A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 17, 06:45 UTC
Delay GPU->CPU sync in sampling (#1337)
Antoni Baum committed
Oct 30, 2023 at 09:01 UTC
15f5632365a98fd43ea42e4948a995aa399e99b5
3 files changed
+66
-47
vllm/model_executor/input_metadata.py
+18
-11
@@ -3,7 +3,7 @@ from typing import Dict, List, Optional, Tuple
3
import torch
4
from xformers.ops import AttentionBias
5
6
-from vllm.sampling_params import SamplingParams
6
+from vllm.sampling_params import SamplingParams, SamplingType
7
from vllm.sequence import SequenceData
8
9
@@ -29,6 +29,8 @@ class InputMetadata:
29
context_lens: torch.Tensor,
30
max_context_len: int,
31
block_tables: torch.Tensor,
32
+ selected_token_indices: torch.Tensor,
33
+ categorized_sample_indices: Dict[SamplingType, torch.Tensor],
34
sliding_window: Optional[int] = None,
35
) -> None:
36
self.seq_groups = seq_groups
@@ -38,6 +40,8 @@ class InputMetadata:
40
self.context_lens = context_lens
41
self.max_context_len = max_context_len
42
self.block_tables = block_tables
43
+ self.selected_token_indices = selected_token_indices
44
+ self.categorized_sample_indices = categorized_sample_indices
45
46
self.max_prompt_len = max(prompt_lens) if prompt_lens else 0
47
self.to_cache = None
@@ -72,13 +76,16 @@ class InputMetadata:
76
77
def __repr__(self) -> str:
78
# Print only useful metadata.
75
- return (f'InputMetadata('
76
- f'num_prompt_tokens={self.num_prompt_tokens}, '
77
- f'num_prompts={self.num_prompts}, '
78
- f'prompt_lens={self.prompt_lens}, '
79
- f'num_generation_tokens={self.num_generation_tokens}, '
80
- f'context_lens={self.context_lens}, '
81
- f'max_context_len={self.max_context_len}, '
82
- f'max_num_blocks_per_seq={self.max_num_blocks_per_seq}, '
83
- f'block_tables={self.block_tables}, '
84
- f'slot_mapping={self.slot_mapping})')
79
+ return (
80
+ f'InputMetadata('
81
+ f'num_prompt_tokens={self.num_prompt_tokens}, '
82
+ f'num_prompts={self.num_prompts}, '
83
+ f'prompt_lens={self.prompt_lens}, '
84
+ f'num_generation_tokens={self.num_generation_tokens}, '
85
+ f'context_lens={self.context_lens}, '
86
+ f'max_context_len={self.max_context_len}), '
87
+ f'max_num_blocks_per_seq={self.max_num_blocks_per_seq}, '
88
+ f'block_tables={self.block_tables}, '
89
+ f'selected_token_indices={self.selected_token_indices}, '
90
+ f'categorized_sample_indices={self.categorized_sample_indices}, '
91
+ f'slot_mapping={self.slot_mapping})')
vllm/model_executor/layers/sampler.py
+3
-34
@@ -109,29 +109,8 @@ def _prune_hidden_states(
109
hidden_states: torch.Tensor,
110
input_metadata: InputMetadata,
111
) -> torch.Tensor:
112
- selected_token_indices: List[int] = []
113
- start_idx = 0
114
- for i, seq_group in enumerate(input_metadata.seq_groups):
115
- seq_ids, sampling_params = seq_group
116
- if i < input_metadata.num_prompts:
117
- assert len(seq_ids) == 1, "Prompt input should have only one seq."
118
- prompt_len = input_metadata.prompt_lens[i]
119
- if sampling_params.prompt_logprobs is not None:
120
- selected_token_indices.extend(
121
- range(start_idx, start_idx + prompt_len - 1))
122
- selected_token_indices.append(start_idx + prompt_len - 1)
123
- start_idx += input_metadata.max_prompt_len
124
- else:
125
- num_seqs = len(seq_ids)
126
- selected_token_indices.extend(
127
- range(start_idx, start_idx + num_seqs))
128
- start_idx += num_seqs
129
-
130
- selected_token_indices = torch.tensor(selected_token_indices,
131
- dtype=torch.long,
132
- device=hidden_states.device)
112
hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
134
- return hidden_states.index_select(0, selected_token_indices)
113
+ return hidden_states.index_select(0, input_metadata.selected_token_indices)
114
115
116
def _get_penalties(
@@ -426,21 +405,11 @@ def _sample(
405
input_metadata: InputMetadata,
406
) -> List[Tuple[List[int], List[int]]]:
407
categorized_seq_group_ids = {t: [] for t in SamplingType}
429
- categorized_sample_indices = {t: [] for t in SamplingType}
430
- start_idx = 0
408
+ categorized_sample_indices = input_metadata.categorized_sample_indices
409
for i, seq_group in enumerate(input_metadata.seq_groups):
432
- seq_ids, sampling_params = seq_group
410
+ _, sampling_params = seq_group
411
sampling_type = sampling_params.sampling_type
434
- if (i < input_metadata.num_prompts
435
- and sampling_params.prompt_logprobs is not None):
436
- # NOTE: prompt token positions do not need sample, skip
437
- prompt_len = input_metadata.prompt_lens[i]
438
- start_idx += prompt_len - 1
412
categorized_seq_group_ids[sampling_type].append(i)
440
- num_seqs = len(seq_ids)
441
- categorized_sample_indices[sampling_type].extend(
442
- range(start_idx, start_idx + num_seqs))
443
- start_idx += num_seqs
413
414
sample_results_dict: Dict[int, Tuple[List[int], List[int]]] = {}
415
for sampling_type in SamplingType:
vllm/worker/worker.py
+45
-2
@@ -10,7 +10,7 @@ from vllm.config import (CacheConfig, ModelConfig, ParallelConfig,
10
from vllm.model_executor import get_model, InputMetadata, set_random_seed
11
from vllm.model_executor.parallel_utils.parallel_state import (
12
initialize_model_parallel)
13
-from vllm.sampling_params import SamplingParams
13
+from vllm.sampling_params import SamplingParams, SamplingType
14
from vllm.sequence import SamplerOutput, SequenceData, SequenceGroupMetadata
15
from vllm.worker.cache_engine import CacheEngine
16
from vllm.utils import get_gpu_memory, get_max_shared_memory_bytes
@@ -161,6 +161,10 @@ class Worker:
161
input_tokens: List[List[int]] = []
162
input_positions: List[List[int]] = []
163
slot_mapping: List[List[int]] = []
164
+ selected_token_indices: List[int] = []
165
+ selected_token_start_idx = 0
166
+ categorized_sample_indices = {t: [] for t in SamplingType}
167
+ categorized_sample_indices_start_idx = 0
168
169
# Add prompt tokens.
170
prompt_lens: List[int] = []
@@ -180,6 +184,14 @@ class Worker:
184
prompt_len = len(prompt_tokens)
185
prompt_lens.append(prompt_len)
186
187
+ if sampling_params.prompt_logprobs is not None:
188
+ # NOTE: prompt token positions do not need sample, skip
189
+ categorized_sample_indices_start_idx += prompt_len - 1
190
+
191
+ categorized_sample_indices[sampling_params.sampling_type].append(
192
+ categorized_sample_indices_start_idx)
193
+ categorized_sample_indices_start_idx += 1
194
+
195
input_tokens.append(prompt_tokens)
196
# NOTE(woosuk): Here we assume that the first token in the prompt
197
# is always the first token in the sequence.
@@ -205,14 +217,37 @@ class Worker:
217
max_num_blocks_per_seq = 0
218
context_lens: List[int] = []
219
generation_block_tables: List[List[int]] = []
220
+ max_seq_len = max(prompt_lens) if prompt_lens else 1
221
for seq_group_metadata in seq_group_metadata_list:
222
if seq_group_metadata.is_prompt:
223
+ # We need to do this in this loop as we need to know max_seq_len
224
+ assert len(
225
+ seq_ids) == 1, "Prompt input should have only one seq."
226
+ sampling_params = seq_group_metadata.sampling_params
227
+ if sampling_params.prompt_logprobs is not None:
228
+ selected_token_indices.extend(
229
+ range(selected_token_start_idx,
230
+ selected_token_start_idx + prompt_len - 1))
231
+ selected_token_indices.append(selected_token_start_idx +
232
+ prompt_len - 1)
233
+ selected_token_start_idx += max_seq_len
234
continue
235
236
seq_ids = list(seq_group_metadata.seq_data.keys())
237
sampling_params = seq_group_metadata.sampling_params
238
seq_groups.append((seq_ids, sampling_params))
239
240
+ num_seqs = len(seq_ids)
241
+ selected_token_indices.extend(
242
+ range(selected_token_start_idx,
243
+ selected_token_start_idx + num_seqs))
244
+ selected_token_start_idx += num_seqs
245
+
246
+ categorized_sample_indices[sampling_params.sampling_type].extend(
247
+ range(categorized_sample_indices_start_idx,
248
+ categorized_sample_indices_start_idx + num_seqs))
249
+ categorized_sample_indices_start_idx += num_seqs
250
+
251
for seq_id in seq_ids:
252
seq_data = seq_group_metadata.seq_data[seq_id]
253
generation_token = seq_data.get_last_token_id()
@@ -242,7 +277,6 @@ class Worker:
277
block_table = block_table[-sliding_window_blocks:]
278
generation_block_tables.append(block_table)
279
245
- max_seq_len = max(prompt_lens) if prompt_lens else 1
280
padded_input_tokens = [
281
_pad_to_max(tokens, max_seq_len, pad=0) for tokens in input_tokens
282
]
@@ -272,6 +306,13 @@ class Worker:
306
context_lens_tensor = torch.tensor(context_lens,
307
dtype=torch.int,
308
device="cuda")
309
+ selected_token_indices = torch.tensor(selected_token_indices,
310
+ dtype=torch.long,
311
+ device="cuda")
312
+ categorized_sample_indices = {
313
+ t: torch.tensor(seq_ids, dtype=torch.int, device="cuda")
314
+ for t, seq_ids in categorized_sample_indices.items()
315
+ }
316
block_tables_tensor = torch.tensor(padded_block_tables,
317
dtype=torch.int,
318
device="cuda")
@@ -288,6 +329,8 @@ class Worker:
329
context_lens=context_lens_tensor,
330
max_context_len=max_context_len,
331
block_tables=block_tables_tensor,
332
+ selected_token_indices=selected_token_indices,
333
+ categorized_sample_indices=categorized_sample_indices,
334
sliding_window=self.sliding_window,
335
)
336
return tokens_tensor, positions_tensor, input_metadata