Add BPE pre-tokenization for Command-R/R+.

[dranger003]

This replaces PR #7033 as a result of merging PR #6511.

Closes #7030 and #7040.

[github-actions]

📈 llama.cpp server for bench-server-baseline on Standard_NC4as_T4_v3 for phi-2-q4_0: 536 iterations 🚀

Expand details for performance related PR only

• Concurrent users: 8, duration: 10m
• HTTP request : avg=8756.89ms p(95)=21734.99ms fails=, finish reason: stop=469 truncated=67
• Prompt processing (pp): avg=103.54tk/s p(95)=469.12tk/s
• Token generation (tg): avg=32.32tk/s p(95)=46.55tk/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=bpe-pretok-command-r-2 commit=f5806b2d09ba2dcf60d8d66046ed5853234f28de

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 536 iterations"
    y-axis "llamacpp:prompt_tokens_seconds"
    x-axis "llamacpp:prompt_tokens_seconds" 1714885613 --> 1714886239
    line [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 675.74, 675.74, 675.74, 675.74, 675.74, 513.3, 513.3, 513.3, 513.3, 513.3, 531.59, 531.59, 531.59, 531.59, 531.59, 577.54, 577.54, 577.54, 577.54, 577.54, 621.43, 621.43, 621.43, 621.43, 621.43, 644.79, 644.79, 644.79, 644.79, 644.79, 647.69, 647.69, 647.69, 647.69, 647.69, 688.4, 688.4, 688.4, 688.4, 688.4, 691.16, 691.16, 691.16, 691.16, 691.16, 709.44, 709.44, 709.44, 709.44, 709.44, 731.08, 731.08, 731.08, 731.08, 731.08, 743.11, 743.11, 743.11, 743.11, 743.11, 724.87, 724.87, 724.87, 724.87, 724.87, 770.53, 770.53, 770.53, 770.53, 770.53, 794.5, 794.5, 794.5, 794.5, 794.5, 789.99, 789.99, 789.99, 789.99, 789.99, 791.02, 791.02, 791.02, 791.02, 791.02, 816.51, 816.51, 816.51, 816.51, 816.51, 813.72, 813.72, 813.72, 813.72, 813.72, 816.15, 816.15, 816.15, 816.15, 816.15, 822.75, 822.75, 822.75, 822.75, 822.75, 826.64, 826.64, 826.64, 826.64, 826.64, 832.54, 832.54, 832.54, 832.54, 832.54, 817.58, 817.58, 817.58, 817.58, 817.58, 820.87, 820.87, 820.87, 820.87, 820.87, 822.5, 822.5, 822.5, 822.5, 822.5, 837.84, 837.84, 837.84, 837.84, 837.84, 835.14, 835.14, 835.14, 835.14, 835.14, 834.02, 834.02, 834.02, 834.02, 834.02, 835.37, 835.37, 835.37, 835.37, 835.37, 840.63, 840.63, 840.63, 840.63, 840.63, 840.19, 840.19, 840.19, 840.19, 840.19, 840.33, 840.33, 840.33, 840.33, 840.33, 842.91, 842.91, 842.91, 842.91, 842.91, 845.73, 845.73, 845.73, 845.73, 845.73, 850.35, 850.35, 850.35, 850.35, 850.35, 861.33, 861.33, 861.33, 861.33, 861.33, 860.72, 860.72, 860.72, 860.72, 860.72, 858.59, 858.59, 858.59, 858.59, 858.59, 861.43, 861.43, 861.43, 861.43, 861.43, 864.45, 864.45, 864.45, 864.45, 864.45, 876.09, 876.09, 876.09, 876.09, 876.09, 860.41, 860.41, 860.41, 860.41, 860.41, 836.69, 836.69, 836.69, 836.69, 836.69, 836.74, 836.74, 836.74, 836.74, 836.74, 834.68, 834.68, 834.68, 834.68, 834.68, 832.0, 832.0, 832.0, 832.0, 832.0, 835.99, 835.99, 835.99, 835.99, 835.99, 838.72, 838.72, 838.72, 838.72, 838.72, 839.72, 839.72, 839.72, 839.72, 839.72, 842.33, 842.33, 842.33, 842.33, 842.33, 844.19, 844.19, 844.19, 844.19, 844.19, 847.05, 847.05, 847.05, 847.05, 847.05, 847.78, 847.78, 847.78, 847.78, 847.78, 849.13, 849.13, 849.13, 849.13, 849.13, 853.68, 853.68, 853.68, 853.68, 853.68, 854.56, 854.56, 854.56, 854.56, 854.56, 854.47, 854.47, 854.47, 854.47, 854.47, 855.58, 855.58, 855.58, 855.58, 855.58, 856.38, 856.38, 856.38, 856.38, 856.38, 856.17, 856.17, 856.17, 856.17]
                    

