[Hardware] using current_platform.seed_everything (#9785)

Signed-off-by: wangshuai09 <391746016@qq.com>
This commit is contained in:
wangshuai09
2024-10-29 14:47:44 +00:00
committed by GitHub
parent 09500f7dde
commit 622b7ab955
27 changed files with 104 additions and 105 deletions
+2 -2
View File
@@ -39,7 +39,7 @@ from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
ParallelLMHead, VocabParallelEmbedding, get_masked_input_and_mask)
from vllm.model_executor.utils import set_random_seed
from vllm.utils import seed_everything
from vllm.platforms import current_platform
from .utils import DummyLoRAManager
@@ -923,7 +923,7 @@ def test_rotary_embedding_long_context(dist_init, num_loras, device,
seq_len) -> None:
dtype = torch.float16
seed = 0
seed_everything(seed)
current_platform.seed_everything(seed)
torch.set_default_device(device)
punica_wrapper = PunicaWrapper(8192, 256, device)
max_loras = 8
+5 -5
View File
@@ -1,5 +1,5 @@
"""
This script is mainly used to tests various hidden_sizes. We have collected the
This script is mainly used to tests various hidden_sizes. We have collected the
hidden_sizes included in the LoRA models currently supported by vLLM. It tests
whether the corresponding Triton kernel can run normally when tensor parallelism
is set to [1, 2, 4, 8, 16, 32, 64].
@@ -15,8 +15,8 @@ from vllm.lora.ops.bgmv_shrink import bgmv_shrink
from vllm.lora.ops.sgmv_expand import sgmv_expand
from vllm.lora.ops.sgmv_expand_slice import sgmv_expand_slice
from vllm.lora.ops.sgmv_shrink import sgmv_shrink
from vllm.platforms import current_platform
from vllm.triton_utils.libentry import LibEntry
from vllm.utils import seed_everything
from .utils import (generate_data, generate_data_for_expand_nslices,
ref_torch_groupgemm)
@@ -146,7 +146,7 @@ def test_punica_sgmv(
device: str,
):
torch.set_default_device(device)
seed_everything(seed)
current_platform.seed_everything(seed)
seq_length = 128
(
@@ -239,7 +239,7 @@ def test_punica_bgmv(
from vllm.lora.ops.bgmv_shrink import _bgmv_shrink_kernel
torch.set_default_device(device)
seed_everything(seed)
current_platform.seed_everything(seed)
seq_length = 1
(
@@ -327,7 +327,7 @@ def test_punica_expand_nslices(
from vllm.lora.ops.bgmv_expand_slice import _bgmv_expand_slice_kernel
torch.set_default_device(device)
seed_everything(seed)
current_platform.seed_everything(seed)
seq_length = 128 if op_type == "sgmv" else 1
(
+6 -6
View File
@@ -1,6 +1,6 @@
"""
This script is mainly used to test whether trtion kernels can run normally
under different conditions, including various batches, numbers of LoRA , and
This script is mainly used to test whether trtion kernels can run normally
under different conditions, including various batches, numbers of LoRA , and
maximum ranks.
"""
from unittest.mock import patch
@@ -14,8 +14,8 @@ from vllm.lora.ops.bgmv_shrink import bgmv_shrink
from vllm.lora.ops.sgmv_expand import sgmv_expand
from vllm.lora.ops.sgmv_expand_slice import sgmv_expand_slice
from vllm.lora.ops.sgmv_shrink import sgmv_shrink
from vllm.platforms import current_platform
from vllm.triton_utils.libentry import LibEntry
from vllm.utils import seed_everything
from .utils import (generate_data, generate_data_for_expand_nslices,
ref_torch_groupgemm)
@@ -61,7 +61,7 @@ def test_punica_sgmv(
device: str,
):
torch.set_default_device(device)
seed_everything(seed)
current_platform.seed_everything(seed)
seq_length = 128
(
@@ -154,7 +154,7 @@ def test_punica_bgmv(
from vllm.lora.ops.bgmv_shrink import _bgmv_shrink_kernel
torch.set_default_device(device)
seed_everything(seed)
current_platform.seed_everything(seed)
seq_length = 1
(
@@ -242,7 +242,7 @@ def test_punica_expand_nslices(
from vllm.lora.ops.bgmv_expand_slice import _bgmv_expand_slice_kernel
torch.set_default_device(device)
seed_everything(seed)
current_platform.seed_everything(seed)
seq_length = 128 if op_type == "sgmv" else 1
(