Skip to content

vulkan: fuse adds #15252

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Open
wants to merge 2 commits into
base: master
Choose a base branch
from
Open

vulkan: fuse adds #15252

wants to merge 2 commits into from
+300 −25

Conversation

jeffbolznv
Copy link
Collaborator

Fuse adds that have the same shape, which are common in MoE models. It will currently fuse up to 6 adds, because we assume no more than 8 descriptors per dispatch. But this could be changed.

5090 before:

Z:\github\jeffbolznv\llama.cpp\build\bin\RelWithDebInfo>llama-bench.exe -fa 1 -n 128 -p 512 -r 10 --prio 1 -m c:\models\bartowski\DeepSeek-Coder-V2-Lite-Instruct-GGUF\DeepSeek-Coder-V2-Lite-Instruct-Q2_K.gguf -m c:\models\Qwen_Qwen3-30B-A3B-Q4_K_M.gguf -m c:\models\gpt-oss-20b-mxfp4.gguf -m c:\models\deepseek-v2-lite-safetensors\deepseek-v2-lite-Q4_K_M.gguf
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce RTX 5090 (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: NV_coopmat2
| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
| deepseek2 16B Q2_K - Medium    |   5.99 GiB |    15.71 B | Vulkan     |  99 |  1 |           pp512 |      7544.08 ± 70.50 |
| deepseek2 16B Q2_K - Medium    |   5.99 GiB |    15.71 B | Vulkan     |  99 |  1 |           tg128 |        249.36 ± 1.58 |
| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |           pp512 |      3971.06 ± 37.61 |
| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |           tg128 |        175.16 ± 0.48 |
| gpt-oss ?B MXFP4 MoE           |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |           pp512 |     6264.83 ± 251.65 |
| gpt-oss ?B MXFP4 MoE           |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |           tg128 |        212.95 ± 0.73 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |           pp512 |     6936.34 ± 183.22 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |           tg128 |        233.17 ± 0.66 |

5090 after:

ggml_vulkan: 0 = NVIDIA GeForce RTX 5090 (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: NV_coopmat2
| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
| deepseek2 16B Q2_K - Medium    |   5.99 GiB |    15.71 B | Vulkan     |  99 |  1 |           pp512 |     7530.23 ± 127.67 |
| deepseek2 16B Q2_K - Medium    |   5.99 GiB |    15.71 B | Vulkan     |  99 |  1 |           tg128 |        263.83 ± 0.93 |
| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |           pp512 |      3999.28 ± 41.87 |
| qwen3moe 30B.A3B Q4_K - Medium |  17.35 GiB |    30.53 B | Vulkan     |  99 |  1 |           tg128 |        188.21 ± 0.95 |
| gpt-oss ?B MXFP4 MoE           |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |           pp512 |     6327.27 ± 161.58 |
| gpt-oss ?B MXFP4 MoE           |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |           tg128 |        218.28 ± 1.83 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |           pp512 |     6916.37 ± 206.58 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |           tg128 |        244.11 ± 0.78 |

4070 before:

Z:\github\jeffbolznv\llama.cpp\build\bin\RelWithDebInfo>llama-bench.exe -fa 1 -n 128 -p 512 -r 10 --prio 1 -m c:\models\bartowski\DeepSeek-Coder-V2-Lite-Instruct-GGUF\DeepSeek-Coder-V2-Lite-Instruct-Q2_K.gguf -m c:\models\gpt-oss-20b-mxfp4.gguf -m c:\models\deepseek-v2-lite-safetensors\deepseek-v2-lite-Q4_K_M.gguf
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce RTX 4070 (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: NV_coopmat2
| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
| deepseek2 16B Q2_K - Medium    |   5.99 GiB |    15.71 B | Vulkan     |  99 |  1 |           pp512 |      2777.76 ± 10.68 |
| deepseek2 16B Q2_K - Medium    |   5.99 GiB |    15.71 B | Vulkan     |  99 |  1 |           tg128 |        181.91 ± 0.37 |
| gpt-oss ?B MXFP4 MoE           |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |           pp512 |      2550.84 ± 25.51 |
| gpt-oss ?B MXFP4 MoE           |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |           tg128 |        120.20 ± 0.20 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |           pp512 |       1983.36 ± 9.40 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |           tg128 |        162.10 ± 0.29 |

4070 after:

ggml_vulkan: 0 = NVIDIA GeForce RTX 4070 (NVIDIA) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: NV_coopmat2
| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
| deepseek2 16B Q2_K - Medium    |   5.99 GiB |    15.71 B | Vulkan     |  99 |  1 |           pp512 |       2790.31 ± 7.72 |
| deepseek2 16B Q2_K - Medium    |   5.99 GiB |    15.71 B | Vulkan     |  99 |  1 |           tg128 |        188.98 ± 0.25 |
| gpt-oss ?B MXFP4 MoE           |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |           pp512 |      2562.82 ± 28.59 |
| gpt-oss ?B MXFP4 MoE           |  11.27 GiB |    20.91 B | Vulkan     |  99 |  1 |           tg128 |        121.95 ± 0.28 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |           pp512 |       1984.36 ± 9.35 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |           tg128 |        166.07 ± 1.04 |

@jeffbolznv
vulkan: fuse adds
8a2e4f7 
Fuse adds that have the same shape, which are common in MoE models.
It will currently fuse up to 6 adds, because we assume no more than
8 descriptors per dispatch. But this could be changed.
@jeffbolznv jeffbolznv requested a review from 0cc4m as a code owner August 11, 2025 21:14
@github-actions github-actions bot added testing Everything test related Vulkan Issues specific to the Vulkan backend ggml changes relating to the ggml tensor library for machine learning labels Aug 11, 2025
0cc4m
0cc4m reviewed Aug 12, 2025
View reviewed changes
Copy link
Collaborator

@0cc4m 0cc4m left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looks good on AMD and Nvidia, but I can't get it to run on Intel.

terminate called after throwing an instance of 'vk::DeviceLostError'
  what():  vk::Device::waitForFences: ErrorDeviceLost

I'll investigate further later.

ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp
Comment on lines +20 to +21
layout (binding = 0) readonly buffer A {A_TYPE data_a[];} a[];
layout (binding = 0) writeonly buffer D {D_TYPE data_d[];} d[];
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Does this not require any support checks on the API side? There seems to be an endless source of obscure GLSL extensions that solve very specific issues. Neat.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It requires the runtimeDescriptorArray feature from the descriptor_indexing extension. This is required to be supported in Vulkan 1.2 so I don't think the check is strictly required, but I'll add it since it's simple.

@jeffbolznv
Copy link
Collaborator Author

Looks good on AMD and Nvidia, but I can't get it to run on Intel.

Strange. Any validation failures? Does the backend test fail, or just in real models?

@jeffbolznv jeffbolznv mentioned this pull request Aug 13, 2025
Draft
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@0cc4m 0cc4m 0cc4m left review comments

At least 1 approving review is required to merge this pull request.

Assignees
No one assigned
Labels
ggml changes relating to the ggml tensor library for machine learning testing Everything test related Vulkan Issues specific to the Vulkan backend
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
2 participants
@jeffbolznv @0cc4m