Skip to content

ggml : drop support for QK_K=64 #7473

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.

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

Merged
merged 2 commits into from
May 23, 2024
Merged

ggml : drop support for QK_K=64 #7473

merged 2 commits into from
May 23, 2024

Conversation

ggerganov
Copy link
Member

Support only super-block size of 256

@github-actions github-actions bot added build Compilation issues Nvidia GPU Issues specific to Nvidia GPUs python python script changes devops improvements to build systems and github actions ggml changes relating to the ggml tensor library for machine learning SYCL https://en.wikipedia.org/wiki/SYCL - GPU programming language labels May 22, 2024
Copy link
Contributor

📈 llama.cpp server for bench-server-baseline on Standard_NC4as_T4_v3 for phi-2-q4_0: 533 iterations 🚀

Expand details for performance related PR only
  • Concurrent users: 8, duration: 10m
  • HTTP request : avg=8765.81ms p(95)=22675.83ms fails=, finish reason: stop=483 truncated=50
  • Prompt processing (pp): avg=104.94tk/s p(95)=449.68tk/s
  • Token generation (tg): avg=38.75tk/s p(95)=49.16tk/s
  • ggml-org/models/phi-2/ggml-model-q4_0.gguf parallel=8 ctx-size=16384 ngl=33 batch-size=2048 ubatch-size=256 pp=1024 pp+tg=2048 branch=gg/drop-qk_k-64 commit=a8dd83b5d8316823033ac961feb9a878700b57bc

prompt_tokens_seconds

More
---
config:
    xyChart:
        titleFontSize: 12
        width: 900
        height: 600
    themeVariables:
        xyChart:
            titleColor: "#000000"
---
xychart-beta
    title "llama.cpp bench-server-baseline on Standard_NC4as_T4_v3
 duration=10m 533 iterations"
    y-axis "llamacpp:prompt_tokens_seconds"
    x-axis "llamacpp:prompt_tokens_seconds" 1716414680 --> 1716415306
    line [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 299.61, 299.61, 299.61, 299.61, 299.61, 748.34, 748.34, 748.34, 748.34, 748.34, 800.86, 800.86, 800.86, 800.86, 800.86, 818.15, 818.15, 818.15, 818.15, 818.15, 890.56, 890.56, 890.56, 890.56, 890.56, 872.81, 872.81, 872.81, 872.81, 872.81, 872.63, 872.63, 872.63, 872.63, 872.63, 885.4, 885.4, 885.4, 885.4, 885.4, 897.48, 897.48, 897.48, 897.48, 897.48, 894.74, 894.74, 894.74, 894.74, 894.74, 915.45, 915.45, 915.45, 915.45, 915.45, 823.11, 823.11, 823.11, 823.11, 823.11, 834.06, 834.06, 834.06, 834.06, 834.06, 853.59, 853.59, 853.59, 853.59, 853.59, 839.24, 839.24, 839.24, 839.24, 839.24, 840.14, 840.14, 840.14, 840.14, 840.14, 840.95, 840.95, 840.95, 840.95, 840.95, 839.18, 839.18, 839.18, 839.18, 839.18, 841.4, 841.4, 841.4, 841.4, 841.4, 840.38, 840.38, 840.38, 840.38, 840.38, 846.88, 846.88, 846.88, 846.88, 846.88, 846.96, 846.96, 846.96, 846.96, 846.96, 849.91, 849.91, 849.91, 849.91, 849.91, 851.95, 851.95, 851.95, 851.95, 851.95, 850.65, 850.65, 850.65, 850.65, 850.65, 852.04, 852.04, 852.04, 852.04, 852.04, 860.63, 860.63, 860.63, 860.63, 860.63, 858.32, 858.32, 858.32, 858.32, 858.32, 856.23, 856.23, 856.23, 856.23, 856.23, 859.17, 859.17, 859.17, 859.17, 859.17, 861.38, 861.38, 861.38, 861.38, 861.38, 859.79, 859.79, 859.79, 859.79, 859.79, 859.71, 859.71, 859.71, 859.71, 859.71, 862.28, 862.28, 862.28, 862.28, 862.28, 867.46, 867.46, 867.46, 867.46, 867.46, 866.27, 866.27, 866.27, 866.27, 866.27, 866.46, 866.46, 866.46, 866.46, 866.46, 869.38, 869.38, 869.38, 869.38, 869.38, 867.08, 867.08, 867.08, 867.08, 867.08, 866.66, 866.66, 866.66, 866.66, 866.66, 869.12, 869.12, 869.12, 869.12, 869.12, 871.59, 871.59, 871.59, 871.59, 871.59, 881.29, 881.29, 881.29, 881.29, 881.29, 885.13, 885.13, 885.13, 885.13, 885.13, 884.32, 884.32, 884.32, 884.32, 884.32, 881.01, 881.01, 881.01, 881.01, 881.01, 878.56, 878.56, 878.56, 878.56, 878.56, 876.27, 876.27, 876.27, 876.27, 876.27, 879.29, 879.29, 879.29, 879.29, 879.29, 881.11, 881.11, 881.11, 881.11, 881.11, 880.17, 880.17, 880.17, 880.17, 880.17, 881.65, 881.65, 881.65, 881.65, 881.65, 881.9, 881.9, 881.9, 881.9, 881.9, 884.25, 884.25, 884.25, 884.25, 884.25, 884.06, 884.06, 884.06, 884.06, 884.06, 885.67, 885.67, 885.67, 885.67, 885.67, 889.36, 889.36, 889.36, 889.36, 889.36, 889.52, 889.52, 889.52, 889.52, 889.52, 889.11, 889.11, 889.11, 889.11, 889.11, 888.88, 888.88, 888.88, 888.88, 888.88, 890.45, 890.45, 890.45, 890.45]
                    
Loading
predicted_tokens_seconds
More
---
config:
    xyChart:
        titleFontSize: 12
        width: 900
        height: 600
    themeVariables:
        xyChart:
            titleColor: "#000000"
---
xychart-beta
    title "llama.cpp bench-server-baseline on Standard_NC4as_T4_v3
 duration=10m 533 iterations"
    y-axis "llamacpp:predicted_tokens_seconds"
    x-axis "llamacpp:predicted_tokens_seconds" 1716414680 --> 1716415306
    line [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 42.17, 42.17, 42.17, 42.17, 42.17, 42.21, 42.21, 42.21, 42.21, 42.21, 28.03, 28.03, 28.03, 28.03, 28.03, 29.45, 29.45, 29.45, 29.45, 29.45, 31.67, 31.67, 31.67, 31.67, 31.67, 32.33, 32.33, 32.33, 32.33, 32.33, 34.1, 34.1, 34.1, 34.1, 34.1, 34.77, 34.77, 34.77, 34.77, 34.77, 35.09, 35.09, 35.09, 35.09, 35.09, 35.35, 35.35, 35.35, 35.35, 35.35, 35.49, 35.49, 35.49, 35.49, 35.49, 35.05, 35.05, 35.05, 35.05, 35.05, 33.66, 33.66, 33.66, 33.66, 33.66, 33.33, 33.33, 33.33, 33.33, 33.33, 31.72, 31.72, 31.72, 31.72, 31.72, 30.1, 30.1, 30.1, 30.1, 30.1, 29.64, 29.64, 29.64, 29.64, 29.64, 29.83, 29.83, 29.83, 29.83, 29.83, 30.15, 30.15, 30.15, 30.15, 30.15, 30.03, 30.03, 30.03, 30.03, 30.03, 30.14, 30.14, 30.14, 30.14, 30.14, 30.13, 30.13, 30.13, 30.13, 30.13, 30.29, 30.29, 30.29, 30.29, 30.29, 30.2, 30.2, 30.2, 30.2, 30.2, 30.3, 30.3, 30.3, 30.3, 30.3, 30.71, 30.71, 30.71, 30.71, 30.71, 30.78, 30.78, 30.78, 30.78, 30.78, 30.55, 30.55, 30.55, 30.55, 30.55, 30.36, 30.36, 30.36, 30.36, 30.36, 30.4, 30.4, 30.4, 30.4, 30.4, 30.64, 30.64, 30.64, 30.64, 30.64, 30.75, 30.75, 30.75, 30.75, 30.75, 30.84, 30.84, 30.84, 30.84, 30.84, 31.0, 31.0, 31.0, 31.0, 31.0, 30.99, 30.99, 30.99, 30.99, 30.99, 30.7, 30.7, 30.7, 30.7, 30.7, 30.56, 30.56, 30.56, 30.56, 30.56, 30.09, 30.09, 30.09, 30.09, 30.09, 29.49, 29.49, 29.49, 29.49, 29.49, 29.5, 29.5, 29.5, 29.5, 29.5, 29.77, 29.77, 29.77, 29.77, 29.77, 29.86, 29.86, 29.86, 29.86, 29.86, 30.1, 30.1, 30.1, 30.1, 30.1, 30.03, 30.03, 30.03, 30.03, 30.03, 29.84, 29.84, 29.84, 29.84, 29.84, 29.47, 29.47, 29.47, 29.47, 29.47, 28.67, 28.67, 28.67, 28.67, 28.67, 28.15, 28.15, 28.15, 28.15, 28.15, 28.1, 28.1, 28.1, 28.1, 28.1, 28.08, 28.08, 28.08, 28.08, 28.08, 28.1, 28.1, 28.1, 28.1, 28.1, 28.23, 28.23, 28.23, 28.23, 28.23, 28.29, 28.29, 28.29, 28.29, 28.29, 28.34, 28.34, 28.34, 28.34, 28.34, 28.32, 28.32, 28.32, 28.32, 28.32, 28.34, 28.34, 28.34, 28.34, 28.34, 28.29, 28.29, 28.29, 28.29, 28.29, 28.38, 28.38, 28.38, 28.38, 28.38, 28.53, 28.53, 28.53, 28.53, 28.53, 28.67, 28.67, 28.67, 28.67, 28.67, 28.78, 28.78, 28.78, 28.78]
                    
Loading

Details

kv_cache_usage_ratio

More
---
config:
    xyChart:
        titleFontSize: 12
        width: 900
        height: 600
    themeVariables:
        xyChart:
            titleColor: "#000000"
---
xychart-beta
    title "llama.cpp bench-server-baseline on Standard_NC4as_T4_v3
 duration=10m 533 iterations"
    y-axis "llamacpp:kv_cache_usage_ratio"
    x-axis "llamacpp:kv_cache_usage_ratio" 1716414680 --> 1716415306
    line [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.14, 0.14, 0.14, 0.14, 0.14, 0.3, 0.3, 0.3, 0.3, 0.3, 0.23, 0.23, 0.23, 0.23, 0.23, 0.14, 0.14, 0.14, 0.14, 0.14, 0.18, 0.18, 0.18, 0.18, 0.18, 0.14, 0.14, 0.14, 0.14, 0.14, 0.14, 0.14, 0.14, 0.14, 0.14, 0.11, 0.11, 0.11, 0.11, 0.11, 0.21, 0.21, 0.21, 0.21, 0.21, 0.16, 0.16, 0.16, 0.16, 0.16, 0.19, 0.19, 0.19, 0.19, 0.19, 0.31, 0.31, 0.31, 0.31, 0.31, 0.28, 0.28, 0.28, 0.28, 0.28, 0.42, 0.42, 0.42, 0.42, 0.42, 0.35, 0.35, 0.35, 0.35, 0.35, 0.38, 0.38, 0.38, 0.38, 0.38, 0.15, 0.15, 0.15, 0.15, 0.15, 0.12, 0.12, 0.12, 0.12, 0.12, 0.32, 0.32, 0.32, 0.32, 0.32, 0.19, 0.19, 0.19, 0.19, 0.19, 0.19, 0.19, 0.19, 0.19, 0.19, 0.19, 0.19, 0.19, 0.19, 0.19, 0.16, 0.16, 0.16, 0.16, 0.16, 0.17, 0.17, 0.17, 0.17, 0.17, 0.14, 0.14, 0.14, 0.14, 0.14, 0.18, 0.18, 0.18, 0.18, 0.18, 0.3, 0.3, 0.3, 0.3, 0.3, 0.21, 0.21, 0.21, 0.21, 0.21, 0.12, 0.12, 0.12, 0.12, 0.12, 0.14, 0.14, 0.14, 0.14, 0.14, 0.15, 0.15, 0.15, 0.15, 0.15, 0.19, 0.19, 0.19, 0.19, 0.19, 0.15, 0.15, 0.15, 0.15, 0.15, 0.14, 0.14, 0.14, 0.14, 0.14, 0.29, 0.29, 0.29, 0.29, 0.29, 0.4, 0.4, 0.4, 0.4, 0.4, 0.4, 0.4, 0.4, 0.4, 0.4, 0.38, 0.38, 0.38, 0.38, 0.38, 0.25, 0.25, 0.25, 0.25, 0.25, 0.09, 0.09, 0.09, 0.09, 0.09, 0.13, 0.13, 0.13, 0.13, 0.13, 0.15, 0.15, 0.15, 0.15, 0.15, 0.31, 0.31, 0.31, 0.31, 0.31, 0.46, 0.46, 0.46, 0.46, 0.46, 0.62, 0.62, 0.62, 0.62, 0.62, 0.59, 0.59, 0.59, 0.59, 0.59, 0.52, 0.52, 0.52, 0.52, 0.52, 0.14, 0.14, 0.14, 0.14, 0.14, 0.25, 0.25, 0.25, 0.25, 0.25, 0.24, 0.24, 0.24, 0.24, 0.24, 0.2, 0.2, 0.2, 0.2, 0.2, 0.15, 0.15, 0.15, 0.15, 0.15, 0.19, 0.19, 0.19, 0.19, 0.19, 0.24, 0.24, 0.24, 0.24, 0.24, 0.16, 0.16, 0.16, 0.16, 0.16, 0.18, 0.18, 0.18, 0.18, 0.18, 0.12, 0.12, 0.12, 0.12, 0.12, 0.1, 0.1, 0.1, 0.1, 0.1, 0.12, 0.12, 0.12, 0.12, 0.12, 0.1, 0.1, 0.1, 0.1, 0.1, 0.16, 0.16, 0.16, 0.16]
                    
