen-settings --model model.onnx -T 20 -S 4 -B 23 -K 24 and get a setting json file:

{"run_args":{"tolerance":{"Abs":{"val":20}},"scale":4,"bits":23,"logrows":24,"batch_size":1,"input_visibility":"Private","output_visibility":"Public","param_visibility":"Private","allocated_constraints":null},"num_constraints":14182400,"model_instance_shapes":[[1,128,8,8]],"module_sizes":{"poseidon":[0,[0]],"elgamal":[0,[0,0,0]]},"required_lookups":[{"ReLU":{"scale":16}}],"check_mode":"SAFE"}

2. Execute ezkl mock --data input.json --model model.onnx --settings-path settings.json to quickly verify the correctness of the model and get succeed:

[D] [849s, ezkl_lib::graph::model] - laying out 16: CONCAT, offset:14174208
[D] [849s, ezkl_lib::graph::model] - model outputs are nodes: [16]
[*] [849s, ezkl_lib::graph::model] - computing...
[*] [1319s, ezkl] - succeeded

3. After executing mock proof, I tried to generate the formal proof and execute ezkl setup --model model.onnx --settings-path settings.json --srs-path perpetual-powers-of-tau-raw-24. Then an error appeared:

[*] [0s, ezkl_lib::pfsys] - loading params from "perpetual-powers-of-tau-raw-24"
[*] [4s, ezkl_lib::execute] - downsizing params to 24 logrows
[*] [4s, ezkl_lib::graph::model] - configuring model
[D] [4s, ezkl_lib::circuit::ops::chip] - assigning lookup input
[D] [4s, ezkl_lib::circuit::ops::chip] - assigning lookup output
thread 'main' panicked at 'assertion failed: `(left == right)`
  left: `16`,
 right: `64`', /.cargo/git/checkouts/halo2-049b997cf7195aea/ac425ba/halo2_proofs/src/poly/domain.rs:101:13
stack backtrace:
   0:     0x556516a31e60 - std::backtrace_rs::backtrace::libunwind::trace::h10a3454512d6724c
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/../../backtrace/src/backtrace/libunwind.rs:93:5
   1:     0x556516a31e60 - std::backtrace_rs::backtrace::trace_unsynchronized::h60c77ea7cb0df229
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/../../backtrace/src/backtrace/mod.rs:66:5
   2:     0x556516a31e60 - std::sys_common::backtrace::_print_fmt::h01643809be5392e9
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/sys_common/backtrace.rs:65:5
   3:     0x556516a31e60 - <std::sys_common::backtrace::_print::DisplayBacktrace as core::fmt::Display>::fmt::hc6486d7b214e20ee
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/sys_common/backtrace.rs:44:22
   4:     0x556516a597fe - core::fmt::write::h396ced1895c2ef11
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/core/src/fmt/mod.rs:1209:17
   5:     0x556516a2bd35 - std::io::Write::write_fmt::haaecb11edc216834
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/io/mod.rs:1682:15
   6:     0x556516a31c25 - std::sys_common::backtrace::_print::h3eebaa973f8e3926
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/sys_common/backtrace.rs:47:5
   7:     0x556516a31c25 - std::sys_common::backtrace::print::h12da160db56dea6e
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/sys_common/backtrace.rs:34:9
   8:     0x556516a334bf - std::panicking::default_hook::{{closure}}::h271b878f2b03712b
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/panicking.rs:267:22
   9:     0x556516a331fa - std::panicking::default_hook::he0bd4fb3722ee06f
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/panicking.rs:286:9
  10:     0x556516a33bbc - std::panicking::rust_panic_with_hook::h0e2a2b7a9776d76b
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/panicking.rs:688:13
  11:     0x556516a33957 - std::panicking::begin_panic_handler::{{closure}}::ha33eff9b02266b02
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/panicking.rs:579:13
  12:     0x556516a3230c - std::sys_common::backtrace::__rust_end_short_backtrace::h35f88c64c7aceda7
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/sys_common/backtrace.rs:137:18
  13:     0x556516a33672 - rust_begin_unwind
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/panicking.rs:575:5
  14:     0x55651562a353 - core::panicking::panic_fmt::h940cb25cf018faef
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/core/src/panicking.rs:65:14
  15:     0x556516a56e5b - core::panicking::assert_failed_inner::h9ca5ddd22a183a92
  16:     0x5565155f51fb - core::panicking::assert_failed::h2ddca21ea3130e05
  17:     0x55651586d7f8 - halo2_proofs::poly::domain::EvaluationDomain<F>::new::h5c7a860e7dd9051e
  18:     0x556515a3c40b - halo2_proofs::plonk::keygen::create_domain::hdda2ce066dd22b8a
  19:     0x556515a3e853 - halo2_proofs::plonk::keygen::keygen_vk::h2375f117e6169df7
  20:     0x5565157d172c - ezkl_lib::pfsys::create_keys::h2a908489321ee526
  21:     0x556515905281 - ezkl_lib::execute::setup::he079da90576d4db7
  22:     0x5565156a7e83 - <core::future::from_generator::GenFuture<T> as core::future::future::Future>::poll::h0e161a371baad51a
  23:     0x55651569b46d - tokio::runtime::scheduler::current_thread::Context::enter::h90b938c164e8ce97
  24:     0x5565156d3014 - tokio::macros::scoped_tls::ScopedKey<T>::set::h53bb5e925955693f
  25:     0x55651569b0d3 - tokio::runtime::scheduler::current_thread::CurrentThread::block_on::h13a56659661966bd
  26:     0x556515687250 - tokio::runtime::runtime::Runtime::block_on::hfc63217a54d0c7a9
  27:     0x556515693441 - ezkl::main::hbe179808ec6d674a
  28:     0x556515653193 - std::sys_common::backtrace::__rust_begin_short_backtrace::h26b62651f8384a60
  29:     0x556515682719 - std::rt::lang_start::{{closure}}::hdd46578b94f7fa87
  30:     0x556516a2503b - core::ops::function::impls::<impl core::ops::function::FnOnce<A> for &F>::call_once::hd38f4fddb0a5cc8c
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/core/src/ops/function.rs:286:13
  31:     0x556516a2503b - std::panicking::try::do_call::h37e120276f68ba22
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/panicking.rs:483:40
  32:     0x556516a2503b - std::panicking::try::h21590e4924b6f178
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/panicking.rs:447:19
  33:     0x556516a2503b - std::panic::catch_unwind::h05a57dd158543eda
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/panic.rs:137:14
  34:     0x556516a2503b - std::rt::lang_start_internal::{{closure}}::h0e7aeee8d0f517e3
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/rt.rs:148:48
  35:     0x556516a2503b - std::panicking::try::do_call::hf73acfa8d08f7180
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/panicking.rs:483:40
  36:     0x556516a2503b - std::panicking::try::h2261d841ace02082
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/panicking.rs:447:19
  37:     0x556516a2503b - std::panic::catch_unwind::hc3cd97821e3e639f
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/panic.rs:137:14
  38:     0x556516a2503b - std::rt::lang_start_internal::hcd4e566636a8bfe9
                               at /rustc/edf0182213a9e30982eb34f3925ddc4cf5ed3471/library/std/src/rt.rs:148:20
  39:     0x556515693525 - main
  40:     0x7f18bd2f5083 - __libc_start_main
                               at /build/glibc-SzIz7B/glibc-2.31/csu/../csu/libc-start.c:308:16
  41:     0x55651562a7fe - _start
  42:                0x0 - <unknown>

Question

1. I try to change the value of the Tolerance parameter, until it is set to be smaller than the scale parameter, and the ezkl setup can run normally. But in this way, I can't set a relatively high Tolerance parameter value for large models.
2. I have tried to use the ezkl forward --model model.onnx --data input.json --output input_new.json --settings-path settings.json command to correct the input file, but input_new.json cannot pass the mock verification with the same setting parameter setting as above(ezkl mock --data input_new.json --model model.onnx --settings-path settings.json).

Any insights are appreciated, thanks for the hard work!

[alexander-camuto]

hey @ExcellentHH my guess is that there's an issue with perpetual-powers-of-tau-raw-24 as seen here. Could you try the following steps in order:

1. Can you try downloading the SRS again and attempting again. -> if that works stop here.
2. Can you try using gen-srs with logrows=24 (this may take a while I apologize) and attempting again with the generated SRS. -> if that works stop here.
3. Can you try with a smaller logrows. -> if that works stop here.
4. Can you try with a larger logrows. -> if that works stop here.

[ExcellentHH]

@alexander-camuto Thank you for your response and suggestions. I will try the solutions you have proposed and keep you updated on any progress.

[alexander-camuto]

Heya ! I spoke to some of the PSE team. With your circuit we're starting to hit the limits in terms of circuit size/constraint order. The issue is that the absolute tolerance parameter when set to a size T increases the circuit order by 2*T -- in this case T=20 so order is 40, which far exceeds the maximum order the curve we're using can accommodate. You should probably use percent tolerance which won't increase the degree of the circuit as much. You can do so simply by changing the tolerance parameter to 20.0 :) We'll probably remove the absolute tolerance parameter. Its easy to misuse. Would percent tolerance be enough for your usecase ?

[ExcellentHH]

@alexander-camuto Sorry for the delayed reply. Thank you very much for your efforts and assistance. I tried out the approach you suggested. The situation is as follows:

For model A, I used the parameter settings of -T 20 -S 4 -B 21 -K 22 and the model verification was successful. For another model B, I used the settings of -T 10.0 -S 11 -B 24 -K 25 and the verification also succeeded.

Unfortunately and with apologies, I have not yet tested setting the Tolerance to relative error and absolute error for the same model.

Once again, thank you for your help!

[alexander-camuto]

Awesome glad that worked :)