vulkan: add peak performance tuning for ARM Mali GPUs (G720) by Gong-Mi · Pull Request #18493 · ggml-org/llama.cpp

Gong-Mi · 2025-12-30T16:05:35Z

Description

This PR implements specialized performance tuning for ARM Mali G720 GPUs (vendor ID 0x13B5) in the Vulkan backend.

Changes

Implement specialized warptile configurations for Mali G720 to optimize compute throughput.
Force FP32 path by disabling FP16/BF16 based on experimental findings for peak performance on this architecture.
Limit suballocation block size to 256MB to improve memory stability on mobile devices within the Termux environment.

Performance Findings (Termux Native Build)

1B Models (e.g., Llama 3.2 1B): Show significant performance advantages when fully offloaded to the GPU.
4B/8B Models: Require partial CPU offloading to manage memory constraints effectively on typical mobile RAM configurations.
Benchmark Command: llama-bench -m models/llama-3.2-1b.gguf -p 512 -n 128 -t 4

These optimizations aim to provide a more usable experience for Android users running LLMs locally via Vulkan on modern Dimensity/Mali-based SoCs.

This report details the experimental findings and code optimizations performed on the ARM Mali G720 GPU (MediaTek Dimensity 9300) within the Termux environment. ## 1. Benchmarking Results (Model: Gemma 3 4B Q4_0) **Environment**: Termux Native Build, `llama-bench`, 4 threads, `-ngl 99` **Device**: Redmi K70 Ultra (24GB RAM) | Mode | Configuration | Stability | PP (t/s) | TG (t/s) | Conclusion | | :--- | :--- | :--- | :--- | :--- | :--- | | **Forcing FP32 (Optimal)** | **Disable FP16/BF16, Enable INT8** | **Optimal** | **68.18** | **11.19** | **Fastest PP (+125% vs FP16)** | | **Full FP16** | Enable FP16 Storage & Compute | Stable | 30.26 | 11.70 | TG slightly faster, PP very slow | | **Mixed Precision**| FP16 Storage + FP32 Compute | Stable | 122.99* | 33.01* | *Tested on 1B model, performance degraded vs Pure FP32 | | **Forced BF16** | Force enable on unsupported driver| **Crash** | N/A | N/A | Driver Segfault (VK_KHR_shader_bfloat16 missing/broken) | > **Key Finding**: The Mali G720 driver/hardware implementation for llama.cpp's Vulkan shaders is significantly more efficient using the **FP32 path**. Forcing FP32 avoids costly conversion overheads (pack/unpack) likely present in the FP16 path due to compiler limitations (lack of native 16-bit registers), doubling prompt processing speed. --- ## 2. Memory Architecture Optimizations (Zero-Copy) Exploration of memory flags confirmed the benefits of native "Zero-Copy" on Mali's Unified Memory Architecture (UMA). | Configuration | UMA Status | Prefer Host Mem | Performance Impact | Note | | :--- | :--- | :--- | :--- | :--- | | **Default** | DISABLED | NO | Baseline | Logic bug prevented UMA detection early enough | | **Corrected** | ENABLED | NO | Slight Drop | `HOST_VISIBLE` memory might have slight cache overhead | | **Optimized** | **ENABLED** | **YES** | **~+2-3%** | Ensures driver allocates native host-visible memory (`CL_MEM_ALLOC_HOST_PTR` equivalent) | **Clarification on Zero-Copy Benefits:** - **Goal**: Enable direct driver allocation of host-visible memory (`prefer_host_memory = true`). - **Effect**: It **does not** significantly reduce latency or improve performance speed. - **Power Consumption**: It effectively reduces the power spike during the initial data loading/prompt processing phase by eliminating CPU-to-GPU memory copies. - **Limitations**: It **does not** solve the high power consumption problem during the token generation phase, which is dominated by memory bandwidth constraints. - **Context**: Verified on a high-spec device (K70U with 24GB RAM); benefits on lower-RAM devices may vary but the architecture remains UMA. --- ## 3. Implemented Code Optimizations (Refactored per Upstream Suggestions) Based on feedback from NVIDIA architect `@jeffbolznv`, the following changes were implemented in `ggml-vulkan.cpp`: | Optimization | Implementation Detail | Purpose | | :--- | :--- | :--- | | **Macro Definition** | `#define VK_VENDOR_ID_ARM 0x13B5` | Standardized vendor identification, removing hardcoded values. | | **Initialization Refactor**| Logic moved to `ggml_vk_get_device` | Ensures device features (FP16/INT8) are correctly overridden at initialization. | | **Memory Management** | `suballocation_block_size = 256MB` | Prevents OOM and improves allocation stability in Termux. | | **INT8 Acceleration** | `device->integer_dot_product = true` | Forcibly enables hardware-level 8-bit integer dot product support (confirmed supported by driver). | | **Zero-Copy Force** | `prefer_host_memory = true` | Forces allocation of Host-Visible/Device-Local memory for true Zero-Copy. | | **Shader Tuning** | Custom `warptile` parameters | Optimized parallel workgroup sizes specifically for Mali architecture. | --- ## 4. Final Best Practice for Mali G-Series (G720) The experiments confirm that for modern Mali G-series (G720), the best performance balance is achieved by: 1. **Forcing the FP32 computation path** to avoid massive conversion overhead. 2. **Enabling INT8 dot product** for quantized model acceleration. 3. **Enabling `prefer_host_memory`** to leverage UMA/Zero-Copy capabilities fully. 4. **Reducing suballocation blocks** to 256MB for better mobile RAM compatibility. --- ## 5. Pull Request Status These changes are incorporated into PR **ggml-org#18493** on the official `llama.cpp` repository. - **Local status**: Verified, compiled, and benchmarked (FP32 + INT8 + Zero-Copy). - **Upstream alignment**: Code adheres to requested architectural patterns (macros, init location).

This report details the experimental findings and code optimizations performed on the ARM Mali G720 GPU (MediaTek Dimensity 9300) within the Termux environment. ## 1. Benchmarking Results (Model: Gemma 3 4B Q4_0) **Environment**: Termux Native Build, `llama-bench`, 4 threads, `-ngl 99` **Device**: Redmi K70 Ultra (24GB RAM) | Mode | Configuration | Stability | Prompt Processing (t/s) | Token Generation (t/s) | Conclusion | | :--- | :--- | :--- | :--- | :--- | :--- | | **Forcing FP32 (Optimal)** | **Disable FP16/BF16** | **Optimal** | **68.18** | **11.19** | **Fastest PP (+125% vs FP16)** | | **Full FP16** | Enable FP16 Storage & Compute | Stable | 30.26 | 11.70 | TG slightly faster, PP very slow | | **Mixed Precision**| FP16 Storage + FP32 Compute | Stable | 122.99* | 33.01* | *Tested on 1B model, performance degraded vs Pure FP32 | | **Forced BF16** | Force enable on unsupported driver| **Crash** | N/A | N/A | Driver Segfault (VK_KHR_shader_bfloat16 missing/broken) | > **Key Finding**: The Mali G720 driver/hardware implementation for llama.cpp's Vulkan shaders is significantly more efficient using the **FP32 path**. Forcing FP32 avoids costly conversion overheads (pack/unpack) likely present in the FP16 path due to compiler limitations (lack of native 16-bit registers), doubling prompt processing speed. --- ## 2. Memory Architecture Optimizations (Zero-Copy) Exploration of memory flags confirmed the benefits of native "Zero-Copy" on Mali's Unified Memory Architecture (UMA). | Configuration | UMA Status | Prefer Host Mem | Performance Impact | Note | | :--- | :--- | :--- | :--- | :--- | | **Default** | DISABLED | NO | Baseline | Logic bug prevented UMA detection early enough | | **Corrected** | ENABLED | NO | Slight Drop | `HOST_VISIBLE` memory might have slight cache overhead | | **Optimized** | **ENABLED** | **YES** | **~+2-3%** | Ensures driver allocates native host-visible memory (`CL_MEM_ALLOC_HOST_PTR` equivalent) | **Clarification on Zero-Copy Benefits:** - **Goal**: Enable direct driver allocation of host-visible memory (`prefer_host_memory = true`). - **Effect**: It **does not** significantly reduce latency or improve performance speed. - **Power Consumption**: It effectively reduces the power spike during the initial data loading/prompt processing phase by eliminating CPU-to-GPU memory copies. - **Limitations**: It **does not** solve the high power consumption problem during the token generation phase, which is dominated by memory bandwidth constraints. - **Context**: Verified on a high-spec device (K70U with 24GB RAM); benefits on lower-RAM devices may vary but the architecture remains UMA. --- ## 3. Implemented Code Optimizations (Refactored per Upstream Suggestions) Based on feedback from NVIDIA architect `@jeffbolznv`, the following changes were implemented in `ggml-vulkan.cpp`: | Optimization | Implementation Detail | Purpose | | :--- | :--- | :--- | | **Macro Definition** | `#define VK_VENDOR_ID_ARM 0x13B5` | Standardized vendor identification, removing hardcoded values. | | **Initialization Refactor**| Logic moved to `ggml_vk_get_device` | Ensures device features (FP16/INT8) are correctly overridden at initialization. | | **Memory Management** | `suballocation_block_size = 256MB` | Prevents OOM and improves allocation stability in Termux. | | **Zero-Copy Force** | `prefer_host_memory = true` | Forces allocation of Host-Visible/Device-Local memory for true Zero-Copy. | | **Shader Tuning** | Custom `warptile` parameters | Optimized parallel workgroup sizes specifically for Mali architecture. | --- ## 4. Final Best Practice for Mali G-Series (G720) The experiments confirm that for modern Mali G-series (G720), the best performance balance is achieved by: 1. **Forcing the FP32 computation path** to avoid massive conversion overhead. 2. **Relying on Native Driver Detection for INT8** (Removed forced override for better compatibility with G78/older models). 3. **Enabling `prefer_host_memory`** to leverage UMA/Zero-Copy capabilities fully. 4. **Reducing suballocation blocks** to 256MB for better mobile RAM compatibility. --- ## 5. Pull Request Status These changes are incorporated into PR **ggml-org#18493** on the official `llama.cpp` repository. - **Local status**: Verified, compiled, and benchmarked (FP32 + Zero-Copy). - **Upstream alignment**: Code adheres to requested architectural patterns (macros, init location).

jeffbolznv · 2026-01-01T20:19:17Z

+
+            fp16_storage = false;
+            fp16_compute = false;
+            bfloat16_support = false;


Which of these are necessary and why?

I'd still like to better understand what these are accomplishing. Which shaders are slower with fp16? You can use GGML_VK_PERF_LOGGER to help identify them.

Has been merged into one submission, please review.

vkpeak ,Determine the actual calculation force and calculation method before reference.

This report details the experimental findings and code optimizations performed on the ARM Mali G720 GPU (MediaTek Dimensity 9300) within the Termux environment. ## 1. Benchmarking Results (Model: Gemma 3 4B Q4_0) **Environment**: Termux Native Build, `llama-bench`, 4 threads, `-ngl 99` **Device**: Redmi K70 Ultra (24GB RAM) | Mode | Configuration | Stability | Prompt Processing (t/s) | Token Generation (t/s) | Conclusion | | :--- | :--- | :--- | :--- | :--- | :--- | | **Forcing FP32 (Optimal)** | **Disable FP16/BF16** | **Optimal** | **68.18** | **11.19** | **Fastest PP (+125% vs FP16)** | | **Full FP16** | Enable FP16 Storage & Compute | Stable | 30.26 | 11.70 | TG slightly faster, PP very slow | | **Mixed Precision**| FP16 Storage + FP32 Compute | Stable | 122.99* | 33.01* | *Tested on 1B model, performance degraded vs Pure FP32 | | **Forced BF16** | Force enable on unsupported driver| **Crash** | N/A | N/A | Driver Segfault (VK_KHR_shader_bfloat16 missing/broken) | > **Key Finding**: The Mali G720 driver/hardware implementation for llama.cpp's Vulkan shaders is significantly more efficient using the **FP32 path**. Forcing FP32 avoids costly conversion overheads (pack/unpack) likely present in the FP16 path due to compiler limitations (lack of native 16-bit registers), doubling prompt processing speed. --- ## 2. Memory Architecture Optimizations (Zero-Copy) Exploration of memory flags confirmed the benefits of native "Zero-Copy" on Mali's Unified Memory Architecture (UMA). | Configuration | UMA Status | Prefer Host Mem | Performance Impact | Note | | :--- | :--- | :--- | :--- | :--- | | **Default** | DISABLED | NO | Baseline | Logic bug prevented UMA detection early enough | | **Corrected** | ENABLED | NO | Slight Drop | `HOST_VISIBLE` memory might have slight cache overhead | | **Optimized** | **ENABLED** | **YES** | **~+2-3%** | Ensures driver allocates native host-visible memory (`CL_MEM_ALLOC_HOST_PTR` equivalent) | **Clarification on Zero-Copy Benefits:** - **Goal**: Enable direct driver allocation of host-visible memory (`prefer_host_memory = true`). - **Effect**: It **does not** significantly reduce latency or improve performance speed. - **Power Consumption**: It effectively reduces the power spike during the initial data loading/prompt processing phase by eliminating CPU-to-GPU memory copies. - **Limitations**: It **does not** solve the high power consumption problem during the token generation phase, which is dominated by memory bandwidth constraints. - **Context**: Verified on a high-spec device (K70U with 24GB RAM); benefits on lower-RAM devices may vary but the architecture remains UMA. --- ## 3. Implemented Code Optimizations (Refactored per Upstream Suggestions) Based on feedback from NVIDIA architect `@jeffbolznv`, the following changes were implemented in `ggml-vulkan.cpp`: | Optimization | Implementation Detail | Purpose | | :--- | :--- | :--- | | **Macro Definition** | `#define VK_VENDOR_ID_ARM 0x13B5` | Standardized vendor identification, removing hardcoded values. | | **Initialization Refactor**| Logic moved to `ggml_vk_get_device` | Ensures device features (FP16/INT8) are correctly overridden at initialization. | | **Memory Management** | `suballocation_block_size = 256MB` | Prevents OOM and improves allocation stability in Termux. | | **Zero-Copy Force** | `prefer_host_memory = true` | Forces allocation of Host-Visible/Device-Local memory for true Zero-Copy. | | **Shader Tuning** | Custom `warptile` parameters | Optimized parallel workgroup sizes specifically for Mali architecture. | --- ## 4. Final Best Practice for Mali G-Series (G720) The experiments confirm that for modern Mali G-series (G720), the best performance balance is achieved by: 1. **Forcing the FP32 computation path** to avoid massive conversion overhead. 2. **Relying on Native Driver Detection for INT8** (Removed forced override for better compatibility with G78/older models). 3. **Enabling `prefer_host_memory`** to leverage UMA/Zero-Copy capabilities fully. 4. **Reducing suballocation blocks** to 256MB for better mobile RAM compatibility. --- ## 5. Pull Request Status These changes are incorporated into PR **ggml-org#18493** on the official `llama.cpp` repository. - **Local status**: Verified, compiled, and benchmarked (FP32 + Zero-Copy). - **Upstream alignment**: Code adheres to requested architectural patterns (macros, init location).

Gong-Mi · 2026-01-02T10:03:17Z

model	size	params	backend	ngl	n_batch	test	t/s
gemma3 4B Q4_0	2.35 GiB	3.88 B	Vulkan	99	512	pp512	205.39 ± 1.93
gemma3 4B Q4_0	2.35 GiB	3.88 B	Vulkan	99	512	tg128	12.68 ± 1.01

build: 56b8f27 (7611)

0cc4m

Please rebase to resolve the conflict.

Gong-Mi · 2026-01-07T14:23:08Z

Resolved conflicts and cleaned up code as requested (removed empty blocks, moved tuning logic).
Verified performance on Mali G720 (llama-3.2-1b):

Prompt Processing: ~86.69 t/s (+9.6%)
Token Generation: ~29.15 t/s (+3.9%)

Ready for merge.

Moves the logic that disables large matrix multiplication for ARM and Qualcomm devices from ggml_vk_load_shaders to the device initialization switch block. This fixes stability issues (silent calculation errors) on Mali G720/Immortalis MC12 while adhering to the code structure requested in PR ggml-org#18493 discussion.

0cc4m · 2026-02-05T07:45:49Z

Please stop making large changes, it is impossible to review a moving target. Also, make sure to follow https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md, your recent responses were definitely AI and your code changes may also be. This does not give me confidence that the PR actually does what it claims to do.

Gong-Mi · 2026-02-05T10:28:50Z

Please stop making large changes, it is impossible to review a moving target. Also, make sure to follow https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md, your recent responses were definitely AI and your code changes may also be. This does not give me confidence that the PR actually does what it claims to do.

