[CI/Build] mypy: Resolve some errors from checking vllm/engine (#9267)
Signed-off-by: Russell Bryant <rbryant@redhat.com>
This commit is contained in:
@@ -1,5 +1,5 @@
|
||||
import functools
|
||||
from typing import Callable, List
|
||||
from typing import Callable, List, cast
|
||||
|
||||
from vllm.core.scheduler import Scheduler
|
||||
from vllm.engine.output_processor.interfaces import (
|
||||
@@ -9,8 +9,10 @@ from vllm.engine.output_processor.single_step import (
|
||||
from vllm.engine.output_processor.stop_checker import StopChecker
|
||||
from vllm.logger import init_logger
|
||||
from vllm.sampling_params import SamplingParams
|
||||
from vllm.sequence import (VLLM_INVALID_TOKEN_ID, Sequence, SequenceGroup,
|
||||
SequenceGroupOutput, SequenceOutput, SequenceStatus)
|
||||
from vllm.sequence import (VLLM_INVALID_TOKEN_ID,
|
||||
CompletionSequenceGroupOutput, Sequence,
|
||||
SequenceGroup, SequenceGroupOutput, SequenceOutput,
|
||||
SequenceStatus)
|
||||
from vllm.transformers_utils.detokenizer import Detokenizer
|
||||
from vllm.transformers_utils.tokenizer import AnyTokenizer
|
||||
from vllm.utils import Counter
|
||||
@@ -57,6 +59,7 @@ class MultiStepOutputProcessor(SequenceGroupOutputProcessor):
|
||||
"""
|
||||
for output in outputs:
|
||||
# Concatenate single-step prompt logprob processing results.
|
||||
assert isinstance(output, CompletionSequenceGroupOutput)
|
||||
single_step_process_prompt_logprob(self, seq_group, output)
|
||||
|
||||
@staticmethod
|
||||
@@ -100,8 +103,18 @@ class MultiStepOutputProcessor(SequenceGroupOutputProcessor):
|
||||
"Beam search not supported in multi-step decoding.")
|
||||
seq = seqs[0]
|
||||
seq_id = seq.seq_id
|
||||
assert all(
|
||||
[seq_id == output.samples[0].parent_seq_id for output in outputs])
|
||||
# This method is defined in the more generic
|
||||
# SequenceGroupOutputProcessor, but here we assume that the outputs are
|
||||
# of a more specific type.
|
||||
assert all([
|
||||
isinstance(output, CompletionSequenceGroupOutput)
|
||||
for output in outputs
|
||||
])
|
||||
compl_outputs = cast(List[CompletionSequenceGroupOutput], outputs)
|
||||
assert all([
|
||||
seq_id == output.samples[0].parent_seq_id
|
||||
for output in compl_outputs
|
||||
])
|
||||
|
||||
if is_async:
|
||||
# Async case: We process tokens one by one. Here, we know the token
|
||||
@@ -113,7 +126,7 @@ class MultiStepOutputProcessor(SequenceGroupOutputProcessor):
|
||||
|
||||
# Since there's only one sequence per sequence group,
|
||||
# we can take the first sample.
|
||||
samples = [output.samples[0] for output in outputs]
|
||||
samples = [output.samples[0] for output in compl_outputs]
