Fix unicode in grammars (fixes #2501) by ejones · Pull Request #2553 · ggml-org/llama.cpp

ejones · 2023-08-08T13:05:20Z

As discussed in #2501, the grammar implementation incorrectly assumes that tokens contain full UTF-8 sequences. In reality, a unicode character may span (the bytes of) several tokens. This patch fixes the issue with the suggestion from @ai-and-i to track partial UTF-8 sequences and match against the range of unicode code points it represents.

Changes

updated decode_utf8 return partial and invalid UTF-8 sequences, and to resume from a partial UTF-8 sequence
added partial UTF-8 sequence to grammar state and sampling candidates
updated grammar candidate selection to check for overlap between char range in grammar and UTF-8 code points in the case of a partial UTF-8 sequence

Testing

Emojis (from #2501)

% ./main -m $LLAMA_30B_Q4_0 -n 32 --grammar 'root ::= [😀-🙏]+'             
main: build = 969 (bedce3c)
main: seed  = 1691499471
llama.cpp: loading model from /Users/evan/llama-models/30B/ggml-model-q4_0.bin
llama_model_load_internal: format     = ggjt v3 (latest)
llama_model_load_internal: n_vocab    = 32000
llama_model_load_internal: n_ctx      = 512
llama_model_load_internal: n_embd     = 6656
llama_model_load_internal: n_mult     = 256
llama_model_load_internal: n_head     = 52
llama_model_load_internal: n_head_kv  = 52
llama_model_load_internal: n_layer    = 60
llama_model_load_internal: n_rot      = 128
llama_model_load_internal: n_gqa      = 1
llama_model_load_internal: rnorm_eps  = 5.0e-06
llama_model_load_internal: n_ff       = 17920
llama_model_load_internal: freq_base  = 10000.0
llama_model_load_internal: freq_scale = 1
llama_model_load_internal: ftype      = 2 (mostly Q4_0)
llama_model_load_internal: model size = 30B
llama_model_load_internal: ggml ctx size =    0.16 MB
llama_model_load_internal: mem required  = 17452.69 MB (+  780.00 MB per state)
llama_new_context_with_model: kv self size  =  780.00 MB
llama_new_context_with_model: compute buffer total size =   97.35 MB

system_info: n_threads = 8 / 12 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 | 
sampling: repeat_last_n = 64, repeat_penalty = 1.100000, presence_penalty = 0.000000, frequency_penalty = 0.000000, top_k = 40, tfs_z = 1.000000, top_p = 0.950000, typical_p = 1.000000, temp = 0.800000, mirostat = 0, mirostat_lr = 0.100000, mirostat_ent = 5.000000
generate: n_ctx = 512, n_batch = 512, n_predict = 32, n_keep = 0


main: grammar:
root ::= root_1 
root_1 ::= [<U+1F600>-<U+1F64F>] root_1 | [<U+1F600>-<U+1F64F>] 

 😊😄🙃🙅🙆🙇🙈🙉
llama_print_timings:        load time =   700.01 ms
llama_print_timings:      sample time =    73.92 ms /    32 runs   (    2.31 ms per token,   432.91 tokens per second)
llama_print_timings: prompt eval time =   270.68 ms /     2 tokens (  135.34 ms per token,     7.39 tokens per second)
llama_print_timings:        eval time =  4851.84 ms /    31 runs   (  156.51 ms per token,     6.39 tokens per second)
llama_print_timings:       total time =  5211.93 ms

Student schema (Jsonformer)

% ./main -m $LLAMA_13B_Q4_0 --grammar "$( python3 examples/json-schema-to-grammar.py ../schemas/student.json --prop-order 'is_student,name,age' )" -p 'Hermione Granger '    
main: build = 969 (bedce3c)
main: seed  = 1691499724
llama.cpp: loading model from /Users/evan/llama-models/13B/ggml-model-q4_0.bin
llama_model_load_internal: format     = ggjt v3 (latest)
llama_model_load_internal: n_vocab    = 32000
llama_model_load_internal: n_ctx      = 512
llama_model_load_internal: n_embd     = 5120
llama_model_load_internal: n_mult     = 256
llama_model_load_internal: n_head     = 40
llama_model_load_internal: n_head_kv  = 40
llama_model_load_internal: n_layer    = 40
llama_model_load_internal: n_rot      = 128
llama_model_load_internal: n_gqa      = 1
llama_model_load_internal: rnorm_eps  = 5.0e-06
llama_model_load_internal: n_ff       = 13824
llama_model_load_internal: freq_base  = 10000.0
llama_model_load_internal: freq_scale = 1
llama_model_load_internal: ftype      = 2 (mostly Q4_0)
llama_model_load_internal: model size = 13B
llama_model_load_internal: ggml ctx size =    0.11 MB
llama_model_load_internal: mem required  = 6983.72 MB (+  400.00 MB per state)
llama_new_context_with_model: kv self size  =  400.00 MB
llama_new_context_with_model: compute buffer total size =   75.35 MB