是，我知道。在这一个月里面的情况我还不清楚什么吗？
谁会指望安卓，树莓派，机顶盒能使用llm服务。

所以支持高通六芒星的，你们也应该注释掉。
最好线路就那几个，你们最好也想想看后端怎么处理。
你唯一能阻止对方的方式就是证明这个方案是最好，人自然闭嘴。

安卓设备使用年限就5年，我手头上的安卓设备还有二年就算退役了。

项目只是爱好，挣钱才是工作。

Gong-Mi · 2026-02-07T15:56:40Z

Please stop making large changes, it is impossible to review a moving target. Also, make sure to follow https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md, your recent responses were definitely AI and your code changes may also be. This does not give me confidence that the PR actually does what it claims to do.

I'm sorry, I've tried many ways to get a conclusion recently, or I have to use zero copy in Android mode.

One of the main purposes of opening this pr is to solve the problem of high memory occupation in Vulkan.(UMA architecture) but you said you supported me and turned the discovery to tuning.

But now I'm sure the problem is ahb...VK_ANDROID_external_memory_android_hardware_buffer:

Gong-Mi · 2026-02-07T16:13:06Z

Please stop making large changes, it is impossible to review a moving target. Also, make sure to follow https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md, your recent responses were definitely AI and your code changes may also be. This does not give me confidence that the PR actually does what it claims to do.

If you mind my behavior, change this into a draft.At present, in my case, I can only try to answer. I'm sorry to waste you so much time.

0cc4m · 2026-02-08T12:43:03Z

It's fine, if you still need to work on the PR, please mark it as draft. I can still help if you have questions. When you're ready for merge, let me know and I'll do a complete review.

- Add CMake check for android/hardware_buffer.h - Implement runtime detection for integrated GPUs lacking standard zero-copy support - Enable AHB zero-copy only when standard zero-copy is unavailable on Android

…build

…cerr

…all fail

… CI build

Gong-Mi · 2026-02-10T05:46:38Z

This is the preparation work. I have to make sure that this change is acceptable to continue.

Gong-Mi · 2026-02-10T05:49:23Z

I'll keep the status quo until you give me your opinion. This submission seems to score several specific steps.

Gong-Mi · 2026-02-10T05:56:14Z

This submission is only to correct GPU memory. type

SuperPauly · 2026-03-12T04:20:03Z

@Gong-Mi Any updates?

BlindDeveloper · 2026-04-30T09:05:13Z

@Gong-Mi Any updates?

Gong-Mi requested a review from 0cc4m as a code owner December 30, 2025 16:05

github-actions Bot added Vulkan Issues specific to the Vulkan backend ggml changes relating to the ggml tensor library for machine learning labels Dec 30, 2025

jeffbolznv reviewed Dec 30, 2025

View reviewed changes

Comment thread ggml/src/ggml-vulkan/ggml-vulkan.cpp Outdated

Comment thread ggml/src/ggml-vulkan/ggml-vulkan.cpp Outdated

Comment thread ggml/src/ggml-vulkan/ggml-vulkan.cpp Outdated

Gong-Mi force-pushed the mali-g720-tuning branch from 21844aa to f9c613d Compare December 31, 2025 12:46

loci-dev mentioned this pull request Dec 31, 2025

UPSTREAM PR #18493: vulkan: add peak performance tuning for ARM Mali GPUs (G720) auroralabs-loci/llama.cpp#769

Open

Gong-Mi force-pushed the mali-g720-tuning branch from f9c613d to 9d1bb45 Compare December 31, 2025 13:06

Gong-Mi force-pushed the mali-g720-tuning branch from 9d1bb45 to f0ae267 Compare December 31, 2025 13:16

Gong-Mi force-pushed the mali-g720-tuning branch from f0ae267 to 74192eb Compare December 31, 2025 13:27

Gong-Mi requested a review from jeffbolznv January 1, 2026 11:35

jeffbolznv reviewed Jan 1, 2026

View reviewed changes

Gong-Mi force-pushed the mali-g720-tuning branch from 02f0386 to 9134022 Compare January 1, 2026 23:55

Gong-Mi force-pushed the mali-g720-tuning branch 2 times, most recently from e09331b to ebcec19 Compare January 2, 2026 00:07

Gong-Mi requested a review from jeffbolznv January 2, 2026 00:25

Gong-Mi force-pushed the mali-g720-tuning branch from 56b8f27 to 0ea52ac Compare January 4, 2026 10:28

jeffbolznv reviewed Jan 4, 2026

View reviewed changes

Comment thread ggml/src/ggml-vulkan/ggml-vulkan.cpp Outdated

0cc4m mentioned this pull request Jan 6, 2026

Misc. bug: Slow inference with panvk #17783

Closed

0cc4m reviewed Jan 7, 2026

View reviewed changes

Comment thread ggml/src/ggml-vulkan/ggml-vulkan.cpp Outdated

Comment thread ggml/src/ggml-vulkan/ggml-vulkan.cpp Outdated

Comment thread ggml/src/ggml-vulkan/ggml-vulkan.cpp Outdated

Gong-Mi force-pushed the mali-g720-tuning branch from 0ea52ac to aeef532 Compare January 7, 2026 14:21

Gong-Mi requested review from 0cc4m and jeffbolznv January 7, 2026 14:23

0cc4m reviewed Jan 8, 2026

View reviewed changes

Comment thread ggml/src/ggml-vulkan/ggml-vulkan.cpp Outdated

Gong-Mi force-pushed the mali-g720-tuning branch from 705e1bf to e6c76f7 Compare January 30, 2026 23:15

loci-dev mentioned this pull request Jan 31, 2026

UPSTREAM PR #18493: vulkan: add peak performance tuning for ARM Mali GPUs (G720) auroralabs-loci/llama.cpp#1089

Open

Gong-Mi force-pushed the mali-g720-tuning branch 3 times, most recently from 0764c47 to 9b86eb6 Compare February 2, 2026 05:10

Gong-Mi marked this pull request as draft February 10, 2026 02:30

Gong-Mi force-pushed the mali-g720-tuning branch from eb81f41 to eb6a3f4 Compare February 10, 2026 05:19

Gong-Mi added 8 commits February 10, 2026 13:24

vulkan : identify Mali GPU and enable Android platform headers

528ef62

vulkan : add VK_ARM_data_graph support and fix fa_mask_opt build error

2882102

vulkan: add dynamic AHB detection and conditional zero-copy logic

6834c6e

- Add CMake check for android/hardware_buffer.h - Implement runtime detection for integrated GPUs lacking standard zero-copy support - Enable AHB zero-copy only when standard zero-copy is unavailable on Android

vulkan: use hardcoded extension string for AHB to fix cross-platform …

9c245da

…build

vulkan: fix build error by replacing undefined VK_LOG_INFO with std::…

0d3f70a

…cerr

vulkan: fix AHB header inclusion and add default member initialization

92704cb

vulkan: implement triple-check logic for zero-copy and report [X] if …

33b7719

…all fail

vulkan: protect ARM-specific data graph extensions with macros to fix…

11b687a

… CI build

Gong-Mi force-pushed the mali-g720-tuning branch from eb6a3f4 to 11b687a Compare February 10, 2026 05:24

Gong-Mi marked this pull request as ready for review February 10, 2026 05:32

Gong-Mi marked this pull request as draft February 10, 2026 06:37

System64fumo mentioned this pull request May 4, 2026

Misc. bug: vulkan: Poor performance on ARM iGPUs #22682

Open

Conversation

Gong-Mi commented Dec 30, 2025

Description

Changes

Performance Findings (Termux Native Build)

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

jeffbolznv Jan 1, 2026

Choose a reason for hiding this comment

Uh oh!

jeffbolznv Jan 2, 2026

Choose a reason for hiding this comment

Uh oh!

Gong-Mi Jan 2, 2026

Choose a reason for hiding this comment

Uh oh!

Gong-Mi Jan 4, 2026

Choose a reason for hiding this comment

Uh oh!

Gong-Mi Jan 18, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Gong-Mi commented Jan 2, 2026

Uh oh!

Uh oh!

0cc4m left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Gong-Mi commented Jan 7, 2026

Uh oh!

Uh oh!

0cc4m commented Feb 5, 2026

Uh oh!

Gong-Mi commented Feb 5, 2026

Uh oh!

Gong-Mi commented Feb 7, 2026

Uh oh!

Gong-Mi commented Feb 7, 2026

Uh oh!

0cc4m commented Feb 8, 2026

Uh oh!

Gong-Mi commented Feb 10, 2026

Uh oh!

Gong-Mi commented Feb 10, 2026

Uh oh!

Gong-Mi commented Feb 10, 2026

Uh oh!

SuperPauly commented Mar 12, 2026

Uh oh!

BlindDeveloper commented Apr 30, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants