add_special option for server tokenize endpoint

[JohanAR]

No description provided.

[phymbert]

I do not use the tokenize endpoint on my end.

But is it worth the effort to add a test case ?

[github-actions]

📈 llama.cpp server for bench-server-baseline on Standard_NC4as_T4_v3 for phi-2-q4_0: 548 iterations 🚀

Expand details for performance related PR only

• Concurrent users: 8, duration: 10m
• HTTP request : avg=8617.66ms p(95)=20828.13ms fails=, finish reason: stop=476 truncated=72
• Prompt processing (pp): avg=97.13tk/s p(95)=406.17tk/s
• Token generation (tg): avg=32.89tk/s p(95)=46.41tk/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=master commit=22fddbb625c7ed309a478cc6c956d067b3856093

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 548 iterations"
    y-axis "llamacpp:prompt_tokens_seconds"
    x-axis "llamacpp:prompt_tokens_seconds" 1714846375 --> 1714847005
    line [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 695.98, 695.98, 695.98, 695.98, 695.98, 723.94, 723.94, 723.94, 723.94, 723.94, 729.66, 729.66, 729.66, 729.66, 729.66, 759.75, 759.75, 759.75, 759.75, 759.75, 806.43, 806.43, 806.43, 806.43, 806.43, 804.36, 804.36, 804.36, 804.36, 804.36, 801.47, 801.47, 801.47, 801.47, 801.47, 834.75, 834.75, 834.75, 834.75, 834.75, 833.49, 833.49, 833.49, 833.49, 833.49, 845.06, 845.06, 845.06, 845.06, 845.06, 862.92, 862.92, 862.92, 862.92, 862.92, 857.67, 857.67, 857.67, 857.67, 857.67, 858.48, 858.48, 858.48, 858.48, 858.48, 882.73, 882.73, 882.73, 882.73, 882.73, 892.25, 892.25, 892.25, 892.25, 892.25, 885.12, 885.12, 885.12, 885.12, 885.12, 891.62, 891.62, 891.62, 891.62, 891.62, 900.44, 900.44, 900.44, 900.44, 900.44, 899.79, 899.79, 899.79, 899.79, 899.79, 893.38, 893.38, 893.38, 893.38, 893.38, 895.62, 895.62, 895.62, 895.62, 895.62, 898.59, 898.59, 898.59, 898.59, 898.59, 890.05, 890.05, 890.05, 890.05, 890.05, 887.62, 887.62, 887.62, 887.62, 887.62, 890.38, 890.38, 890.38, 890.38, 890.38, 881.0, 881.0, 881.0, 881.0, 881.0, 877.45, 877.45, 877.45, 877.45, 877.45, 874.73, 874.73, 874.73, 874.73, 874.73, 871.15, 871.15, 871.15, 871.15, 871.15, 876.77, 876.77, 876.77, 876.77, 876.77, 876.34, 876.34, 876.34, 876.34, 876.34, 873.85, 873.85, 873.85, 873.85, 873.85, 876.76, 876.76, 876.76, 876.76, 876.76, 886.6, 886.6, 886.6, 886.6, 886.6, 892.94, 892.94, 892.94, 892.94, 892.94, 891.58, 891.58, 891.58, 891.58, 891.58, 890.32, 890.32, 890.32, 890.32, 890.32, 888.24, 888.24, 888.24, 888.24, 888.24, 889.46, 889.46, 889.46, 889.46, 889.46, 889.85, 889.85, 889.85, 889.85, 889.85, 889.53, 889.53, 889.53, 889.53, 889.53, 865.4, 865.4, 865.4, 865.4, 865.4, 861.97, 861.97, 861.97, 861.97, 861.97, 838.54, 838.54, 838.54, 838.54, 838.54, 837.56, 837.56, 837.56, 837.56, 837.56, 835.05, 835.05, 835.05, 835.05, 835.05, 844.19, 844.19, 844.19, 844.19, 844.19, 843.83, 843.83, 843.83, 843.83, 843.83, 842.14, 842.14, 842.14, 842.14, 842.14, 838.65, 838.65, 838.65, 838.65, 838.65, 843.0, 843.0, 843.0, 843.0, 843.0, 847.29, 847.29, 847.29, 847.29, 847.29, 845.88, 845.88, 845.88, 845.88, 845.88, 852.93, 852.93, 852.93, 852.93, 852.93, 853.8, 853.8, 853.8, 853.8, 853.8, 853.51, 853.51, 853.51, 853.51, 853.51, 853.1, 853.1, 853.1, 853.1, 853.1, 852.17, 852.17, 852.17, 852.17, 852.17, 853.43, 853.43, 853.43, 853.43, 853.43, 853.77, 853.77, 853.77, 853.77, 853.77, 856.13, 856.13, 856.13, 856.13, 856.13, 856.13]
                    

