from gguf build chat_template and tokenizer
This commit is contained in:
@@ -66,11 +66,12 @@ impl<'a> GenerateModel for Qwen3GenerateModel<'a> {
|
||||
let seed = mes.seed.unwrap_or(34562) as u64;
|
||||
let mut logit_processor =
|
||||
get_logit_processor(Some(temperature), Some(top_p), Some(top_k), seed);
|
||||
let enable_thinking = extract_metadata_value::<bool>(&mes.metadata, "enable_thinking");
|
||||
// let mes_render = self.chat_template.apply_chat_template(&mes)?;
|
||||
let mes_render = self
|
||||
.chat_template
|
||||
.apply_chat_temp_think(&mes, enable_thinking)?;
|
||||
|
||||
let mes_render = self.chat_template.apply_chat_template(&mes)?;
|
||||
// let enable_thinking = extract_metadata_value::<bool>(&mes.metadata, "enable_thinking");
|
||||
// let mes_render = self
|
||||
// .chat_template
|
||||
// .apply_chat_temp_think(&mes, enable_thinking)?;
|
||||
let mut input_ids = self.tokenizer.text_encode(mes_render, &self.device)?;
|
||||
let mut seq_len = input_ids.dim(1)?;
|
||||
let prompt_tokens = seq_len as u32;
|
||||
@@ -115,10 +116,11 @@ impl<'a> GenerateModel for Qwen3GenerateModel<'a> {
|
||||
let seed = mes.seed.unwrap_or(34562) as u64;
|
||||
let mut logit_processor =
|
||||
get_logit_processor(Some(temperature), Some(top_p), Some(top_k), seed);
|
||||
let enable_thinking = extract_metadata_value::<bool>(&mes.metadata, "enable_thinking");
|
||||
let mes_render = self
|
||||
.chat_template
|
||||
.apply_chat_temp_think(&mes, enable_thinking)?;
|
||||
let mes_render = self.chat_template.apply_chat_template(&mes)?;
|
||||
// let enable_thinking = extract_metadata_value::<bool>(&mes.metadata, "enable_thinking");
|
||||
// let mes_render = self
|
||||
// .chat_template
|
||||
// .apply_chat_temp_think(&mes, enable_thinking)?;
|
||||
let mut input_ids = self.tokenizer.text_encode(mes_render, &self.device)?;
|
||||
let mut seq_len = input_ids.dim(1)?;
|
||||
let mut seqlen_offset = 0;
|
||||
|
||||
Reference in New Issue
Block a user