llama : adds llama-grammar memoization stacks (#4218) #9833

clarismiranda · 2024-10-11T01:39:51Z

I brought back @HanClinto idea on memoize. You might have forgotten to add the <stack, stacks> pair in the llama_grammar_stacks_cache.

Here is a helpful case of memoization to avoid exponential increments on the stacks.

Grammar

root ::= digit ((letter (letter (letter (letter (letter)?)?)?)?)? digit )*
digit ::= [0-9]
letter ::= mega-rule-a | mega-rule-b
mega-rule-b ::= l-b  | l-ba
mega-rule-a ::= l-a | l-ab | l-aa | l-aaa | l-aaaa | l-aaaaa | l-aaaaaa | l-aaaaaaa | l-aaaaaaaa | l-aaaaaaaaa | l-aaaaaaaa | l-aaaaaaaaa | l-aaaaaaaaaa
l-a ::= "a"
l-b ::= "b"
l-ab ::= "ab"
l-ba ::= "ba"
l-aa ::= "aa"
l-aaa ::= "aaa"
l-aaaa ::= "aaaa"
l-aaaaa ::= "aaaaa"
l-aaaaaa ::= "aaaaaa"
l-aaaaaaa ::= "aaaaaaa"
l-aaaaaaaa ::= "aaaaaaaa"
l-aaaaaaaaa ::= "aaaaaaaaa"
l-aaaaaaaaaa ::= "a"+

long_test.txt

Current llama-gbnf-validator profiling:

Memoization llama-gbnf-validator profiling:

I have read the contributing guidelines
Self-reported review complexity:
- Low
- Medium
- High

src/llama-grammar.cpp

clarismiranda · 2024-10-14T06:20:23Z

src/llama-grammar.h

@@ -3,6 +3,7 @@
 #include "llama-impl.h"

 #include <map>
+#include <unordered_map>


Should we use unordered_map vs map? Is there any benefit or necessity of a sorted key? @ggerganov

I can't think of any need to have a sorted key for this -- feels like unordered_map would be my default way to go on this one. If you're curious, a good set of profiling runs to test both options wouldn't be a bad exercise.

ggerganov · 2024-10-14T06:29:28Z

src/llama-grammar.cpp

@@ -1127,9 +1225,10 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
    const auto & code_points = decoded.first;

    llama_grammar_stacks stacks_new;
+    llama_grammar_stacks_cache stacks_cache;


I haven't yet understood the details of the implementation, but just want to confirm that recreating the cache for each token is really what we want? Looks like before moving the cache from global to local state, it was initialized during grammar init and maintained across multiple tokens. Now it is recreated for each token (IIUC).

Yes, is it better to initialize the cache in grammar_init and return it? This implementation adds the cache in the llama_grammar; however, this can be a valid assumption as the grammar doesn't change. @ggerganov

Yes, is it better to initialize the cache in grammar_init and return it?

I don't know yet. I was just looking at the lifetime of the cache and noticed this discrepancy. I will need more time to review the logic of the implementation.

IMO it's correct put the cache into llama_grammar

If we consider each sequence to have one llama_grammar, then every time we generate a new sequence, we need to create a new llama_grammar.

In other words, the cache will reset whenever we done sampling the sequence (i.e. stop generation)

I have added stacks_cache into llama_grammar :)

ggerganov

Need to copy the cache upon grammar clone:

diff --git a/src/llama-grammar.cpp b/src/llama-grammar.cpp
index 21482079..5476976b 100644
--- a/src/llama-grammar.cpp
+++ b/src/llama-grammar.cpp
@@ -1153,7 +1153,7 @@ void llama_grammar_free_impl(struct llama_grammar * grammar) {
 }
 
 struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & grammar) {
-    llama_grammar * result = new llama_grammar { grammar.vocab, grammar.rules, grammar.stacks, grammar.partial_utf8, };
+    llama_grammar * result = new llama_grammar { grammar.vocab, grammar.rules, grammar.stacks, grammar.partial_utf8, grammar.stacks_cache, };
 
     // redirect elements in stacks to point to new rules
     for (size_t is = 0; is < result->stacks.size(); is++) {

ggerganov · 2024-10-17T08:29:52Z

src/llama-grammar.cpp

+    if (it != stacks_cache.end()) {
+           advanced_stacks = it->second;
+    } else {
+        // Advance stacks with memorization 


Suggested change

// Advance stacks with memorization

// Advance stacks with memorization

Suggested change

// Advance stacks with memorization

// Advance stacks with memoization

ggml-ci

ggerganov · 2024-10-17T09:25:21Z

The llama_grammar implementation is not great because it expose a lot of internal logic in order to make the tests work. In this PR, instead of exposing the stacks_cache as well, we should put some efforts to contain things better. I've opened a small refactoring PR here: clarismiranda#1

Please give it a thorough test and merge it. After that, we'll review the PR again.