Loading

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 548 iterations"
    y-axis "llamacpp:predicted_tokens_seconds"
    x-axis "llamacpp:predicted_tokens_seconds" 1714846375 --> 1714847005
    line [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 46.25, 46.25, 46.25, 46.25, 46.25, 43.25, 43.25, 43.25, 43.25, 43.25, 32.39, 32.39, 32.39, 32.39, 32.39, 30.19, 30.19, 30.19, 30.19, 30.19, 31.32, 31.32, 31.32, 31.32, 31.32, 32.23, 32.23, 32.23, 32.23, 32.23, 33.97, 33.97, 33.97, 33.97, 33.97, 34.39, 34.39, 34.39, 34.39, 34.39, 34.44, 34.44, 34.44, 34.44, 34.44, 34.31, 34.31, 34.31, 34.31, 34.31, 34.51, 34.51, 34.51, 34.51, 34.51, 34.4, 34.4, 34.4, 34.4, 34.4, 33.54, 33.54, 33.54, 33.54, 33.54, 32.43, 32.43, 32.43, 32.43, 32.43, 32.16, 32.16, 32.16, 32.16, 32.16, 32.5, 32.5, 32.5, 32.5, 32.5, 32.66, 32.66, 32.66, 32.66, 32.66, 32.22, 32.22, 32.22, 32.22, 32.22, 32.03, 32.03, 32.03, 32.03, 32.03, 31.82, 31.82, 31.82, 31.82, 31.82, 31.67, 31.67, 31.67, 31.67, 31.67, 31.79, 31.79, 31.79, 31.79, 31.79, 31.84, 31.84, 31.84, 31.84, 31.84, 31.71, 31.71, 31.71, 31.71, 31.71, 31.76, 31.76, 31.76, 31.76, 31.76, 31.91, 31.91, 31.91, 31.91, 31.91, 31.54, 31.54, 31.54, 31.54, 31.54, 31.07, 31.07, 31.07, 31.07, 31.07, 31.1, 31.1, 31.1, 31.1, 31.1, 31.34, 31.34, 31.34, 31.34, 31.34, 31.52, 31.52, 31.52, 31.52, 31.52, 31.59, 31.59, 31.59, 31.59, 31.59, 31.66, 31.66, 31.66, 31.66, 31.66, 31.62, 31.62, 31.62, 31.62, 31.62, 31.56, 31.56, 31.56, 31.56, 31.56, 31.46, 31.46, 31.46, 31.46, 31.46, 31.28, 31.28, 31.28, 31.28, 31.28, 31.22, 31.22, 31.22, 31.22, 31.22, 31.35, 31.35, 31.35, 31.35, 31.35, 31.56, 31.56, 31.56, 31.56, 31.56, 31.65, 31.65, 31.65, 31.65, 31.65, 31.69, 31.69, 31.69, 31.69, 31.69, 31.59, 31.59, 31.59, 31.59, 31.59, 30.83, 30.83, 30.83, 30.83, 30.83, 30.79, 30.79, 30.79, 30.79, 30.79, 29.83, 29.83, 29.83, 29.83, 29.83, 29.83, 29.83, 29.83, 29.83, 29.83, 29.84, 29.84, 29.84, 29.84, 29.84, 29.96, 29.96, 29.96, 29.96, 29.96, 30.05, 30.05, 30.05, 30.05, 30.05, 30.15, 30.15, 30.15, 30.15, 30.15, 30.15, 30.15, 30.15, 30.15, 30.15, 30.06, 30.06, 30.06, 30.06, 30.06, 29.9, 29.9, 29.9, 29.9, 29.9, 29.96, 29.96, 29.96, 29.96, 29.96, 29.99, 29.99, 29.99, 29.99, 29.99, 30.17, 30.17, 30.17, 30.17, 30.17, 30.25, 30.25, 30.25, 30.25, 30.25, 30.36, 30.36, 30.36, 30.36, 30.36, 30.43, 30.43, 30.43, 30.43, 30.43, 30.43, 30.43, 30.43, 30.43, 30.43, 30.46]
                    

Loading

Details

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 548 iterations"
    y-axis "llamacpp:kv_cache_usage_ratio"
    x-axis "llamacpp:kv_cache_usage_ratio" 1714846375 --> 1714847005
    line [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.18, 0.18, 0.18, 0.18, 0.18, 0.4, 0.4, 0.4, 0.4, 0.4, 0.29, 0.29, 0.29, 0.29, 0.29, 0.14, 0.14, 0.14, 0.14, 0.14, 0.15, 0.15, 0.15, 0.15, 0.15, 0.13, 0.13, 0.13, 0.13, 0.13, 0.12, 0.12, 0.12, 0.12, 0.12, 0.14, 0.14, 0.14, 0.14, 0.14, 0.18, 0.18, 0.18, 0.18, 0.18, 0.13, 0.13, 0.13, 0.13, 0.13, 0.24, 0.24, 0.24, 0.24, 0.24, 0.15, 0.15, 0.15, 0.15, 0.15, 0.18, 0.18, 0.18, 0.18, 0.18, 0.27, 0.27, 0.27, 0.27, 0.27, 0.14, 0.14, 0.14, 0.14, 0.14, 0.11, 0.11, 0.11, 0.11, 0.11, 0.24, 0.24, 0.24, 0.24, 0.24, 0.25, 0.25, 0.25, 0.25, 0.25, 0.27, 0.27, 0.27, 0.27, 0.27, 0.17, 0.17, 0.17, 0.17, 0.17, 0.14, 0.14, 0.14, 0.14, 0.14, 0.15, 0.15, 0.15, 0.15, 0.15, 0.31, 0.31, 0.31, 0.31, 0.31, 0.1, 0.1, 0.1, 0.1, 0.1, 0.17, 0.17, 0.17, 0.17, 0.17, 0.31, 0.31, 0.31, 0.31, 0.31, 0.37, 0.37, 0.37, 0.37, 0.37, 0.22, 0.22, 0.22, 0.22, 0.22, 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, 0.15, 0.15, 0.15, 0.15, 0.15, 0.19, 0.19, 0.19, 0.19, 0.19, 0.23, 0.23, 0.23, 0.23, 0.23, 0.18, 0.18, 0.18, 0.18, 0.18, 0.34, 0.34, 0.34, 0.34, 0.34, 0.16, 0.16, 0.16, 0.16, 0.16, 0.21, 0.21, 0.21, 0.21, 0.21, 0.1, 0.1, 0.1, 0.1, 0.1, 0.11, 0.11, 0.11, 0.11, 0.11, 0.1, 0.1, 0.1, 0.1, 0.1, 0.38, 0.38, 0.38, 0.38, 0.38, 0.52, 0.52, 0.52, 0.52, 0.52, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.09, 0.09, 0.09, 0.09, 0.09, 0.17, 0.17, 0.17, 0.17, 0.17, 0.11, 0.11, 0.11, 0.11, 0.11, 0.14, 0.14, 0.14, 0.14, 0.14, 0.12, 0.12, 0.12, 0.12, 0.12, 0.16, 0.16, 0.16, 0.16, 0.16, 0.31, 0.31, 0.31, 0.31, 0.31, 0.2, 0.2, 0.2, 0.2, 0.2, 0.26, 0.26, 0.26, 0.26, 0.26, 0.15, 0.15, 0.15, 0.15, 0.15, 0.11, 0.11, 0.11, 0.11, 0.11, 0.1, 0.1, 0.1, 0.1, 0.1, 0.16, 0.16, 0.16, 0.16, 0.16, 0.13, 0.13, 0.13, 0.13, 0.13, 0.15, 0.15, 0.15, 0.15, 0.15, 0.21, 0.21, 0.21, 0.21, 0.21, 0.31]
                    

Loading

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 548 iterations"
    y-axis "llamacpp:requests_processing"
    x-axis "llamacpp:requests_processing" 1714846375 --> 1714847005
    line [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 4.0, 4.0, 4.0, 4.0, 4.0, 8.0, 8.0, 8.0, 8.0, 8.0, 5.0, 5.0, 5.0, 5.0, 5.0, 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, 1.0, 1.0, 1.0, 1.0, 1.0, 4.0, 4.0, 4.0, 4.0, 4.0, 8.0, 8.0, 8.0, 8.0, 8.0, 4.0, 4.0, 4.0, 4.0, 4.0, 6.0, 6.0, 6.0, 6.0, 6.0, 7.0, 7.0, 7.0, 7.0, 7.0, 1.0, 1.0, 1.0, 1.0, 1.0, 2.0, 2.0, 2.0, 2.0, 2.0, 5.0, 5.0, 5.0, 5.0, 5.0, 4.0, 4.0, 4.0, 4.0, 4.0, 5.0, 5.0, 5.0, 5.0, 5.0, 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, 6.0, 6.0, 6.0, 6.0, 6.0, 5.0, 5.0, 5.0, 5.0, 5.0, 7.0, 7.0, 7.0, 7.0, 7.0, 1.0, 1.0, 1.0, 1.0, 1.0, 7.0, 7.0, 7.0, 7.0, 7.0, 8.0, 8.0, 8.0, 8.0, 8.0, 5.0, 5.0, 5.0, 5.0, 5.0, 3.0, 3.0, 3.0, 3.0, 3.0, 8.0, 8.0, 8.0, 8.0, 8.0, 3.0, 3.0, 3.0, 3.0, 3.0, 4.0, 4.0, 4.0, 4.0, 4.0, 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, 8.0, 8.0, 8.0, 8.0, 8.0, 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, 5.0, 5.0, 5.0, 5.0, 5.0, 2.0, 2.0, 2.0, 2.0, 2.0, 7.0, 7.0, 7.0, 7.0, 7.0, 8.0, 8.0, 8.0, 8.0, 8.0, 7.0, 7.0, 7.0, 7.0, 7.0, 7.0, 7.0, 7.0, 7.0, 7.0, 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, 7.0, 7.0, 7.0, 7.0, 7.0, 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, 4.0, 4.0, 4.0, 4.0, 4.0, 2.0, 2.0, 2.0, 2.0, 2.0, 4.0, 4.0, 4.0, 4.0, 4.0, 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, 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, 3.0]
                    

Loading

[JohanAR]

I do not use the tokenize endpoint on my end.

But is it worth the effort to add a test case ?

I'm using the /completion endpoint with lists of tokens instead of strings, and I thought it would be nice if the client didn't need to know how the model represents the BOS token, now that there's a bunch of new models which don't use <s>

It's a pretty small feature, but I can look into adding a test if you prefer.

[teleprint-me]

As an FYI, train-text-from-scratch uses the BOS by default. I've been depending on these tokens because I use sentencepiece models most of the time and sentencepiece uses these tokens out of the box. Not a big deal, I just think it's worth mentioning.

[ngxson]

Yeah this seems to be an useful thing, since sometimes we want to tokenize a part of text (BOS should not be added in this case)

For extra safe, I'd suggest a test case like this: Tokenize a text with add_special = true and then with add_special = false, then compare if the output is different or not (with add_special = true we should see more tokens)

[JohanAR]

@ngxson currently trying to understand the python behave framework and how to formulate the tests :) I'm not sure how to express something like "when the following string is tokenized, the resulting length must be longer than when this other string is tokenized" in behave, but perhaps I could figure it out.

Was thinking I would look up the BOS token for the test model (I'm guessing it's <s>) and then verify that /tokenize with {content="<s>"}, {content="<s>", add_special=False} and {content="", add_special=True} all tokenize to the same sequence. Though perhaps it's not a good idea to hard-code a dependence to a particular tokenizer, i.e. making a case which fails if the test model is substituted for one which has <|begin_of_text|> or <BOS_TOKEN> instead.

Update: The above did not work, the tokenizer in tinyllamas/stories260K.gguf behaves strange. I'll hard-code the BOS token as [1] in the tests and check for that instead, not really worse than hard-coding it to <s> IMO

[ngxson]

Update: The above did not work, the tokenizer in tinyllamas/stories260K.gguf behaves strange. I'll hard-code the BOS token as [1] in the tests and check for that instead, not really worse than hard-coding it to <s> IMO

Instead of hard-coding, another idea would be to input the BOS as a behave step, something like:

    Given a server listening on localhost:8080
    And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
    And   a model file test-model.gguf
    And   a model alias tinyllama-2
    ...
    And BOS token is 1

The handler for that is defined at the beginning of steps.py

[JohanAR]

@ngxson yea, perhaps that is better. Changed it to do that.