huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-04-27 13:02:12 +00:00
Code Issues Packages Projects Releases Wiki Activity

Modify the default for max_new_tokens. (#1097)

Browse Source 
# What does this PR do?

Now clients which do not specify a max_length will be implying
`max_new_tokens = max_total_tokens - input_length`.
This is a serious change, but which seems more in line with what users
expect from standing server.

<!--
Congratulations! You've made it this far! You're not quite done yet
though.

Once merged, your PR is going to appear in the release notes with the
title you set, so make sure it's a great title that fully reflects the
extent of your awesome contribution.

Then, please replace this with a description of the change and which
issue is fixed (if applicable). Please also include relevant motivation
and context. List any dependencies (if any) that are required for this
change.

Once you're done, someone will review your PR shortly (see the section
"Who can review?" below to tag some potential reviewers). They may
suggest changes to make the code even better. If no one reviewed your PR
after a week has passed, don't hesitate to post a new comment
@-mentioning the same persons---sometimes notifications get lost.
-->

<!-- Remove if not applicable -->

Fixes # (issue)


## Before submitting
- [ ] This PR fixes a typo or improves the docs (you can dismiss the
other checks if that's the case).
- [ ] Did you read the [contributor
guideline](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#start-contributing-pull-requests),
      Pull Request section?
- [ ] Was this discussed/approved via a Github issue or the
[forum](https://discuss.huggingface.co/)? Please add a link
      to it if that's the case.
- [ ] Did you make sure to update the documentation with your changes?
Here are the
[documentation
guidelines](https://github.com/huggingface/transformers/tree/main/docs),
and
[here are tips on formatting
docstrings](https://github.com/huggingface/transformers/tree/main/docs#writing-source-documentation).
- [ ] Did you write any new necessary tests?


## Who can review?

Anyone in the community is free to review the PR once the tests have
passed. Feel free to tag
members/contributors who may be interested in your PR.

<!-- Your PR will be replied to more quickly if you can figure out the
right person to tag with @


@OlivierDehaene OR @Narsil

 -->

---------

Co-authored-by: OlivierDehaene <olivier@huggingface.co>
This commit is contained in:
Nicolas Patry 2023-10-04 17:38:42 +02:00 committed by GitHub
parent 66ce2fa7c1
commit 6df43da0a4
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 88 additions and 49 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

98
docs/openapi.json
View File

@ -367,7 +367,7 @@
"type": "integer", "type": "integer",
"format": "int32", "format": "int32",
"example": 1, "example": 1,
"minimum": 0.0 "minimum": 0
}, },
"prefill": { "prefill": {
"type": "array", "type": "array",
@ -380,13 +380,22 @@
"format": "int64", "format": "int64",
"example": 42, "example": 42,
"nullable": true, "nullable": true,
"minimum": 0.0 "minimum": 0
}, },
"tokens": { "tokens": {
"type": "array", "type": "array",
"items": { "items": {
"$ref": "#/components/schemas/Token" "$ref": "#/components/schemas/Token"
} }
},
"top_tokens": {
"type": "array",
"items": {
"type": "array",
"items": {
"$ref": "#/components/schemas/Token"
}
}
} }
} }
}, },
@ -432,7 +441,7 @@
"type": "integer", "type": "integer",
"format": "int32", "format": "int32",
"example": 1, "example": 1,
"minimum": 0.0 "minimum": 0
}, },
"prefill": { "prefill": {
"type": "array", "type": "array",
@ -445,13 +454,22 @@
"format": "int64", "format": "int64",
"example": 42, "example": 42,
"nullable": true, "nullable": true,
"minimum": 0.0 "minimum": 0
}, },
"tokens": { "tokens": {
"type": "array", "type": "array",
"items": { "items": {
"$ref": "#/components/schemas/Token" "$ref": "#/components/schemas/Token"
} }
},
"top_tokens": {
"type": "array",
"items": {
"type": "array",
"items": {
"$ref": "#/components/schemas/Token"
}
}
} }
} }
}, },
@ -486,8 +504,8 @@
"default": "null", "default": "null",
"example": 1, "example": 1,
"nullable": true, "nullable": true,
"minimum": 0.0, "minimum": 0,
"exclusiveMinimum": 0.0 "exclusiveMinimum": 0
}, },
"decoder_input_details": { "decoder_input_details": {
"type": "boolean", "type": "boolean",
@ -505,10 +523,10 @@
"max_new_tokens": { "max_new_tokens": {
"type": "integer", "type": "integer",
"format": "int32", "format": "int32",
"default": "20", "default": "null",
"minimum": 0.0, "example": "20",
"exclusiveMaximum": 512.0, "nullable": true,
"exclusiveMinimum": 0.0 "minimum": 0
}, },
"repetition_penalty": { "repetition_penalty": {
"type": "number", "type": "number",
@ -516,7 +534,7 @@
"default": "null", "default": "null",
"example": 1.03, "example": 1.03,
"nullable": true, "nullable": true,
"exclusiveMinimum": 0.0 "exclusiveMinimum": 0
}, },
"return_full_text": { "return_full_text": {
"type": "boolean", "type": "boolean",
@ -530,8 +548,8 @@
"default": "null", "default": "null",
"example": "null", "example": "null",
"nullable": true, "nullable": true,
"minimum": 0.0, "minimum": 0,
"exclusiveMinimum": 0.0 "exclusiveMinimum": 0
}, },
"stop": { "stop": {
"type": "array", "type": "array",
@ -549,7 +567,7 @@
"default": "null", "default": "null",
"example": 0.5, "example": 0.5,
"nullable": true, "nullable": true,
"exclusiveMinimum": 0.0 "exclusiveMinimum": 0
}, },
"top_k": { "top_k": {
"type": "integer", "type": "integer",
@ -557,7 +575,16 @@
"default": "null", "default": "null",
"example": 10, "example": 10,
"nullable": true, "nullable": true,
"exclusiveMinimum": 0.0 "exclusiveMinimum": 0
},
"top_n_tokens": {
"type": "integer",
"format": "int32",
"default": "null",
"example": 5,
"nullable": true,
"minimum": 0,
"exclusiveMinimum": 0
}, },
"top_p": { "top_p": {
"type": "number", "type": "number",
@ -565,15 +592,15 @@
"default": "null", "default": "null",
"example": 0.95, "example": 0.95,
"nullable": true, "nullable": true,
"maximum": 1.0, "maximum": 1,
