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, 15:45 UTC
[CI/Build] Avoid CUDA initialization (#8534)
Cyrus Leung committed
Sep 18, 2024 at 18:38 UTC
6ffa3f314c59e42238f1c5f923ff2839e0af9698
55 files changed
+256
-256
benchmarks/kernels/benchmark_layernorm.py
+3
-6
@@ -1,10 +1,10 @@
1
-import random
1
import time
2
3
import torch
4
5
from vllm.model_executor.layers.layernorm import RMSNorm
7
-from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser
6
+from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
7
+ seed_everything)
8
9
10
@torch.inference_mode()
@@ -16,10 +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
- random.seed(seed)
20
- torch.random.manual_seed(seed)
21
- if torch.cuda.is_available():
22
- torch.cuda.manual_seed(seed)
19
+ seed_everything(seed)
20
torch.set_default_device("cuda")
21
22
layer = RMSNorm(hidden_size).to(dtype=dtype)
benchmarks/kernels/benchmark_moe.py
+3
-3
@@ -10,7 +10,7 @@ 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
13
+from vllm.utils import FlexibleArgumentParser, seed_everything
14
15
16
class BenchmarkConfig(TypedDict):
@@ -166,7 +166,7 @@ class BenchmarkWorker:
166
167
def __init__(self, seed: int) -> None:
168
torch.set_default_device("cuda")
169
- torch.cuda.manual_seed_all(seed)
169
+ seed_everything(seed)
170
self.seed = seed
171
172
def benchmark(
@@ -180,7 +180,7 @@ class BenchmarkWorker:
180
use_fp8_w8a8: bool,
181
use_int8_w8a16: bool,
182
) -> Tuple[Dict[str, int], float]:
183
- torch.cuda.manual_seed_all(self.seed)
183
+ seed_everything(self.seed)
184
dtype_str = get_config_dtype_str(dtype,
185
use_int8_w8a16=use_int8_w8a16,
186
use_fp8_w8a8=use_fp8_w8a8)
benchmarks/kernels/benchmark_paged_attention.py
+2
-5
@@ -6,7 +6,7 @@ import torch
6
7
from vllm import _custom_ops as ops
8
from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
9
- create_kv_caches_with_random)
9
+ create_kv_caches_with_random, seed_everything)
10
11
NUM_BLOCKS = 1024
12
PARTITION_SIZE = 512
@@ -28,10 +28,7 @@ def main(
28
device: str = "cuda",
29
kv_cache_dtype: Optional[str] = None,
30
) -> None:
31
- random.seed(seed)
32
- torch.random.manual_seed(seed)
33
- if torch.cuda.is_available():
34
- torch.cuda.manual_seed(seed)
31
+ seed_everything(seed)
32
33
scale = float(1.0 / (head_size**0.5))
34
query = torch.empty(num_seqs,
benchmarks/kernels/benchmark_quant.py
+3
-6
@@ -1,10 +1,10 @@
1
-import random
1
import time
2
3
import torch
4
5
from vllm import _custom_ops as ops
7
-from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser
6
+from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
7
+ seed_everything)
8
9
10
@torch.inference_mode()
@@ -17,10 +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
- random.seed(seed)
21
- torch.random.manual_seed(seed)
22
- if torch.cuda.is_available():
23
- torch.cuda.manual_seed(seed)
20
+ 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
+2
-4
@@ -6,7 +6,7 @@ import torch
6
7
from vllm.model_executor.layers.rotary_embedding import (RotaryEmbedding,
8
get_rope)
9
-from vllm.utils import FlexibleArgumentParser
9
+from vllm.utils import FlexibleArgumentParser, seed_everything
10
11
12
def benchmark_rope_kernels_multi_lora(
@@ -22,9 +22,7 @@ def benchmark_rope_kernels_multi_lora(
22
max_position: int = 8192,
23
base: int = 10000,
24
) -> None:
25
- torch.random.manual_seed(seed)
26
- if torch.cuda.is_available():
27
- torch.cuda.manual_seed(seed)
25
+ seed_everything(seed)
26
torch.set_default_device(device)
27
if rotary_dim is None:
28
rotary_dim = head_size
tests/kernels/test_activation.py
+3
-6
@@ -7,6 +7,7 @@ from tests.kernels.utils import opcheck
7
from vllm.model_executor.layers.activation import (FastGELU, GeluAndMul,
8
NewGELU, QuickGELU,
9
SiluAndMul)
10
+from vllm.utils import seed_everything
11
12
from .allclose_default import get_default_atol, get_default_rtol
13
@@ -34,9 +35,7 @@ def test_act_and_mul(
35
seed: int,
36
device: str,
37
) -> None:
37
- torch.random.manual_seed(seed)
38
- if torch.cuda.is_available():
39
- torch.cuda.manual_seed(seed)
38
+ seed_everything(seed)
39
torch.set_default_device(device)
40
x = torch.randn(num_tokens, 2 * d, dtype=dtype)
41
if activation == "silu":
@@ -77,9 +76,7 @@ def test_activation(
76
seed: int,
77
device: str,
78
) -> None:
80
- torch.random.manual_seed(seed)
81
- if torch.cuda.is_available():
82
- torch.cuda.manual_seed(seed)
79
+ seed_everything(seed)
80
torch.set_default_device(device)
81
x = torch.randn(num_tokens, d, dtype=dtype)
82
layer = activation[0]()
tests/kernels/test_attention.py
+5
-13
@@ -6,7 +6,7 @@ import torch
6
7
from tests.kernels.utils import opcheck
8
from vllm import _custom_ops as ops
9
-from vllm.utils import get_max_shared_memory_bytes, is_hip
9
+from vllm.utils import get_max_shared_memory_bytes, is_hip, seed_everything
10
11
from .allclose_default import get_default_atol, get_default_rtol
12
@@ -139,10 +139,8 @@ def test_paged_attention(
139
) -> None:
140
if kv_cache_dtype == "fp8" and head_size % 16:
141
pytest.skip()
142
- random.seed(seed)
143
- torch.random.manual_seed(seed)
144
- if torch.cuda.is_available():
145
- torch.cuda.manual_seed(seed)
142
+
143
+ seed_everything(seed)
144
torch.set_default_device(device)
145
scale = float(1.0 / (head_size**0.5))
146
num_query_heads, num_kv_heads = num_heads
@@ -354,10 +352,7 @@ def test_paged_attention_rocm(
352
seed: int,
353
device: str,
354
) -> None:
357
- random.seed(seed)
358
- torch.random.manual_seed(seed)
359
- if torch.cuda.is_available():
360
- torch.cuda.manual_seed(seed)
355
+ seed_everything(seed)
356
torch.set_default_device(device)
357
scale = float(1.0 / (head_size**0.5))
358
num_query_heads, num_kv_heads = num_heads
@@ -506,10 +501,7 @@ def test_multi_query_kv_attention(
501
seed: int,
502
device: str,
503
) -> None:
509
- random.seed(seed)
510
- torch.random.manual_seed(seed)
511
- if torch.cuda.is_available():
512
- torch.cuda.manual_seed(seed)
504
+ seed_everything(seed)
505
torch.set_default_device(device)
506
# MAX_SEQ_LEN sometimes causes OOM in the reference implementation.
507
# As the xformers library is already tested with its own tests, we can use
tests/kernels/test_attention_selector.py
+1
-1
@@ -45,7 +45,7 @@ def test_flash_attn(monkeypatch):
45
override_backend_env_variable(monkeypatch, STR_FLASH_ATTN_VAL)
46
47
# Unsupported CUDA arch
48
- with patch("torch.cuda.get_device_capability", return_value=[7, 5]):
48
+ with patch("torch.cuda.get_device_capability", return_value=(7, 5)):
49
backend = which_attn_to_use(8, 16, 8, None, torch.float16, None, 16)
50
assert backend.name != STR_FLASH_ATTN_VAL
51
tests/kernels/test_awq_triton.py
+3
-2
@@ -7,6 +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
11
12
device = "cuda"
13
@@ -79,7 +80,7 @@ def test_dequantize(qweight_rows, qweight_cols, group_size):
80
zeros_cols = qweight_cols
81
zeros_dtype = torch.int32
82
82
- torch.manual_seed(0)
83
+ seed_everything(0)
84
85
qweight = torch.randint(0,
86
torch.iinfo(torch.int32).max,
@@ -133,7 +134,7 @@ def test_gemm(N, K, M, splitK, group_size):
134
qzeros_rows = scales_rows
135
qzeros_cols = qweight_cols
136
136
- torch.manual_seed(0)
137
+ seed_everything(0)
138
139
input = torch.rand((input_rows, input_cols),
140
dtype=input_dtype,
tests/kernels/test_blocksparse_attention.py
+3
-9
@@ -7,7 +7,7 @@ import torch
7
from vllm import _custom_ops as ops
8
from vllm.attention.ops.blocksparse_attention.interface import (
9
LocalStridedBlockSparseAttn)
10
-from vllm.utils import get_max_shared_memory_bytes, is_hip
10
+from vllm.utils import get_max_shared_memory_bytes, is_hip, seed_everything
11
12
from .allclose_default import get_default_atol, get_default_rtol
13
@@ -172,10 +172,7 @@ def test_paged_attention(
172
blocksparse_block_size: int,
173
blocksparse_head_sliding_step: int,
174
) -> None:
175
- random.seed(seed)
176
- torch.random.manual_seed(seed)
177
- if torch.cuda.is_available():
178
- torch.cuda.manual_seed(seed)
175
+ seed_everything(seed)
176
torch.set_default_device(device)
177
scale = float(1.0 / (head_size**0.5))
178
num_query_heads, num_kv_heads = num_heads
@@ -386,10 +383,7 @@ def test_varlen_blocksparse_attention_prefill(
383
seed: int,
384
device: str,
385
) -> None:
389
- random.seed(seed)
390
- torch.random.manual_seed(seed)
391
- if torch.cuda.is_available():
392
- torch.cuda.manual_seed(seed)
386
+ seed_everything(seed)
387
torch.set_default_device(device)
388
# MAX_SEQ_LEN sometimes causes OOM in the reference implementation.
389
# As the xformers library is already tested with its own tests, we can use
tests/kernels/test_cache.py
+7
-18
@@ -6,6 +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
10
11
COPYING_DIRECTION = [('cuda', 'cpu'), ('cuda', 'cuda'), ('cpu', 'cuda')]
12
DTYPES = [torch.half, torch.bfloat16, torch.float]
@@ -55,10 +56,7 @@ def test_copy_blocks(
56
) -> None:
57
if kv_cache_dtype == "fp8" and head_size % 16:
58
pytest.skip()
58
- random.seed(seed)
59
- torch.random.manual_seed(seed)
60
- if torch.cuda.is_available():
61
- torch.cuda.manual_seed(seed)
59
+ 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.
@@ -134,10 +132,7 @@ def test_reshape_and_cache(
132
) -> None:
133
if kv_cache_dtype == "fp8" and head_size % 16:
134
pytest.skip()
137
- random.seed(seed)
138
- torch.random.manual_seed(seed)
139
- if torch.cuda.is_available():
140
- torch.cuda.manual_seed(seed)
135
+ seed_everything(seed)
136
torch.set_default_device(device)
137
# Create a random slot mapping.
138
num_slots = block_size * num_blocks
@@ -229,9 +224,7 @@ def test_reshape_and_cache_flash(
224
device: str,
225
kv_cache_dtype: str,
226
) -> None:
232
- random.seed(seed)
233
- torch.random.manual_seed(seed)
234
- torch.cuda.manual_seed(seed)
227
+ seed_everything(seed)
228
torch.set_default_device(device)
229
230
# Create a random slot mapping.
@@ -345,10 +338,8 @@ def test_swap_blocks(
338
pytest.skip()
339
if kv_cache_dtype == "fp8" and head_size % 16:
340
pytest.skip()
348
- random.seed(seed)
349
- torch.random.manual_seed(seed)
350
- if torch.cuda.is_available():
351
- torch.cuda.manual_seed(seed)
341
+
342
+ seed_everything(seed)
343
344
src_device = device if direction[0] == "cuda" else 'cpu'
345
dst_device = device if direction[1] == "cuda" else 'cpu'
@@ -417,9 +408,7 @@ def test_fp8_e4m3_conversion(
408
seed: int,
409
device: str,
410
) -> None:
420
- random.seed(seed)
421
- torch.random.manual_seed(seed)
422
- torch.cuda.manual_seed(seed)
411
+ seed_everything(seed)
412
413
low = -224.0
414
high = 224.0
tests/kernels/test_causal_conv1d.py
+3
-2
@@ -7,6 +7,7 @@ from einops import rearrange
7
8
from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
9
causal_conv1d_fn, causal_conv1d_update)
10
+from vllm.utils import seed_everything
11
12
13
def causal_conv1d_ref(
@@ -104,7 +105,7 @@ def test_causal_conv1d(batch, dim, seqlen, width, has_bias, silu_activation,
105
if itype == torch.bfloat16:
106
rtol, atol = 1e-2, 5e-2
107
# set seed
107
- torch.random.manual_seed(0)
108
+ seed_everything(0)
109
if not channel_last:
110
x = torch.randn(batch,
111
4096 + dim + 64,
@@ -175,7 +176,7 @@ def test_causal_conv1d_update(batch, dim, width, has_bias, silu_activation,
176
if itype == torch.bfloat16:
177
rtol, atol = 1e-2, 5e-2
178
# set seed
178
- torch.random.manual_seed(0)
179
+ seed_everything(0)
180
batch = 2
181
x = torch.randn(batch, dim, device=device, dtype=itype)
182
conv_state = torch.randn(batch, dim, width, device=device, dtype=itype)
tests/kernels/test_cutlass.py
+4
-7
@@ -15,9 +15,6 @@ CUDA_DEVICES = [
15
f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)
16
]
17
18
-capability = current_platform.get_device_capability()
19
-capability = capability[0] * 10 + capability[1]
20
-
18
19
def to_fp8(tensor: torch.Tensor):
20
finfo = torch.finfo(torch.float8_e4m3fn)
@@ -119,7 +116,7 @@ def cutlass_int8_gemm_helper(m: int,
116
@pytest.mark.parametrize("per_act_token", [True, False])
117
@pytest.mark.parametrize("per_out_ch", [True, False])
118
@pytest.mark.parametrize("use_bias", [True, False])
122
-@pytest.mark.skipif(capability < 89,
119
+@pytest.mark.skipif(not current_platform.has_device_capability(89),
120
reason="FP8 is not supported on this GPU type.")