HanClinto · 2024-10-17T10:27:20Z

The llama_grammar implementation is not great because it expose a lot of internal logic in order to make the tests work.

The tests that I think are most valuable are the integration tests, and of the three (integration, unit, and gbnf-validator), ideally those are the ones that expose the least amount of internal logic. Should we maybe disable the unit tests and the gbnf-validator example for now -- until we do more of this refactoring? It's always frustrating when tests become a hindrance to reorganization.

I like the change you made in clarismiranda#1 -- that feels very good.

src/llama-grammar.cpp

llama : minor llama_grammar refactoring

Co-authored-by: Clint Herron <hanclinto@gmail.com>

ggerganov · 2024-10-17T21:08:57Z

Should we maybe disable the unit tests and the gbnf-validator example for now -- until we do more of this refactoring?

I guess we can do that. It's just that I am still not very intimately familiar with how the grammar implementation works in general (i.e. using it a bit as a black box) and not very sure what is the difference between the existing tests. So I just try to refactor the code in simple ways that preserve the original behaviour, but at least make it a bit more clear and easier to read. Ideally, the implementation has to be rewritten to avoid exposing the internal state and the tests should work using only the public libllama API.

HanClinto · 2024-10-18T18:09:05Z

Should we maybe disable the unit tests and the gbnf-validator example for now -- until we do more of this refactoring?

I guess we can do that. It's just that I am still not very intimately familiar with how the grammar implementation works in general (i.e. using it a bit as a black box) and not very sure what is the difference between the existing tests.

This maybe could be documented better elsewhere, but here's the landscape of grammar tests as I see it. I could be misremembering / misinformed about any of this, so anyone feel free to jump in and correct if you see something:

test-grammar-parser.cpp - The "original", this is the oldest set of tests around the grammar parser, and I've (personally) always found it the most inscrutable (though Ochafik did a lot to help the readability). Given a particular grammar, this test ensures that the internal rule structure is constructed as originally designed. This is a unit-level test (meaning: tests how the code does what it does, but doesn't test what it does). Especially the original version was pretty hard-coded for a few particular test-cases, and if someone wanted to optimize the code to change the way grammar rule trees are generated (such as Ochafik did in grammars: x{min,max} repetition operator #6640), then one has to break the tests to change any implementation. I imagine this was very useful in the early days of first developing the grammar engine when one has a very clear vision of how the engine should work internally, but IMO, unit tests like this are of limited use (and often a hindrance) during the refactoring or optimization stages of development.
test-llama-grammar.cpp - Was written soon after the grammar-parser test, and functions very similarly. Logic-wise, it picks up after the previous test (takes in a tree of pre-parsed grammar rules, organized as a vector of vectors) and now advances the stacks one token at a time and tests that the internal state is still as-expected.
test-grammar-integration.cpp - As @ejones noted in a comment re: the above unit tests, the early unit tests were implementation-level tests, and he thought that a test that only leverages the public interface would perhaps be a better focus. I didn't do my implementation quite like Evan suggested (I sidestepped the tokenization and logit portion), but I did try to focus more on the public API. As part of digging into llama : speed-up grammar sampling #4218, we needed a way to ensure that the functionality of the grammar engine stayed the same (the "what" of the library) while we made changes to the internals (the "how"). Unit tests don't usually help in this situation, so we created the end-to-end / integration tests to not care how the grammar engine works, only that it works. The aim is to let us change the internals all around -- completely remaking internal structures or organization (or even move away from a stack-based parser altogether!) -- while ensuring that end-user functionality remains unchanged. This is the only kind of test that is useful in a refactoring.

