huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-04-24 08:22:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

feat: improve temperature logic in chat (#1749)

Browse Source 
This PR adds support for `do_sample` to chat to enable greedy sampling

---------

Co-authored-by: Nicolas Patry <patry.nicolas@protonmail.com>
This commit is contained in:
drbh 2024-04-25 09:31:35 -04:00 committed by Karol Damaszke
parent 5d36a5e368
commit ab59a5e346
2 changed files with 12 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
router/src/server.rs
View File

@ -1002,6 +1002,7 @@ async fn chat_completions(
tools, tools,
tool_choice, tool_choice,
tool_prompt, tool_prompt,
temperature,
.. ..
} = req; } = req;
@ -1010,6 +1011,11 @@ async fn chat_completions(
let logprobs = logprobs.unwrap_or(false); let logprobs = logprobs.unwrap_or(false);
let tool_prompt = tool_prompt.unwrap_or_default(); let tool_prompt = tool_prompt.unwrap_or_default();
let stop = stop.unwrap_or_default(); let stop = stop.unwrap_or_default();
// enable greedy only when temperature is 0
let (do_sample, temperature) = match temperature {
Some(temperature) if temperature == 0.0 => (false, None),
other => (true, other),
};
// extract tool grammar if present // extract tool grammar if present
let tool_grammar = match ToolGrammar::apply(tools, tool_choice) { let tool_grammar = match ToolGrammar::apply(tools, tool_choice) {
@ -1056,13 +1062,13 @@ async fn chat_completions(
inputs: inputs.to_string(), inputs: inputs.to_string(),
parameters: GenerateParameters { parameters: GenerateParameters {
best_of: None, best_of: None,
temperature: req.temperature, temperature,
repetition_penalty, repetition_penalty,
frequency_penalty: req.frequency_penalty, frequency_penalty: req.frequency_penalty,
top_k: None, top_k: None,
top_p: req.top_p, top_p: req.top_p,
typical_p: None, typical_p: None,
do_sample: true, do_sample,
max_new_tokens, max_new_tokens,
return_full_text: None, return_full_text: None,
stop, stop,

8
server/text_generation_server/utils/tokens.py
View File

@ -274,7 +274,7 @@ class HeterogeneousNextTokenChooser:
else None else None
) )
if any([x != 1.0 for x in temperature]): if any(x != 1.0 for x in temperature):
do_sample = [ do_sample = [
sample or x != 1.0 for x, sample in zip(temperature, do_sample) sample or x != 1.0 for x, sample in zip(temperature, do_sample)
] ]
@ -282,15 +282,15 @@ class HeterogeneousNextTokenChooser:
HeterogeneousTemperatureLogitsWarper(temperature, dtype, device) HeterogeneousTemperatureLogitsWarper(temperature, dtype, device)
) )
if any([x != 0 for x in top_k]): if any(x != 0 for x in top_k):
do_sample = [sample or x != 0 for x, sample in zip(top_k, do_sample)] do_sample = [sample or x != 0 for x, sample in zip(top_k, do_sample)]
warpers.append(HeterogeneousTopKLogitsWarper(top_k, device)) warpers.append(HeterogeneousTopKLogitsWarper(top_k, device))
if any([x < 1.0 for x in top_p]): if any(x < 1.0 for x in top_p):
do_sample = [sample or x < 1.0 for x, sample in zip(top_p, do_sample)] do_sample = [sample or x < 1.0 for x, sample in zip(top_p, do_sample)]
warpers.append(HeterogeneousTopPLogitsWarper(top_p, dtype, device)) warpers.append(HeterogeneousTopPLogitsWarper(top_p, dtype, device))
if any([x < 1.0 for x in typical_p]): if any(x < 1.0 for x in typical_p):
do_sample = [sample or x < 1.0 for x, sample in zip(typical_p, do_sample)] do_sample = [sample or x < 1.0 for x, sample in zip(typical_p, do_sample)]
warpers.append(HeterogeneousTypicalLogitsWarper(typical_p, dtype, device)) warpers.append(HeterogeneousTypicalLogitsWarper(typical_p, dtype, device))