"exclusiveMinimum": 0.0 "exclusiveMinimum": 0
}, },
"truncate": { "truncate": {
"type": "integer", "type": "integer",
"default": "null", "default": "null",
"example": "null", "example": "null",
"nullable": true, "nullable": true,
"minimum": 0.0 "minimum": 0
}, },
"typical_p": { "typical_p": {
"type": "number", "type": "number",
@ -581,8 +608,8 @@
"default": "null", "default": "null",
"example": 0.95, "example": 0.95,
"nullable": true, "nullable": true,
"maximum": 1.0, "maximum": 1,
"exclusiveMinimum": 0.0 "exclusiveMinimum": 0
}, },
"watermark": { "watermark": {
"type": "boolean", "type": "boolean",
@ -653,38 +680,38 @@
"type": "integer", "type": "integer",
"format": "int32", "format": "int32",
"example": "32000", "example": "32000",
"minimum": 0.0 "minimum": 0
}, },
"max_best_of": { "max_best_of": {
"type": "integer", "type": "integer",
"example": "2", "example": "2",
"minimum": 0.0 "minimum": 0
}, },
"max_concurrent_requests": { "max_concurrent_requests": {
"type": "integer", "type": "integer",
"description": "Router Parameters", "description": "Router Parameters",
"example": "128", "example": "128",
"minimum": 0.0 "minimum": 0
}, },
"max_input_length": { "max_input_length": {
"type": "integer", "type": "integer",
"example": "1024", "example": "1024",
"minimum": 0.0 "minimum": 0
}, },
"max_stop_sequences": { "max_stop_sequences": {
"type": "integer", "type": "integer",
"example": "4", "example": "4",
"minimum": 0.0 "minimum": 0
}, },
"max_total_tokens": { "max_total_tokens": {
"type": "integer", "type": "integer",
"example": "2048", "example": "2048",
"minimum": 0.0 "minimum": 0
}, },
"max_waiting_tokens": { "max_waiting_tokens": {
"type": "integer", "type": "integer",
"example": "20", "example": "20",
"minimum": 0.0 "minimum": 0
}, },
"model_device_type": { "model_device_type": {
"type": "string", "type": "string",
@ -717,7 +744,7 @@
"validation_workers": { "validation_workers": {
"type": "integer", "type": "integer",
"example": "2", "example": "2",
"minimum": 0.0 "minimum": 0
}, },
"version": { "version": {
"type": "string", "type": "string",
@ -743,7 +770,7 @@
"type": "integer", "type": "integer",
"format": "int32", "format": "int32",
"example": 0, "example": 0,
"minimum": 0.0 "minimum": 0
}, },
"logprob": { "logprob": {
"type": "number", "type": "number",
@ -771,14 +798,14 @@
"type": "integer", "type": "integer",
"format": "int32", "format": "int32",
"example": 1, "example": 1,
"minimum": 0.0 "minimum": 0
}, },
"seed": { "seed": {
"type": "integer", "type": "integer",
"format": "int64", "format": "int64",
"example": 42, "example": 42,
"nullable": true, "nullable": true,
"minimum": 0.0 "minimum": 0
} }
} }
}, },
@ -794,6 +821,7 @@
"$ref": "#/components/schemas/StreamDetails" "$ref": "#/components/schemas/StreamDetails"
} }
], ],
"default": "null",
"nullable": true "nullable": true
}, },
"generated_text": { "generated_text": {
@ -804,6 +832,12 @@
}, },
"token": { "token": {
"$ref": "#/components/schemas/Token" "$ref": "#/components/schemas/Token"
},
"top_tokens": {
"type": "array",
"items": {
"$ref": "#/components/schemas/Token"
}
} }
} }
}, },
@ -820,7 +854,7 @@
"type": "integer", "type": "integer",
"format": "int32", "format": "int32",
"example": 0, "example": 0,
"minimum": 0.0 "minimum": 0
}, },
"logprob": { "logprob": {
"type": "number", "type": "number",

8
router/src/lib.rs
View File

@ -107,8 +107,8 @@ pub(crate) struct GenerateParameters {
#[schema(default = "false", example = true)] #[schema(default = "false", example = true)]
pub do_sample: bool, pub do_sample: bool,
#[serde(default = "default_max_new_tokens")] #[serde(default = "default_max_new_tokens")]
#[schema(exclusive_minimum = 0, exclusive_maximum = 512, default = "20")] #[schema(nullable = true, default = "null", example = "20")]
pub max_new_tokens: u32, pub max_new_tokens: Option<u32>,
#[serde(default)] #[serde(default)]
#[schema(nullable = true, default = "null", example = false)] #[schema(nullable = true, default = "null", example = false)]
pub return_full_text: Option<bool>, pub return_full_text: Option<bool>,
@ -140,8 +140,8 @@ pub(crate) struct GenerateParameters {
pub top_n_tokens: Option<u32>, pub top_n_tokens: Option<u32>,
} }
fn default_max_new_tokens() -> u32 { fn default_max_new_tokens() -> Option<u32> {
20 None
} }
fn default_parameters() -> GenerateParameters { fn default_parameters() -> GenerateParameters {

31
router/src/validation.rs
View File

@ -67,8 +67,8 @@ impl Validation {
&self, &self,
inputs: String, inputs: String,
truncate: Option<usize>, truncate: Option<usize>,
max_new_tokens: u32, max_new_tokens: Option<u32>,
) -> Result<(String, usize), ValidationError> { ) -> Result<(String, usize, u32), ValidationError> {
// If we have a fast tokenizer // If we have a fast tokenizer
if let Some(sender) = &self.sender { if let Some(sender) = &self.sender {
// Create response channel // Create response channel
@ -84,6 +84,11 @@ impl Validation {
let (inputs, input_length) = response_receiver.await.unwrap()?; let (inputs, input_length) = response_receiver.await.unwrap()?;
// Get total tokens // Get total tokens
let max_new_tokens: u32 = if let Some(max_new_tokens) = max_new_tokens {
max_new_tokens
} else {
self.max_total_tokens.saturating_sub(input_length) as u32
};
let total_tokens = input_length + max_new_tokens as usize; let total_tokens = input_length + max_new_tokens as usize;
// Validate MaxTotalTokens // Validate MaxTotalTokens
@ -104,7 +109,7 @@ impl Validation {
} }
metrics::histogram!("tgi_request_input_length", input_length as f64); metrics::histogram!("tgi_request_input_length", input_length as f64);
Ok((inputs, input_length)) Ok((inputs, input_length, max_new_tokens))
} }
// Return inputs without validation // Return inputs without validation
else { else {
@ -112,6 +117,11 @@ impl Validation {
// However, the inputs will be truncated by the python servers // However, the inputs will be truncated by the python servers
// We make sure that truncate + max_new_tokens <= self.max_total_tokens // We make sure that truncate + max_new_tokens <= self.max_total_tokens
let input_length = truncate.unwrap_or(self.max_input_length); let input_length = truncate.unwrap_or(self.max_input_length);
let max_new_tokens: u32 = if let Some(max_new_tokens) = max_new_tokens {
max_new_tokens
} else {
self.max_total_tokens.saturating_sub(input_length) as u32
};
// Validate MaxNewTokens // Validate MaxNewTokens
if (input_length as u32 + max_new_tokens) > self.max_total_tokens as u32 { if (input_length as u32 + max_new_tokens) > self.max_total_tokens as u32 {
@ -121,7 +131,7 @@ impl Validation {
)); ));
} }
Ok((inputs, input_length)) Ok((inputs, input_length, max_new_tokens))
} }
} }
@ -200,7 +210,7 @@ impl Validation {
}) })
.unwrap_or(Ok(0))?; .unwrap_or(Ok(0))?;
if max_new_tokens == 0 { if max_new_tokens == Some(0) {
return Err(ValidationError::NegativeMaxNewTokens); return Err(ValidationError::NegativeMaxNewTokens);
} }
@ -247,7 +257,7 @@ impl Validation {
.unwrap_or(Ok(None))?; .unwrap_or(Ok(None))?;
// Validate inputs // Validate inputs
let (inputs, input_length) = self let (inputs, input_length, max_new_tokens) = self
.validate_input(request.inputs, truncate, max_new_tokens) .validate_input(request.inputs, truncate, max_new_tokens)
.await?; .await?;
@ -426,7 +436,7 @@ mod tests {
let max_new_tokens = 10; let max_new_tokens = 10;
match validation match validation
.validate_input("Hello".to_string(), None, max_new_tokens) .validate_input("Hello".to_string(), None, Some(max_new_tokens))
.await .await
{ {
Err(ValidationError::MaxNewTokens(1, 10)) => (), Err(ValidationError::MaxNewTokens(1, 10)) => (),
@ -455,7 +465,7 @@ mod tests {
let max_new_tokens = 10; let max_new_tokens = 10;
match validation match validation
.validate_input("Hello".to_string(), None, max_new_tokens) .validate_input("Hello".to_string(), None, Some(max_new_tokens))
.await .await
{ {
Err(ValidationError::MaxTotalTokens(6, 1, 10)) => (), Err(ValidationError::MaxTotalTokens(6, 1, 10)) => (),
@ -534,7 +544,6 @@ mod tests {
inputs: "Hello".to_string(), inputs: "Hello".to_string(),
parameters: GenerateParameters { parameters: GenerateParameters {
top_p: Some(0.99), top_p: Some(0.99),
max_new_tokens: 1,
..default_parameters() ..default_parameters()
}, },
}) })
@ -549,7 +558,6 @@ mod tests {
inputs: "Hello".to_string(), inputs: "Hello".to_string(),
parameters: GenerateParameters { parameters: GenerateParameters {
top_p: None, top_p: None,
max_new_tokens: 1,
..default_parameters() ..default_parameters()
}, },
}) })
@ -596,7 +604,6 @@ mod tests {
inputs: "Hello".to_string(), inputs: "Hello".to_string(),
parameters: GenerateParameters { parameters: GenerateParameters {
top_n_tokens: Some(4), top_n_tokens: Some(4),
max_new_tokens: 1,
..default_parameters() ..default_parameters()
}, },
}) })
@ -608,7 +615,6 @@ mod tests {
inputs: "Hello".to_string(), inputs: "Hello".to_string(),
parameters: GenerateParameters { parameters: GenerateParameters {
top_n_tokens: Some(0), top_n_tokens: Some(0),
max_new_tokens: 1,
..default_parameters() ..default_parameters()
}, },
}) })
@ -620,7 +626,6 @@ mod tests {
inputs: "Hello".to_string(), inputs: "Hello".to_string(),
parameters: GenerateParameters { parameters: GenerateParameters {
top_n_tokens: None, top_n_tokens: None,
max_new_tokens: 1,
..default_parameters() ..default_parameters()
}, },
}) })