Loading
requests_processing
More
---
config:
    xyChart:
        titleFontSize: 12
        width: 900
        height: 600
    themeVariables:
        xyChart:
            titleColor: "#000000"
---
xychart-beta
    title "llama.cpp bench-server-baseline on Standard_NC4as_T4_v3
 duration=10m 533 iterations"
    y-axis "llamacpp:requests_processing"
    x-axis "llamacpp:requests_processing" 1716414680 --> 1716415306
    line [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 8.0, 8.0, 8.0, 8.0, 8.0, 4.0, 4.0, 4.0, 4.0, 4.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0, 3.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 5.0, 5.0, 5.0, 5.0, 5.0, 6.0, 6.0, 6.0, 6.0, 6.0, 2.0, 2.0, 2.0, 2.0, 2.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 5.0, 5.0, 5.0, 5.0, 5.0, 8.0, 8.0, 8.0, 8.0, 8.0, 7.0, 7.0, 7.0, 7.0, 7.0, 5.0, 5.0, 5.0, 5.0, 5.0, 6.0, 6.0, 6.0, 6.0, 6.0, 7.0, 7.0, 7.0, 7.0, 7.0, 5.0, 5.0, 5.0, 5.0, 5.0, 4.0, 4.0, 4.0, 4.0, 4.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 1.0, 1.0, 1.0, 1.0, 1.0, 6.0, 6.0, 6.0, 6.0, 6.0, 7.0, 7.0, 7.0, 7.0, 7.0, 5.0, 5.0, 5.0, 5.0, 5.0, 7.0, 7.0, 7.0, 7.0, 7.0, 8.0, 8.0, 8.0, 8.0, 8.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 5.0, 5.0, 5.0, 5.0, 5.0, 3.0, 3.0, 3.0, 3.0, 3.0, 6.0, 6.0, 6.0, 6.0, 6.0, 2.0, 2.0, 2.0, 2.0, 2.0, 8.0, 8.0, 8.0, 8.0, 8.0, 5.0, 5.0, 5.0, 5.0, 5.0, 8.0, 8.0, 8.0, 8.0, 8.0, 7.0, 7.0, 7.0, 7.0, 7.0, 3.0, 3.0, 3.0, 3.0, 3.0, 7.0, 7.0, 7.0, 7.0, 7.0, 3.0, 3.0, 3.0, 3.0, 3.0, 5.0, 5.0, 5.0, 5.0, 5.0, 6.0, 6.0, 6.0, 6.0, 6.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 6.0, 6.0, 6.0, 6.0, 6.0, 7.0, 7.0, 7.0, 7.0, 7.0, 6.0, 6.0, 6.0, 6.0, 6.0, 3.0, 3.0, 3.0, 3.0, 3.0, 6.0, 6.0, 6.0, 6.0, 6.0, 4.0, 4.0, 4.0, 4.0, 4.0, 5.0, 5.0, 5.0, 5.0, 5.0, 4.0, 4.0, 4.0, 4.0, 4.0, 7.0, 7.0, 7.0, 7.0, 7.0, 7.0, 7.0, 7.0, 7.0, 7.0, 1.0, 1.0, 1.0, 1.0, 1.0, 3.0, 3.0, 3.0, 3.0, 3.0, 5.0, 5.0, 5.0, 5.0, 5.0, 7.0, 7.0, 7.0, 7.0, 7.0, 3.0, 3.0, 3.0, 3.0]
                    
Loading

@mofosyne mofosyne added refactoring Refactoring Review Complexity : Medium Generally require more time to grok but manageable by beginner to medium expertise level labels May 23, 2024
@ggerganov ggerganov merged commit e84b71c into master May 23, 2024
64 of 75 checks passed
@ggerganov ggerganov deleted the gg/drop-qk_k-64 branch May 23, 2024 07:00
teleprint-me pushed a commit to teleprint-me/llama.cpp that referenced this pull request May 23, 2024
* ggml : drop support for QK_K=64

ggml-ci

* opencl : restore QK_K=256 define
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
build Compilation issues devops improvements to build systems and github actions ggml changes relating to the ggml tensor library for machine learning Nvidia GPU Issues specific to Nvidia GPUs python python script changes refactoring Refactoring Review Complexity : Medium Generally require more time to grok but manageable by beginner to medium expertise level SYCL https://en.wikipedia.org/wiki/SYCL - GPU programming language
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants