Delay GPU->CPU sync in sampling (#1337)

This commit is contained in:
Antoni Baum
2023-10-30 09:01:34 -07:00
committed by GitHub
parent aa9af07cac
commit 15f5632365
3 changed files with 66 additions and 47 deletions
+18 -11
View File
@@ -3,7 +3,7 @@ from typing import Dict, List, Optional, Tuple
import torch
from xformers.ops import AttentionBias
from vllm.sampling_params import SamplingParams
from vllm.sampling_params import SamplingParams, SamplingType
from vllm.sequence import SequenceData
@@ -29,6 +29,8 @@ class InputMetadata:
context_lens: torch.Tensor,
max_context_len: int,
block_tables: torch.Tensor,
selected_token_indices: torch.Tensor,
categorized_sample_indices: Dict[SamplingType, torch.Tensor],
sliding_window: Optional[int] = None,
) -> None:
self.seq_groups = seq_groups
@@ -38,6 +40,8 @@ class InputMetadata:
self.context_lens = context_lens
self.max_context_len = max_context_len
self.block_tables = block_tables
self.selected_token_indices = selected_token_indices
self.categorized_sample_indices = categorized_sample_indices
self.max_prompt_len = max(prompt_lens) if prompt_lens else 0
self.to_cache = None
@@ -72,13 +76,16 @@ class InputMetadata:
def __repr__(self) -> str:
# Print only useful metadata.
return (f'InputMetadata('
f'num_prompt_tokens={self.num_prompt_tokens}, '
f'num_prompts={self.num_prompts}, '
f'prompt_lens={self.prompt_lens}, '
f'num_generation_tokens={self.num_generation_tokens}, '
f'context_lens={self.context_lens}, '
f'max_context_len={self.max_context_len}, '
f'max_num_blocks_per_seq={self.max_num_blocks_per_seq}, '
f'block_tables={self.block_tables}, '
f'slot_mapping={self.slot_mapping})')
return (
f'InputMetadata('
f'num_prompt_tokens={self.num_prompt_tokens}, '
f'num_prompts={self.num_prompts}, '
f'prompt_lens={self.prompt_lens}, '
f'num_generation_tokens={self.num_generation_tokens}, '
f'context_lens={self.context_lens}, '
f'max_context_len={self.max_context_len}), '
f'max_num_blocks_per_seq={self.max_num_blocks_per_seq}, '
f'block_tables={self.block_tables}, '
f'selected_token_indices={self.selected_token_indices}, '
f'categorized_sample_indices={self.categorized_sample_indices}, '
f'slot_mapping={self.slot_mapping})')
+3 -34
View File
@@ -109,29 +109,8 @@ def _prune_hidden_states(
hidden_states: torch.Tensor,
input_metadata: InputMetadata,
) -> torch.Tensor:
selected_token_indices: List[int] = []
start_idx = 0
for i, seq_group in enumerate(input_metadata.seq_groups):
seq_ids, sampling_params = seq_group
if i < input_metadata.num_prompts:
assert len(seq_ids) == 1, "Prompt input should have only one seq."
prompt_len = input_metadata.prompt_lens[i]
if sampling_params.prompt_logprobs is not None:
selected_token_indices.extend(
range(start_idx, start_idx + prompt_len - 1))
selected_token_indices.append(start_idx + prompt_len - 1)
start_idx += input_metadata.max_prompt_len
else:
num_seqs = len(seq_ids)
selected_token_indices.extend(
range(start_idx, start_idx + num_seqs))
start_idx += num_seqs
selected_token_indices = torch.tensor(selected_token_indices,
dtype=torch.long,
device=hidden_states.device)
hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
return hidden_states.index_select(0, selected_token_indices)
return hidden_states.index_select(0, input_metadata.selected_token_indices)
def _get_penalties(
@@ -426,21 +405,11 @@ def _sample(
input_metadata: InputMetadata,
) -> List[Tuple[List[int], List[int]]]:
categorized_seq_group_ids = {t: [] for t in SamplingType}
categorized_sample_indices = {t: [] for t in SamplingType}
start_idx = 0
categorized_sample_indices = input_metadata.categorized_sample_indices
for i, seq_group in enumerate(input_metadata.seq_groups):
seq_ids, sampling_params = seq_group
_, sampling_params = seq_group
sampling_type = sampling_params.sampling_type
if (i < input_metadata.num_prompts
and sampling_params.prompt_logprobs is not None):
# NOTE: prompt token positions do not need sample, skip
prompt_len = input_metadata.prompt_lens[i]
start_idx += prompt_len - 1
categorized_seq_group_ids[sampling_type].append(i)
num_seqs = len(seq_ids)
categorized_sample_indices[sampling_type].extend(
range(start_idx, start_idx + num_seqs))
start_idx += num_seqs
sample_results_dict: Dict[int, Tuple[List[int], List[int]]] = {}
for sampling_type in SamplingType: