| Name | grammar-guide JSON |
| Version |
0.0.11
JSON |
| download |
| home_page | None |
| Summary | Speculative grammar backtracking algorithm for LLM decoding conforming to some lark context-free grammar (CFG) |
| upload_time | 2024-10-16 16:30:04 |
| maintainer | None |
| docs_url | None |
| author | None |
| requires_python | >=3.9 |
| license | Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 1. Definitions. "License" shall mean the terms and conditions for use, reproduction, and distribution as defined by Sections 1 through 9 of this document. "Licensor" shall mean the copyright owner or entity authorized by the copyright owner that is granting the License. "Legal Entity" shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, "control" means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity. "You" (or "Your") shall mean an individual or Legal Entity exercising permissions granted by this License. "Source" form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files. "Object" form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types. "Work" shall mean the work of authorship, whether in Source or Object form, made available under the License, as indicated by a copyright notice that is included in or attached to the work (an example is provided in the Appendix below). "Derivative Works" shall mean any work, whether in Source or Object form, that is based on (or derived from) the Work and for which the editorial revisions, annotations, elaborations, or other modifications represent, as a whole, an original work of authorship. For the purposes of this License, Derivative Works shall not include works that remain separable from, or merely link (or bind by name) to the interfaces of, the Work and Derivative Works thereof. "Contribution" shall mean any work of authorship, including the original version of the Work and any modifications or additions to that Work or Derivative Works thereof, that is intentionally submitted to Licensor for inclusion in the Work by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Work, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution." "Contributor" shall mean Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Work. 2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form. 3. Grant of Patent License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Work to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Work or a Contribution incorporated within the Work constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for that Work shall terminate as of the date such litigation is filed. 4. Redistribution. You may reproduce and distribute copies of the Work or Derivative Works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions: (a) You must give any other recipients of the Work or Derivative Works a copy of this License; and (b) You must cause any modified files to carry prominent notices stating that You changed the files; and (c) You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works; and (d) If the Work includes a "NOTICE" text file as part of its distribution, then any Derivative Works that You distribute must include a readable copy of the attribution notices contained within such NOTICE file, excluding those notices that do not pertain to any part of the Derivative Works, in at least one of the following places: within a NOTICE text file distributed as part of the Derivative Works; within the Source form or documentation, if provided along with the Derivative Works; or, within a display generated by the Derivative Works, if and wherever such third-party notices normally appear. The contents of the NOTICE file are for informational purposes only and do not modify the License. You may add Your own attribution notices within Derivative Works that You distribute, alongside or as an addendum to the NOTICE text from the Work, provided that such additional attribution notices cannot be construed as modifying the License. You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Derivative Works as a whole, provided Your use, reproduction, and distribution of the Work otherwise complies with the conditions stated in this License. 5. Submission of Contributions. Unless You explicitly state otherwise, any Contribution intentionally submitted for inclusion in the Work by You to the Licensor shall be under the terms and conditions of this License, without any additional terms or conditions. Notwithstanding the above, nothing herein shall supersede or modify the terms of any separate license agreement you may have executed with Licensor regarding such Contributions. 6. Trademarks. This License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor, except as required for reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file. 7. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Work (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Work and assume any risks associated with Your exercise of permissions under this License. 8. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Work (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages. 9. Accepting Warranty or Additional Liability. While redistributing the Work or Derivative Works thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability. END OF TERMS AND CONDITIONS APPENDIX: How to apply the Apache License to your work. To apply the Apache License to your work, attach the following boilerplate notice, with the fields enclosed by brackets "[]" replaced with your own identifying information. (Don't include the brackets!) The text should be enclosed in the appropriate comment syntax for the file format. We also recommend that a file or class name and description of purpose be included on the same "printed page" as the copyright notice for easier identification within third-party archives. Copyright [yyyy] [name of copyright owner] Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. |
| keywords |
language models
structured generation
constrained decoding
formal grammars
|
| VCS |
 |
| bugtrack_url |
|
| requirements |
No requirements were recorded.
|
| Travis-CI |
No Travis.
|
| coveralls test coverage |
No coveralls.
|
<div align="center"><h1> Grammar Guide </h1>
Speculative grammar backtracking algorithm to perform grammar-constrained decoding with <i>any</i> text generation function (OpenAI, Anthropic, etc.)
</div>
---
This repo is a slightly modified implementation of the decoding mechanism described in Section 3.2 of [Grammar Prompting for Domain-Specific Language
Generation with Large Language Models](https://arxiv.org/pdf/2305.19234) by [@berlino](https://github.com/berlino). I refer to the general algorithm as **speculative grammar backtracking**.
It is a form of constrained decoding, and can be used to guide even proprietary, black-box LLM APIs according to some context-free grammar. It's rooted in the idea that as LLMs get better, not *all* steps of the decoding process need to apply a strict logit mask - like a good teacher, we let the student give an answer, and only intervene and correct when necessary.
When using local Transformer models, we can efficiently backtrack the KV cache according to a given Lark CFG. Below is a benchmark showing tokens/sec when generating a JSON object with [10, 20, 30, 40] keys using [HuggingFaceTB/SmolLM-135M](https://huggingface.co/HuggingFaceTB/SmolLM-135M) on my Macbook M1.
### Features
- Compatible with *any* text generation function
- OpenAI, Anthropic etc. - as long as you can provide some `generate(prompt: str) -> str` function!
- Efficient re-use of KV cache for all CausalLM Transformer models
- Optimistic, speculative decoding = no need to manually update to support new tokenizers
- Visualization and logging of grammar corrections
- Token healing to ensure high probability continuations
```
pip install grammar-guide
```
## Examples
- [With Anthropic](./examples/anthropic-gg.ipynb)
- [With OpenAI](./examples/openai-gg.ipynb)
- [With Transformers](./examples/transformers-gg.ipynb)
### With Transformer Models
When using HuggingFace Transformer models, we get an extra speed boost by leveraging efficient caching and backtracking of the KV cache. When a grammar correction is made, we backtrack to the state of the KV cache aligned to the longest prefix that is valid under our Lark context-free grammar.
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import guidance
import grammar_guide as gg
model_name_or_path = "HuggingFaceTB/SmolLM-135M"
model = AutoModelForCausalLM.from_pretrained(model_name_or_path)
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
parser = gg.load_parser("../grammars/json_five_values_string_only.lark")
res = gg.guide(
draft_model=model,
tokenizer=tokenizer,
parser=parser,
prompt="Here's a JSON object with only string values:",
target_model=guidance.models.Transformers(
model_name_or_path, echo=False
),
stop_at=['```'],
token_lookahead=20,
max_grammar_corrections=20,
temperature=0.0
)
```
<img src="./img/grammar-guide-recording.gif" width="100%" height="100%" />
In the visualization below:
- <span style="background-color: lightgreen;">Green highlights</span> = text generated by the draft model
- <span style="background-color: lightblue;">Blue highlights</span> = candidates selected by our target model
- <span style="color: red;">Red text</span> = backtracked text that violated the context-free grammar
- <span style="color: orange;">Orange text</span> = tokens that were fed through the token healing logits processor
### With General API-based Providers
```python
import os
from openai import OpenAI
import guidance
import grammar_guide as gg
client = OpenAI(api_key=os.environ.get("OPENAI_API_KEY"))
parser = gg.load_parser("../grammars/json_five_values_string_only.lark")
# Define our core completion predict function
# This just needs to follow the `fn(s: str) -> str` contract
# so we can use any black-box API provider.
def openai_generate(prefix: str, prompt: str, max_new_tokens: int) -> str:
messages = [
{
"role": "system",
"content": prompt
}
]
if prefix:
messages += [
{
"role": "assistant",
"content": prefix
}
]
chat_completion = client.chat.completions.create(
messages=messages,
model="gpt-4o-mini",
max_tokens=max_new_tokens,
temperature=0.0
)
return chat_completion.choices[0].message.content
res = gg.guide(
draft_model=openai_generate,
parser=parser,
prompt="Here's a JSON object with only string values:",
target_model=guidance.models.Transformers(
"HuggingFaceTB/SmolLM-135M", echo=False
),
max_grammar_corrections=20,
verbose=True,
)
```
## Documentation
All calls to `gg.guide` take the following arguments. When `draft_model` is of type `AutoModelForCausalLM`, we have a bit of extra control, hence the 'Transformers only' arguments.