|
||||
|
||||
# entries in sample tokens may be invalid (eg. due to spec decode
|
||||
# rejecting tokens).
|
||||
|
||||
@@ -6,8 +6,9 @@ from vllm.engine.output_processor.interfaces import (
|
||||
SequenceGroupOutputProcessor)
|
||||
from vllm.engine.output_processor.stop_checker import StopChecker
|
||||
from vllm.logger import init_logger
|
||||
from vllm.sequence import (Sequence, SequenceGroup, SequenceGroupOutput,
|
||||
SequenceOutput, SequenceStatus)
|
||||
from vllm.sequence import (CompletionSequenceGroupOutput, Sequence,
|
||||
SequenceGroup, SequenceGroupOutput, SequenceOutput,
|
||||
SequenceStatus)
|
||||
from vllm.transformers_utils.detokenizer import Detokenizer
|
||||
from vllm.utils import Counter
|
||||
|
||||
@@ -16,7 +17,7 @@ logger = init_logger(__name__)
|
||||
|
||||
def single_step_process_prompt_logprob(
|
||||
sg_output_proc: SequenceGroupOutputProcessor, seq_group: SequenceGroup,
|
||||
output: SequenceGroupOutput) -> None:
|
||||
output: CompletionSequenceGroupOutput) -> None:
|
||||
"""Process prompt logprobs associated with the :class:`SequenceGroupOutput`
|
||||
for a given step.
|
||||
|
||||
@@ -106,6 +107,7 @@ class SingleStepOutputProcessor(SequenceGroupOutputProcessor):
|
||||
"""
|
||||
assert len(outputs) == 1, ("Single step should only has 1 output.")
|
||||
output = outputs[0]
|
||||
assert isinstance(output, CompletionSequenceGroupOutput)
|
||||
single_step_process_prompt_logprob(self, seq_group, output)
|
||||
|
||||
def _process_sequence_group_outputs(self, seq_group: SequenceGroup,
|
||||
|
||||
@@ -57,7 +57,7 @@ class StopChecker:
|
||||
# Check if a stop token was encountered.
|
||||
# This assumes a single token produced per step.
|
||||
last_token_id = seq.get_last_token_id()
|
||||
if last_token_id in sampling_params.stop_token_ids:
|
||||
if last_token_id in (sampling_params.stop_token_ids or ()):
|
||||
if new_char_count and (
|
||||
not sampling_params.include_stop_str_in_output):
|
||||
# Remove last token
|
||||
@@ -92,7 +92,7 @@ class StopChecker:
|
||||
|
||||
Returns the stop string if matched or else None.
|
||||
"""
|
||||
if not new_char_count:
|
||||
if not new_char_count or not sampling_params.stop:
|
||||
return None
|
||||
|
||||
for stop_str in sampling_params.stop:
|
||||
|
||||
@@ -1,22 +1,25 @@
|
||||
from typing import List
|
||||
from typing import Sequence as GenericSequence
|
||||
from typing import Union
|
||||
from typing import cast
|
||||
|
||||
from vllm.model_executor.layers.sampler import SamplerOutput
|
||||
from vllm.sequence import PoolerOutput, SequenceGroupOutput
|
||||
from vllm.sequence import CompletionSequenceGroupOutput, SequenceGroupOutput
|
||||
|
||||
|
||||
def create_output_by_sequence_group(
|
||||
outputs: GenericSequence[Union[SamplerOutput, PoolerOutput]],
|
||||
outputs: GenericSequence[SamplerOutput],
|
||||
num_seq_groups: int) -> List[List[SequenceGroupOutput]]:
|
||||
"""Helper method which transforms a 2d list organized by
|
||||
[step][sequence group] into [sequence group][step].
|
||||
"""
|
||||
output_by_sequence_group: List[List[SequenceGroupOutput]] = [
|
||||
output_by_sequence_group: List[List[CompletionSequenceGroupOutput]] = [
|
||||
[] for _ in range(num_seq_groups)
|
||||
]
|
||||
for step in outputs:
|
||||
sequence_group_output: CompletionSequenceGroupOutput
|
||||
for i, sequence_group_output in enumerate(step):
|
||||
output_by_sequence_group[i].append(sequence_group_output)
|
||||
|
||||
return output_by_sequence_group
|
||||
# Cast to the more generic type that CompletionSequenceGroupOutput
|
||||
# inherits from.
|
||||
return cast(List[List[SequenceGroupOutput]], output_by_sequence_group)
|
||||
|
||||
Reference in New Issue
Block a user