Mirrors
/
hf_text-generation-inference
mirror of https://github.com/huggingface/text-generation-inference.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

fix: remove continue_final_message chat request param

This commit is contained in:
David Holtz 2024-11-19 21:24:18 +00:00
parent 1895d7b745
commit fa577c9be2
8 changed files with 33 additions and 62 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
docs/openapi.json
View File

@ -984,12 +984,6 @@
"messages"
],
"properties": {
"continue_final_message": {
"type": "boolean",
"description": "Whether to continue the final message in the next request.",
"default": "false",
"example": true
},
"frequency_penalty": {
"type": "number",
"format": "float",

8
integration-tests/models/__snapshots__/test_continue_final_message/test_llama_completion_single_prompt.json
View File

@ -5,19 +5,19 @@
"index": 0,
"logprobs": null,
"message": {
"content": "Hi, I hope this is the right place for your written question. Please provide the maximum possible length to help me complete the message for you! Based",
"content": "\nGenerate according to: It is an elephant's one year old baby or a mouse's one year old baby. It is",
"role": "assistant"
}
}
],
"created": 1731082056,
"created": 1732050325,
"id": "",
"model": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
"object": "chat.completion",
"system_fingerprint": "2.4.1-dev0-native",
"usage": {
"completion_tokens": 30,
"prompt_tokens": 57,
"total_tokens": 87
"prompt_tokens": 37,
"total_tokens": 67
}
}

8
integration-tests/models/__snapshots__/test_continue_final_message/test_llama_completion_single_prompt_continue.json
View File

@ -5,19 +5,19 @@
"index": 0,
"logprobs": null,
"message": {
"content": ": Thanks for the awesome slides, they were just what we needed to produce the presentation we needed to deliver for our company's budgeting system",
"content": " Shere Kohan's fantastic exploits? written by David Shimomura & illustrated by Sarah Stern\n\nTitle: Elephant",
"role": "assistant"
}
}
],
"created": 1731082129,
"created": 1732050326,
"id": "",
"model": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
"object": "chat.completion",
"system_fingerprint": "2.4.1-dev0-native",
"usage": {
"completion_tokens": 30,
"prompt_tokens": 44,
"total_tokens": 74
"prompt_tokens": 61,
"total_tokens": 91
}
}

16
integration-tests/models/test_continue_final_message.py
View File

@ -28,13 +28,11 @@ def test_llama_completion_single_prompt(
"model": "tgi",
"messages": [
{"role": "system", "content": "system message"},
{"role": "user", "content": "user message"},
{"role": "assistant", "content": "assistant message"},
{"role": "user", "content": "Which is bigger an elephant or a mouse?"},
],
"max_tokens": 30,
"stream": False,
"seed": 1337,
"continue_final_message": False,
},
headers=llama_continue_final_message.headers,
stream=False,
@ -45,7 +43,7 @@ def test_llama_completion_single_prompt(
content = response["choices"][0]["message"]["content"]
assert (
content
== "Hi, I hope this is the right place for your written question. Please provide the maximum possible length to help me complete the message for you! Based"
== "\nGenerate according to: It is an elephant's one year old baby or a mouse's one year old baby. It is"
)
assert response == response_snapshot
@ -59,13 +57,15 @@ def test_llama_completion_single_prompt_continue(
"model": "tgi",
"messages": [
{"role": "system", "content": "system message"},
{"role": "user", "content": "user message"},
{"role": "assistant", "content": "assistant message"},
{"role": "user", "content": "Which is bigger an elephant or a mouse?"},
{
"role": "assistant",
"content": "the elephant, but have you heard about",
},
],
"max_tokens": 30,
"stream": False,
"seed": 1337,
"continue_final_message": True,
},
headers=llama_continue_final_message.headers,
stream=False,
@ -76,6 +76,6 @@ def test_llama_completion_single_prompt_continue(
content = response["choices"][0]["message"]["content"]
assert (
content
== ": Thanks for the awesome slides, they were just what we needed to produce the presentation we needed to deliver for our company's budgeting system"
== " Shere Kohan's fantastic exploits? written by David Shimomura & illustrated by Sarah Stern\n\nTitle: Elephant"
)
assert response == response_snapshot