Ideally, the integration tests are the ones that should dig the least into the internals of the grammar engine, and this is the bare-minimum of what I think the public grammar API support. I could also see an argument to be made for rehanging these integration tests to incorporate the tokenization and logit-modification portion, but that discussion may be best-left for another day (IMO, the grammar engine is confusing enough without having to get overly concerned with tokenization).

test-json-schema-to-grammar.cpp - This doesn't deal with the grammar engine itself, but instead the translator from JSON Schema to GBNF. Outside of the scope of this discussion, and it should not be changed / removed.
GBNF Validator - A utility program I wrote to let me feed in a grammar and see if it matched an input like I expected. Kinda' like a GBNF imitation of something like Regex101 or Regex Buddy to help me: A) Ensure that my grammars are syntactically correct, and B) Understand why a particular input does or doesn't match a grammar as-expected. I honestly have no idea if this utility script is useful to anyone else, but I think it's unfortunate how much of the private API it uses. If we need to break this in order to refactor the grammar engine, I say we should do it -- this is not an important utility -- just disable / remove it. The cleaner we make the grammar API, the easier it will be to rewrite this sort of tool in the future.

In short, the only one I really care about right now is the integration tests -- the unit tests and validator are the most expendable, because they're the most-tied to the particulars of the grammar internals.

So I just try to refactor the code in simple ways that preserve the original behaviour, but at least make it a bit more clear and easier to read. Ideally, the implementation has to be rewritten to avoid exposing the internal state and the tests should work using only the public libllama API.

If we (generally) want the tests to work using only the public libllama API, then I vote we cull the unit tests / validator and focus only on the integration tests. I think those unit tests have served their purpose, and can be safely removed.

HanClinto · 2024-10-23T04:41:03Z

I tested the speed of the PR against master (at 7eee341), and I'm not seeing an appreciable speedup. This is just on my M1 Macbook Pro (while simultaneously running VS Code and Firefox with about a million open tabs) so it's far from a controlled test, but it's still eyebrow raising.

Hyperfine Command:

hyperfine \
    --warmup 1 --runs 5 \
    -L branch grammar-memo,7eee341bee09957139789c2d828995953f0fc7ff \
    --setup 'git checkout {branch} && make clean && make -j LLAMA_CURL=1 llama-cli' \
    './llama-cli \
        -mu https://huggingface.co/TheBloke/phi-2-GGUF/resolve/main/phi-2.Q4_K_M.gguf \
        --grammar-file grammars/big_stacks.gbnf \
        -p "List of numbers and assortments of letters composed of a and b:" \
        --seed 12344'

Grammar is nearly the same as the one @clarismiranda posted, with the exception that it is limited to exactly 1000 entries (without this, the command never seemed to want to finish -- even after I put the limit in, it still took roughly a minute for each iteration to complete):

big_stacks.gbnf

root ::= digit ((letter (letter (letter (letter (letter)?)?)?)?)? digit ){1000}
digit ::= [0-9]
letter ::= mega-rule-a | mega-rule-b
mega-rule-b ::= l-b  | l-ba
mega-rule-a ::= l-a | l-ab | l-aa | l-aaa | l-aaaa | l-aaaaa | l-aaaaaa | l-aaaaaaa | l-aaaaaaaa | l-aaaaaaaaa | l-aaaaaaaa | l-aaaaaaaaa | l-aaaaaaaaaa
l-a ::= "a"
l-b ::= "b"
l-ab ::= "ab"
l-ba ::= "ba"
l-aa ::= "aa"
l-aaa ::= "aaa"
l-aaaa ::= "aaaa"
l-aaaaa ::= "aaaaa"
l-aaaaaa ::= "aaaaaa"
l-aaaaaaa ::= "aaaaaaa"
l-aaaaaaaa ::= "aaaaaaaa"
l-aaaaaaaaa ::= "aaaaaaaaa"
l-aaaaaaaaaa ::= "a"+

Results:

`7eee341bee09957139789c2d828995953f0fc7ff` ran 1.03 ± 0.04 times faster than `grammar-memo`

Benchmark 1: ./llama-cli \
        -mu https://huggingface.co/TheBloke/phi-2-GGUF/resolve/main/phi-2.Q4_K_M.gguf \
        --grammar-file grammars/big_stacks.gbnf \
        -p "List of numbers and assortments of letters composed of a and b:" \
        --seed 12344 (branch = grammar-memo)
  Time (mean ± σ):     56.460 s ±  1.874 s    [User: 2.239 s, System: 0.666 s]
  Range (min … max):   54.704 s … 59.347 s    5 runs

Benchmark 2: ./llama-cli \
        -mu https://huggingface.co/TheBloke/phi-2-GGUF/resolve/main/phi-2.Q4_K_M.gguf \
        --grammar-file grammars/big_stacks.gbnf \
        -p "List of numbers and assortments of letters composed of a and b:" \
        --seed 12344 (branch = 7eee341bee09957139789c2d828995953f0fc7ff)
  Time (mean ± σ):     54.675 s ±  0.517 s    [User: 2.174 s, System: 0.624 s]
  Range (min … max):   54.196 s … 55.420 s    5 runs

Summary
  ./llama-cli \
        -mu https://huggingface.co/TheBloke/phi-2-GGUF/resolve/main/phi-2.Q4_K_M.gguf \
        --grammar-file grammars/big_stacks.gbnf \
        -p "List of numbers and assortments of letters composed of a and b:" \
        --seed 12344 (branch = 7eee341bee09957139789c2d828995953f0fc7ff) ran
    1.03 ± 0.04 times faster than ./llama-cli \
        -mu https://huggingface.co/TheBloke/phi-2-GGUF/resolve/main/phi-2.Q4_K_M.gguf \
        --grammar-file grammars/big_stacks.gbnf \
        -p "List of numbers and assortments of letters composed of a and b:" \
        --seed 12344 (branch = grammar-memo)

I ran a few different times with different grammars (such as with {0,1000} instead of {1000}, and testing against current master instead of the commit that this PR branched off of), and none of these changes seemed to make much of a difference -- everything was nearly identical in runtime:

`master` ran 1.00 ± 0.00 times faster than `grammar-memo`

Benchmark 1: ./llama-cli \
        -mu https://huggingface.co/TheBloke/phi-2-GGUF/resolve/main/phi-2.Q4_K_M.gguf \
        --grammar-file grammars/big_stacks.gbnf \
        -p "List of numbers and assortments of letters composed of a and b:" \
        --seed 12344 (branch = grammar-memo)
  Time (mean ± σ):     54.734 s ±  0.040 s    [User: 1.884 s, System: 0.602 s]
  Range (min … max):   54.688 s … 54.786 s    5 runs

Benchmark 2: ./llama-cli \
        -mu https://huggingface.co/TheBloke/phi-2-GGUF/resolve/main/phi-2.Q4_K_M.gguf \
        --grammar-file grammars/big_stacks.gbnf \
        -p "List of numbers and assortments of letters composed of a and b:" \
        --seed 12344 (branch = master)
  Time (mean ± σ):     54.668 s ±  0.097 s    [User: 1.806 s, System: 0.605 s]
  Range (min … max):   54.561 s … 54.789 s    5 runs

Summary
  ./llama-cli \
        -mu https://huggingface.co/TheBloke/phi-2-GGUF/resolve/main/phi-2.Q4_K_M.gguf \
        --grammar-file grammars/big_stacks.gbnf \
        -p "List of numbers and assortments of letters composed of a and b:" \
        --seed 12344 (branch = master) ran
    1.00 ± 0.00 times faster than ./llama-cli \
        -mu https://huggingface.co/TheBloke/phi-2-GGUF/resolve/main/phi-2.Q4_K_M.gguf \
        --grammar-file grammars/big_stacks.gbnf \
        -p "List of numbers and assortments of letters composed of a and b:" \
        --seed 12344 (branch = grammar-memo)

`master` ran 1.00 ± 0.01 times faster than `grammar-memo`

Benchmark 1: ./llama-cli \
        -mu https://huggingface.co/TheBloke/phi-2-GGUF/resolve/main/phi-2.Q4_K_M.gguf \
        --grammar-file grammars/big_stacks.gbnf \
        -p "List of numbers and assortments of letters composed of a and b:" \
        --seed 12344 (branch = grammar-memo)
  Time (mean ± σ):     54.905 s ±  0.239 s    [User: 2.007 s, System: 0.625 s]
  Range (min … max):   54.722 s … 55.322 s    5 runs

Benchmark 2: ./llama-cli \
        -mu https://huggingface.co/TheBloke/phi-2-GGUF/resolve/main/phi-2.Q4_K_M.gguf \
        --grammar-file grammars/big_stacks.gbnf \
        -p "List of numbers and assortments of letters composed of a and b:" \
        --seed 12344 (branch = master)
  Time (mean ± σ):     54.763 s ±  0.346 s    [User: 1.966 s, System: 0.616 s]
  Range (min … max):   54.560 s … 55.380 s    5 runs

  Warning: Statistical outliers were detected. Consider re-running this benchmark on a quiet system without any interferences from other programs. It might help to use the '--warmup' or '--prepare' options.

Summary
  ./llama-cli \
        -mu https://huggingface.co/TheBloke/phi-2-GGUF/resolve/main/phi-2.Q4_K_M.gguf \
        --grammar-file grammars/big_stacks.gbnf \
        -p "List of numbers and assortments of letters composed of a and b:" \
        --seed 12344 (branch = master) ran
    1.00 ± 0.01 times faster than ./llama-cli \
        -mu https://huggingface.co/TheBloke/phi-2-GGUF/resolve/main/phi-2.Q4_K_M.gguf \
        --grammar-file grammars/big_stacks.gbnf \
        -p "List of numbers and assortments of letters composed of a and b:" \
        --seed 12344 (branch = grammar-memo)

I haven't yet tested in the harness of gbnf-validator, but at least in terms of inference in the main CLI, I suspect something is going on that is preventing the cache speedup from taking effect.

@clarismiranda I realize I'm not using the same test setup as you used, but I'm curious to know if your benchmarking suite is still showing the same improvements you were seeing originally.

…anov#9833 Grammar memo Co-Authored-By: Clarissa Miranda <80654285+clarissamiranda@users.noreply.github.com>

ggerganov#9833" This reverts commit 4cbf5c392af62252a69e17143e8a81d771ca6f8a.

clarismiranda · 2024-10-28T00:18:15Z

Hi, @HanClinto. I will do more generation tests since I only profiled the llama-gbnf-validator (llama_grammar_accept and llama_grammar_advance_stack). I will be curious to see how generation works in larger contexts—for example, asking the LLM to repeat a valid long string and complete it with the following sequence.

HanClinto · 2024-10-28T19:35:23Z

Hi, @HanClinto. I will do more generation tests since I only profiled the llama-gbnf-validator (llama_grammar_accept and llama_grammar_advance_stack). I will be curious to see how generation works in larger contexts—for example, asking the LLM to repeat a valid long string and complete it with the following sequence.

Thanks for clarifying!

I haven't tested in that context yet, but I should.

With the changes of moving the cache location, do your profile tests on llama-gbnf-validator still hold up?

clarismiranda · 2024-10-29T21:37:14Z

Hi @HanClinto, here is the latest version of this branch profiling:

Captura de Pantalla 2024-10-30 a la(s) 8 34 07

And here is master profiling:

Captura de Pantalla 2024-10-30 a la(s) 8 34 26

It still shows a time benefit and overall memory benefit.

HanClinto · 2024-12-14T14:32:41Z

Hi @HanClinto, here is the latest version of this branch profiling:
...
It still shows a time benefit and overall memory benefit.

Hi @clarismiranda -- sorry for my delay in getting back to this.

It's good that there is still a performance benefit here in llama-gbnf-validator, but I'm leery of adding complexity if it doesn't help with the primary path. Do you think there is anything you can do with this PR to help llama-cli to get a similar speedup?

clarismiranda · 2024-12-16T00:50:19Z

Hi @HanClinto, I will give it a try :)

llama : adds llama-grammar memorization stacks (ggerganov#4218)

cb1632b

clarismiranda changed the title ~~llama : adds llama-grammar memorization stacks (#4218)~~ llama : adds llama-grammar memoization stacks (#4218) Oct 11, 2024

clarismiranda marked this pull request as ready for review October 11, 2024 03:21

ggerganov reviewed Oct 11, 2024

View reviewed changes

src/llama-grammar.cpp Outdated Show resolved Hide resolved

move cache stack to advance stack

901a347

github-actions bot added testing Everything test related examples labels Oct 14, 2024

clarismiranda commented Oct 14, 2024

View reviewed changes

ggerganov reviewed Oct 14, 2024

View reviewed changes

add stacks cache into llama_grammar

2aa6dd2

clarismiranda requested review from ngxson and ggerganov October 17, 2024 06:40

ggerganov reviewed Oct 17, 2024

View reviewed changes

llama : minor llama_grammar refactoring

17b3a3e

ggml-ci

ggerganov mentioned this pull request Oct 17, 2024

llama : minor llama_grammar refactoring clarismiranda/llama.cpp#1

Merged

HanClinto reviewed Oct 17, 2024

View reviewed changes

src/llama-grammar.cpp Outdated Show resolved Hide resolved

clarismiranda and others added 2 commits October 17, 2024 21:41

Merge pull request #1 from ggerganov/gg/grammar-refactor

34fc44d

llama : minor llama_grammar refactoring

Update spelling in memoize

a33fbbe

Co-authored-by: Clint Herron <hanclinto@gmail.com>

clarismiranda requested review from HanClinto and ggerganov October 17, 2024 10:44

update spelling

dc68a59

Nexesenex added a commit to Nexesenex/croco.cpp that referenced this pull request Oct 24, 2024

llama : adds llama-grammar memoization stacks (ggerganov#4218) ggerg…

e78bed5

…anov#9833 Grammar memo Co-Authored-By: Clarissa Miranda <80654285+clarissamiranda@users.noreply.github.com>

Nexesenex added a commit to Nexesenex/croco.cpp that referenced this pull request Oct 24, 2024

Revert " llama : adds llama-grammar memoization stacks (ggerganov#4218)

6a83ecb

ggerganov#9833" This reverts commit 4cbf5c392af62252a69e17143e8a81d771ca6f8a.

ggerganov mentioned this pull request Dec 19, 2024

llama : minor grammar refactor #10897

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

llama : adds llama-grammar memoization stacks (#4218) #9833

llama : adds llama-grammar memoization stacks (#4218) #9833

clarismiranda commented Oct 11, 2024

clarismiranda Oct 14, 2024

HanClinto Oct 15, 2024

ggerganov Oct 14, 2024

clarismiranda Oct 14, 2024 •

edited

Loading

ggerganov Oct 14, 2024

ngxson Oct 16, 2024 •

edited

Loading

clarismiranda Oct 17, 2024

ggerganov left a comment

ggerganov Oct 17, 2024

HanClinto Oct 17, 2024

ggerganov commented Oct 17, 2024

HanClinto commented Oct 17, 2024

ggerganov commented Oct 17, 2024

HanClinto commented Oct 18, 2024 •

edited

Loading

HanClinto commented Oct 23, 2024 •

edited

Loading

clarismiranda commented Oct 28, 2024 •

edited

Loading

HanClinto commented Oct 28, 2024

clarismiranda commented Oct 29, 2024

HanClinto commented Dec 14, 2024

clarismiranda commented Dec 16, 2024

	// Advance stacks with memorization
	// Advance stacks with memorization

	// Advance stacks with memorization
	// Advance stacks with memoization

llama : adds llama-grammar memoization stacks (#4218) #9833

Are you sure you want to change the base?

llama : adds llama-grammar memoization stacks (#4218) #9833

Conversation

clarismiranda commented Oct 11, 2024

clarismiranda Oct 14, 2024

Choose a reason for hiding this comment

HanClinto Oct 15, 2024

Choose a reason for hiding this comment

ggerganov Oct 14, 2024

Choose a reason for hiding this comment

clarismiranda Oct 14, 2024 • edited Loading

Choose a reason for hiding this comment

ggerganov Oct 14, 2024

Choose a reason for hiding this comment

ngxson Oct 16, 2024 • edited Loading

Choose a reason for hiding this comment

clarismiranda Oct 17, 2024

Choose a reason for hiding this comment

ggerganov left a comment

Choose a reason for hiding this comment

ggerganov Oct 17, 2024

Choose a reason for hiding this comment

HanClinto Oct 17, 2024

Choose a reason for hiding this comment

ggerganov commented Oct 17, 2024

HanClinto commented Oct 17, 2024

ggerganov commented Oct 17, 2024

HanClinto commented Oct 18, 2024 • edited Loading

HanClinto commented Oct 23, 2024 • edited Loading

clarismiranda commented Oct 28, 2024 • edited Loading

HanClinto commented Oct 28, 2024

clarismiranda commented Oct 29, 2024

HanClinto commented Dec 14, 2024

clarismiranda commented Dec 16, 2024

clarismiranda Oct 14, 2024 •

edited

Loading

ngxson Oct 16, 2024 •

edited

Loading

HanClinto commented Oct 18, 2024 •

edited

Loading

HanClinto commented Oct 23, 2024 •

edited

Loading

clarismiranda commented Oct 28, 2024 •

edited

Loading