[core] remove beam search from the core (#9105)

This commit is contained in:
youkaichao
2024-10-07 05:47:04 +00:00
committed by GitHub
parent c8f26bb636
commit 18b296fdb2
25 changed files with 98 additions and 596 deletions
+10 -3
View File
@@ -28,7 +28,8 @@ from vllm.transformers_utils.tokenizer import (AnyTokenizer, MistralTokenizer,
get_cached_tokenizer)
from vllm.transformers_utils.tokenizer_group import TokenizerGroup
from vllm.usage.usage_lib import UsageContext
from vllm.utils import Counter, deprecate_kwargs, is_list_of
from vllm.utils import (Counter, deprecate_kwargs, get_beam_search_score,
is_list_of)
logger = init_logger(__name__)
@@ -404,6 +405,12 @@ class LLM:
max_tokens = params.max_tokens
temperature = params.temperature
ignore_eos = params.ignore_eos
length_penalty = params.length_penalty
def sort_beams_key(x: BeamSearchSequence) -> float:
return get_beam_search_score(x.tokens, x.cum_logprob,
tokenizer.eos_token_id,
length_penalty)
tokenizer = self.get_tokenizer()
# generate 2 * beam_width candidates at each step
@@ -466,7 +473,7 @@ class LLM:
else:
instance_new_beams.append(new_beam)
sorted_beams = sorted(instance_new_beams,
key=lambda x: x.cum_logprob,
key=sort_beams_key,
reverse=True)
instance.beams = sorted_beams[:beam_width]
@@ -474,7 +481,7 @@ class LLM:
for instance in instances:
instance.completed.extend(instance.beams)
sorted_completed = sorted(instance.completed,
key=lambda x: x.cum_logprob,
key=sort_beams_key,
reverse=True)
best_beams = sorted_completed[:beam_width]
+2 -8
View File
@@ -184,7 +184,6 @@ class ChatCompletionRequest(OpenAIBaseModel):
min_p: float = 0.0
repetition_penalty: float = 1.0
length_penalty: float = 1.0
early_stopping: bool = False
stop_token_ids: Optional[List[int]] = Field(default_factory=list)
include_stop_str_in_output: bool = False
ignore_eos: bool = False
@@ -302,6 +301,7 @@ class ChatCompletionRequest(OpenAIBaseModel):
max_tokens=max_tokens,
ignore_eos=self.ignore_eos,
temperature=temperature,
length_penalty=self.length_penalty,
)
def to_sampling_params(self, default_max_tokens: int) -> SamplingParams:
@@ -345,12 +345,9 @@ class ChatCompletionRequest(OpenAIBaseModel):
ignore_eos=self.ignore_eos,
max_tokens=max_tokens,
min_tokens=self.min_tokens,
use_beam_search=self.use_beam_search,
early_stopping=self.early_stopping,
skip_special_tokens=self.skip_special_tokens,
spaces_between_special_tokens=self.spaces_between_special_tokens,
include_stop_str_in_output=self.include_stop_str_in_output,
length_penalty=self.length_penalty,
truncate_prompt_tokens=self.truncate_prompt_tokens,
output_kind=RequestOutputKind.DELTA if self.stream \
else RequestOutputKind.FINAL_ONLY,
@@ -518,7 +515,6 @@ class CompletionRequest(OpenAIBaseModel):
min_p: float = 0.0
repetition_penalty: float = 1.0
length_penalty: float = 1.0
early_stopping: bool = False
stop_token_ids: Optional[List[int]] = Field(default_factory=list)
include_stop_str_in_output: bool = False
ignore_eos: bool = False
@@ -597,6 +593,7 @@ class CompletionRequest(OpenAIBaseModel):
max_tokens=max_tokens,
ignore_eos=self.ignore_eos,
temperature=temperature,
length_penalty=self.length_penalty,
)
def to_sampling_params(self, default_max_tokens: int) -> SamplingParams:
@@ -641,13 +638,10 @@ class CompletionRequest(OpenAIBaseModel):
ignore_eos=self.ignore_eos,
max_tokens=max_tokens if not echo_without_generation else 1,
min_tokens=self.min_tokens,
use_beam_search=self.use_beam_search,
early_stopping=self.early_stopping,
prompt_logprobs=prompt_logprobs,
skip_special_tokens=self.skip_special_tokens,
spaces_between_special_tokens=self.spaces_between_special_tokens,
include_stop_str_in_output=self.include_stop_str_in_output,
length_penalty=self.length_penalty,
truncate_prompt_tokens=self.truncate_prompt_tokens,
output_kind=RequestOutputKind.DELTA if self.stream \
else RequestOutputKind.FINAL_ONLY,