db318326a5 · tduncan/Nvidia-Optimization-Stack-VLLM

benchmarks/kernels/benchmark_activation.py

+2 -3

@@ -8,10 +8,9 @@ import torch

8

9 import vllm.model_executor.layers.activation # noqa F401

10 from vllm.model_executor.custom_op import CustomOp

11 -from vllm.platforms import current_platform

11 from vllm.triton_utils import triton

12 from vllm.utils.argparse_utils import FlexibleArgumentParser

14 -from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE

13 +from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE, set_random_seed

14

15 batch_size_range = [1, 16, 128]

16 seq_len_range = [1, 16, 64, 1024, 4096]

@@ -30,7 +29,7 @@ def benchmark_activation(

29 device = "cuda"

30 num_tokens = batch_size * seq_len

31 dim = intermediate_size

33 - current_platform.seed_everything(42)

32 + set_random_seed(42)

33 torch.set_default_device(device)

34

35 if func_name == "gelu_and_mul":

benchmarks/kernels/benchmark_layernorm.py

+2 -3

@@ -6,9 +6,8 @@ import time

6 import torch

7

8 from vllm.model_executor.layers.layernorm import RMSNorm

9 -from vllm.platforms import current_platform

9 from vllm.utils.argparse_utils import FlexibleArgumentParser

11 -from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE

10 +from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE, set_random_seed

11

12

13 @torch.inference_mode()

@@ -22,7 +21,7 @@ def main(

21 num_warmup_iters: int = 5,

22 num_iters: int = 100,

23 ) -> None:

25 - current_platform.seed_everything(seed)

24 + set_random_seed(seed)

25 torch.set_default_device("cuda")

26

27 layer = RMSNorm(hidden_size).to(dtype=dtype)

benchmarks/kernels/benchmark_moe.py

+3 -2

@@ -24,6 +24,7 @@ from vllm.platforms import current_platform

24 from vllm.transformers_utils.config import get_config

25 from vllm.triton_utils import triton

26 from vllm.utils.argparse_utils import FlexibleArgumentParser

27 +from vllm.utils.torch_utils import set_random_seed

28

29 FP8_DTYPE = current_platform.fp8_dtype()

30

@@ -431,7 +432,7 @@ def merge_unique_dicts(list1, list2):

432 class BenchmarkWorker:

433 def __init__(self, seed: int) -> None:

434 torch.set_default_device("cuda")

434 - current_platform.seed_everything(seed)

435 + set_random_seed(seed)

436 self.seed = seed

437 # Get the device ID to allocate tensors and kernels

438 # on the respective GPU. This is required for Ray to work

@@ -451,7 +452,7 @@ class BenchmarkWorker:

452 block_quant_shape: list[int] = None,

453 use_deep_gemm: bool = False,

454 ) -> tuple[dict[str, int], float]:

454 - current_platform.seed_everything(self.seed)

455 + set_random_seed(self.seed)

456 dtype_str = _get_config_dtype_str(

457 dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8

458 )

benchmarks/kernels/benchmark_moe_permute_unpermute.py

+3 -2

@@ -18,6 +18,7 @@ from vllm.model_executor.layers.fused_moe.moe_permute_unpermute import (

18 from vllm.model_executor.layers.fused_moe.utils import _fp8_quantize

19 from vllm.platforms import current_platform

20 from vllm.utils.argparse_utils import FlexibleArgumentParser

21 +from vllm.utils.torch_utils import set_random_seed

22

23 FP8_DTYPE = current_platform.fp8_dtype()

24

@@ -261,7 +262,7 @@ def benchmark_unpermute(

262 class BenchmarkWorker:

263 def __init__(self, seed: int) -> None:

264 torch.set_default_device("cuda")

264 - current_platform.seed_everything(seed)

265 + set_random_seed(seed)

266 self.seed = seed

267 # Get the device ID to allocate tensors and kernels

268 # on the respective GPU. This is required for Ray to work

@@ -279,7 +280,7 @@ class BenchmarkWorker:

