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, 09:15 UTC
Support top-k sampling (#94)
Woosuk Kwon committed
May 10, 2023 at 12:51 UTC
9f88db35da012544a9cd9450c68b8df2e6509b92
3 files changed
+78
-43
cacheflow/model_executor/layers/sampler.py
+37
-16
@@ -46,12 +46,13 @@ class Sampler(nn.Module):
46
# Compute the log probabilities (before applying top-p).
47
logprobs = torch.log(probs)
48
49
- # Apply top-p truncation.
50
- top_ps = _get_top_ps(input_metadata)
51
- assert len(top_ps) == probs.shape[0]
52
- if any(p < 1.0 for p in top_ps):
49
+ # Apply top-p and top-k truncation.
50
+ top_ps, top_ks = _get_top_p_top_k(input_metadata, self.vocab_size)
51
+ assert len(top_ps) == len(top_ks) == probs.shape[0]
52
+ if any(p < 1.0 for p in top_ps) or any(k != -1 for k in top_ks):
53
p = torch.tensor(top_ps, dtype=probs.dtype, device=probs.device)
54
- probs = _apply_top_p(probs, p)
54
+ k = torch.tensor(top_ks, dtype=torch.int, device=probs.device)
55
+ probs = _apply_top_p_top_k(probs, p, k)
56
57
# Sample the next tokens.
58
return _sample(probs, logprobs, input_metadata)
@@ -94,31 +95,51 @@ def _get_temperatures(
95
return temperatures
96
97
97
-def _get_top_ps(
98
+def _get_top_p_top_k(
99
input_metadata: InputMetadata,
99
-) -> List[float]:
100
+ vocab_size: int,
101
+) -> Tuple[List[float], List[int]]:
102
top_ps: List[float] = []
103
+ top_ks: List[int] = []
104
for i, seq_group in enumerate(input_metadata.seq_groups):
105
seq_ids, sampling_params = seq_group
106
+ top_p = sampling_params.top_p
107
+ # k should not be greater than the vocab size.
108
+ top_k = min(sampling_params.top_k, vocab_size)
109
+ # k=-1 means no truncation.
110
+ top_k = vocab_size if top_k == -1 else top_k
111
if i < input_metadata.num_prompts:
112
# A prompt input.
105
- top_ps.append(sampling_params.top_p)
113
+ top_ps.append(top_p)
114
+ top_ks.append(top_k)
115
else:
116
# A generation token.
108
- top_ps += [sampling_params.top_p] * len(seq_ids)
109
- return top_ps
117
+ top_ps += [top_p] * len(seq_ids)
118
+ top_ks += [top_k] * len(seq_ids)
119
+ return top_ps, top_ks
120
121
112
-def _apply_top_p(
122
+def _apply_top_p_top_k(
123
probs: torch.Tensor,
124
p: torch.Tensor,
125
+ k: torch.Tensor,
126
) -> torch.Tensor:
127
# TODO(woosuk): Optimize.
128
probs_sort, probs_idx = probs.sort(dim=-1, descending=True)
129
+
130
+ # Apply top-p.
131
probs_sum = torch.cumsum(probs_sort, dim=-1)
119
- mask = (probs_sum - probs_sort) > p.unsqueeze(dim=1)
120
- probs_sort[mask] = 0.0
121
- probs_sort.div_(probs_sort.sum(dim=-1, keepdim=True))
132
+ top_p_mask = (probs_sum - probs_sort) > p.unsqueeze(dim=1)
133
+ probs_sort[top_p_mask] = 0.0
134
+
135
+ # Apply top-k.
136
+ # Create a mask for the top-k elements.
137
+ top_k_mask = torch.arange(probs_idx.shape[-1], device=probs_idx.device)
138
+ top_k_mask = top_k_mask.expand(probs_idx.shape[0], -1)
139
+ top_k_mask = top_k_mask >= k.unsqueeze(dim=1)
140
+ probs_sort[top_k_mask] = 0.0
141
+
142
+ # Re-sort the probabilities.
143
probs = torch.gather(
144
probs_sort, dim=-1, index=torch.argsort(probs_idx, dim=-1))
145
return probs
@@ -160,7 +181,7 @@ def _sample_from_prompt(
181
next_token_id = torch.argmax(prob)
182
next_token_ids = [next_token_id.item()]
183
else:
163
- # Neucleus sampling.
184
+ # Random sampling.
185
# Sample n tokens for the prompt.
186
n = sampling_params.n
187
next_token_ids = torch.multinomial(
@@ -218,7 +239,7 @@ def _sample_from_generation_tokens(
239
next_token_ids = [next_token_id.item()]
240
parent_seq_ids = seq_ids
241
else:
221
- # Neucleus sampling.
242
+ # Random sampling.
243
# Sample 1 token for each sequence in the group.
244
next_token_ids = torch.multinomial(
245
probs, num_samples=1, replacement=True)
cacheflow/sampling_params.py
+38
-25
@@ -8,69 +8,82 @@ class SamplingParams:
8
n: int,
9
temperature: float,
10
top_p: float,
11
+ top_k: int,
12
use_beam_search: bool,
13
stop_token_ids: Set[int],
14
max_num_steps: int,
15
num_logprobs: int,
16
) -> None:
17
if n < 1:
17
- raise ValueError(f'n must be at least 1, got {n}.')
18
+ raise ValueError(f"n must be at least 1, got {n}.")
19
if temperature < 0.0:
20
raise ValueError(
20
- f'temperature must be non-negative, got {temperature}.')
21
+ f"temperature must be non-negative, got {temperature}.")
22
if not 0.0 < top_p <= 1.0:
22
- raise ValueError(f'top_p must be in (0, 1], got {top_p}.')
23
+ raise ValueError(f"top_p must be in (0, 1], got {top_p}.")
24
+ if top_k < -1 or top_k == 0:
25
+ raise ValueError(f"top_k must be -1 (disable), or at least 1, "
26
+ f"got {top_k}.")
27
if max_num_steps < 1:
28
raise ValueError(
25
- f'max_num_steps must be at least 1, got {max_num_steps}.')
29
+ f"max_num_steps must be at least 1, got {max_num_steps}.")
30
if num_logprobs < 0:
31
raise ValueError(
28
- f'num_logprobs must be non-negative, got {num_logprobs}.')
32
+ f"num_logprobs must be non-negative, got {num_logprobs}.")
33
34
if use_beam_search:
35
if n == 1:
36
raise ValueError(
33
- 'n must be greater than 1 when using beam search.')
37
+ "n must be greater than 1 when using beam search.")
38
if temperature > 0.0:
39
raise ValueError(
36
- 'temperature must be 0 when using beam search.')
40
+ "temperature must be 0 when using beam search.")
41
if top_p < 1.0:
42
raise ValueError(
39
- 'top_p must be 1 when using beam search.')
43
+ "top_p must be 1 when using beam search.")
44
+ if top_k != -1:
45
+ raise ValueError(
46
+ "top_k must be -1 when using beam search.")
47
elif temperature == 0.0:
48
# Zero temperature means greedy sampling.
49
if n > 1:
50
raise ValueError(
44
- 'n must be 1 when using greedy sampling.')
51
+ "n must be 1 when using greedy sampling.")
52
if top_p < 1.0:
53
raise ValueError(
47
- 'top_p must be 1 when using greedy sampling.')
54
+ "top_p must be 1 when using greedy sampling.")
55
+ if top_k != -1:
56
+ raise ValueError(
57
+ "top_k must be -1 when using greedy sampling.")
58
59
self.n = n
60
self.temperature = temperature
61
self.top_p = top_p
62
+ self.top_k = top_k
63
self.use_beam_search = use_beam_search
64
self.stop_token_ids = stop_token_ids
65
self.max_num_steps = max_num_steps
66
self.num_logprobs = num_logprobs
67
68
def __repr__(self) -> str:
58
- return (f'SamplingParams(n={self.n}, '
59
- f'temperature={self.temperature}, '
60
- f'top_p={self.top_p}, '
61
- f'use_beam_search={self.use_beam_search}, '
62
- f'stop_token_ids={self.stop_token_ids}, '
63
- f'max_num_steps={self.max_num_steps}, '
64
- f'num_logprobs={self.num_logprobs}')
69
+ return (f"SamplingParams(n={self.n}, "
70
+ f"temperature={self.temperature}, "
71
+ f"top_p={self.top_p}, "
72
+ f"top_k={self.top_k},"
73
+ f"use_beam_search={self.use_beam_search}, "
74
+ f"stop_token_ids={self.stop_token_ids}, "
75
+ f"max_num_steps={self.max_num_steps}, "
76
+ f"num_logprobs={self.num_logprobs}")
77
78
@classmethod
67
- def from_dict(cls, d: Dict) -> 'SamplingParams':
79
+ def from_dict(cls, d: Dict) -> "SamplingParams":
80
return cls(
69
- n=d.get('n', 1),
70
- temperature=d.get('temperature', 1.0),
71
- top_p=d.get('top_p', 1.0),
72
- use_beam_search=d.get('use_beam_search', False),
73
- stop_token_ids=set(d.get('stop_token_ids', set())),
74
- max_num_steps=d.get('max_num_steps', 16),
75
- num_logprobs=d.get('num_logprobs', 0),
81
+ n=d.get("n", 1),
82
+ temperature=d.get("temperature", 1.0),
83
+ top_p=d.get("top_p", 1.0),
84
+ top_k=d.get("top_k", -1),
85
+ use_beam_search=d.get("use_beam_search", False),
86
+ stop_token_ids=set(d.get("stop_token_ids", set())),
87
+ max_num_steps=d.get("max_num_steps", 16),
88
+ num_logprobs=d.get("num_logprobs", 0),
89
)
simple_server.py
+3
-2
@@ -11,8 +11,9 @@ def main(args: argparse.Namespace):
11
# Test the following inputs.
12
test_inputs = [
13
("A robot may not injure a human being", {}), # Use default parameters.
14
- ("What is the meaning of life?", {"n": 3, "temperature": 0.8, "top_p": 0.99}),
15
- ("It is only with the heart that one can see rightly", {"n": 4, "use_beam_search": True, "temperature": 0.0}),
14
+ ("To be or not to be,", {"temperature": 0.8, "top_k": 5}),
15
+ ("What is the meaning of life?", {"n": 2, "temperature": 0.8, "top_p": 0.95}),
16
+ ("It is only with the heart that one can see rightly", {"n": 3, "use_beam_search": True, "temperature": 0.0}),
17
]
18
while True:
19
if test_inputs: