Skip to content

vulkan: optimizations for deepseek prompt processing #14555

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
Merged
merged 4 commits into from
Jul 12, 2025
Merged

vulkan: optimizations for deepseek prompt processing #14555

merged 4 commits into from
Jul 12, 2025

Conversation

jeffbolznv
Copy link
Collaborator

Some optimizations for mul_mat_id, and flash attention with large head size. See commit messages for more detail.

before:

coopmat2:

Z:\github\jeffbolznv\llama.cpp\build\bin\RelWithDebInfo>llama-bench -m C:\models\deepseek-v2-lite-safetensors\deepseek-v2-lite-Q4_K_M.gguf -fa 1 -n 128 -p 512 -d 512,8192 --prio 1 -r 5
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce RTX 5090 (NVIDIA) | uma: 0 | fp16: 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 Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |    pp512 @ d512 |      4061.98 ± 17.67 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |    tg128 @ d512 |        210.15 ± 0.88 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |   pp512 @ d8192 |      2392.49 ± 16.93 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |   tg128 @ d8192 |        116.90 ± 0.26 |

coopmat1:

Z:\github\jeffbolznv\llama.cpp\build\bin\RelWithDebInfo>llama-bench -m C:\models\deepseek-v2-lite-safetensors\deepseek-v2-lite-Q4_K_M.gguf -fa 1 -n 128 -p 512 -d 512,8192 --prio 1 -r 1
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce RTX 5090 (NVIDIA) | uma: 0 | fp16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: KHR_coopmat
| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |    pp512 @ d512 |       1013.83 ± 0.00 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |    tg128 @ d512 |        203.44 ± 0.00 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |   pp512 @ d8192 |        159.70 ± 0.00 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |   tg128 @ d8192 |        151.71 ± 0.00 |

scalar:

Z:\github\jeffbolznv\llama.cpp\build\bin\RelWithDebInfo>llama-bench -m C:\models\deepseek-v2-lite-safetensors\deepseek-v2-lite-Q4_K_M.gguf -fa 1 -n 128 -p 512 -d 512,8192 --prio 1 -r 1
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce RTX 5090 (NVIDIA) | uma: 0 | fp16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: none
| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |    pp512 @ d512 |        868.16 ± 0.00 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |    tg128 @ d512 |        202.75 ± 0.00 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |   pp512 @ d8192 |        161.14 ± 0.00 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |   tg128 @ d8192 |        151.54 ± 0.00 |

after:

coopmat2:

Z:\github\jeffbolznv\llama.cpp\build\bin\RelWithDebInfo>llama-bench -m C:\models\deepseek-v2-lite-safetensors\deepseek-v2-lite-Q4_K_M.gguf -fa 1 -n 128 -p 512 -d 512,8192 --prio 1 -r 5
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce RTX 5090 (NVIDIA) | uma: 0 | fp16: 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 Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |    pp512 @ d512 |      6419.30 ± 17.34 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |    tg128 @ d512 |        209.36 ± 0.86 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |   pp512 @ d8192 |      3606.39 ± 10.05 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |   tg128 @ d8192 |        116.51 ± 0.28 |

coopmat1:

Z:\github\jeffbolznv\llama.cpp\build\bin\RelWithDebInfo>llama-bench -m C:\models\deepseek-v2-lite-safetensors\deepseek-v2-lite-Q4_K_M.gguf -fa 1 -n 128 -p 512 -d 512,8192 --prio 1 -r 5
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce RTX 5090 (NVIDIA) | uma: 0 | fp16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: KHR_coopmat
| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |    pp512 @ d512 |      5412.19 ± 58.07 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |    tg128 @ d512 |        202.60 ± 0.77 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |   pp512 @ d8192 |       1630.19 ± 3.38 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |   tg128 @ d8192 |        151.86 ± 0.46 |

scalar:

Z:\github\jeffbolznv\llama.cpp\build\bin\RelWithDebInfo>llama-bench -m C:\models\deepseek-v2-lite-safetensors\deepseek-v2-lite-Q4_K_M.gguf -fa 1 -n 128 -p 512 -d 512,8192 --prio 1 -r 5
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce RTX 5090 (NVIDIA) | uma: 0 | fp16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: none
| model                          |       size |     params | backend    | ngl | fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | --------------: | -------------------: |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |    pp512 @ d512 |       1833.69 ± 7.71 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |    tg128 @ d512 |        204.14 ± 0.45 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |   pp512 @ d8192 |       1034.34 ± 1.45 |
| deepseek2 16B Q4_K - Medium    |   9.65 GiB |    15.71 B | Vulkan     |  99 |  1 |   tg128 @ d8192 |        151.88 ± 0.27 |

@jeffbolznv jeffbolznv requested a review from 0cc4m July 6, 2025 22:36
@github-actions github-actions bot added Vulkan Issues specific to the Vulkan backend ggml changes relating to the ggml tensor library for machine learning labels Jul 6, 2025
0cc4m
0cc4m approved these changes Jul 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.

LGTM

@0cc4m 0cc4m merged commit 98197e5 into ggml-org:master Jul 12, 2025
47 of 48 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@0cc4m 0cc4m 0cc4m approved these changes

Assignees
No one assigned
Labels
ggml changes relating to the ggml tensor library for machine learning 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