add hunyuan moe support for 561 #565

ubergarm · 2025-06-30T18:13:39Z

Based this PR on mainline ggml-org/llama.cpp#14425. Didn't merge any python stuff (used mainline convert script). Tested with bf16 on hybrid CUDA+CPU.

model=/mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/Hunyuan-A13B-Instruct-BF16-00001-of-00004.gguf
./build/bin/llama-server \
  --model "$model" \
  --alias ubergarm/Hunyuan-A13B-Instruct-bf16 \
  -fa \
  -ctk q8_0 -ctv q8_0 \
  -c 8192 \
  --temp 0.6 \
  --presence-penalty 0.7 \
  --min-p 0.1 \
  -ts 48,48 \
  -ngl 16 \
  --threads 24 \
  --host 127.0.0.1 \
  --port 8080

Would be great if anyone else could test e.g. @Downtown-Case as per #561

I haven't yet made imatrix nor tried to quantize further.

Might be able to use one of the following if was converted recently enough:

The behavior seems a bit odd and will answer in chinese if I don't use some kind of system prompt or explicitly say speak in english. Mainline seems to use some kind of --jinja thing which isn't supported here psure. So ymmv.

ubergarm · 2025-06-30T18:28:48Z

I'm currently processing an imatrix and noticed that it requires -fa or will have very large numbers.

This seems to be working so far, though still seems a higher than I expected which could be indicative of an problem:

./build/bin/llama-imatrix \
    --verbosity 1 \
    --layer-similarity \
    -m /mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/Hunyuan-A13B-Instruct-BF16-00001-of-00004.gguf \
    -f ubergarm-imatrix-calibration-corpus-v02.txt \
    -o /mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/imatrix-Hunyuan-A13B-Instruct-BF16.dat \
    -fa \
    --ctx-size 512 \
    -ts 48,48 \
    -ngl 18 \
    --threads 24

system_info: n_threads = 24 / 48 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 |

compute_imatrix: tokenizing the input ..
compute_imatrix: tokenization took 709.171 ms
compute_imatrix: computing over 865 chunks with batch_size 512
compute_imatrix: 4.37 seconds per pass - ETA 1 hours 3.07 minutes
[1]12.7104,[2]14.8010,[3]14.3374,[4]30.5778,[5]17.4738,[6]14.5285,[7]20.2402,[8]14.9318,[9]11.7604,
save_imatrix: stored collected data after 10 chunks in /mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/imatrix-Hunyuan-A13B-Instruct-BF16.dat
[10]12.0205,[11]10.2799,[12]12.3863,[13]14.9808,[14]16.1885,[15]16.6677,[16]20.9547,[17]19.1613,[18]17.4531,[19]15.5200,
save_imatrix: stored collected data after 20 chunks in /mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/imatrix-Hunyuan-A13B-Instruct-BF16.dat
[20]14.7222,[21]13.4574,[22]12.5603,[23]11.8334,[24]11.1943,[25]10.7840,[26]10.5614,[27]10.8168,[28]11.2630,[29]11.9753,
save_imatrix: stored collected data after 30 chunks in /mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/imatrix-Hunyuan-A13B-Instruct-BF16.dat
[30]12.7904,[31]12.8568,[32]12.7520,[33]13.2066,[34]13.7438,[35]14.3701,[36]15.2825,[37]16.4474,[38]17.2615,[39]17.7246,
save_imatrix: stored collected data after 40 chunks in /mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/imatrix-Hunyuan-A13B-Instruct-BF16.dat
[40]20.3797,[41]22.3074,[42]22.9196,[43]23.5967,[44]24.9652,[45]26.3450,[46]28.0728,[47]28.1975,[48]27.9526,[49]31.3467,
save_imatrix: stored collected data after 50 chunks in /mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/imatrix-Hunyuan-A13B-Instruct-BF16.dat
[50]30.1730,[51]31.2195,[52]30.6089,[53]30.0938,[54]29.5127,[55]29.9680,[56]29.2944,[57]28.2416,[58]27.2467,[59]26.2110,
save_imatrix: stored collected data after 60 chunks in /mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/imatrix-Hunyuan-A13B-Instruct-BF16.dat
[60]25.3394,[61]24.4437,[62]23.7538,[63]25.8637,[64]27.0096,[65]28.0507,[66]27.7521,[67]29.0344,[68]29.8659,[69]30.3886,
save_imatrix: stored collected data after 70 chunks in /mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/imatrix-Hunyuan-A13B-Instruct-BF16.dat
[70]31.4350,[71]31.8531,[72]31.7906,[73]31.7912,[74]32.9230,[75]34.9214,[76]37.0384,[77]38.7590,[78]38.9847,[79]40.2656,
save_imatrix: stored collected data after 80 chunks in /mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/imatrix-Hunyuan-A13B-Instruct-BF16.dat
[80]41.5627,[81]41.0075,[82]42.5855,[83]44.5075,[84]43.9110,[85]43.3078,[86]42.7130,[87]41.7924,[88]41.2850,[89]41.5686,
save_imatrix: stored collected data after 90 chunks in /mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/imatrix-Hunyuan-A13B-Instruct-BF16.dat
[90]40.8182,[91]41.2610,[92]42.4782,[93]44.0758,[94]43.5943,[95]43.7613,[96]43.0079,[97]42.6615,[98]43.6499,[99]43.1762,
save_imatrix: stored collected data after 100 chunks in /mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/imatrix-Hunyuan-A13B-Instruct-BF16.dat
[100]42.4092,[101]43.1918,[102]44.5605,[103]44.1737,[104]44.2998,[105]45.3024,[106]45.5803,[107]45.3388,[108]45.5154,[109]45.8490,
save_imatrix: stored collected data after 110 chunks in /mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/imatrix-Hunyuan-A13B-Instruct-BF16.dat
[110]45.6819,[111]46.1607,[112]46.8070,[113]47.5833,[114]48.5492,[115]48.9797,[116]49.6842,[117]49.8659,[118]51.1640,[119]51.3824,
save_imatrix: stored collected data after 120 chunks in /mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/imatrix-Hunyuan-A13B-Instruct-BF16.dat
[120]52.0141,[121]53.6073,[122]55.3684,[123]56.2596,[124]56.0548,[125]56.1662,[126]56.3532,[127]57.2403,[128]56.6770,[129]58.3851,
save_imatrix: stored collected data after 130 chunks in /mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/imatrix-Hunyuan-A13B-Instruct-BF16.dat
[130]58.2333,[131]59.2614,[132]60.7497,[133]62.4619,[134]63.7352,[135]64.8522,[136]66.5478,[137]64.9457,[138]63.5455,[139]63.2199,
save_imatrix: stored collected data after 140 chunks in /mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/imatrix-Hunyuan-A13B-Instruct-BF16.dat

...

EDIT: I also tried adding this model to the list for ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32); but the imatrix PPL values look about the same. Seems to be about the same with or without -fmoe as well.

fwiw this seems to be similar numbers as I'm getting using mainline llama-imatrix:

[1]12.7998,[2]14.9052,[3]14.4276,[4]30.9156,[5]17.5724,[6]14.6579,[7]20.3671,[8]15.0254,[9]11.8121,[10]12.0809,[11]10.3416,[12]12.4422,[13]15.1108,^C

ikawrakow · 2025-06-30T20:20:40Z

No FA and FA giving very different PPL values is not a good sign.

PPL of 60 is not a good sign either, especially for a model of that size.

ubergarm · 2025-06-30T20:36:19Z

I'm going to leave an endpoint up for a little bit if anyone wants to try the first experimental quant.. No promises lol

Endpoint

WebUI: https://llm.ubergarm.com/
APIEndpoint: https://llm.ubergarm.com/ (it is llama-server API endpoint with no API key)

There are 8 concurrent slots each with 64k prompt limit.

Test Quant

I just rolled an imatrix.dat and made my first quant for testing.

llm_load_print_meta: model type       = 80B.A13B
llm_load_print_meta: model ftype      = IQ4_K - 4.5 bpw
llm_load_print_meta: model params     = 80.393 B
llm_load_print_meta: model size       = 48.581 GiB (5.191 BPW)
llm_load_print_meta: general.name     = Hunyuan A13B Instruct

blk\..*\.attn_k.*=iq6_k
blk\..*\.attn_v.*=iq6_k

blk\..*\.attn_q.*=iq5_k
blk\..*\.attn_o.*=iq5_k

# 1x Shared Expert
blk\..*\.ffn_(gate|up)_shexp.*=iq6_k
blk\..*\.ffn_(down)_shexp.*=iq5_k

# 64x Routed Experts
blk\..*\.ffn_(gate|up)_exps.*=iq5_k
blk\..*\.ffn_(down)_exps.*=iq4_k

# Token Embedding
token_embd\.weight=iq4_k

How I ran it:

model=/mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/Hunyuan-A13B-Instruct-IQ4_K.gguf
./build/bin/llama-server \
  --model "$model" \
  --alias ubergarm/Hunyuan-A13B-Instruct-IQ4_K \
  -fa \
  -ctk q8_0 -ctv q8_0 \
  -c 524288 \
  --temp 0.6 \
  --presence-penalty 0.7 \
  --min-p 0.1 \
  -ts 48,48 \
  -ngl 99 \
  --parallel 8 \
  --threads 1 \
  --host 127.0.0.1 \
  --port 8080

ikawrakow · 2025-07-01T06:00:36Z

src/llama.cpp

+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);


If you check your previous PR about GLM4 you will see that you had to remove the Vcur reshaping. It is the same here. Remove this line and it is likely the difference between FA and no FA will go away.

Yup, thanks for the reminder! The two trickiest parts of porting an architecture is remembering to:

Remove the Vcur reshaping.

On mainline build_attn() the argument order goes Qcur, Kcur, Vcur,, but here with llm_build_kv() the order goes Kcur, Vcur, Qcur,.

Just re-downloaded the new .safetensors, converted, and built a fresh quant to test:

FA=1 Final estimate: PPL = 522.7473 +/- 5.68072

FA=0 Final estimate: PPL = 527.6625 +/- 5.73144

So looks "good" now haha... I didn't wait to find the bf16's PPL but this lines up in the ball-park with what mainline is seeing around ~500.

Of course I couldn't help myself and had to try out the new IQ3_KS quant as well lol...

So far so good!

llm_load_print_meta: model type = 80B.A13B llm_load_print_meta: model ftype = IQ3_KS - 3.1875 bpw llm_load_print_meta: model params = 80.393 B llm_load_print_meta: model size = 34.088 GiB (3.642 BPW) llm_load_print_meta: general.name = Hunyuan A13B Instruct # Attention blk\..*\.attn_k.*=iq6_k blk\..*\.attn_v.*=iq6_k blk\..*\.attn_q.*=iq5_k blk\..*\.attn_o.*=iq5_k # 1x Shared Expert blk\..*\.ffn_(down)_shexp.*=iq6_k blk\..*\.ffn_(gate|up)_shexp.*=iq5_k # 64x Routed Experts blk\..*\.ffn_(down)_exps.*=iq4_ks blk\..*\.ffn_(gate|up)_exps.*=iq3_ks # let's live dangerously # Token Embedding token_embd\.weight=iq6_k # splurged here a bit as this model's tokenization seems wierd

kiron111 · 2025-07-02T02:42:43Z

I'm going to leave an endpoint up for a little bit if anyone wants to try the first experimental quant.. No promises lol

Endpoint

WebUI: https://llm.ubergarm.com/ APIEndpoint: https://llm.ubergarm.com/ (it is llama-server API endpoint with no API key)

There are 8 concurrent slots each with 64k prompt limit.

Test Quant

I just rolled an imatrix.dat and made my first quant for testing.

llm_load_print_meta: model type       = 80B.A13B
llm_load_print_meta: model ftype      = IQ4_K - 4.5 bpw
llm_load_print_meta: model params     = 80.393 B
llm_load_print_meta: model size       = 48.581 GiB (5.191 BPW)
llm_load_print_meta: general.name     = Hunyuan A13B Instruct

blk\..*\.attn_k.*=iq6_k
blk\..*\.attn_v.*=iq6_k

blk\..*\.attn_q.*=iq5_k
blk\..*\.attn_o.*=iq5_k

# 1x Shared Expert
blk\..*\.ffn_(gate|up)_shexp.*=iq6_k
blk\..*\.ffn_(down)_shexp.*=iq5_k

# 64x Routed Experts
blk\..*\.ffn_(gate|up)_exps.*=iq5_k
blk\..*\.ffn_(down)_exps.*=iq4_k

# Token Embedding
token_embd\.weight=iq4_k

How I ran it:

model=/mnt/raid/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/Hunyuan-A13B-Instruct-IQ4_K.gguf
./build/bin/llama-server \
  --model "$model" \
  --alias ubergarm/Hunyuan-A13B-Instruct-IQ4_K \
  -fa \
  -ctk q8_0 -ctv q8_0 \
  -c 524288 \
  --temp 0.6 \
  --presence-penalty 0.7 \
  --min-p 0.1 \
  -ts 48,48 \
  -ngl 99 \
  --parallel 8 \
  --threads 1 \
  --host 127.0.0.1 \
  --port 8080

tested on your api, it works for Chinese Q&A.

ubergarm · 2025-07-02T02:44:39Z

tested on your api, it works for Chinese Q&A.

Ahh very good, thanks you. Tonight I was running my updated experimental IQ3_KS which I went ahead and released prematurely because oh well it seems okay lol...

Thanks for testing!

It can fit 256k context in under 24GB VRAM when not offloading additional exps and with -ub 1024 get over 500 tok/sec PP and about 17 tok/sec TG. So quite a flexible model in terms of size at least.

kiron111 · 2025-07-02T03:54:01Z

run on wsl I got a error: Floating point exception (core dumped), in the initial procress of ik_llama.cpp

model=/mnt/g/lm-studio/models/ubergarm/Hunyuan-A13B-Instruct-GGUF/Hunyuan-A13B-Instruct-IQ3_KS.gguf
./build/bin/llama-server \
  --model "$model" \
  --alias ubergarm/Hunyuan-A13B-Instruct-IQ4_K \
  -fa \
  -ctk q8_0 -ctv q8_0 \
  -c 4096 \
  --temp 0.6 \
  --presence-penalty 0.7 \
  --min-p 0.1 \
  -ts 48,48 \
  -ngl 99 \
  --parallel 8 \
  --threads 1 \
  --host 127.0.0.1 \
  --port 8080
INFO [                    main] build info | tid="140543167610880" timestamp=1751427934 build=3776 commit="c6c23fa4"
INFO [                    main] system info | tid="140543167610880" timestamp=1751427934 n_threads=24 n_threads_batch=-1 total_threads=16 system_info="AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 | "
Floating point exception (core dumped)

OS: Win 11 +WSL
branch: Uberarm/huanyuan-moe2 ( I use this command to download the source code: git clone --branch ug/hunyuan-moe-2 https://github.com/ubergarm/ik_llama.cpp.git)
Cmake config: cmake -B ./build -DGGML_CUDA=ON -DGGML_BLAS=OFF -DGGML_SCHED_MAX_COPIES=1 -DGGML_CUDA_IQK_FORCE_BF16=1 -DGGML_CUDA_F16=ON -DGGML_CCACHE=OFF
Hardware: AMD Ryzen 5700X, 4070 12GB, WSL ram limit = 42 ( physical ram 48 Gb in total)

ubergarm · 2025-07-02T04:03:30Z

run on wsl I got a error: Floating point exception (core dumped), in the initial procress of ik_llama.cpp

Its becase I'm a madman and released a quant depending on two unmerged PRs. Check here for instructions how to get the IQ3_KS PR here: https://huggingface.co/ubergarm/Hunyuan-A13B-Instruct-GGUF#note-building-experimental-prs

Also @kiron111 look at the examples on the model card you will need to use -ngl 99 -ot exps=CPU remove that -ts stuff that was specific to my test rig.

This model is great for low VRAM machines and can probably run in 6GB VRAM with some usable context.

kiron111 · 2025-07-02T04:08:50Z

run on wsl I got a error: Floating point exception (core dumped), in the initial procress of ik_llama.cpp

Its becase I'm a madman and released a quant depending on two unmerged PRs. Check here for instructions how to get the IQ3_KS PR here: https://huggingface.co/ubergarm/Hunyuan-A13B-Instruct-GGUF#note-building-experimental-prs

Also @kiron111 look at the examples on the model card you will need to use -ngl 99 -ot exps=CPU remove that -ts stuff that was specific to my test rig.

This model is great for low VRAM machines and can probably run in 6GB VRAM with some usable context.

thankyou
oh....I miss so many points...let me redownload and recompile first

ikawrakow · 2025-07-02T07:34:31Z

The PPL of 500+ is not very promising. I suspect this is because of the not implemented technique to reduce the importance of recently used experts, which completely modifies the inference compared to how the model was trained, that was discussed in the mainline PR. Hence still wondering if to merge. They have merged as is in mainline, but ik_llama.cpp tries to be better than mainline.

ubergarm · 2025-07-02T18:58:03Z

The PPL of 500+ is not very promising. I suspect this is because of the not implemented technique to reduce the importance of recently used experts, which completely modifies the inference compared to how the model was trained, that was discussed in the mainline PR

Looking more closely, yes I see that the official pytorch reference MoE routing "capacity" mechanism is not seem implemented in the build_moe_ffn() code.