system_info: n_threads = 8 / 12 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 | 
sampling: repeat_last_n = 64, repeat_penalty = 1.100000, presence_penalty = 0.000000, frequency_penalty = 0.000000, top_k = 40, tfs_z = 1.000000, top_p = 0.950000, typical_p = 1.000000, temp = 0.800000, mirostat = 0, mirostat_lr = 0.100000, mirostat_ent = 5.000000
generate: n_ctx = 512, n_batch = 512, n_predict = -1, n_keep = 0


main: grammar:
space ::= space_1 
space_1 ::= [ ] | 
boolean ::= boolean_3 space 
boolean_3 ::= [t] [r] [u] [e] | [f] [a] [l] [s] [e] 
string ::= ["] string_7 ["] space 
string_5 ::= [^"\] | [\] string_6 
string_6 ::= ["\/bfnrt] | [u] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] 
string_7 ::= string_5 string_7 | 
number ::= number_9 number_15 number_19 space 
number_9 ::= number_10 number_11 
number_10 ::= [-] | 
number_11 ::= [0-9] | [1-9] number_12 
number_12 ::= [0-9] number_12 | 
number_13 ::= [.] number_14 
number_14 ::= [0-9] number_14 | [0-9] 
number_15 ::= number_13 | 
number_16 ::= [eE] number_17 number_18 
number_17 ::= [-+] | 
number_18 ::= [0-9] number_18 | [0-9] 
number_19 ::= number_16 | 
courses ::= [[] space courses_24 []] space 
courses_21 ::= string courses_23 
courses_22 ::= [,] space string 
courses_23 ::= courses_22 courses_23 | 
courses_24 ::= courses_21 | 
root ::= [{] space ["] [i] [s] [_] [s] [t] [u] [d] [e] [n] [t] ["] space [:] space boolean [,] space ["] [n] [a] [m] [e] ["] space [:] space string [,] space ["] [a] [g] [e] ["] space [:] space number [,] space ["] [c] [o] [u] [r] [s] [e] [s] ["] space [:] space courses [}] space 

 Hermione Granger {"is_student":true,"name":"Hermione Granger","age":15,"courses":["Transfiguration","Charms","Defence Against the Dark Arts","Herbology","Potions","History of Magic","Arithmancy","Divination","Care of Magical Creatures","Astronomy","Muggle Studies","Ancient Runes"]} [end of text]

llama_print_timings:        load time =   357.06 ms
llama_print_timings:      sample time =   386.62 ms /    86 runs   (    4.50 ms per token,   222.44 tokens per second)
llama_print_timings: prompt eval time =   421.07 ms /     6 tokens (   70.18 ms per token,    14.25 tokens per second)
llama_print_timings:        eval time =  5607.78 ms /    85 runs   (   65.97 ms per token,    15.16 tokens per second)
llama_print_timings:       total time =  6466.07 ms

CJK

% ./main -m $LLAMA_30B_Q4_0  -p $'Creating a website in 5 steps (Chinese):\n\n' --grammar-file grammars/japanese.gbnf -n 32                                              
main: build = 969 (bedce3c)
main: seed  = 1691499821
llama.cpp: loading model from /Users/evan/llama-models/30B/ggml-model-q4_0.bin
llama_model_load_internal: format     = ggjt v3 (latest)
llama_model_load_internal: n_vocab    = 32000
llama_model_load_internal: n_ctx      = 512
llama_model_load_internal: n_embd     = 6656
llama_model_load_internal: n_mult     = 256
llama_model_load_internal: n_head     = 52
llama_model_load_internal: n_head_kv  = 52
llama_model_load_internal: n_layer    = 60
llama_model_load_internal: n_rot      = 128
llama_model_load_internal: n_gqa      = 1
llama_model_load_internal: rnorm_eps  = 5.0e-06
llama_model_load_internal: n_ff       = 17920
llama_model_load_internal: freq_base  = 10000.0
llama_model_load_internal: freq_scale = 1
llama_model_load_internal: ftype      = 2 (mostly Q4_0)
llama_model_load_internal: model size = 30B
llama_model_load_internal: ggml ctx size =    0.16 MB
llama_model_load_internal: mem required  = 17452.69 MB (+  780.00 MB per state)
llama_new_context_with_model: kv self size  =  780.00 MB
llama_new_context_with_model: compute buffer total size =   97.35 MB

system_info: n_threads = 8 / 12 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 | 
sampling: repeat_last_n = 64, repeat_penalty = 1.100000, presence_penalty = 0.000000, frequency_penalty = 0.000000, top_k = 40, tfs_z = 1.000000, top_p = 0.950000, typical_p = 1.000000, temp = 0.800000, mirostat = 0, mirostat_lr = 0.100000, mirostat_ent = 5.000000
generate: n_ctx = 512, n_batch = 512, n_predict = 32, n_keep = 0


main: grammar:
root ::= root_2 root_5 
jp-char ::= hiragana | katakana | punctuation | cjk 
root_2 ::= jp-char root_2 | jp-char 
root_3 ::= [ <U+0009><U+000A>] root_4 
root_4 ::= jp-char root_4 | jp-char 
root_5 ::= root_3 root_5 | 
hiragana ::= [<U+3041>-<U+309F>] 
katakana ::= [<U+30A1>-<U+30FF>] 
punctuation ::= [<U+3001>-<U+303E>] 
cjk ::= [<U+4E00>-<U+9FFF>] 

 Creating a website in 5 steps (Chinese):

從電腦開始輕鬆建立網站組件。
使用不同的
llama_print_timings:        load time =   500.31 ms
llama_print_timings:      sample time =   133.83 ms /    32 runs   (    4.18 ms per token,   239.11 tokens per second)
llama_print_timings: prompt eval time =  1669.25 ms /    14 tokens (  119.23 ms per token,     8.39 tokens per second)
llama_print_timings:        eval time =  4848.56 ms /    31 runs   (  156.41 ms per token,     6.39 tokens per second)
llama_print_timings:       total time =  6668.48 ms

SlyEcho

Seems to work!

./main -m $MODEL -ngl 99 --grammar 'root ::= [🐀-🐿]+' -n 32 -p Rodents:

output:

 Rodents:🐀🐿🐁🐁🐁🐦🐤🐜

Fix unicode in grammars (fixes ggml-org#2501)

961f2ab

ejones mentioned this pull request Aug 8, 2023

Emojis in grammar cause assertion failure #2501

Closed

4 tasks

klosax mentioned this pull request Aug 8, 2023

llama : fix tokenizer #2315

Closed

add more comments

236194f

ejones marked this pull request as ready for review August 14, 2023 08:59

ejones requested a review from SlyEcho August 16, 2023 00:29

drbh mentioned this pull request Aug 16, 2023

adds simple llama grammar tests #2618

Merged

SlyEcho approved these changes Aug 17, 2023

View reviewed changes

ggerganov mentioned this pull request Aug 17, 2023

GGUF #2398

Merged

34 tasks

ejones added 2 commits August 17, 2023 19:01

Merge remote-tracking branch 'upstream/master' into fix-unicode-2

e029b50

fix test-llama-grammar

af2cd7f

ejones merged commit 604b8bd into ggml-org:master Aug 17, 2023

ejones deleted the fix-unicode-2 branch August 17, 2023 23:54

ejones mentioned this pull request Aug 24, 2023

llama : fix grammar sometimes generating null char #2756

Merged

Saibo-creator mentioned this pull request Oct 22, 2024

json() function in guidance library generates content only in English, can't output in other languages guidance-ai/guidance#1055

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix unicode in grammars (fixes #2501)#2553

Fix unicode in grammars (fixes #2501)#2553
ejones merged 4 commits intoggml-org:masterfrom
ejones:fix-unicode-2

ejones commented Aug 8, 2023 •

edited

Loading

Uh oh!

SlyEcho left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

ejones commented Aug 8, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Changes

Testing

Uh oh!

SlyEcho left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

ejones commented Aug 8, 2023 •

edited

Loading