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 536 iterations"
    y-axis "llamacpp:predicted_tokens_seconds"
    x-axis "llamacpp:predicted_tokens_seconds" 1714885613 --> 1714886239
    line [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 43.8, 43.8, 43.8, 43.8, 43.8, 40.63, 40.63, 40.63, 40.63, 40.63, 34.12, 34.12, 34.12, 34.12, 34.12, 33.15, 33.15, 33.15, 33.15, 33.15, 32.78, 32.78, 32.78, 32.78, 32.78, 32.89, 32.89, 32.89, 32.89, 32.89, 33.74, 33.74, 33.74, 33.74, 33.74, 34.58, 34.58, 34.58, 34.58, 34.58, 34.84, 34.84, 34.84, 34.84, 34.84, 34.7, 34.7, 34.7, 34.7, 34.7, 34.54, 34.54, 34.54, 34.54, 34.54, 34.41, 34.41, 34.41, 34.41, 34.41, 33.57, 33.57, 33.57, 33.57, 33.57, 33.47, 33.47, 33.47, 33.47, 33.47, 32.14, 32.14, 32.14, 32.14, 32.14, 31.47, 31.47, 31.47, 31.47, 31.47, 31.87, 31.87, 31.87, 31.87, 31.87, 31.98, 31.98, 31.98, 31.98, 31.98, 31.28, 31.28, 31.28, 31.28, 31.28, 30.99, 30.99, 30.99, 30.99, 30.99, 30.96, 30.96, 30.96, 30.96, 30.96, 31.09, 31.09, 31.09, 31.09, 31.09, 31.31, 31.31, 31.31, 31.31, 31.31, 31.17, 31.17, 31.17, 31.17, 31.17, 31.2, 31.2, 31.2, 31.2, 31.2, 31.38, 31.38, 31.38, 31.38, 31.38, 31.37, 31.37, 31.37, 31.37, 31.37, 30.79, 30.79, 30.79, 30.79, 30.79, 30.52, 30.52, 30.52, 30.52, 30.52, 30.71, 30.71, 30.71, 30.71, 30.71, 30.86, 30.86, 30.86, 30.86, 30.86, 31.05, 31.05, 31.05, 31.05, 31.05, 31.27, 31.27, 31.27, 31.27, 31.27, 31.31, 31.31, 31.31, 31.31, 31.31, 31.26, 31.26, 31.26, 31.26, 31.26, 31.19, 31.19, 31.19, 31.19, 31.19, 31.09, 31.09, 31.09, 31.09, 31.09, 30.87, 30.87, 30.87, 30.87, 30.87, 30.88, 30.88, 30.88, 30.88, 30.88, 31.08, 31.08, 31.08, 31.08, 31.08, 31.22, 31.22, 31.22, 31.22, 31.22, 31.26, 31.26, 31.26, 31.26, 31.26, 31.23, 31.23, 31.23, 31.23, 31.23, 31.14, 31.14, 31.14, 31.14, 31.14, 30.9, 30.9, 30.9, 30.9, 30.9, 29.61, 29.61, 29.61, 29.61, 29.61, 29.6, 29.6, 29.6, 29.6, 29.6, 29.56, 29.56, 29.56, 29.56, 29.56, 29.55, 29.55, 29.55, 29.55, 29.55, 29.69, 29.69, 29.69, 29.69, 29.69, 29.7, 29.7, 29.7, 29.7, 29.7, 29.89, 29.89, 29.89, 29.89, 29.89, 29.88, 29.88, 29.88, 29.88, 29.88, 29.87, 29.87, 29.87, 29.87, 29.87, 29.69, 29.69, 29.69, 29.69, 29.69, 29.62, 29.62, 29.62, 29.62, 29.62, 29.67, 29.67, 29.67, 29.67, 29.67, 29.81, 29.81, 29.81, 29.81, 29.81, 29.92, 29.92, 29.92, 29.92, 29.92, 30.03, 30.03, 30.03, 30.03, 30.03, 30.07, 30.07, 30.07, 30.07]
                    

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 536 iterations"
    y-axis "llamacpp:kv_cache_usage_ratio"
    x-axis "llamacpp:kv_cache_usage_ratio" 1714885613 --> 1714886239
    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, 0.39, 0.39, 0.39, 0.39, 0.39, 0.18, 0.18, 0.18, 0.18, 0.18, 0.16, 0.16, 0.16, 0.16, 0.16, 0.2, 0.2, 0.2, 0.2, 0.2, 0.14, 0.14, 0.14, 0.14, 0.14, 0.07, 0.07, 0.07, 0.07, 0.07, 0.15, 0.15, 0.15, 0.15, 0.15, 0.2, 0.2, 0.2, 0.2, 0.2, 0.2, 0.2, 0.2, 0.2, 0.2, 0.2, 0.2, 0.2, 0.2, 0.2, 0.2, 0.2, 0.2, 0.2, 0.2, 0.15, 0.15, 0.15, 0.15, 0.15, 0.32, 0.32, 0.32, 0.32, 0.32, 0.17, 0.17, 0.17, 0.17, 0.17, 0.17, 0.17, 0.17, 0.17, 0.17, 0.13, 0.13, 0.13, 0.13, 0.13, 0.29, 0.29, 0.29, 0.29, 0.29, 0.29, 0.29, 0.29, 0.29, 0.29, 0.24, 0.24, 0.24, 0.24, 0.24, 0.14, 0.14, 0.14, 0.14, 0.14, 0.17, 0.17, 0.17, 0.17, 0.17, 0.26, 0.26, 0.26, 0.26, 0.26, 0.26, 0.26, 0.26, 0.26, 0.26, 0.12, 0.12, 0.12, 0.12, 0.12, 0.18, 0.18, 0.18, 0.18, 0.18, 0.3, 0.3, 0.3, 0.3, 0.3, 0.26, 0.26, 0.26, 0.26, 0.26, 0.09, 0.09, 0.09, 0.09, 0.09, 0.16, 0.16, 0.16, 0.16, 0.16, 0.13, 0.13, 0.13, 0.13, 0.13, 0.12, 0.12, 0.12, 0.12, 0.12, 0.15, 0.15, 0.15, 0.15, 0.15, 0.15, 0.15, 0.15, 0.15, 0.15, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.33, 0.33, 0.33, 0.33, 0.33, 0.12, 0.12, 0.12, 0.12, 0.12, 0.14, 0.14, 0.14, 0.14, 0.14, 0.16, 0.16, 0.16, 0.16, 0.16, 0.09, 0.09, 0.09, 0.09, 0.09, 0.33, 0.33, 0.33, 0.33, 0.33, 0.45, 0.45, 0.45, 0.45, 0.45, 0.5, 0.5, 0.5, 0.5, 0.5, 0.6, 0.6, 0.6, 0.6, 0.6, 0.46, 0.46, 0.46, 0.46, 0.46, 0.13, 0.13, 0.13, 0.13, 0.13, 0.17, 0.17, 0.17, 0.17, 0.17, 0.11, 0.11, 0.11, 0.11, 0.11, 0.13, 0.13, 0.13, 0.13, 0.13, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.3, 0.3, 0.3, 0.3, 0.3, 0.27, 0.27, 0.27, 0.27, 0.27, 0.15, 0.15, 0.15, 0.15, 0.15, 0.14, 0.14, 0.14, 0.14, 0.14, 0.14, 0.14, 0.14, 0.14, 0.14, 0.14, 0.14, 0.14, 0.14, 0.14, 0.17, 0.17, 0.17, 0.17, 0.17, 0.14, 0.14, 0.14, 0.14, 0.14, 0.17, 0.17, 0.17, 0.17]
                    

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 536 iterations"
    y-axis "llamacpp:requests_processing"
    x-axis "llamacpp:requests_processing" 1714885613 --> 1714886239
    line [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 5.0, 5.0, 5.0, 5.0, 5.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 8.0, 7.0, 7.0, 7.0, 7.0, 7.0, 1.0, 1.0, 1.0, 1.0, 1.0, 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, 6.0, 6.0, 6.0, 6.0, 6.0, 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, 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, 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, 4.0, 4.0, 4.0, 4.0, 4.0, 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, 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, 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, 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, 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, 5.0, 5.0, 5.0, 5.0, 5.0, 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, 6.0, 6.0, 6.0, 6.0, 6.0, 5.0, 5.0, 5.0, 5.0, 5.0, 2.0, 2.0, 2.0, 2.0, 2.0, 4.0, 4.0, 4.0, 4.0, 4.0, 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, 4.0, 4.0, 4.0, 4.0, 4.0, 8.0, 8.0, 8.0, 8.0, 8.0, 3.0, 3.0, 3.0, 3.0, 3.0, 6.0, 6.0, 6.0, 6.0, 6.0, 7.0, 7.0, 7.0, 7.0, 7.0, 7.0, 7.0, 7.0, 7.0, 7.0, 5.0, 5.0, 5.0, 5.0, 5.0, 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, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 4.0, 2.0, 2.0, 2.0, 2.0, 2.0, 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]
                    

Loading

[slaren]

This and #7041 have different regex. Which one is correct?

[dranger003]

also has 'Digits' and individual_digits=True, so making an assumption there now.

@slaren There is mention of an assumption about digits, which I haven't included but I can include if needed. The regex in this PR has been tested with test-tokenizer-0 which I presume does not cover all scenarios?

[araleza]

Hi, does this mean that Command-R was always running at reduced quality, and we just didn't know until recently? Or have the recent Llama 3 changes to the llama.cpp tokenizer resulted in this update being needed to get it back to where it was before the Llama 3 changes went in?

[eskeletor97]

There is mention of an assumption about digits, which I haven't included but I can include if needed. The regex in this PR has been tested with test-tokenizer-0 which I presume does not cover all scenarios?

I haven't really tested command-r before with any math or numbers, but isn't it a similar issue to llama3 where digits were grouped and tokenized incorrectly?

[ggerganov]

I had to update to new transformers:

diff --git a/requirements/requirements-convert.txt b/requirements/requirements-convert.txt
index a3d6ecec..5520ba73 100644
--- a/requirements/requirements-convert.txt
+++ b/requirements/requirements-convert.txt
@@ -1,5 +1,5 @@
 numpy~=1.24.4
 sentencepiece~=0.1.98
-transformers>=4.35.2,<5.0.0
+transformers>=4.40.1,<5.0.0
 gguf>=0.1.0
 protobuf>=4.21.0,<5.0.0

Else, I got this error:

python3 convert-hf-to-gguf-update.py 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
Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
Traceback (most recent call last):
  File "/Users/ggerganov/development/github/llama.cpp/convert-hf-to-gguf-update.py", line 135, in <module>
    tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/homebrew/lib/python3.11/site-packages/transformers/models/auto/tokenization_auto.py", line 784, in from_pretrained
    raise ValueError(
ValueError: Tokenizer class CohereTokenizer does not exist or is not currently imported.

[ggerganov]

Let's rebase on latest master and I will run some extra tests to check if the regexes are correct

[dranger003]

@ggerganov Thanks, the PR has been rebased and I added the transformers change.