ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[data] optimize qwen3 loss computation (#7923)

Browse Source
This commit is contained in:
hoshi-hiyouga
2025-04-30 16:18:00 +08:00
committed by GitHub
parent 73198a6645
commit 052ca871bd
11 changed files with 205 additions and 39 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
src/llamafactory/chat/hf_engine.py
View File

@@ -103,9 +103,11 @@ class HuggingfaceEngine(BaseEngine):
messages = template.mm_plugin.process_messages(
messages, mm_input_dict["images"], mm_input_dict["videos"], mm_input_dict["audios"], processor
)
paired_messages = messages + [{"role": "assistant", "content": ""}]
# add thought words to avoid skipping thinking
paired_messages = messages + [{"role": "assistant", "content": template.add_thought("")}]
system = system or generating_args["default_system"]
prompt_ids, _ = template.encode_oneturn(tokenizer, paired_messages, system, tools)
enable_thinking = input_kwargs.pop("enable_thinking", True)
prompt_ids, _ = template.encode_oneturn(tokenizer, paired_messages, system, tools, enable_thinking)
prompt_ids, _ = template.mm_plugin.process_token_ids(
prompt_ids,
None,

6
src/llamafactory/chat/sglang_engine.py
View File

@@ -146,9 +146,11 @@ class SGLangEngine(BaseEngine):
messages = self.template.mm_plugin.process_messages(
messages, images or [], videos or [], audios or [], self.processor
)
paired_messages = messages + [{"role": "assistant", "content": ""}]
# add thought words to avoid skipping thinking
paired_messages = messages + [{"role": "assistant", "content": self.template.add_thought("")}]
system = system or self.generating_args["default_system"]
prompt_ids, _ = self.template.encode_oneturn(self.tokenizer, paired_messages, system, tools)
enable_thinking = input_kwargs.pop("enable_thinking", True)
prompt_ids, _ = self.template.encode_oneturn(self.tokenizer, paired_messages, system, tools, enable_thinking)
prompt_length = len(prompt_ids)
temperature: Optional[float] = input_kwargs.pop("temperature", None)

6
src/llamafactory/chat/vllm_engine.py
View File

@@ -123,9 +123,11 @@ class VllmEngine(BaseEngine):
messages = self.template.mm_plugin.process_messages(
messages, images or [], videos or [], audios or [], self.processor
)
paired_messages = messages + [{"role": "assistant", "content": ""}]
# add thought words to avoid skipping thinking
paired_messages = messages + [{"role": "assistant", "content": self.template.add_thought("")}]
system = system or self.generating_args["default_system"]
prompt_ids, _ = self.template.encode_oneturn(self.tokenizer, paired_messages, system, tools)
enable_thinking = input_kwargs.pop("enable_thinking", True)
prompt_ids, _ = self.template.encode_oneturn(self.tokenizer, paired_messages, system, tools, enable_thinking)
prompt_length = len(prompt_ids)
temperature: Optional[float] = input_kwargs.pop("temperature", None)