| Argument | Type | Description |
|----------|------|-------------|
| `draft_model` | `Union[AutoModelForCausalLM, Callable[[str], str]]` | A transformer model or callable to use for text generation. |
| `tokenizer` | `AutoTokenizer` | Transformers only, the tokenizer associated with the model. |
| `parser` | `EarleyParser` | The parser used for grammar checking. |
| `prompt` | `str` | The initial prompt for text generation. |
| `target_model` | `guidance.models.Model` | The guidance model to use for constrained grammar correction. See [guidance-ai/guidance](https://github.com/guidance-ai/guidance) |
| `seed_str` | `Optional[str]` | An optional seed string to start the generation. |
| `max_grammar_corrections` | `int` | Maximum number of grammar corrections to attempt. |
| `stop_at` | `Collection[str]` | Collection of strings to stop generation at. |
| `token_healing` | `Optional[bool]` | Transformers only, whether to use token healing during generation. |
| `top_p` | `float` | Transformers only, the cumulative probability for top-p sampling. |
| `temperature` | `float` | Transformers only, the temperature for controlling randomness in generation. |
| `token_lookahead` | `int` | Maximum number of new tokens to generate using draft model. Essentially the $K$ parameter in speculative decoding. |
| `save_html` | `bool` | Whether to save the generation process as HTML. |
| `verbose` | `bool` | Whether to print verbose output. |
| `debug` | `bool` | Whether to run in debug mode with additional checks. |
As described in [the paper](https://arxiv.org/pdf/2305.19234), one way many existing libraries achieve constrained decoding is by enforcing some constraint at each decoding timestep. For local models, it is possible to pre-process the logit masks such that this is relatively efficient. However, for closed models (think OpenAI, Anthropic, etc.), this can be 'prohitively expensive', since it would require calling the API at each timestep with the full prompt and valid continuation tokens.
Instead, this library takes an optimistic approach to constrained decoding. Autoregressive language models are only going to get better, and often times the overhead of strict, mask-driven constrained decoding isn't worth it.
For example, if we want gpt-4o to generate some SQLite query, chances are, it'll generate a valid query without any constraints.
If there is a mistake, though, we use our grammar to parse the longest prefix that abides by our grammar definition.
```python
prediction = "SELECT * FROM students WHERE name SIMILAR TO 'Dan%';"
# Oops! `SIMILAR TO` works in PostgreSQL, but not SQLite
prefix, candidates = obtain_correction_pairs(prediction, parser)
print(prefix)
# SELECT * FROM students WHERE name
print(candidates)
# ['IN', '>', '=', 'NOT', 'BETWEEN', 'LIKE', ...]
```
Once we have a list of candidates, we can use our target model to select a valid continuation. In the above example, our candidates are fairly simple strings. However, our grammar may define regular expression continuations as well (e.g. `(?:(?:[A-Z]|[a-z])|_)(?:(?:(?:[A-Z]|[a-z])|[0-9]|_))*`).
This is powered by the library [guidance](https://github.com/guidance-ai/guidance).
Once the target model has selected a valid continuation, we are free to pass the new prefix back to the draft lanugage model to complete the prediction.
```python
selected_candidate = choose_candidate(candidates, prefix, target_model)
print(selected_candidate)
# 'LIKE'
# Now, pass back to the main model to continue its prediction from this new breakpoint
draft_model.predict(prefix + selected_candidate)
```
> [!TIP]
> We borrow "Draft" and "Target" terminology from one of the original speculative decoding papers ([1](https://arxiv.org/pdf/2302.01318)). However, in our case, we consider the model constrained by the grammar which generates very small bits of text to be the 'target' model, since these generations will always be accepted. The draft model, then, is the often larger model that generates unconstrained up until we tell it to stop (governed by the `token_lookahead` parameter)
### Benchmarks
The below benchmarks are done on my Macbook M1, with the command `python -m examples.benchmarks.run`.
They measure the tokens/sec for the respective methods to generate a JSON with exactly *n* string key-value pairs, using [HuggingFaceTB/SmolLM-135M](https://huggingface.co/HuggingFaceTB/SmolLM-135M) and the below prompt.
> Here is a JSON object, with {n} keys, using only string values:\n\n```json\n
For most general usecases when using local Transformers models, I highly recommend the library [transformers-CFG](https://github.com/epfl-dlab/transformers-CFG)!
Raw data
{
"_id": null,
"home_page": null,
"name": "grammar-guide",
"maintainer": null,
"docs_url": null,
"requires_python": ">=3.9",
"maintainer_email": null,
"keywords": "language models, structured generation, constrained decoding, formal grammars",
"author": null,
"author_email": "Parker Glenn <parkervg5@gmail.com>",
"download_url": "https://files.pythonhosted.org/packages/7e/e0/eaa7dd5c7160a7390b4d1e24fbbcd6ce39b01477e4d07a1c42bf3759e9d0/grammar_guide-0.0.11.tar.gz",
"platform": null,
"description": "<div align=\"center\"><h1> Grammar Guide </h1>\n Speculative grammar backtracking algorithm to perform grammar-constrained decoding with <i>any</i> text generation function (OpenAI, Anthropic, etc.)\n</div>\n\n---\n\nThis repo is a slightly modified implementation of the decoding mechanism described in Section 3.2 of [Grammar Prompting for Domain-Specific Language\nGeneration with Large Language Models](https://arxiv.org/pdf/2305.19234) by [@berlino](https://github.com/berlino). I refer to the general algorithm as **speculative grammar backtracking**.\n\nIt is a form of constrained decoding, and can be used to guide even proprietary, black-box LLM APIs according to some context-free grammar. It's rooted in the idea that as LLMs get better, not *all* steps of the decoding process need to apply a strict logit mask - like a good teacher, we let the student give an answer, and only intervene and correct when necessary.\n\nWhen using local Transformer models, we can efficiently backtrack the KV cache according to a given Lark CFG. Below is a benchmark showing tokens/sec when generating a JSON object with [10, 20, 30, 40] keys using [HuggingFaceTB/SmolLM-135M](https://huggingface.co/HuggingFaceTB/SmolLM-135M) on my Macbook M1.\n\n\n### Features\n- Compatible with *any* text generation function\n - OpenAI, Anthropic etc. - as long as you can provide some `generate(prompt: str) -> str` function!\n- Efficient re-use of KV cache for all CausalLM Transformer models\n - Optimistic, speculative decoding = no need to manually update to support new tokenizers\n- Visualization and logging of grammar corrections\n- Token healing to ensure high probability continuations\n\n```\npip install grammar-guide\n```\n## Examples\n- [With Anthropic](./examples/anthropic-gg.ipynb)\n- [With OpenAI](./examples/openai-gg.ipynb)\n- [With Transformers](./examples/transformers-gg.ipynb)\n\n### With Transformer Models\nWhen using HuggingFace Transformer models, we get an extra speed boost by leveraging efficient caching and backtracking of the KV cache. When a grammar correction is made, we backtrack to the state of the KV cache aligned to the longest prefix that is valid under our Lark context-free grammar.\n\n```python\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\nimport guidance\n\nimport grammar_guide as gg\n\nmodel_name_or_path = \"HuggingFaceTB/SmolLM-135M\"\nmodel = AutoModelForCausalLM.from_pretrained(model_name_or_path)\ntokenizer = AutoTokenizer.from_pretrained(model_name_or_path)\nparser = gg.load_parser(\"../grammars/json_five_values_string_only.lark\")\n\nres = gg.guide(\n draft_model=model,\n tokenizer=tokenizer,\n parser=parser,\n prompt=\"Here's a JSON object with only string values:\",\n target_model=guidance.models.Transformers(\n model_name_or_path, echo=False\n ),\n stop_at=['```'],\n token_lookahead=20,\n max_grammar_corrections=20,\n temperature=0.0\n)\n```\n<img src=\"./img/grammar-guide-recording.gif\" width=\"100%\" height=\"100%\" />\n\nIn the visualization below:\n- <span style=\"background-color: lightgreen;\">Green highlights</span> = text generated by the draft model\n- <span style=\"background-color: lightblue;\">Blue highlights</span> = candidates selected by our target model\n- <span style=\"color: red;\">Red text</span> = backtracked text that violated the context-free grammar\n- <span style=\"color: orange;\">Orange text</span> = tokens that were fed through the token healing logits processor\n\n\n\n\n### With General API-based Providers\n```python\nimport os\nfrom openai import OpenAI\nimport guidance\n\nimport grammar_guide as gg\n\nclient = OpenAI(api_key=os.environ.get(\"OPENAI_API_KEY\"))\nparser = gg.load_parser(\"../grammars/json_five_values_string_only.lark\")\n\n# Define our core completion predict function\n# This just needs to follow the `fn(s: str) -> str` contract\n# so we can use any black-box API provider.\ndef openai_generate(prefix: str, prompt: str, max_new_tokens: int) -> str:\n messages = [\n {\n \"role\": \"system\",\n \"content\": prompt\n }\n ]\n if prefix:\n messages += [\n {\n \"role\": \"assistant\",\n \"content\": prefix\n }\n ]\n chat_completion = client.chat.completions.create(\n messages=messages,\n model=\"gpt-4o-mini\",\n max_tokens=max_new_tokens,\n temperature=0.0\n )\n return chat_completion.choices[0].message.content\n\nres = gg.guide(\n draft_model=openai_generate,\n parser=parser,\n prompt=\"Here's a JSON object with only string values:\",\n target_model=guidance.models.Transformers(\n \"HuggingFaceTB/SmolLM-135M\", echo=False\n ),\n max_grammar_corrections=20,\n verbose=True,\n)\n```\n\n## Documentation\n\nAll calls to `gg.guide` take the following arguments. When `draft_model` is of type `AutoModelForCausalLM`, we have a bit of extra control, hence the 'Transformers only' arguments.\n\n| Argument | Type | Description |\n|----------|------|-------------|\n| `draft_model` | `Union[AutoModelForCausalLM, Callable[[str], str]]` | A transformer model or callable to use for text generation. |\n| `tokenizer` | `AutoTokenizer` | Transformers only, the tokenizer associated with the model. |\n| `parser` | `EarleyParser` | The parser used for grammar checking. |\n| `prompt` | `str` | The initial prompt for text generation. |\n| `target_model` | `guidance.models.Model` | The guidance model to use for constrained grammar correction. See [guidance-ai/guidance](https://github.com/guidance-ai/guidance) |\n| `seed_str` | `Optional[str]` | An optional seed string to start the generation. |\n| `max_grammar_corrections` | `int` | Maximum number of grammar corrections to attempt. |\n| `stop_at` | `Collection[str]` | Collection of strings to stop generation at. |\n| `token_healing` | `Optional[bool]` | Transformers only, whether to use token healing during generation. |\n| `top_p` | `float` | Transformers only, the cumulative probability for top-p sampling. |\n| `temperature` | `float` | Transformers only, the temperature for controlling randomness in generation. |\n| `token_lookahead` | `int` | Maximum number of new tokens to generate using draft model. Essentially the $K$ parameter in speculative decoding. |\n| `save_html` | `bool` | Whether to save the generation process as HTML. |\n| `verbose` | `bool` | Whether to print verbose output. |\n| `debug` | `bool` | Whether to run in debug mode with additional checks. |\n\nAs described in [the paper](https://arxiv.org/pdf/2305.19234), one way many existing libraries achieve constrained decoding is by enforcing some constraint at each decoding timestep. For local models, it is possible to pre-process the logit masks such that this is relatively efficient. However, for closed models (think OpenAI, Anthropic, etc.), this can be 'prohitively expensive', since it would require calling the API at each timestep with the full prompt and valid continuation tokens.\n\nInstead, this library takes an optimistic approach to constrained decoding. Autoregressive language models are only going to get better, and often times the overhead of strict, mask-driven constrained decoding isn't worth it.\n\nFor example, if we want gpt-4o to generate some SQLite query, chances are, it'll generate a valid query without any constraints.\n\nIf there is a mistake, though, we use our grammar to parse the longest prefix that abides by our grammar definition.\n\n```python\nprediction = \"SELECT * FROM students WHERE name SIMILAR TO 'Dan%';\"\n# Oops! `SIMILAR TO` works in PostgreSQL, but not SQLite\nprefix, candidates = obtain_correction_pairs(prediction, parser)\nprint(prefix)\n# SELECT * FROM students WHERE name\nprint(candidates)\n# ['IN', '>', '=', 'NOT', 'BETWEEN', 'LIKE', ...]\n```\nOnce we have a list of candidates, we can use our target model to select a valid continuation. In the above example, our candidates are fairly simple strings. However, our grammar may define regular expression continuations as well (e.g. `(?:(?:[A-Z]|[a-z])|_)(?:(?:(?:[A-Z]|[a-z])|[0-9]|_))*`).\nThis is powered by the library [guidance](https://github.com/guidance-ai/guidance).\n\nOnce the target model has selected a valid continuation, we are free to pass the new prefix back to the draft lanugage model to complete the prediction.\n\n```python\nselected_candidate = choose_candidate(candidates, prefix, target_model)\nprint(selected_candidate)\n# 'LIKE'\n# Now, pass back to the main model to continue its prediction from this new breakpoint\ndraft_model.predict(prefix + selected_candidate)\n```\n\n> [!TIP]\n> We borrow \"Draft\" and \"Target\" terminology from one of the original speculative decoding papers ([1](https://arxiv.org/pdf/2302.01318)). However, in our case, we consider the model constrained by the grammar which generates very small bits of text to be the 'target' model, since these generations will always be accepted. The draft model, then, is the often larger model that generates unconstrained up until we tell it to stop (governed by the `token_lookahead` parameter)\n\n### Benchmarks\nThe below benchmarks are done on my Macbook M1, with the command `python -m examples.benchmarks.run`.\n\nThey measure the tokens/sec for the respective methods to generate a JSON with exactly *n* string key-value pairs, using [HuggingFaceTB/SmolLM-135M](https://huggingface.co/HuggingFaceTB/SmolLM-135M) and the below prompt.\n> Here is a JSON object, with {n} keys, using only string values:\\n\\n```json\\n\n\nFor most general usecases when using local Transformers models, I highly recommend the library [transformers-CFG](https://github.com/epfl-dlab/transformers-CFG)!\n\n\n",
"bugtrack_url": null,
"license": "Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 1. Definitions. \"License\" shall mean the terms and conditions for use, reproduction, and distribution as defined by Sections 1 through 9 of this document. \"Licensor\" shall mean the copyright owner or entity authorized by the copyright owner that is granting the License. \"Legal Entity\" shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, \"control\" means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity. \"You\" (or \"Your\") shall mean an individual or Legal Entity exercising permissions granted by this License. \"Source\" form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files. \"Object\" form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types. \"Work\" shall mean the work of authorship, whether in Source or Object form, made available under the License, as indicated by a copyright notice that is included in or attached to the work (an example is provided in the Appendix below). \"Derivative Works\" shall mean any work, whether in Source or Object form, that is based on (or derived from) the Work and for which the editorial revisions, annotations, elaborations, or other modifications represent, as a whole, an original work of authorship. For the purposes of this License, Derivative Works shall not include works that remain separable from, or merely link (or bind by name) to the interfaces of, the Work and Derivative Works thereof. \"Contribution\" shall mean any work of authorship, including the original version of the Work and any modifications or additions to that Work or Derivative Works thereof, that is intentionally submitted to Licensor for inclusion in the Work by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, \"submitted\" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Work, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as \"Not a Contribution.\" \"Contributor\" shall mean Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Work. 2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form. 3. Grant of Patent License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Work to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Work or a Contribution incorporated within the Work constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for that Work shall terminate as of the date such litigation is filed. 4. Redistribution. You may reproduce and distribute copies of the Work or Derivative Works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions: (a) You must give any other recipients of the Work or Derivative Works a copy of this License; and (b) You must cause any modified files to carry prominent notices stating that You changed the files; and (c) You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works; and (d) If the Work includes a \"NOTICE\" text file as part of its distribution, then any Derivative Works that You distribute must include a readable copy of the attribution notices contained within such NOTICE file, excluding those notices that do not pertain to any part of the Derivative Works, in at least one of the following places: within a NOTICE text file distributed as part of the Derivative Works; within the Source form or documentation, if provided along with the Derivative Works; or, within a display generated by the Derivative Works, if and wherever such third-party notices normally appear. The contents of the NOTICE file are for informational purposes only and do not modify the License. You may add Your own attribution notices within Derivative Works that You distribute, alongside or as an addendum to the NOTICE text from the Work, provided that such additional attribution notices cannot be construed as modifying the License. You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Derivative Works as a whole, provided Your use, reproduction, and distribution of the Work otherwise complies with the conditions stated in this License. 5. Submission of Contributions. Unless You explicitly state otherwise, any Contribution intentionally submitted for inclusion in the Work by You to the Licensor shall be under the terms and conditions of this License, without any additional terms or conditions. Notwithstanding the above, nothing herein shall supersede or modify the terms of any separate license agreement you may have executed with Licensor regarding such Contributions. 6. Trademarks. This License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor, except as required for reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file. 7. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Work (and each Contributor provides its Contributions) on an \"AS IS\" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Work and assume any risks associated with Your exercise of permissions under this License. 8. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Work (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages. 9. Accepting Warranty or Additional Liability. While redistributing the Work or Derivative Works thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability. END OF TERMS AND CONDITIONS APPENDIX: How to apply the Apache License to your work. To apply the Apache License to your work, attach the following boilerplate notice, with the fields enclosed by brackets \"[]\" replaced with your own identifying information. (Don't include the brackets!) The text should be enclosed in the appropriate comment syntax for the file format. We also recommend that a file or class name and description of purpose be included on the same \"printed page\" as the copyright notice for easier identification within third-party archives. Copyright [yyyy] [name of copyright owner] Licensed under the Apache License, Version 2.0 (the \"License\"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an \"AS IS\" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ",
"summary": "Speculative grammar backtracking algorithm for LLM decoding conforming to some lark context-free grammar (CFG)",
"version": "0.0.11",
"project_urls": {
"Documentation": "https://github.com/parkervg/grammar-guide"
},
"split_keywords": [
"language models",
" structured generation",
" constrained decoding",
" formal grammars"
],
"urls": [
{
"comment_text": "",
"digests": {
"blake2b_256": "13257d43ac20bc8d3880a79c92076fe2256b1a20405746721eed973ae2814779",
"md5": "7a32697f8c3d26ae037fe76280ea70c7",
"sha256": "110e797fb1fd45596cf77ea77d0b6c53c56dd2d439866b661fc5ff2b47074bfe"
},
"downloads": -1,
"filename": "grammar_guide-0.0.11-py3-none-any.whl",
"has_sig": false,
"md5_digest": "7a32697f8c3d26ae037fe76280ea70c7",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": ">=3.9",
"size": 49392,
"upload_time": "2024-10-16T16:30:02",
"upload_time_iso_8601": "2024-10-16T16:30:02.516904Z",
"url": "https://files.pythonhosted.org/packages/13/25/7d43ac20bc8d3880a79c92076fe2256b1a20405746721eed973ae2814779/grammar_guide-0.0.11-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "7ee0eaa7dd5c7160a7390b4d1e24fbbcd6ce39b01477e4d07a1c42bf3759e9d0",
"md5": "1c19b10ae43ed548b6524d8bcc6f813a",
"sha256": "ac3c5c40dee9b835e3f7222daf57634b6f4749c38d07362c81c793252971eb4f"
},
"downloads": -1,
"filename": "grammar_guide-0.0.11.tar.gz",
"has_sig": false,
"md5_digest": "1c19b10ae43ed548b6524d8bcc6f813a",
"packagetype": "sdist",
"python_version": "source",
"requires_python": ">=3.9",
"size": 50589,
"upload_time": "2024-10-16T16:30:04",
"upload_time_iso_8601": "2024-10-16T16:30:04.366277Z",
"url": "https://files.pythonhosted.org/packages/7e/e0/eaa7dd5c7160a7390b4d1e24fbbcd6ce39b01477e4d07a1c42bf3759e9d0/grammar_guide-0.0.11.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2024-10-16 16:30:04",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "parkervg",
"github_project": "grammar-guide",
"travis_ci": false,
"coveralls": false,
"github_actions": true,
"lcname": "grammar-guide"
}
