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 16, 09:30 UTC
[Hardware] using current_platform.seed_everything (#9785)
Signed-off-by: wangshuai09 <391746016@qq.com>
wangshuai09 committed
Oct 29, 2024 at 22:47 UTC
622b7ab955186f37879208d7a30e9faf985be220
27 files changed
+104
-105
benchmarks/kernels/benchmark_layernorm.py
+3
-3
@@ -3,8 +3,8 @@ import time
3
import torch
4
5
from vllm.model_executor.layers.layernorm import RMSNorm
6
-from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
7
- seed_everything)
6
+from vllm.platforms import current_platform
7
+from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser
8
9
10
@torch.inference_mode()
@@ -16,7 +16,7 @@ def main(num_tokens: int,
16
do_profile: bool = False,
17
num_warmup_iters: int = 5,
18
num_iters: int = 100) -> None:
19
- seed_everything(seed)
19
+ current_platform.seed_everything(seed)
20
torch.set_default_device("cuda")
21
22
layer = RMSNorm(hidden_size).to(dtype=dtype)
benchmarks/kernels/benchmark_moe.py
+4
-3
@@ -10,7 +10,8 @@ from ray.experimental.tqdm_ray import tqdm
10
from transformers import AutoConfig
11
12
from vllm.model_executor.layers.fused_moe.fused_moe import *
13
-from vllm.utils import FlexibleArgumentParser, seed_everything
13
+from vllm.platforms import current_platform
14
+from vllm.utils import FlexibleArgumentParser
15
16
17
class BenchmarkConfig(TypedDict):
@@ -167,7 +168,7 @@ class BenchmarkWorker:
168
169
def __init__(self, seed: int) -> None:
170
torch.set_default_device("cuda")
170
- seed_everything(seed)
171
+ current_platform.seed_everything(seed)
172
self.seed = seed
173
174
def benchmark(
@@ -181,7 +182,7 @@ class BenchmarkWorker:
182
use_fp8_w8a8: bool,
183
use_int8_w8a16: bool,
184
) -> Tuple[Dict[str, int], float]:
184
- seed_everything(self.seed)
185
+ current_platform.seed_everything(self.seed)
186
dtype_str = get_config_dtype_str(dtype,
187
use_int8_w8a16=use_int8_w8a16,
188
use_fp8_w8a8=use_fp8_w8a8)
benchmarks/kernels/benchmark_paged_attention.py
+3
-2
@@ -5,8 +5,9 @@ from typing import List, Optional
5
import torch
6
7
from vllm import _custom_ops as ops
8
+from vllm.platforms import current_platform
9
from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
9
- create_kv_caches_with_random, seed_everything)
10
+ create_kv_caches_with_random)
11
12
NUM_BLOCKS = 1024
13
PARTITION_SIZE = 512
@@ -28,7 +29,7 @@ def main(
29
device: str = "cuda",
30
kv_cache_dtype: Optional[str] = None,
31
) -> None:
31
- seed_everything(seed)
32
+ current_platform.seed_everything(seed)
33
34
scale = float(1.0 / (head_size**0.5))
35
query = torch.empty(num_seqs,
benchmarks/kernels/benchmark_quant.py
+3
-3
@@ -3,8 +3,8 @@ import time
3
import torch
4
5
from vllm import _custom_ops as ops
6
-from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
7
- seed_everything)
6
+from vllm.platforms import current_platform
7
+from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser
8
9
10
@torch.inference_mode()
@@ -17,7 +17,7 @@ def main(num_tokens: int,
17
do_profile: bool = False,
18
num_warmup_iters: int = 5,
19
num_iters: int = 100) -> None:
20
- seed_everything(seed)
20
+ current_platform.seed_everything(seed)
21
torch.set_default_device("cuda")
22
23
x = torch.randn(num_tokens, hidden_size, dtype=dtype)
benchmarks/kernels/benchmark_rope.py
+3
-2
@@ -6,7 +6,8 @@ import torch
6
7
from vllm.model_executor.layers.rotary_embedding import (RotaryEmbedding,
8
get_rope)
9
-from vllm.utils import FlexibleArgumentParser, seed_everything
9
+from vllm.platforms import current_platform
10
+from vllm.utils import FlexibleArgumentParser
11
12
13
def benchmark_rope_kernels_multi_lora(
@@ -22,7 +23,7 @@ def benchmark_rope_kernels_multi_lora(
23
max_position: int = 8192,
24
base: int = 10000,
25
) -> None:
25
- seed_everything(seed)
26
+ current_platform.seed_everything(seed)
27
torch.set_default_device(device)
28
if rotary_dim is None:
29
rotary_dim = head_size
tests/kernels/test_activation.py
+3
-3
@@ -8,7 +8,7 @@ from tests.kernels.utils import opcheck
8
from vllm.model_executor.layers.activation import (FastGELU, FatreluAndMul,
9
GeluAndMul, NewGELU,
10
QuickGELU, SiluAndMul)
11
-from vllm.utils import seed_everything
11
+from vllm.platforms import current_platform
12
13
from .allclose_default import get_default_atol, get_default_rtol
14
@@ -37,7 +37,7 @@ def test_act_and_mul(
37
seed: int,
38
device: str,
39
) -> None:
40
- seed_everything(seed)
40
+ current_platform.seed_everything(seed)
41
torch.set_default_device(device)
42
x = torch.randn(num_tokens, 2 * d, dtype=dtype)
43
if activation == "silu":
@@ -85,7 +85,7 @@ def test_activation(
85
seed: int,
86
device: str,
87
) -> None:
88
- seed_everything(seed)
88
+ current_platform.seed_everything(seed)
89
torch.set_default_device(device)
90
x = torch.randn(num_tokens, d, dtype=dtype)
91
layer = activation[0]()
tests/kernels/test_attention.py
+3
-3
@@ -7,7 +7,7 @@ import torch
7
from tests.kernels.utils import opcheck
8
from vllm import _custom_ops as ops
9
from vllm.platforms import current_platform
10
-from vllm.utils import get_max_shared_memory_bytes, seed_everything
10
+from vllm.utils import get_max_shared_memory_bytes
11
12
from .allclose_default import get_default_atol, get_default_rtol
13
@@ -144,7 +144,7 @@ def test_paged_attention(
144
or (version == "rocm" and head_size not in (64, 128))):
145
pytest.skip()
146
147
- seed_everything(seed)
147
+ current_platform.seed_everything(seed)
148
torch.set_default_device(device)
149
scale = float(1.0 / (head_size**0.5))
150
num_query_heads, num_kv_heads = num_heads
@@ -382,7 +382,7 @@ def test_multi_query_kv_attention(
382
seed: int,
383
device: str,
384
) -> None:
385
- seed_everything(seed)
385
+ current_platform.seed_everything(seed)
386
torch.set_default_device(device)
387
# MAX_SEQ_LEN sometimes causes OOM in the reference implementation.
388
# As the xformers library is already tested with its own tests, we can use
tests/kernels/test_awq_triton.py
+3
-3
@@ -7,7 +7,7 @@ import torch
7
8
from vllm.model_executor.layers.quantization.awq_triton import (
9
AWQ_TRITON_SUPPORTED_GROUP_SIZES, awq_dequantize_triton, awq_gemm_triton)
10
-from vllm.utils import seed_everything
10
+from vllm.platforms import current_platform
11
12
device = "cuda"
13
@@ -80,7 +80,7 @@ def test_dequantize(qweight_rows, qweight_cols, group_size):
80
zeros_cols = qweight_cols
81
zeros_dtype = torch.int32
82
83
- seed_everything(0)
83
+ current_platform.seed_everything(0)
84
85
qweight = torch.randint(0,
86
torch.iinfo(torch.int32).max,
@@ -134,7 +134,7 @@ def test_gemm(N, K, M, splitK, group_size):
134
qzeros_rows = scales_rows
135
qzeros_cols = qweight_cols
136
137
- seed_everything(0)
137
+ current_platform.seed_everything(0)
138
139
input = torch.rand((input_rows, input_cols),
140
dtype=input_dtype,
tests/kernels/test_blocksparse_attention.py
+3
-3
@@ -8,7 +8,7 @@ from vllm import _custom_ops as ops
8
from vllm.attention.ops.blocksparse_attention.interface import (
9
LocalStridedBlockSparseAttn)
10
from vllm.platforms import current_platform
11
-from vllm.utils import get_max_shared_memory_bytes, seed_everything
11
+from vllm.utils import get_max_shared_memory_bytes
12
13
from .allclose_default import get_default_atol, get_default_rtol
14
@@ -173,7 +173,7 @@ def test_paged_attention(
173
blocksparse_block_size: int,
174
blocksparse_head_sliding_step: int,
175
) -> None:
176
- seed_everything(seed)
176
+ current_platform.seed_everything(seed)
177
torch.set_default_device(device)
178
scale = float(1.0 / (head_size**0.5))
179
num_query_heads, num_kv_heads = num_heads
@@ -384,7 +384,7 @@ def test_varlen_blocksparse_attention_prefill(
384
seed: int,
385
device: str,
386
) -> None:
387
- seed_everything(seed)
387
+ current_platform.seed_everything(seed)
388
torch.set_default_device(device)
389
# MAX_SEQ_LEN sometimes causes OOM in the reference implementation.
390
# As the xformers library is already tested with its own tests, we can use
tests/kernels/test_cache.py
+6
-6
@@ -6,7 +6,7 @@ import torch
6
7
from tests.kernels.utils import DEFAULT_OPCHECK_TEST_UTILS, opcheck
8
from vllm import _custom_ops as ops
9
-from vllm.utils import seed_everything
9
+from vllm.platforms import current_platform
10
11
COPYING_DIRECTION = [('cuda', 'cpu'), ('cuda', 'cuda'), ('cpu', 'cuda')]
12
DTYPES = [torch.half, torch.bfloat16, torch.float]
@@ -56,7 +56,7 @@ def test_copy_blocks(
56
) -> None:
57
if kv_cache_dtype == "fp8" and head_size % 16:
58
pytest.skip()
59
- seed_everything(seed)
59
+ current_platform.seed_everything(seed)
60
torch.set_default_device(device)
61
# Generate random block mappings where each source block is mapped to two
62
# destination blocks.
@@ -132,7 +132,7 @@ def test_reshape_and_cache(
132
) -> None:
133
if kv_cache_dtype == "fp8" and head_size % 16:
134
pytest.skip()
135
- seed_everything(seed)
135
+ current_platform.seed_everything(seed)
136
torch.set_default_device(device)
137
# Create a random slot mapping.
138
num_slots = block_size * num_blocks
@@ -224,7 +224,7 @@ def test_reshape_and_cache_flash(
224
device: str,
225
kv_cache_dtype: str,
226
) -> None:
227
- seed_everything(seed)
227
+ current_platform.seed_everything(seed)
228
torch.set_default_device(device)
229
230
# Create a random slot mapping.
@@ -339,7 +339,7 @@ def test_swap_blocks(
339
if kv_cache_dtype == "fp8" and head_size % 16:
340
pytest.skip()
341
342
- seed_everything(seed)
342
+ current_platform.seed_everything(seed)
343
344
src_device = device if direction[0] == "cuda" else 'cpu'
345
dst_device = device if direction[1] == "cuda" else 'cpu'
@@ -408,7 +408,7 @@ def test_fp8_e4m3_conversion(
408
seed: int,
409
device: str,
410
) -> None:
411
- seed_everything(seed)
411
+ current_platform.seed_everything(seed)
412
413
low = -224.0
414
high = 224.0
tests/kernels/test_causal_conv1d.py
+6
-6
@@ -9,7 +9,7 @@ from vllm import _custom_ops as ops # noqa: F401
9
from vllm.attention.backends.utils import PAD_SLOT_ID
10
from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
11
causal_conv1d_fn, causal_conv1d_update)
12
-from vllm.utils import seed_everything
12
+from vllm.platforms import current_platform
13
14
15
def causal_conv1d_ref(
@@ -70,7 +70,7 @@ def causal_conv1d_update_ref(x,
70
bias: (dim,)
71
cache_seqlens: (batch,), dtype int32.
72
If not None, the conv_state is treated as a circular buffer.
73
- The conv_state will be updated by copying x to the
73
+ The conv_state will be updated by copying x to the
74
conv_state starting at the index
75
@cache_seqlens % state_len before performing the convolution.
76
@@ -161,7 +161,7 @@ def test_causal_conv1d(batch, dim, seqlen, width, has_bias, silu_activation,
161
if itype == torch.bfloat16:
162
rtol, atol = 1e-2, 5e-2
163
# set seed
164
- seed_everything(0)
164
+ current_platform.seed_everything(0)
165
x = torch.randn(batch, dim, seqlen, device=device,
166
dtype=itype).contiguous()
167
@@ -223,7 +223,7 @@ def test_causal_conv1d_update(dim, width, seqlen, has_bias, silu_activation,
223
if itype == torch.bfloat16:
224
rtol, atol = 1e-2, 5e-2
225
# set seed
226
- seed_everything(0)
226
+ current_platform.seed_everything(0)
227
batch = 2
228
x = torch.randn(batch, dim, seqlen, device=device, dtype=itype)
229
x_ref = x.clone()
@@ -270,7 +270,7 @@ def test_causal_conv1d_update_with_batch_gather(with_padding, dim, width,
270
rtol, atol = 1e-2, 5e-2
271
272
# set seed
273
- seed_everything(0)
273
+ current_platform.seed_everything(0)
274
275
batch_size = 3
276
padding = 5 if with_padding else 0
@@ -343,7 +343,7 @@ def test_causal_conv1d_varlen(with_padding, dim, seqlen, width, has_bias,
343
if itype == torch.bfloat16:
344
rtol, atol = 1e-2, 5e-2
345
# set seed
346
- seed_everything(0)
346
+ current_platform.seed_everything(0)
347
seqlens = []
348
batch_size = 4
349
if seqlen < 10:
tests/kernels/test_flash_attn.py
+3
-3
@@ -3,7 +3,7 @@ from typing import List, Optional, Tuple
3
import pytest
4
import torch
5
6
-from vllm.utils import seed_everything
6
+from vllm.platforms import current_platform
7
from vllm.vllm_flash_attn import (flash_attn_varlen_func,
8
flash_attn_with_kvcache)
9
@@ -91,7 +91,7 @@ def test_flash_attn_with_paged_kv(
91
sliding_window: Optional[int],
92
) -> None:
93
torch.set_default_device("cuda")
94
- seed_everything(0)
94
+ current_platform.seed_everything(0)
95
num_seqs = len(kv_lens)
96
num_query_heads = num_heads[0]
97
num_kv_heads = num_heads[1]
@@ -161,7 +161,7 @@ def test_varlen_with_paged_kv(
161
num_blocks: int,
162
) -> None:
163
torch.set_default_device("cuda")
164
- seed_everything(0)
164
+ current_platform.seed_everything(0)
165
num_seqs = len(seq_lens)
166
query_lens = [x[0] for x in seq_lens]
167
kv_lens = [x[1] for x in seq_lens]
tests/kernels/test_flashinfer.py
+5
-5
@@ -4,7 +4,7 @@ import flashinfer
4
import pytest
5
import torch
6
7
-from vllm.utils import seed_everything
7
+from vllm.platforms import current_platform
8
9
NUM_HEADS = [(16, 16), (32, 8), (64, 8), (6, 1)]
10
HEAD_SIZES = [128, 256]
@@ -84,7 +84,7 @@ def test_flashinfer_decode_with_paged_kv(
84
soft_cap: Optional[float],
85
) -> None:
86
torch.set_default_device("cuda")
87
- seed_everything(0)
87
+ current_platform.seed_everything(0)
88
num_seqs = len(kv_lens)
89
num_query_heads = num_heads[0]
90
num_kv_heads = num_heads[1]
@@ -170,7 +170,7 @@ def test_flashinfer_prefill_with_paged_kv(seq_lens: List[Tuple[int, int]],
170
block_size: int,
171
soft_cap: Optional[float]) -> None:
172
torch.set_default_device("cuda")
173
- seed_everything(0)
173
+ current_platform.seed_everything(0)
174
num_seqs = len(seq_lens)
175
query_lens = [x[0] for x in seq_lens]
176
kv_lens = [x[1] for x in seq_lens]
@@ -268,7 +268,7 @@ def test_flashinfer_prefill_with_paged_fp8_kv(
268
head_size: int, dtype: torch.dtype, block_size: int,
269
soft_cap: Optional[float]) -> None:
270
torch.set_default_device("cuda")
271
- seed_everything(0)
271
+ current_platform.seed_everything(0)
272
num_seqs = len(seq_lens)
273
query_lens = [x[0] for x in seq_lens]
274
kv_lens = [x[1] for x in seq_lens]
@@ -381,7 +381,7 @@ def test_flashinfer_decode_with_paged_fp8_kv(
381
) -> None:
382
# test doesn't work for num_heads = (16,16)
383
torch.set_default_device("cuda")
384
- seed_everything(0)
384
+ current_platform.seed_everything(0)
385
num_seqs = len(kv_lens)
386
num_query_heads = num_heads[0]
387
num_kv_heads = num_heads[1]
tests/kernels/test_fp8_quant.py
+4
-4
@@ -6,7 +6,7 @@ from tests.kernels.quant_utils import (FP8_DTYPE,
6
ref_dynamic_per_tensor_fp8_quant,
7
ref_dynamic_per_token_quant)
8
from tests.kernels.utils import opcheck
9
-from vllm.utils import seed_everything
9
+from vllm.platforms import current_platform
10
11
DTYPES = [torch.half, torch.bfloat16, torch.float]
12
HIDDEN_SIZES = [1, 2, 3, 4, 16, 67, 768, 2048, 5120, 5137, 8192,
@@ -46,7 +46,7 @@ def opcheck_fp8_quant(output,
46
def test_dynamic_per_token_fp8_quant(num_tokens: int, hidden_size: int,
47
dtype: torch.dtype, scale_ub: bool,
48
seed: int) -> None:
49
- seed_everything(seed)
49
+ current_platform.seed_everything(seed)
50
51
x = torch.rand(num_tokens, hidden_size, dtype=dtype,
52
device="cuda") + 1e-6 # avoid nans
@@ -76,7 +76,7 @@ def test_dynamic_per_token_fp8_quant(num_tokens: int, hidden_size: int,
76
@torch.inference_mode()
77
def test_dynamic_per_tensor_fp8_quant(num_tokens: int, hidden_size: int,
78
dtype: torch.dtype, seed: int) -> None:
79
- seed_everything(seed)
79
+ current_platform.seed_everything(seed)
80
81
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda")
82
@@ -95,7 +95,7 @@ def test_dynamic_per_tensor_fp8_quant(num_tokens: int, hidden_size: int,
95
@torch.inference_mode()
96
@pytest.mark.parametrize("seed", SEEDS)
97
def test_fp8_quant_large(seed: int) -> None:
98
- seed_everything(seed)
98
+ current_platform.seed_everything(seed)
99
100
num_tokens = 1024000 # Mistral-Nemo's max_position_embeddings
101
hidden_size = 1152 # Smallest hidden_size to reproduce the error
tests/kernels/test_gguf.py
+3
-3
@@ -7,7 +7,7 @@ from gguf import GGMLQuantizationType, GGUFReader, ReaderTensor, dequantize
7
from huggingface_hub import snapshot_download
8
9
import vllm._custom_ops as ops
10
-from vllm.utils import seed_everything
10
+from vllm.platforms import current_platform
11
12
GGUF_SAMPLE = snapshot_download("Isotr0py/test-gguf-sample")
13
@@ -75,7 +75,7 @@ def test_dequantize(hidden_size: int, dtype: torch.dtype,
75
@torch.inference_mode()
76
def test_mmvq(hidden_size: int, dtype: torch.dtype,
77
quant_type: GGMLQuantizationType):
78
- seed_everything(0)
78
+ current_platform.seed_everything(0)
79
80
tensors = get_gguf_sample_tensors(hidden_size, quant_type)
81
x = torch.rand((1, hidden_size), dtype=dtype, device="cuda")
@@ -111,7 +111,7 @@ def test_mmvq(hidden_size: int, dtype: torch.dtype,
111
@torch.inference_mode()
112
def test_mmq(num_tokens: int, hidden_size: int, dtype: torch.dtype,
113
quant_type: GGMLQuantizationType):
114
- seed_everything(0)
114
+ current_platform.seed_everything(0)
115
116
tensors = get_gguf_sample_tensors(hidden_size, quant_type)
117
x = torch.rand((num_tokens, hidden_size), dtype=dtype, device="cuda")
tests/kernels/test_int8_quant.py
+5
-5
@@ -4,7 +4,7 @@ import torch
4
from tests.kernels.quant_utils import ref_dynamic_per_token_quant
5
from tests.kernels.utils import opcheck
6
from vllm._custom_ops import scaled_int8_quant
7
-from vllm.utils import seed_everything
7
+from vllm.platforms import current_platform
8
9
DTYPES = [torch.half, torch.bfloat16, torch.float]
10
HIDDEN_SIZES = [16, 67, 768, 2048, 5120, 5137, 8192,
@@ -45,7 +45,7 @@ def opcheck_int8_quant_dynamic(output, input, symmetric=True):
45
@torch.inference_mode()
46
def test_dynamic_scaled_int8_quant(num_tokens: int, hidden_size: int,
47
dtype: torch.dtype, seed: int) -> None:
48
- seed_everything(seed)
48
+ current_platform.seed_everything(seed)
49
50
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda") * 1000
51
@@ -68,7 +68,7 @@ def test_dynamic_scaled_int8_quant(num_tokens: int, hidden_size: int,
68
@torch.inference_mode()
69
def test_dynamic_scaled_int8_azp_quant(num_tokens: int, hidden_size: int,
70
dtype: torch.dtype, seed: int) -> None:
71
- seed_everything(seed)
71
+ current_platform.seed_everything(seed)
72
int8_traits = torch.iinfo(torch.int8)
73
74
x = torch.rand(num_tokens, hidden_size, dtype=dtype,
@@ -112,7 +112,7 @@ def test_dynamic_scaled_int8_azp_quant(num_tokens: int, hidden_size: int,
112
def test_static_scaled_int8_quant(num_tokens: int, hidden_size: int,
113
dtype: torch.dtype, seed: int,
114
scale: float) -> None:
115
- seed_everything(seed)
115
+ current_platform.seed_everything(seed)
116
int8_traits = torch.iinfo(torch.int8)
117
118
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda") * 1000
@@ -138,7 +138,7 @@ def test_static_scaled_int8_quant(num_tokens: int, hidden_size: int,
138
def test_static_scaled_int8_azp_quant(num_tokens: int, hidden_size: int,
139
dtype: torch.dtype, seed: int,
140
scale: float, azp: int) -> None:
141
- seed_everything(seed)
141
+ current_platform.seed_everything(seed)
142
int8_traits = torch.iinfo(torch.int8)
143
144
x = torch.rand(num_tokens, hidden_size, dtype=dtype,
tests/kernels/test_layernorm.py
+2
-2
@@ -3,7 +3,7 @@ import torch
3
4
from tests.kernels.utils import opcheck
5
from vllm.model_executor.layers.layernorm import RMSNorm
6
-from vllm.utils import seed_everything
6
+from vllm.platforms import current_platform
7
8
DTYPES = [torch.half, torch.bfloat16, torch.float]
9
NUM_TOKENS = [7, 83, 4096] # Arbitrary values for testing
@@ -31,7 +31,7 @@ def test_rms_norm(
31
seed: int,
32
device: str,
33
) -> None:
34
- seed_everything(seed)
34
+ current_platform.seed_everything(seed)
35
torch.set_default_device(device)
36
layer = RMSNorm(hidden_size).to(dtype=dtype)
37
layer.weight.data.normal_(mean=1.0, std=0.1)
tests/kernels/test_mamba_ssm.py
+3
-3
@@ -8,7 +8,7 @@ from vllm import _custom_ops as ops # noqa: F401
8
from vllm.attention.backends.utils import PAD_SLOT_ID
9
from vllm.model_executor.layers.mamba.ops.mamba_ssm import (
10
selective_scan_fn, selective_state_update)
11
-from vllm.utils import seed_everything
11
+from vllm.platforms import current_platform
12
13
14
def selective_state_update_ref(state,
@@ -235,7 +235,7 @@ def test_selective_scan(is_variable_B, is_variable_C, varBC_groups, has_D,
235
rtolw = max(rtolw, rtol)
236
atolw = max(atolw, atol)
237
# set seed
238
- seed_everything(0)
238
+ current_platform.seed_everything(0)
239
batch_size = 1
240
dim = 4
241
dstate = 8
@@ -358,7 +358,7 @@ def test_selective_state_update(dim, dstate, has_z, itype):
358
if torch.version.hip:
359
atol *= 2
360
# set seed
361
- seed_everything(0)
361
+ current_platform.seed_everything(0)
362
batch_size = 1
363
state = torch.randn(batch_size, dim, dstate, dtype=itype, device=device)
364
x = torch.randn(batch_size, dim, device=device, dtype=itype)
tests/kernels/test_moe.py
+1
-2
@@ -19,7 +19,6 @@ from vllm.model_executor.layers.quantization.utils.marlin_utils_test import (
19
from vllm.model_executor.models.mixtral import MixtralMoE
20
from vllm.platforms import current_platform
21
from vllm.scalar_type import scalar_types
22
-from vllm.utils import seed_everything
22
23
24
@pytest.mark.parametrize("m", [1024 * 128, 512, 222, 33, 1])
@@ -115,7 +114,7 @@ def test_fused_marlin_moe(
114
num_bits: int,
115
is_k_full: bool,
116
):
118
- seed_everything(7)
117
+ current_platform.seed_everything(7)
118
119
# Filter act_order
120
if act_order:
tests/kernels/test_pos_encoding.py
+4
-4
@@ -5,7 +5,7 @@ import pytest
5
import torch
6
7
from vllm.model_executor.layers.rotary_embedding import get_rope
8
-from vllm.utils import seed_everything
8
+from vllm.platforms import current_platform
9
10
from .allclose_default import get_default_atol, get_default_rtol
11
@@ -48,7 +48,7 @@ def test_rotary_embedding(
48
if rotary_dim is None:
49
rotary_dim = head_size
50
51
- seed_everything(seed)
51
+ current_platform.seed_everything(seed)
52
torch.set_default_device(device)
53
if rotary_dim is None:
54
rotary_dim = head_size
@@ -100,7 +100,7 @@ def test_batched_rotary_embedding(
100
max_position: int = 8192,
101
base: int = 10000,
102
) -> None:
103
- seed_everything(seed)
103
+ current_platform.seed_everything(seed)
104
torch.set_default_device(device)
105
if rotary_dim is None:
106
rotary_dim = head_size
@@ -160,7 +160,7 @@ def test_batched_rotary_embedding_multi_lora(
160
max_position: int = 8192,
161
base: int = 10000,
162
) -> None:
163
- seed_everything(seed)
163
+ current_platform.seed_everything(seed)
164
torch.set_default_device(device)
165
if rotary_dim is None:
166
rotary_dim = head_size
tests/kernels/test_prefix_prefill.py
+4
-3
@@ -9,7 +9,8 @@ from xformers.ops.fmha.attn_bias import BlockDiagonalCausalFromBottomRightMask
9
10
from vllm.attention.backends.xformers import _make_alibi_bias
11
from vllm.attention.ops.prefix_prefill import context_attention_fwd
12
-from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, seed_everything
12
+from vllm.platforms import current_platform
13
+from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE
14
15
NUM_HEADS = [64]
16
NUM_QUERIES_PER_KV = [1, 8, 64]
@@ -39,7 +40,7 @@ def test_contexted_kv_attention(
40
kv_cache_dtype: str,
41
device: str,
42
) -> None:
42
- seed_everything(0)
43
+ current_platform.seed_everything(0)
44
torch.set_default_device(device)
45
46
# Need this, otherwise when we capture the graph the process
@@ -234,7 +235,7 @@ def test_contexted_kv_attention_alibi(
235
kv_cache_dtype: str,
236
device: str,
237
) -> None:
237
- seed_everything(0)
238
+ current_platform.seed_everything(0)
239
torch.set_default_device(device)
240
241
# Need this, otherwise when we capture the graph the process
tests/lora/test_layers.py
+2
-2
@@ -39,7 +39,7 @@ from vllm.model_executor.layers.rotary_embedding import get_rope
39
from vllm.model_executor.layers.vocab_parallel_embedding import (
40
ParallelLMHead, VocabParallelEmbedding, get_masked_input_and_mask)
41
from vllm.model_executor.utils import set_random_seed
42
-from vllm.utils import seed_everything
42
+from vllm.platforms import current_platform
43
44
from .utils import DummyLoRAManager
45
@@ -923,7 +923,7 @@ def test_rotary_embedding_long_context(dist_init, num_loras, device,
923
seq_len) -> None:
924
dtype = torch.float16
925
seed = 0
926
- seed_everything(seed)
926
+ current_platform.seed_everything(seed)
927
torch.set_default_device(device)
928
punica_wrapper = PunicaWrapper(8192, 256, device)
929
max_loras = 8
tests/lora/test_punica_sizes.py
+5
-5
@@ -1,5 +1,5 @@
1
"""
2
-This script is mainly used to tests various hidden_sizes. We have collected the
2
+This script is mainly used to tests various hidden_sizes. We have collected the
3
hidden_sizes included in the LoRA models currently supported by vLLM. It tests
4
whether the corresponding Triton kernel can run normally when tensor parallelism
5
is set to [1, 2, 4, 8, 16, 32, 64].
@@ -15,8 +15,8 @@ from vllm.lora.ops.bgmv_shrink import bgmv_shrink
15
from vllm.lora.ops.sgmv_expand import sgmv_expand
16
from vllm.lora.ops.sgmv_expand_slice import sgmv_expand_slice
17
from vllm.lora.ops.sgmv_shrink import sgmv_shrink
18
+from vllm.platforms import current_platform
19
from vllm.triton_utils.libentry import LibEntry
19
-from vllm.utils import seed_everything
20
21
from .utils import (generate_data, generate_data_for_expand_nslices,
22
ref_torch_groupgemm)
@@ -146,7 +146,7 @@ def test_punica_sgmv(
146
device: str,
147
):
148
torch.set_default_device(device)
149
- seed_everything(seed)
149
+ current_platform.seed_everything(seed)
150
151
seq_length = 128
152
(
@@ -239,7 +239,7 @@ def test_punica_bgmv(
239
from vllm.lora.ops.bgmv_shrink import _bgmv_shrink_kernel
240
241
torch.set_default_device(device)
242
- seed_everything(seed)
242
+ current_platform.seed_everything(seed)
243
244
seq_length = 1
245
(
@@ -327,7 +327,7 @@ def test_punica_expand_nslices(
327
from vllm.lora.ops.bgmv_expand_slice import _bgmv_expand_slice_kernel
328
329
torch.set_default_device(device)
330
- seed_everything(seed)
330
+ current_platform.seed_everything(seed)
331
332
seq_length = 128 if op_type == "sgmv" else 1
333
(
tests/lora/test_punica_variation.py
+6
-6
@@ -1,6 +1,6 @@
1
"""
2
-This script is mainly used to test whether trtion kernels can run normally
3
-under different conditions, including various batches, numbers of LoRA , and
2
+This script is mainly used to test whether trtion kernels can run normally
3
+under different conditions, including various batches, numbers of LoRA , and
4
maximum ranks.
5
"""
6
from unittest.mock import patch
@@ -14,8 +14,8 @@ from vllm.lora.ops.bgmv_shrink import bgmv_shrink
14
from vllm.lora.ops.sgmv_expand import sgmv_expand
15
from vllm.lora.ops.sgmv_expand_slice import sgmv_expand_slice
16
from vllm.lora.ops.sgmv_shrink import sgmv_shrink
17
+from vllm.platforms import current_platform
18
from vllm.triton_utils.libentry import LibEntry
18
-from vllm.utils import seed_everything
19
20
from .utils import (generate_data, generate_data_for_expand_nslices,
21
ref_torch_groupgemm)
@@ -61,7 +61,7 @@ def test_punica_sgmv(
61
device: str,
62
):
63
torch.set_default_device(device)
64
- seed_everything(seed)
64
+ current_platform.seed_everything(seed)
65
66
seq_length = 128
67
(
@@ -154,7 +154,7 @@ def test_punica_bgmv(
154
from vllm.lora.ops.bgmv_shrink import _bgmv_shrink_kernel
155
156
torch.set_default_device(device)
157
- seed_everything(seed)
157
+ current_platform.seed_everything(seed)
158
159
seq_length = 1
160
(
@@ -242,7 +242,7 @@ def test_punica_expand_nslices(
242
from vllm.lora.ops.bgmv_expand_slice import _bgmv_expand_slice_kernel
243
244
torch.set_default_device(device)
245
- seed_everything(seed)
245
+ current_platform.seed_everything(seed)
246
247
seq_length = 128 if op_type == "sgmv" else 1
248
(
vllm/model_executor/utils.py
+1
-2
@@ -4,11 +4,10 @@ from typing import Any, Dict, Optional
4
import torch
5
6
from vllm.platforms import current_platform
7
-from vllm.utils import seed_everything
7
8
9
def set_random_seed(seed: int) -> None:
11
- seed_everything(seed)
10
+ current_platform.seed_everything(seed)
11
12
13
def set_weight_attrs(
vllm/platforms/interface.py
+14
@@ -1,6 +1,8 @@
1
import enum
2
+import random
3
from typing import NamedTuple, Optional, Tuple, Union
4
5
+import numpy as np
6
import torch
7
8
@@ -111,6 +113,18 @@ class Platform:
113
"""
114
return torch.inference_mode(mode=True)
115
116
+ @classmethod
117
+ def seed_everything(cls, seed: int) -> None:
118
+ """
119
+ Set the seed of each random module.
120
+ `torch.manual_seed` will set seed on all devices.
121
+
122
+ Loosely based on: https://github.com/Lightning-AI/pytorch-lightning/blob/2.4.0/src/lightning/fabric/utilities/seed.py#L20
123
+ """
124
+ random.seed(seed)
125
+ np.random.seed(seed)
126
+ torch.manual_seed(seed)
127
+
128
129
class UnspecifiedPlatform(Platform):
130
_enum = PlatformEnum.UNSPECIFIED
vllm/utils.py
+2
-19
@@ -7,7 +7,6 @@ import gc
7
import inspect
8
import ipaddress
9
import os
10
-import random
10
import socket
11
import subprocess
12
import sys
@@ -331,22 +330,6 @@ def get_cpu_memory() -> int:
330
return psutil.virtual_memory().total
331
332
334
-def seed_everything(seed: int) -> None:
335
- """
336
- Set the seed of each random module.
337
-
338
- Loosely based on: https://github.com/Lightning-AI/pytorch-lightning/blob/2.4.0/src/lightning/fabric/utilities/seed.py#L20
339
- """
340
- random.seed(seed)
341
- np.random.seed(seed)
342
-
343
- if current_platform.is_cuda_alike():
344
- torch.cuda.manual_seed_all(seed)
345
-
346
- if current_platform.is_xpu():
347
- torch.xpu.manual_seed_all(seed)
348
-
349
-
333
def random_uuid() -> str:
334
return str(uuid.uuid4().hex)
335
@@ -643,7 +626,7 @@ def create_kv_caches_with_random_flash(
626
seed: int = 0,
627
device: Optional[str] = "cuda",
628
) -> Tuple[List[torch.Tensor], List[torch.Tensor]]:
646
- seed_everything(seed)
629
+ current_platform.seed_everything(seed)
630
631
torch_dtype = get_kv_cache_torch_dtype(cache_dtype, model_dtype)
632
key_value_cache_shape = (num_blocks, 2, block_size, num_heads, head_size)
@@ -685,7 +668,7 @@ def create_kv_caches_with_random(
668
f"Does not support key cache of type fp8 with head_size {head_size}"
669
)
670
688
- seed_everything(seed)
671
+ current_platform.seed_everything(seed)
672
673
torch_dtype = get_kv_cache_torch_dtype(cache_dtype, model_dtype)
674