121
def test_cutlass_fp8_gemm(m: int, n: int, k: int, per_act_token: bool,
122
per_out_ch: bool, use_bias: bool):
@@ -157,7 +154,7 @@ def test_cutlass_int8_gemm_output_dtype(per_act_token: bool, per_out_ch: bool,
154
@pytest.mark.parametrize("per_out_ch", [True, False])
155
@pytest.mark.parametrize("out_dtype", [torch.bfloat16, torch.float16])
156
@pytest.mark.parametrize("use_bias", [True, False])
160
-@pytest.mark.skipif(capability < 89,
157
+@pytest.mark.skipif(not current_platform.has_device_capability(89),
158
reason="FP8 is not supported on this GPU type.")
159
def test_cutlass_fp8_gemm_output_dtype(per_act_token: bool, per_out_ch: bool,
160
out_dtype: Type[torch.dtype],
@@ -175,7 +172,7 @@ def test_cutlass_fp8_gemm_output_dtype(per_act_token: bool, per_out_ch: bool,
172
@pytest.mark.parametrize("per_out_ch", [True, False])
173
@pytest.mark.parametrize("use_bias", [True, False])
174
@pytest.mark.parametrize("device", CUDA_DEVICES)
178
-@pytest.mark.skipif(capability < 89,
175
+@pytest.mark.skipif(not current_platform.has_device_capability(89),
176
reason="FP8 is not supported on this GPU type.")
177
def test_cutlass_fp8_gemm_devices(per_act_token: bool, per_out_ch: bool,
178
use_bias: bool, device: str):
@@ -207,7 +204,7 @@ def test_cutlass_int8_gemm_devices(per_act_token: bool, per_out_ch: bool,
204
@pytest.mark.parametrize("per_act_token", [True, False])
205
@pytest.mark.parametrize("per_out_ch", [True, False])
206
@pytest.mark.parametrize("use_bias", [True, False])
210
-@pytest.mark.skipif(capability < 89,
207
+@pytest.mark.skipif(not current_platform.has_device_capability(89),
208
reason="FP8 is not supported on this GPU type.")
209
def test_cutlass_fp8_gemm_m_sweep(per_act_token: bool, per_out_ch: bool,
210
use_bias: bool):
tests/kernels/test_flash_attn.py
+3
-2
@@ -4,6 +4,7 @@ import pytest
4
import torch
5
6
import vllm.attention.backends.flash_attn # noqa: F401
7
+from vllm.utils import seed_everything
8
9
NUM_HEADS = [(4, 4), (8, 2), (16, 2)]
10
HEAD_SIZES = [128, 256]
@@ -87,7 +88,7 @@ def test_flash_attn_with_paged_kv(
88
num_blocks: int,
89
) -> None:
90
torch.set_default_device("cuda")
90
- torch.cuda.manual_seed_all(0)
91
+ seed_everything(0)
92
num_seqs = len(kv_lens)
93
num_query_heads = num_heads[0]
94
num_kv_heads = num_heads[1]
@@ -174,7 +175,7 @@ def test_varlen_with_paged_kv(
175
num_blocks: int,
176
) -> None:
177
torch.set_default_device("cuda")
177
- torch.cuda.manual_seed_all(0)
178
+ seed_everything(0)
179
num_seqs = len(seq_lens)
180
query_lens = [x[0] for x in seq_lens]
181
kv_lens = [x[1] for x in seq_lens]
tests/kernels/test_flashinfer.py
+6
-4
@@ -4,6 +4,8 @@ import flashinfer
4
import pytest
5
import torch
6
7
+from vllm.utils import seed_everything
8
+
9
NUM_HEADS = [(16, 16), (32, 8), (64, 8), (6, 1)]
10
HEAD_SIZES = [128, 256]
11
BLOCK_SIZES = [16, 32]
@@ -82,7 +84,7 @@ def test_flashinfer_decode_with_paged_kv(
84
soft_cap: Optional[float],
85
) -> None:
86
torch.set_default_device("cuda")
85
- torch.cuda.manual_seed_all(0)
87
+ seed_everything(0)
88
num_seqs = len(kv_lens)
89
num_query_heads = num_heads[0]
90
num_kv_heads = num_heads[1]
@@ -168,7 +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")
171
- torch.cuda.manual_seed_all(0)
173
+ 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]
@@ -266,7 +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")
269
- torch.cuda.manual_seed_all(0)
271
+ 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]
@@ -379,7 +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")
382
- torch.cuda.manual_seed_all(0)
384
+ 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
-6
@@ -5,6 +5,7 @@ import vllm._custom_ops as ops
5
from tests.kernels.quant_utils import (FP8_DTYPE,
6
ref_dynamic_per_tensor_fp8_quant,
7
ref_dynamic_per_token_quant)
8
+from vllm.utils import seed_everything
9
10
DTYPES = [torch.half, torch.bfloat16, torch.float]
11
HIDDEN_SIZES = [1, 2, 3, 4, 16, 67, 768, 2048, 5120, 5137, 8192,
@@ -24,8 +25,7 @@ SEEDS = [0]
25
def test_dynamic_per_token_fp8_quant(num_tokens: int, hidden_size: int,
26
dtype: torch.dtype, scale_ub: bool,
27
seed: int) -> None:
27
- torch.random.manual_seed(seed)
28
- torch.cuda.manual_seed(seed)
28
+ seed_everything(seed)
29
30
x = torch.rand(num_tokens, hidden_size, dtype=dtype,
31
device="cuda") + 1e-6 # avoid nans
@@ -49,8 +49,7 @@ def test_dynamic_per_token_fp8_quant(num_tokens: int, hidden_size: int,
49
@torch.inference_mode()
50
def test_dynamic_per_tensor_fp8_quant(num_tokens: int, hidden_size: int,
51
dtype: torch.dtype, seed: int) -> None:
52
- torch.random.manual_seed(seed)
53
- torch.cuda.manual_seed(seed)
52
+ seed_everything(seed)
53
54
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda")
55
@@ -67,8 +66,7 @@ def test_dynamic_per_tensor_fp8_quant(num_tokens: int, hidden_size: int,
66
@torch.inference_mode()
67
@pytest.mark.parametrize("seed", SEEDS)
68
def test_fp8_quant_large(seed: int) -> None:
70
- torch.random.manual_seed(seed)
71
- torch.cuda.manual_seed(seed)
69
+ seed_everything(seed)
70
71
num_tokens = 1024000 # Mistral-Nemo's max_position_embeddings
72
hidden_size = 1152 # Smallest hidden_size to reproduce the error
tests/kernels/test_gguf.py
+3
-2
@@ -7,6 +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
11
12
GGUF_SAMPLE = snapshot_download("Isotr0py/test-gguf-sample")
13
@@ -74,7 +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):
77
- torch.cuda.manual_seed_all(0)
78
+ 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")
@@ -110,7 +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):
113
- torch.cuda.manual_seed_all(0)
114
+ 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
-8
@@ -4,6 +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
8
9
DTYPES = [torch.half, torch.bfloat16, torch.float]
10
HIDDEN_SIZES = [16, 67, 768, 2048, 5120, 5137, 8192,
@@ -44,8 +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:
47
- torch.random.manual_seed(seed)
48
- torch.cuda.manual_seed(seed)
48
+ seed_everything(seed)
49
50
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda") * 1000
51
@@ -68,8 +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
- torch.random.manual_seed(seed)
72
- torch.cuda.manual_seed(seed)
71
+ seed_everything(seed)
72
int8_traits = torch.iinfo(torch.int8)
73
74
x = torch.rand(num_tokens, hidden_size, dtype=dtype,
@@ -113,8 +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:
116
- torch.random.manual_seed(seed)
117
- torch.cuda.manual_seed(seed)
115
+ seed_everything(seed)
116
int8_traits = torch.iinfo(torch.int8)
117
118
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda") * 1000
@@ -140,8 +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:
143
- torch.random.manual_seed(seed)
144
- torch.cuda.manual_seed(seed)
141
+ 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
-3
@@ -3,6 +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
7
8
DTYPES = [torch.half, torch.bfloat16, torch.float]
9
NUM_TOKENS = [7, 83, 4096] # Arbitrary values for testing
@@ -30,9 +31,7 @@ def test_rms_norm(
31
seed: int,
32
device: str,
33
) -> None:
33
- torch.random.manual_seed(seed)
34
- if torch.cuda.is_available():
35
- torch.cuda.manual_seed(seed)
34
+ 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_machete_gemm.py
+1
-1
@@ -48,7 +48,7 @@ WTYPE_ZEROPOINTS = [
48
# `is_quant_method_supported` conflates kernels with quantization methods
49
# an assumption which is breaking down as quantizations methods can have
50
# have kernels and some kernels support multiple quantization methods.
51
-IS_SUPPORTED_BY_GPU = current_platform.get_device_capability()[0] >= 9
51
+IS_SUPPORTED_BY_GPU = current_platform.has_device_capability(90)
52
53
54
def rand_data(shape, dtype=torch.float16):
tests/kernels/test_mamba_ssm.py
+3
-2
@@ -5,6 +5,7 @@ from einops import rearrange, repeat
5
6
from vllm.model_executor.layers.mamba.ops.mamba_ssm import (
7
selective_scan_fn, selective_state_update)
8
+from vllm.utils import seed_everything
9
10
11
def selective_state_update_ref(state,
@@ -186,7 +187,7 @@ def test_selective_scan(is_variable_B, is_variable_C, varBC_groups, has_D,
187
rtolw = max(rtolw, rtol)
188
atolw = max(atolw, atol)
189
# set seed
189
- torch.random.manual_seed(0)
190
+ seed_everything(0)
191
batch_size = 2
192
dim = 4
193
dstate = 8
@@ -287,7 +288,7 @@ def test_selective_state_update(dim, dstate, has_z, itype):
288
if torch.version.hip:
289
atol *= 2
290
# set seed
290
- torch.random.manual_seed(0)
291
+ seed_everything(0)
292
batch_size = 1
293
state = torch.randn(batch_size, dim, dstate, dtype=itype, device=device)
294
x = torch.randn(batch_size, dim, device=device, dtype=itype)
tests/kernels/test_moe.py
+2
-1
@@ -18,6 +18,7 @@ from vllm.model_executor.layers.quantization.utils.marlin_utils_test import (
18
marlin_quantize)
19
from vllm.model_executor.models.mixtral import MixtralMoE
20
from vllm.scalar_type import scalar_types
21
+from vllm.utils import seed_everything
22
23
24
def torch_moe(a, w1, w2, score, topk):
@@ -151,7 +152,7 @@ def test_fused_marlin_moe(
152
act_order: bool,
153
num_bits: int,
154
):
154
- torch.manual_seed(7)
155
+ seed_everything(7)
156
157
if topk > e:
158
return
tests/kernels/test_pos_encoding.py
+5
-9
@@ -5,6 +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
9
10
from .allclose_default import get_default_atol, get_default_rtol
11
@@ -46,9 +47,8 @@ def test_rotary_embedding(
47
) -> None:
48
if rotary_dim is None:
49
rotary_dim = head_size
49
- torch.random.manual_seed(seed)
50
- if torch.cuda.is_available():
51
- torch.cuda.manual_seed(seed)
50
+
51
+ seed_everything(seed)
52
torch.set_default_device(device)
53
if rotary_dim is None:
54
rotary_dim = head_size
@@ -100,9 +100,7 @@ def test_batched_rotary_embedding(
100
max_position: int = 8192,
101
base: int = 10000,
102
) -> None:
103
- torch.random.manual_seed(seed)
104
- if torch.cuda.is_available():
105
- torch.cuda.manual_seed(seed)
103
+ seed_everything(seed)
104
torch.set_default_device(device)
105
if rotary_dim is None:
106
rotary_dim = head_size
@@ -162,9 +160,7 @@ def test_batched_rotary_embedding_multi_lora(
160
max_position: int = 8192,
161
base: int = 10000,
162
) -> None:
165
- torch.random.manual_seed(seed)
166
- if torch.cuda.is_available():
167
- torch.cuda.manual_seed(seed)
163
+ 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
+3
-9
@@ -9,7 +9,7 @@ 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
12
+from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, seed_everything
13
14
NUM_HEADS = [64]
15
NUM_QUERIES_PER_KV = [1, 8, 64]
@@ -39,10 +39,7 @@ def test_contexted_kv_attention(
39
kv_cache_dtype: str,
40
device: str,
41
) -> None:
42
- random.seed(0)
43
- torch.manual_seed(0)
44
- if torch.cuda.is_available():
45
- torch.cuda.manual_seed(0)
42
+ seed_everything(0)
43
torch.set_default_device(device)
44
45
# Need this, otherwise when we capture the graph the process
@@ -237,10 +234,7 @@ def test_contexted_kv_attention_alibi(
234
kv_cache_dtype: str,
235
device: str,
236
) -> None:
240
- random.seed(0)
241
- torch.manual_seed(0)
242
- if torch.cuda.is_available():
243
- torch.cuda.manual_seed(0)
237
+ seed_everything(0)
238
torch.set_default_device(device)
239
240
# Need this, otherwise when we capture the graph the process
tests/lora/test_layers.py
+2
-3
@@ -39,6 +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
43
44
from .utils import DummyLoRAManager
45
@@ -922,9 +923,7 @@ def test_rotary_embedding_long_context(dist_init, num_loras, device,
923
seq_len) -> None:
924
dtype = torch.float16
925
seed = 0
925
- torch.random.manual_seed(seed)
926
- if torch.cuda.is_available():
927
- torch.cuda.manual_seed(seed)
926
+ 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
-13
@@ -4,7 +4,6 @@ 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].
6
"""
7
-import random
7
from unittest.mock import patch
8
9
import pytest
@@ -17,6 +16,7 @@ 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.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)
@@ -145,11 +145,8 @@ def test_punica_sgmv(
145
seed: int,
146
device: str,
147
):
148
- random.seed(seed)
148
torch.set_default_device(device)
150
- torch.random.manual_seed(seed)
151
- if torch.cuda.is_available():
152
- torch.cuda.manual_seed(seed)
149
+ seed_everything(seed)
150
151
seq_length = 128
152
(
@@ -238,11 +235,8 @@ def test_punica_bgmv(
235
from vllm.lora.ops.bgmv_expand import _bgmv_expand_kernel
236
from vllm.lora.ops.bgmv_shrink import _bgmv_shrink_kernel
237
241
- random.seed(seed)
238
torch.set_default_device(device)
243
- torch.random.manual_seed(seed)
244
- if torch.cuda.is_available():
245
- torch.cuda.manual_seed(seed)
239
+ seed_everything(seed)
240
241
seq_length = 1
242
(
@@ -329,11 +323,9 @@ def test_punica_expand_nslices(
323
):
324
from vllm.lora.ops.bgmv_expand_slice import _bgmv_expand_slice_kernel
325
332
- random.seed(seed)
326
torch.set_default_device(device)
334
- torch.random.manual_seed(seed)
335
- if torch.cuda.is_available():
336
- torch.cuda.manual_seed(seed)
327
+ seed_everything(seed)
328
+
329
seq_length = 128 if op_type == "sgmv" else 1
330
(
331
inputs_tensor,
tests/lora/test_punica_variation.py
+5
-13
@@ -3,7 +3,6 @@ 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
-import random
6
from unittest.mock import patch
7
8
import pytest
@@ -16,6 +15,7 @@ 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.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)
@@ -60,11 +60,8 @@ def test_punica_sgmv(
60
seed: int,
61
device: str,
62
):
63
- random.seed(seed)
63
torch.set_default_device(device)
65
- torch.random.manual_seed(seed)
66
- if torch.cuda.is_available():
67
- torch.cuda.manual_seed(seed)
64
+ seed_everything(seed)
65
66
seq_length = 128
67
(
@@ -153,11 +150,8 @@ def test_punica_bgmv(
150
from vllm.lora.ops.bgmv_expand import _bgmv_expand_kernel
151
from vllm.lora.ops.bgmv_shrink import _bgmv_shrink_kernel
152
156
- random.seed(seed)
153
torch.set_default_device(device)
158
- torch.random.manual_seed(seed)
159
- if torch.cuda.is_available():
160
- torch.cuda.manual_seed(seed)
154
+ seed_everything(seed)
155
156
seq_length = 1
157
(
@@ -244,11 +238,9 @@ def test_punica_expand_nslices(
238
):
239
from vllm.lora.ops.bgmv_expand_slice import _bgmv_expand_slice_kernel
240
247
- random.seed(seed)
241
torch.set_default_device(device)
249
- torch.random.manual_seed(seed)
250
- if torch.cuda.is_available():
251
- torch.cuda.manual_seed(seed)
242
+ seed_everything(seed)
243
+
244
seq_length = 128 if op_type == "sgmv" else 1
245
(
246
inputs_tensor,
tests/models/decoder_only/language/test_granite.py
+2
-7
@@ -2,23 +2,18 @@
2
3
Run `pytest tests/models/test_granite.py`.
4
"""
5
-import importlib.metadata
6
-
5
import pytest
6
+import transformers
7
8
from ...utils import check_logprobs_close
9
11
-TRANSFORMERS_VERSION = tuple(
12
- map(int,
13
- importlib.metadata.version("transformers").split(".")))
14
-
10
MODELS = [
11
"ibm/PowerLM-3b",
12
]
13
14
15
# GraniteForCausalLM will be in transformers >= 4.45
21
-@pytest.mark.skipif(TRANSFORMERS_VERSION < (4, 45),
16
+@pytest.mark.skipif(transformers.__version__ < "4.45",
17
reason="granite model test requires transformers >= 4.45")
18
@pytest.mark.parametrize("model", MODELS)
19
@pytest.mark.parametrize("dtype", ["bfloat16"])
tests/quantization/test_fp8.py
+1
-3
@@ -86,9 +86,7 @@ def test_load_fp16_model(vllm_runner, kv_cache_dtype: str, force_marlin: bool,
86
assert attn._k_scale == 1.0
87
assert attn._v_scale == 1.0
88
89
- capability = current_platform.get_device_capability()
90
- capability = capability[0] * 10 + capability[1]
91
- if capability >= 89 and not force_marlin:
89
+ if current_platform.has_device_capability(89) and not force_marlin:
90
# For GPUs with hardware support, we keep weights in fp8
91
assert fc1.weight.dtype == torch.float8_e4m3fn
92
else:
tests/quantization/utils.py
+5
-3
@@ -8,6 +8,8 @@ def is_quant_method_supported(quant_method: str) -> bool:
8
return False
9
10
capability = current_platform.get_device_capability()
11
- capability = capability[0] * 10 + capability[1]
12
- return (capability >=
13
- QUANTIZATION_METHODS[quant_method].get_min_capability())
11
+ assert capability is not None
12
+
13
+ min_capability = QUANTIZATION_METHODS[quant_method].get_min_capability()
14
+
15
+ return capability.to_int() >= min_capability
vllm/attention/backends/rocm_flash_attn.py
+2
-1
@@ -13,6 +13,7 @@ from vllm.attention.backends.utils import (CommonAttentionState,
13
from vllm.attention.ops.paged_attn import (PagedAttention,
14
PagedAttentionMetadata)
15
from vllm.logger import init_logger
16
+from vllm.platforms import current_platform
17
18
logger = init_logger(__name__)
19
@@ -299,7 +300,7 @@ class ROCmFlashAttentionImpl(AttentionImpl):
300
else:
301
# if not using triton, navi3x/navi21/navi10 do not use flash-attn
302
# either
302
- if torch.cuda.get_device_capability()[0] != 9:
303
+ if not current_platform.has_device_capability(90):
304
self.use_naive_attn = True
305
else:
306
try:
vllm/attention/ops/blocksparse_attention/interface.py
+2
-3
@@ -8,8 +8,7 @@ from vllm.utils import is_cpu, is_hip
8
from .utils import (dense_to_crow_col, get_head_sliding_step,
9
get_sparse_attn_mask)
10
11
-IS_COMPUTE_8_OR_ABOVE = (torch.cuda.is_available()
12
- and current_platform.get_device_capability()[0] >= 8)
11
+IS_COMPUTE_8_OR_ABOVE = current_platform.has_device_capability(80)
12
13
if IS_COMPUTE_8_OR_ABOVE:
14
from .blocksparse_attention_kernel import blocksparse_flash_attn_varlen_fwd
@@ -36,7 +35,7 @@ class LocalStridedBlockSparseAttn(torch.nn.Module):
35
use_spda = is_hip() or is_cpu() or not \
36
IS_COMPUTE_8_OR_ABOVE
37
device = device or (torch.cuda.current_device()
39
- if torch.cuda.is_available() else "cpu")
38
+ if current_platform.is_cuda_alike() else "cpu")
39
device = torch.device(device)
40
# NOTE: vllm CPU backend support BF16 instead of FP16.
41
dtype = dtype or (torch.bfloat16 if IS_COMPUTE_8_OR_ABOVE
vllm/attention/ops/prefix_prefill.py
+1
-2
@@ -709,8 +709,7 @@ if triton.__version__ >= "2.1.0":
709
alibi_slopes=None,
710
sliding_window=None):
711
712
- cap = current_platform.get_device_capability()
713
- BLOCK = 128 if cap[0] >= 8 else 64
712
+ BLOCK = 128 if current_platform.has_device_capability(80) else 64
713
NUM_WARPS = 8
714
715
# need to reduce num. blocks when using fp32
vllm/attention/selector.py
+2
-2
@@ -203,7 +203,7 @@ def which_attn_to_use(
203
selected_backend = (_Backend.ROCM_FLASH if selected_backend
204
== _Backend.FLASH_ATTN else selected_backend)
205
if selected_backend == _Backend.ROCM_FLASH:
206
- if current_platform.get_device_capability()[0] != 9:
206
+ if not current_platform.has_device_capability(90):
207
# not Instinct series GPUs.
208
logger.info("flash_attn is not supported on NAVI GPUs.")
209
else:
@@ -212,7 +212,7 @@ def which_attn_to_use(
212
213
# FlashAttn in NVIDIA GPUs.
214
if selected_backend == _Backend.FLASH_ATTN:
215
- if current_platform.get_device_capability()[0] < 8:
215
+ if not current_platform.has_device_capability(80):
216
# Volta and Turing NVIDIA GPUs.
217
logger.info(
218
"Cannot use FlashAttention-2 backend for Volta and Turing "
vllm/config.py
+6
-6
@@ -17,7 +17,7 @@ from vllm.transformers_utils.config import (ConfigFormat, get_config,
17
get_hf_image_processor_config,
18
get_hf_text_config)
19
from vllm.utils import (GiB_bytes, cuda_device_count_stateless, get_cpu_memory,
20
- is_cpu, is_hip, is_neuron, is_openvino, is_xpu,
20
+ is_hip, is_neuron, is_openvino, is_xpu,
21
print_warning_once)
22
23
if TYPE_CHECKING:
@@ -1035,20 +1035,20 @@ class DeviceConfig:
1035
def __init__(self, device: str = "auto") -> None:
1036
if device == "auto":
1037
# Automated device type detection
1038
- if is_neuron():
1038
+ if current_platform.is_cuda_alike():
1039
+ self.device_type = "cuda"
1040
+ elif is_neuron():
1041
self.device_type = "neuron"
1042
elif is_openvino():
1043
self.device_type = "openvino"
1044
elif current_platform.is_tpu():
1045
self.device_type = "tpu"
1044
- elif is_cpu():
1046
+ elif current_platform.is_cpu():
1047
self.device_type = "cpu"
1048
elif is_xpu():
1049
self.device_type = "xpu"
1050
else:
1049
- # We don't call torch.cuda.is_available() here to
1050
- # avoid initializing CUDA before workers are forked
1051
- self.device_type = "cuda"
1051
+ raise RuntimeError("Failed to infer device type")
1052
else:
1053
# Device type is assigned explicitly
1054
self.device_type = device
vllm/distributed/parallel_state.py
+2
-1
@@ -35,6 +35,7 @@ from torch.distributed import Backend, ProcessGroup
35
36
import vllm.envs as envs
37
from vllm.logger import init_logger
38
+from vllm.platforms import current_platform
39
40
41
@dataclass
@@ -191,7 +192,7 @@ class GroupCoordinator:
192
assert self.cpu_group is not None
193
assert self.device_group is not None
194
194
- if torch.cuda.is_available():
195
+ if current_platform.is_cuda_alike():
196
self.device = torch.device(f"cuda:{local_rank}")
197
else:
198
self.device = torch.device("cpu")
vllm/envs.py
+1
@@ -60,6 +60,7 @@ if TYPE_CHECKING:
60
VLLM_RPC_GET_DATA_TIMEOUT_MS: int = 5000
61
VLLM_PLUGINS: Optional[List[str]] = None
62
VLLM_TORCH_PROFILER_DIR: Optional[str] = None
63
+ VLLM_USE_TRITON_AWQ: bool = False
64
VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
65
66
vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py
+3
-3
@@ -116,10 +116,10 @@ class CompressedTensorsConfig(QuantizationConfig):
116
def _check_scheme_supported(self,
117
min_capability: int,
118
error: bool = True) -> bool:
119
- capability = current_platform.get_device_capability() # type: ignore
119
+ capability_tuple = current_platform.get_device_capability()
120
121
- if capability is not None:
122
- capability = capability[0] * 10 + capability[1]
121
+ if capability_tuple is not None:
122
+ capability = capability_tuple.to_int()
123
supported = capability >= min_capability
124
if error and not supported:
125
raise RuntimeError(
vllm/model_executor/layers/quantization/fbgemm_fp8.py
+1
-3
@@ -32,9 +32,7 @@ class FBGEMMFp8Config(QuantizationConfig):
32
33
# For GPUs that lack FP8 hardware support, we can leverage the Marlin
34
# kernel for fast weight-only FP8 quantization
35
- capability = current_platform.get_device_capability()
36
- capability = capability[0] * 10 + capability[1]
37
- self.use_marlin = capability < 89
35
+ self.use_marlin = not current_platform.has_device_capability(89)
36
37
@classmethod
38
def get_name(cls) -> str:
vllm/model_executor/layers/quantization/fp8.py
+2
-3
@@ -120,9 +120,8 @@ class Fp8LinearMethod(LinearMethodBase):
120
121
# For GPUs that lack FP8 hardware support, we can leverage the Marlin
122
# kernel for fast weight-only FP8 quantization
123
- capability = current_platform.get_device_capability()
124
- capability = capability[0] * 10 + capability[1]
125
- self.use_marlin = capability < 89 or envs.VLLM_TEST_FORCE_FP8_MARLIN
123
+ self.use_marlin = (not current_platform.has_device_capability(89)
124
+ or envs.VLLM_TEST_FORCE_FP8_MARLIN)
125
# Disable marlin for rocm
126
if is_hip():
127
self.use_marlin = False
vllm/model_executor/layers/quantization/utils/marlin_utils.py
+6
-4
@@ -29,8 +29,9 @@ def query_marlin_supported_quant_types(has_zp: bool,
29
device_capability: Optional[int] = None
30
):
31
if device_capability is None:
32
- major, minor = current_platform.get_device_capability()
33
- device_capability = major * 10 + minor
32
+ capability_tuple = current_platform.get_device_capability()
33
+ device_capability = (-1 if capability_tuple is None else
34
+ capability_tuple.to_int())
35
36
if device_capability < 80:
37
return []
@@ -52,8 +53,9 @@ def _check_marlin_supported(
53
device_capability: Optional[int] = None) -> Tuple[bool, Optional[str]]:
54
55
if device_capability is None:
55
- major, minor = current_platform.get_device_capability()
56
- device_capability = major * 10 + minor
56
+ capability_tuple = current_platform.get_device_capability()
57
+ device_capability = (-1 if capability_tuple is None else
58
+ capability_tuple.to_int())
59
60
supported_types = query_marlin_supported_quant_types(
61
has_zp, device_capability)
vllm/model_executor/layers/quantization/utils/marlin_utils_fp8.py
+1
-2
@@ -10,8 +10,7 @@ from .marlin_utils import marlin_make_workspace, marlin_permute_scales
10
11
12
def is_fp8_marlin_supported():
13
- capability = current_platform.get_device_capability()
14
- return capability[0] >= 8
13
+ return current_platform.has_device_capability(80)
14
15
16
def apply_fp8_marlin_linear(
vllm/model_executor/layers/quantization/utils/w8a8_utils.py
+3
-2
@@ -17,8 +17,9 @@ def cutlass_fp8_supported() -> bool:
17
# cutlass is not supported on Rocm
18
if is_hip():
19
return False
20
- capability = current_platform.get_device_capability()
21
- capability = capability[0] * 10 + capability[1]
20
+
21
+ capability_tuple = current_platform.get_device_capability()
22
+ capability = -1 if capability_tuple is None else capability_tuple.to_int()
23
24
return ops.cutlass_scaled_mm_supports_fp8(capability)
25
vllm/model_executor/model_loader/loader.py
+3
-3
@@ -97,10 +97,10 @@ def _get_quantization_config(
97
"""Get the quantization config."""
98
if model_config.quantization is not None:
99
quant_config = get_quant_config(model_config, load_config)
100
- capability = current_platform.get_device_capability() # type: ignore
100
+ capability_tuple = current_platform.get_device_capability()
101
102
- if capability is not None:
103
- capability = capability[0] * 10 + capability[1]
102
+ if capability_tuple is not None:
103
+ capability = capability_tuple.to_int()
104
if capability < quant_config.get_min_capability():
105
raise ValueError(
106
f"The quantization method {model_config.quantization} "
vllm/model_executor/models/qwen2_vl.py
+1
-1
@@ -207,7 +207,7 @@ class Qwen2VisionAttention(nn.Module):
207
selected_backend = backend_name_to_enum(backend_by_env_var)
208
if selected_backend is None:
209
# For Volta and Turing GPUs, use xformers instead.
210
- device_available = current_platform.get_device_capability()[0] >= 8
210
+ device_available = current_platform.has_device_capability(80)
211
if device_available:
212
from transformers.utils import is_flash_attn_2_available
213
vllm/model_executor/utils.py
+3
-7
@@ -1,17 +1,13 @@
1
"""Utils for model executor."""
2
-import random
2
from typing import Any, Dict, Optional
3
5
-import numpy as np
4
import torch
5
6
+from vllm.utils import seed_everything
7
+
8
9
def set_random_seed(seed: int) -> None:
10
- random.seed(seed)
11
- np.random.seed(seed)
12
- torch.manual_seed(seed)
13
- if torch.cuda.is_available():
14
- torch.cuda.manual_seed_all(seed)
10
+ seed_everything(seed)
11
12
13
def set_weight_attrs(
vllm/platforms/cpu.py
+4
-4
@@ -6,10 +6,10 @@ from .interface import Platform, PlatformEnum
6
class CpuPlatform(Platform):
7
_enum = PlatformEnum.CPU
8
9
- @staticmethod
10
- def get_device_name(device_id: int = 0) -> str:
9
+ @classmethod
10
+ def get_device_name(cls, device_id: int = 0) -> str:
11
return "cpu"
12
13
- @staticmethod
14
- def inference_mode():
13
+ @classmethod
14
+ def inference_mode(cls):
15
return torch.no_grad()
vllm/platforms/cuda.py
+9
-8
@@ -11,7 +11,7 @@ from typing_extensions import ParamSpec
11
12
from vllm.logger import init_logger
13
14
-from .interface import Platform, PlatformEnum
14
+from .interface import DeviceCapability, Platform, PlatformEnum
15
16
logger = init_logger(__name__)
17
@@ -96,19 +96,20 @@ def device_id_to_physical_device_id(device_id: int) -> int:
96
class CudaPlatform(Platform):
97
_enum = PlatformEnum.CUDA
98
99
- @staticmethod
100
- def get_device_capability(device_id: int = 0) -> Tuple[int, int]:
99
+ @classmethod
100
+ def get_device_capability(cls, device_id: int = 0) -> DeviceCapability:
101
physical_device_id = device_id_to_physical_device_id(device_id)
102
- return get_physical_device_capability(physical_device_id)
102
+ major, minor = get_physical_device_capability(physical_device_id)
103
+ return DeviceCapability(major=major, minor=minor)
104
104
- @staticmethod
105
- def get_device_name(device_id: int = 0) -> str:
105
+ @classmethod
106
+ def get_device_name(cls, device_id: int = 0) -> str:
107
physical_device_id = device_id_to_physical_device_id(device_id)
108
return get_physical_device_name(physical_device_id)
109
109
- @staticmethod
110
+ @classmethod
111
@with_nvml_context
111
- def is_full_nvlink(physical_device_ids: List[int]) -> bool:
112
+ def is_full_nvlink(cls, physical_device_ids: List[int]) -> bool:
113
"""
114
query if the set of gpus are fully connected by nvlink (1 hop)
115
"""
vllm/platforms/interface.py
+55
-7
@@ -1,5 +1,5 @@
1
import enum
2
-from typing import Optional, Tuple
2
+from typing import NamedTuple, Optional, Tuple, Union
3
4
import torch
5
@@ -12,6 +12,23 @@ class PlatformEnum(enum.Enum):
12
UNSPECIFIED = enum.auto()
13
14
15
+class DeviceCapability(NamedTuple):
16
+ major: int
17
+ minor: int
18
+
19
+ def as_version_str(self) -> str:
20
+ return f"{self.major}.{self.minor}"
21
+
22
+ def to_int(self) -> int:
23
+ """
24
+ Express device capability as an integer ``<major><minor>``.
25
+
26
+ It is assumed that the minor version is always a single digit.
27
+ """
28
+ assert 0 <= self.minor < 10
29
+ return self.major * 10 + self.minor
30
+
31
+
32
class Platform:
33
_enum: PlatformEnum
34
@@ -27,16 +44,47 @@ class Platform:
44
def is_cpu(self) -> bool:
45
return self._enum == PlatformEnum.CPU
46
30
- @staticmethod
31
- def get_device_capability(device_id: int = 0) -> Optional[Tuple[int, int]]:
47
+ def is_cuda_alike(self) -> bool:
48
+ """Stateless version of :func:`torch.cuda.is_available`."""
49
+ return self._enum in (PlatformEnum.CUDA, PlatformEnum.ROCM)
50
+
51
+ @classmethod
52
+ def get_device_capability(
53
+ cls,
54
+ device_id: int = 0,
55
+ ) -> Optional[DeviceCapability]:
56
+ """Stateless version of :func:`torch.cuda.get_device_capability`."""
57
return None
58
34
- @staticmethod
35
- def get_device_name(device_id: int = 0) -> str:
59
+ @classmethod
60
+ def has_device_capability(
61
+ cls,
62
+ capability: Union[Tuple[int, int], int],
63
+ device_id: int = 0,
64
+ ) -> bool:
65
+ """
66
+ Test whether this platform is compatible with a device capability.
67
+
68
+ The ``capability`` argument can either be:
69
+
70
+ - A tuple ``(major, minor)``.
71
+ - An integer ``<major><minor>``. (See :meth:`DeviceCapability.to_int`)
72
+ """
73
+ current_capability = cls.get_device_capability(device_id=device_id)
74
+ if current_capability is None:
75
+ return False
76
+
77
+ if isinstance(capability, tuple):
78
+ return current_capability >= capability
79
+
80
+ return current_capability.to_int() >= capability
81
+
82
+ @classmethod
83
+ def get_device_name(cls, device_id: int = 0) -> str:
84
raise NotImplementedError
85
38
- @staticmethod
39
- def inference_mode():
86
+ @classmethod
87
+ def inference_mode(cls):
88
"""A device-specific wrapper of `torch.inference_mode`.
89
90
This wrapper is recommended because some hardware backends such as TPU
vllm/platforms/rocm.py
+7
-7
@@ -1,12 +1,11 @@
1
import os
2
from functools import lru_cache
3
-from typing import Tuple
3
4
import torch
5
6
from vllm.logger import init_logger
7
9
-from .interface import Platform, PlatformEnum
8
+from .interface import DeviceCapability, Platform, PlatformEnum
9
10
logger = init_logger(__name__)
11
@@ -20,12 +19,13 @@ if os.environ.get("VLLM_WORKER_MULTIPROC_METHOD", None) in ["fork", None]:
19
class RocmPlatform(Platform):
20
_enum = PlatformEnum.ROCM
21
23
- @staticmethod
22
+ @classmethod
23
@lru_cache(maxsize=8)
25
- def get_device_capability(device_id: int = 0) -> Tuple[int, int]:
26
- return torch.cuda.get_device_capability(device_id)
24
+ def get_device_capability(cls, device_id: int = 0) -> DeviceCapability:
25
+ major, minor = torch.cuda.get_device_capability(device_id)
26
+ return DeviceCapability(major=major, minor=minor)
27
28
- @staticmethod
28
+ @classmethod
29
@lru_cache(maxsize=8)
30
- def get_device_name(device_id: int = 0) -> str:
30
+ def get_device_name(cls, device_id: int = 0) -> str:
31
return torch.cuda.get_device_name(device_id)
vllm/platforms/tpu.py
+6
-2
@@ -6,6 +6,10 @@ from .interface import Platform, PlatformEnum
6
class TpuPlatform(Platform):
7
_enum = PlatformEnum.TPU
8
9
- @staticmethod
10
- def inference_mode():
9
+ @classmethod
10
+ def get_device_name(cls, device_id: int = 0) -> str:
11
+ raise NotImplementedError
12
+
13
+ @classmethod
14
+ def inference_mode(cls):
15
return torch.no_grad()
vllm/prompt_adapter/utils.py
+3
-1
@@ -8,13 +8,15 @@ from huggingface_hub import file_exists, hf_hub_download
8
from huggingface_hub.utils import EntryNotFoundError
9
from safetensors.torch import load_file as safe_load_file
10
11
+from vllm.platforms import current_platform
12
+
13
WEIGHTS_NAME = "adapter_model.bin"
14
SAFETENSORS_WEIGHTS_NAME = "adapter_model.safetensors"
15
16
17
# Get current device name based on available devices
18
def infer_device() -> str:
17
- if torch.cuda.is_available():
19
+ if current_platform.is_cuda_alike():
20
return "cuda"
21
return "cpu"
22
vllm/usage/usage_lib.py
+2
-1
@@ -17,6 +17,7 @@ import torch
17
18
import vllm.envs as envs
19
from vllm.connections import global_http_connection
20
+from vllm.platforms import current_platform
21
from vllm.version import __version__ as VLLM_VERSION
22
23
_config_home = envs.VLLM_CONFIG_ROOT
@@ -151,7 +152,7 @@ class UsageMessage:
152
usage_context: UsageContext,
153
extra_kvs: Dict[str, Any]) -> None:
154
# Platform information
154
- if torch.cuda.is_available():
155
+ if current_platform.is_cuda_alike():
156
device_property = torch.cuda.get_device_properties(0)
157
self.gpu_count = torch.cuda.device_count()
158
self.gpu_type = device_property.name
vllm/utils.py
+21
-7
@@ -5,6 +5,7 @@ import datetime
5
import enum
6
import gc
7
import os
8
+import random
9
import socket
10
import subprocess
11
import sys
@@ -32,6 +33,7 @@ from typing_extensions import ParamSpec, TypeIs, assert_never
33
34
import vllm.envs as envs
35
from vllm.logger import enable_trace_function_call, init_logger
36
+from vllm.platforms import current_platform
37
38
logger = init_logger(__name__)
39
@@ -373,6 +375,22 @@ def get_cpu_memory() -> int:
375
return psutil.virtual_memory().total
376
377
378
+def seed_everything(seed: int) -> None:
379
+ """
380
+ Set the seed of each random module.
381
+
382
+ Loosely based on: https://github.com/Lightning-AI/pytorch-lightning/blob/2.4.0/src/lightning/fabric/utilities/seed.py#L20
383
+ """
384
+ random.seed(seed)
385
+ np.random.seed(seed)
386
+
387
+ if current_platform.is_cuda_alike():
388
+ torch.cuda.manual_seed_all(seed)
389
+
390
+ if is_xpu():
391
+ torch.xpu.manual_seed_all(seed)
392
+
393
+
394
def random_uuid() -> str:
395
return str(uuid.uuid4().hex)
396
@@ -634,9 +652,7 @@ def create_kv_caches_with_random_flash(
652
seed: int = 0,
653
device: Optional[str] = "cuda",
654
) -> Tuple[List[torch.Tensor], List[torch.Tensor]]:
637
- torch.random.manual_seed(seed)
638
- if torch.cuda.is_available():
639
- torch.cuda.manual_seed(seed)
655
+ seed_everything(seed)
656
657
torch_dtype = get_kv_cache_torch_dtype(cache_dtype, model_dtype)
658
key_value_cache_shape = (num_blocks, 2, block_size, num_heads, head_size)
@@ -678,9 +694,7 @@ def create_kv_caches_with_random(
694
f"Does not support key cache of type fp8 with head_size {head_size}"
695
)
696
681
- torch.random.manual_seed(seed)
682
- if torch.cuda.is_available():
683
- torch.cuda.manual_seed(seed)
697
+ seed_everything(seed)
698
699
torch_dtype = get_kv_cache_torch_dtype(cache_dtype, model_dtype)
700
@@ -750,7 +764,7 @@ class CudaMemoryProfiler:
764
765
def current_memory_usage(self) -> float:
766
# Return the memory usage in bytes.
753
- if torch.cuda.is_available():
767
+ if current_platform.is_cuda_alike():
768
torch.cuda.reset_peak_memory_stats(self.device)
769
mem = torch.cuda.max_memory_allocated(self.device)
770
elif is_xpu():
vllm/worker/worker.py
+11
-5
@@ -454,14 +454,20 @@ def init_worker_distributed_environment(
454
455
def _check_if_gpu_supports_dtype(torch_dtype: torch.dtype):
456
# Check if the GPU supports the dtype.
457
- if torch_dtype == torch.bfloat16:
458
- compute_capability = current_platform.get_device_capability()
459
- if compute_capability[0] < 8:
457
+ if torch_dtype == torch.bfloat16: # noqa: SIM102
458
+ if not current_platform.has_device_capability(80):
459
+ capability = current_platform.get_device_capability()
460
gpu_name = current_platform.get_device_name()
461
+
462
+ if capability is None:
463
+ compute_str = "does not have a compute capability"
464
+ else:
465
+ version_str = capability.as_version_str()
466
+ compute_str = f"has compute capability {version_str}"
467
+
468
raise ValueError(
469
"Bfloat16 is only supported on GPUs with compute capability "
463
- f"of at least 8.0. Your {gpu_name} GPU has compute capability "
464
- f"{compute_capability[0]}.{compute_capability[1]}. "
470
+ f"of at least 8.0. Your {gpu_name} GPU {compute_str}. "
471
"You can use float16 instead by explicitly setting the"
472
"`dtype` flag in CLI, for example: --dtype=half.")
473