32
router/src/infer/chat_template.rs
View File

@ -54,7 +54,6 @@ impl ChatTemplate {
pub(crate) fn apply(
&self,
guideline: Option<&str>,
continue_final_message: bool,
mut messages: Vec<Message>,
tools_and_prompt: Option<(Vec<Tool>, String)>,
) -> Result<String, InferError> {
@ -85,7 +84,7 @@ impl ChatTemplate {
};
let messages: Vec<TextMessage> = messages.into_iter().map(|c| c.into()).collect();
let final_message_content = messages.last().map(|m| m.content.clone());
let final_message = messages.last().cloned();
let mut rendered_template = self
.template
.render(ChatTemplateInputs {
@ -98,20 +97,20 @@ impl ChatTemplate {
})
.map_err(InferError::TemplateError)?;
if continue_final_message {
// find the last occurrence of the final message in the rendered chat
if let Some(final_message) = final_message_content {
rendered_template = if let Some(index) = rendered_template.rfind(&final_message) {
// if the last message is from the assistant, continue the generation prompt
rendered_template = match final_message {
Some(msg) if msg.role == "assistant" => {
match rendered_template.rfind(msg.content.as_str()) {
// implementation based on feature in transformers pipeline
// https://github.com/huggingface/transformers/blob/1cf17077bf2d4affed31387c0943251a4ba8fab7/src/transformers/pipelines/text_generation.py#L418
rendered_template[..index + final_message.len()]
Some(index) => rendered_template[..index + msg.content.len()]
.trim_end()
.to_string()
} else {
rendered_template
.to_string(),
None => rendered_template,
}
}
_ => rendered_template,
};
}
}
Ok(rendered_template)
}
@ -843,9 +842,8 @@ mod tests {
content: MessageContent::SingleText("Hello, how are you?".to_string()),
},
];
let continue_final_message = false;
let result = ct.apply(None, continue_final_message, msgs, None);
let result = ct.apply(None, msgs, None);
match result {
Ok(_) => panic!("Should have failed since no guideline is provided"),
@ -885,10 +883,9 @@ mod tests {
];
let tools_string = r#"[{"type": "function","function": {"name": "get_current_weather","description": "Get the current weather","parameters": {"type": "object","properties": {"location": {"type": "string","description": "The city and state, e.g. San Francisco, CA"},"format": {"type": "string","enum": ["celsius", "fahrenheit"],"description": "The temperature unit to use. Infer this from the users location."}},"required": ["location", "format"]}}}]"#.to_string();
let tools: Vec<Tool> = serde_json::from_str(&tools_string).unwrap();
let continue_final_message = false;
let tool_prompt = "This default prompt will be used".to_string();
let tools_and_prompt = Some((tools, tool_prompt));
let result = ct.apply(None, continue_final_message, msgs, tools_and_prompt);
let result = ct.apply(None, msgs, tools_and_prompt);
let expected = "<s>[INST] I'd like to show off how chat templating works! [/INST]Great! How can I help you today?</s> [INST] Just testing\n---\n[{\"type\":\"function\",\"function\":{\"description\":\"Get the current weather\",\"name\":\"get_current_weather\",\"arguments\":{\"properties\":{\"format\":{\"description\":\"The temperature unit to use. Infer this from the users location.\",\"enum\":[\"celsius\",\"fahrenheit\"],\"type\":\"string\"},\"location\":{\"description\":\"The city and state, e.g. San Francisco, CA\",\"type\":\"string\"}},\"required\":[\"location\",\"format\"],\"type\":\"object\"}}}]\nThis default prompt will be used [/INST]".to_string();
assert_eq!(result.unwrap(), expected);
}
@ -920,10 +917,9 @@ mod tests {
];
let tools_string = r#"[{"type": "function","function": {"name": "get_current_weather","description": "Get the current weather","parameters": {"type": "object","properties": {"location": {"type": "string","description": "The city and state, e.g. San Francisco, CA"},"format": {"type": "string","enum": ["celsius", "fahrenheit"],"description": "The temperature unit to use. Infer this from the users location."}},"required": ["location", "format"]}}}]"#.to_string();
let tools: Vec<Tool> = serde_json::from_str(&tools_string).unwrap();
let continue_final_message = false;
let tool_prompt = "This default prompt will be used".to_string();
let tools_and_prompt = Some((tools, tool_prompt));
let result = ct.apply(None, continue_final_message, msgs, tools_and_prompt);
let result = ct.apply(None, msgs, tools_and_prompt);
let expected = "<s><|start_header_id|>system<|end_header_id|>\n\nEnvironment: ipython\nCutting Knowledge Date: December 2023\nToday Date: 26 Jul 2024\n\nYoure a helpful assistant! Answer the users question best you can.<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nGiven the following functions, please respond with a JSON for a function call with its proper arguments that best answers the given prompt.\n\nRespond in the format {\"name\": function name, \"parameters\": dictionary of argument name and its value}.Do not use variables.\n\n{\n \"function\": {\n \"arguments\": {\n \"properties\": {\n \"format\": {\n \"description\": \"The temperature unit to use. Infer this from the users location.\",\n \"enum\": [\n \"celsius\",\n \"fahrenheit\"\n ],\n \"type\": \"string\"\n },\n \"location\": {\n \"description\": \"The city and state, e.g. San Francisco, CA\",\n \"type\": \"string\"\n }\n },\n \"required\": [\n \"location\",\n \"format\"\n ],\n \"type\": \"object\"\n },\n \"description\": \"Get the current weather\",\n \"name\": \"get_current_weather\"\n },\n \"type\": \"function\"\n}\n\nWhat is the weather like in Brooklyn, New York?\n---\nThis default prompt will be used<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n".to_string();
assert_eq!(result.unwrap(), expected);
}

8
router/src/infer/mod.rs
View File

@ -158,19 +158,13 @@ impl Infer {
pub(crate) fn apply_chat_template(
&self,
guideline: Option<String>,
continue_final_message: bool,
messages: Vec<Message>,
tools_and_prompt: Option<(Vec<Tool>, String)>,
) -> Result<String, InferError> {
self.chat_template
.as_ref()
.ok_or_else(|| InferError::TemplateError(ErrorKind::TemplateNotFound.into()))?
.apply(
guideline.as_deref(),
continue_final_message,
messages,
tools_and_prompt,
)
.apply(guideline.as_deref(), messages, tools_and_prompt)
.map_err(|e| {
metrics::counter!("tgi_request_failure", "err" => "template").increment(1);
tracing::error!("{e}");

7
router/src/lib.rs
View File

@ -917,11 +917,6 @@ pub(crate) struct ChatRequest {
#[serde(default)]
#[schema(nullable = true, example = "null")]
pub stream_options: Option<StreamOptions>,
/// Whether to continue the final message in the next request.
#[serde(default)]
#[schema(default = "false", example = true)]
pub continue_final_message: bool,
}
impl ChatRequest {
@ -943,7 +938,6 @@ impl ChatRequest {
frequency_penalty,
top_p,
top_logprobs,
continue_final_message,
..
} = self;
@ -966,7 +960,6 @@ impl ChatRequest {
&tool_prompt,
guideline,
messages,
continue_final_message,
)?;
Ok((

10
router/src/server.rs
View File

@ -2525,7 +2525,6 @@ pub enum WebServerError {
type PreparedInput = (String, Option<GrammarType>, bool);
#[allow(clippy::too_many_arguments)]
pub(crate) fn prepare_chat_input(
infer: &Infer,
response_format: Option<GrammarType>,
@ -2534,7 +2533,6 @@ pub(crate) fn prepare_chat_input(
tool_prompt: &str,
guideline: Option<String>,
messages: Vec<Message>,
continue_final_message: bool,
) -> Result<PreparedInput, InferError> {
if response_format.is_some() && tools.is_some() {
return Err(InferError::ToolError(
@ -2544,8 +2542,7 @@ pub(crate) fn prepare_chat_input(
// when response_format is set, tools are not included when applying the chat template to generate inputs
if let Some(format) = response_format {
let inputs =
infer.apply_chat_template(guideline, continue_final_message, messages, None)?;
let inputs = infer.apply_chat_template(guideline, messages, None)?;
return Ok((inputs, Some(format), false));
}
@ -2560,7 +2557,6 @@ pub(crate) fn prepare_chat_input(
let inputs: String = infer.apply_chat_template(
guideline,
continue_final_message,
messages,
Some((updated_tools, tool_prompt.into())),
)?;
@ -2568,7 +2564,7 @@ pub(crate) fn prepare_chat_input(
}
// if no response_format or tools are set simply apply the chat template to generate inputs
let inputs = infer.apply_chat_template(guideline, continue_final_message, messages, None)?;
let inputs = infer.apply_chat_template(guideline, messages, None)?;
Ok((inputs, None, false))
}
@ -2666,7 +2662,6 @@ mod tests {
"What is the weather like in New York?".to_string(),
),
}];
let continue_final_message = false;
let result = prepare_chat_input(
&infer,
@ -2676,7 +2671,6 @@ mod tests {
tool_prompt,
guideline,
messages,
continue_final_message,
);
assert!(result.is_ok());