280 use_int8_w8a16: bool,

281 use_customized_permute: bool = False,

282 ) -> tuple[dict[str, int], float]:

282 - current_platform.seed_everything(self.seed)

283 + set_random_seed(self.seed)

284

285 permute_time = benchmark_permute(

286 num_tokens,

benchmarks/kernels/benchmark_mrope.py

+2 -2

@@ -37,9 +37,9 @@ import numpy as np

37 import torch

38

39 from vllm.model_executor.layers.rotary_embedding import get_rope

40 -from vllm.platforms import current_platform

40 from vllm.transformers_utils.config import get_config

41 from vllm.utils.argparse_utils import FlexibleArgumentParser

42 +from vllm.utils.torch_utils import set_random_seed

43

44 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

45

@@ -94,7 +94,7 @@ def benchmark_mrope(

94 benchmark_iter: int = 100,

95 csv_writer=None,

96 ):

97 - current_platform.seed_everything(seed)

97 + set_random_seed(seed)

98 torch.set_default_device(device)

99 # the parameters to compute the q k v size based on tp_size

100 mrope_helper_class = get_rope(

benchmarks/kernels/benchmark_paged_attention.py

+2 -1

@@ -13,6 +13,7 @@ from vllm.utils.argparse_utils import FlexibleArgumentParser

13 from vllm.utils.torch_utils import (

14 STR_DTYPE_TO_TORCH_DTYPE,

15 create_kv_caches_with_random,

16 + set_random_seed,

17 )

18

19 logger = init_logger(__name__)

@@ -38,7 +39,7 @@ def main(

39 device: str = "cuda",

40 kv_cache_dtype: str | None = None,

41 ) -> None:

41 - current_platform.seed_everything(seed)

42 + set_random_seed(seed)

43

44 scale = float(1.0 / (head_size**0.5))

45 query = torch.empty(

benchmarks/kernels/benchmark_quant.py

+2 -3

@@ -6,9 +6,8 @@ import time

6 import torch

7

8 from vllm import _custom_ops as ops

9 -from vllm.platforms import current_platform

9 from vllm.utils.argparse_utils import FlexibleArgumentParser

11 -from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE

10 +from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE, set_random_seed

11

12

13 @torch.inference_mode()

@@ -23,7 +22,7 @@ def main(

22 num_warmup_iters: int = 5,

23 num_iters: int = 100,

24 ) -> None:

26 - current_platform.seed_everything(seed)

25 + set_random_seed(seed)

26 torch.set_default_device("cuda")

27

28 x = torch.randn(num_tokens, hidden_size, dtype=dtype)

benchmarks/kernels/benchmark_reshape_and_cache.py

+2 -2

@@ -8,11 +8,11 @@ from tabulate import tabulate

8

9 from vllm import _custom_ops as ops

10 from vllm.logger import init_logger

11 -from vllm.platforms import current_platform

11 from vllm.utils.argparse_utils import FlexibleArgumentParser

12 from vllm.utils.torch_utils import (

13 STR_DTYPE_TO_TORCH_DTYPE,

14 create_kv_caches_with_random,

15 + set_random_seed,

16 )

17

18 logger = init_logger(__name__)

@@ -36,7 +36,7 @@ def run_benchmark(

36 if kv_cache_dtype == "fp8" and head_size % 16:

37 raise ValueError("fp8 kv-cache requires head_size to be a multiple of 16.")

38

39 - current_platform.seed_everything(42)

39 + set_random_seed(42)

40 torch.set_default_device(device)

41

42 # create random key / value tensors [T, H, D].

benchmarks/kernels/benchmark_reshape_and_cache_flash.py

+2 -2

@@ -11,11 +11,11 @@ from vllm.attention.ops.triton_reshape_and_cache_flash import (

11 triton_reshape_and_cache_flash,

12 )

13 from vllm.logger import init_logger

14 -from vllm.platforms import current_platform

14 from vllm.utils.argparse_utils import FlexibleArgumentParser

15 from vllm.utils.torch_utils import (

16 STR_DTYPE_TO_TORCH_DTYPE,

17 create_kv_caches_with_random_flash,

18 + set_random_seed,

19 )

20

21 logger = init_logger(__name__)

@@ -49,7 +49,7 @@ def run_benchmark(

49 if implementation == "triton" and kv_cache_layout == "HND":

50 return float("nan") # Triton does not support HND layout yet.

51

52 - current_platform.seed_everything(42)

52 + set_random_seed(42)

53 torch.set_default_device(device)

54

55 # create random key / value tensors [T, H, D].

benchmarks/kernels/benchmark_silu_mul_fp8_quant.py

+2 -2

@@ -23,9 +23,9 @@ import torch

23 from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (

24 persistent_masked_m_silu_mul_quant,

25 )

26 -from vllm.platforms import current_platform

26 from vllm.triton_utils import tl, triton

27 from vllm.utils.deep_gemm import is_deep_gemm_e8m0_used

28 +from vllm.utils.torch_utils import set_random_seed

29

30

31 @triton.jit

@@ -207,7 +207,7 @@ def benchmark(

207 ):

208 def generate_data(seed_offset=0):

209 """Generate input data with given seed offset"""

210 - current_platform.seed_everything(42 + seed_offset)

210 + set_random_seed(42 + seed_offset)

211 y = torch.rand((E, T, 2 * H), dtype=torch.bfloat16, device="cuda").contiguous()

212

213 if gen_strategy == "random_imbalanced":

docs/design/plugin_system.md

+1 -1

@@ -154,4 +154,4 @@ The interface for the model/module may change during vLLM's development. If you

154 !!! warning "Deprecations"

155 - `use_v1` parameter in `Platform.get_attn_backend_cls` is deprecated. It has been removed in v0.13.0.

156 - `_Backend` in `vllm.attention` is deprecated. It has been removed in v0.13.0. Please use `vllm.attention.backends.registry.register_backend` to add new attention backend to `AttentionBackendEnum` instead.

157 - - `seed_everything` platform interface is deprecated. It will be removed in v0.14.0 or later. Please use `vllm.utils.torch_utils.set_random_seed` instead.

157 + - `seed_everything` platform interface is deprecated. It will be removed in v0.15.0 or later. Please use `vllm.utils.torch_utils.set_random_seed` instead.

vllm/platforms/interface.py

+5 -4

@@ -11,6 +11,7 @@ from typing import TYPE_CHECKING, Any, NamedTuple, Optional

11

12 import numpy as np

13 import torch

14 +from typing_extensions import deprecated

15

16 from vllm.attention.backends.registry import AttentionBackendEnum

17 from vllm.logger import init_logger

@@ -365,6 +366,10 @@ class Platform:

366 return torch.inference_mode(mode=True)

367

368 @classmethod

369 + @deprecated(

370 + "`seed_everything` is deprecated. It will be removed in v0.15.0 or later. "

371 + "Please use `vllm.utils.torch_utils.set_random_seed` instead."

372 + )

373 def seed_everything(cls, seed: int | None = None) -> None:

374 """

375 Set the seed of each random module.

@@ -372,10 +377,6 @@ class Platform:

377

378 Loosely based on: https://github.com/Lightning-AI/pytorch-lightning/blob/2.4.0/src/lightning/fabric/utilities/seed.py#L20

379 """

375 - logger.info_once(

376 - "`seed_everything` is deprecated. It will be removed in v0.14.0 or later. "

377 - "Please use `vllm.utils.torch_utils.set_random_seed` instead."

378 - )

380 if seed is not None:

381 random.seed(seed)

382 np.random.seed(seed)

[Misc] Use `deprecated` for `seed_everything` (#31780)