The mainline PR https://github.com/ggml-org/llama.cpp/pull/14425 seems still open for now, and yes no rush to merge this. (I've updated instructions on the hugginface model if any brave souls want to test the current implementation.)

I'll try quanting from the Pretrain version just to see how it performs, given that bf16 scores much lower PPL oddly enough:

model=Hunyuan-A13B-Pretrain-BF16-00001-of-00004.gguf
./build/bin/llama-perplexity \
        --model "$model" \
        -f wiki.test.raw \
        --seed 1337 \
        -ts 48,48 \
        -ngl 18 \
        --threads 24

Final estimate: PPL = 5.2880 +/- 0.03236

EDIT

model=Hunyuan-A13B-Pretrain-IQ3_KS.gguf
# model type       = 80B.A13B
# model ftype      = IQ3_KS - 3.1875 bpw
# model params     = 80.393 B
# model size       = 34.088 GiB (3.642 BPW)
# general.name     = Hunyuan A13B Pretrain

Final estimate: PPL = 5.4382 +/- 0.03349

ikawrakow

OK, lets merge this.

ikawrakow reviewed Jul 1, 2025

View reviewed changes

ubergarm force-pushed the ug/hunyuan-moe-2 branch from c6c23fa to 89c3210 Compare July 2, 2025 15:35

ubergarm added 3 commits July 8, 2025 17:35

add hunyuan moe

a3f0cbd

Don't reshape Vcur

a5a20a5

Apply chat template fix from mainline PR14584

90bc417

ubergarm force-pushed the ug/hunyuan-moe-2 branch from 89c3210 to 90bc417 Compare July 8, 2025 21:36

ubergarm mentioned this pull request Jul 8, 2025

Feature Request: Tencent Hunyuan-A13B model support #561

Closed

ikawrakow approved these changes Jul 9, 2025

View reviewed changes

ikawrakow merged commit 5446ccc into ikawrakow:main Jul 9, 2025

ubergarm deleted the ug/hunyuan-moe-2 branch July 15, 2025 17:57

ubergarm mentioned this pull request Aug 1, 2025

Glm 4.5 #662

Closed

4 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

add hunyuan moe support for 561 #565

add hunyuan moe support for 561 #565

ubergarm commented Jun 30, 2025 •

edited

Loading

Uh oh!

ubergarm commented Jun 30, 2025 •

edited

Loading

Uh oh!

ikawrakow commented Jun 30, 2025

Uh oh!

ubergarm commented Jun 30, 2025

Uh oh!

ikawrakow Jul 1, 2025

Uh oh!

ubergarm Jul 1, 2025

Uh oh!

kiron111 commented Jul 2, 2025

Endpoint

Test Quant

Uh oh!

ubergarm commented Jul 2, 2025 •

edited

Loading

Uh oh!

kiron111 commented Jul 2, 2025 •

edited

Loading

Uh oh!

ubergarm commented Jul 2, 2025 •

edited

Loading

Uh oh!

kiron111 commented Jul 2, 2025

Uh oh!

ikawrakow commented Jul 2, 2025

Uh oh!

ubergarm commented Jul 2, 2025 •

edited

Loading

Uh oh!

ikawrakow left a comment

Uh oh!

Uh oh!

add hunyuan moe support for 561 #565

add hunyuan moe support for 561 #565

Conversation

ubergarm commented Jun 30, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ubergarm commented Jun 30, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ikawrakow commented Jun 30, 2025

Uh oh!

ubergarm commented Jun 30, 2025

Endpoint

Test Quant

Uh oh!

ikawrakow Jul 1, 2025

Choose a reason for hiding this comment

Uh oh!

ubergarm Jul 1, 2025

Choose a reason for hiding this comment

Uh oh!

kiron111 commented Jul 2, 2025

Endpoint

Test Quant

Uh oh!

ubergarm commented Jul 2, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

kiron111 commented Jul 2, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ubergarm commented Jul 2, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

kiron111 commented Jul 2, 2025

Uh oh!

ikawrakow commented Jul 2, 2025

Uh oh!

ubergarm commented Jul 2, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ikawrakow left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

ubergarm commented Jun 30, 2025 •

edited

Loading

ubergarm commented Jun 30, 2025 •

edited

Loading

ubergarm commented Jul 2, 2025 •

edited

Loading

kiron111 commented Jul 2, 2025 •

edited

Loading

ubergarm commented Jul 2, 2025 •

edited

Loading

ubergarm commented Jul 2, 2025 •

edited

Loading