[Platform] Deprecate seed_everything (#31659)
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
@@ -154,3 +154,4 @@ The interface for the model/module may change during vLLM's development. If you
|
|||||||
!!! warning "Deprecations"
|
!!! warning "Deprecations"
|
||||||
- `use_v1` parameter in `Platform.get_attn_backend_cls` is deprecated. It has been removed in v0.13.0.
|
- `use_v1` parameter in `Platform.get_attn_backend_cls` is deprecated. It has been removed in v0.13.0.
|
||||||
- `_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.
|
- `_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.
|
||||||
|
- `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.
|
||||||
|
|||||||
@@ -26,6 +26,7 @@ from vllm.distributed.parallel_state import (
|
|||||||
)
|
)
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.utils.system_utils import update_environment_variables
|
from vllm.utils.system_utils import update_environment_variables
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
from ...models.registry import HF_EXAMPLE_MODELS
|
from ...models.registry import HF_EXAMPLE_MODELS
|
||||||
from ...utils import (
|
from ...utils import (
|
||||||
@@ -301,7 +302,7 @@ def async_tp_pass_on_test_model(
|
|||||||
dtype: torch.dtype,
|
dtype: torch.dtype,
|
||||||
dynamic: bool,
|
dynamic: bool,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
|
|
||||||
device = torch.device(f"cuda:{local_rank}")
|
device = torch.device(f"cuda:{local_rank}")
|
||||||
torch.cuda.set_device(device)
|
torch.cuda.set_device(device)
|
||||||
|
|||||||
@@ -32,6 +32,7 @@ from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
|
|||||||
)
|
)
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.utils.system_utils import update_environment_variables
|
from vllm.utils.system_utils import update_environment_variables
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
from ...utils import has_module_attribute, multi_gpu_test
|
from ...utils import has_module_attribute, multi_gpu_test
|
||||||
from ..backend import TestBackend
|
from ..backend import TestBackend
|
||||||
@@ -263,7 +264,7 @@ def all_reduce_fusion_pass_on_test_model(
|
|||||||
enable_rms_norm_custom_op,
|
enable_rms_norm_custom_op,
|
||||||
enable_quant_fp8_custom_op,
|
enable_quant_fp8_custom_op,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
|
|
||||||
device = torch.device(f"cuda:{local_rank}")
|
device = torch.device(f"cuda:{local_rank}")
|
||||||
torch.cuda.set_device(device)
|
torch.cuda.set_device(device)
|
||||||
|
|||||||
@@ -31,6 +31,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
|
|||||||
from vllm.model_executor.layers.quantization.utils.w8a8_utils import Fp8LinearOp
|
from vllm.model_executor.layers.quantization.utils.w8a8_utils import Fp8LinearOp
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.utils.system_utils import update_environment_variables
|
from vllm.utils.system_utils import update_environment_variables
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
from ...utils import multi_gpu_test
|
from ...utils import multi_gpu_test
|
||||||
from ..backend import TestBackend
|
from ..backend import TestBackend
|
||||||
@@ -232,7 +233,7 @@ def sequence_parallelism_pass_on_test_model(
|
|||||||
fuse_norm_quant: bool,
|
fuse_norm_quant: bool,
|
||||||
dynamic: bool,
|
dynamic: bool,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
|
|
||||||
device = torch.device(f"cuda:{local_rank}")
|
device = torch.device(f"cuda:{local_rank}")
|
||||||
torch.cuda.set_device(device)
|
torch.cuda.set_device(device)
|
||||||
|
|||||||
@@ -8,6 +8,7 @@ import torch
|
|||||||
import vllm.v1.attention.backends.rocm_aiter_fa # noqa: F401
|
import vllm.v1.attention.backends.rocm_aiter_fa # noqa: F401
|
||||||
from vllm.attention.utils.fa_utils import is_flash_attn_varlen_func_available
|
from vllm.attention.utils.fa_utils import is_flash_attn_varlen_func_available
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
NUM_HEADS = [(4, 4), (8, 2)]
|
NUM_HEADS = [(4, 4), (8, 2)]
|
||||||
HEAD_SIZES = [128, 256]
|
HEAD_SIZES = [128, 256]
|
||||||
@@ -104,7 +105,7 @@ def test_varlen_with_paged_kv(
|
|||||||
if not is_flash_attn_varlen_func_available():
|
if not is_flash_attn_varlen_func_available():
|
||||||
pytest.skip("flash_attn_varlen_func required to run this test.")
|
pytest.skip("flash_attn_varlen_func required to run this test.")
|
||||||
torch.set_default_device("cuda")
|
torch.set_default_device("cuda")
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
num_seqs = len(seq_lens)
|
num_seqs = len(seq_lens)
|
||||||
query_lens = [x[0] for x in seq_lens]
|
query_lens = [x[0] for x in seq_lens]
|
||||||
kv_lens = [x[1] for x in seq_lens]
|
kv_lens = [x[1] for x in seq_lens]
|
||||||
|
|||||||
@@ -13,6 +13,7 @@ from vllm.attention.layer import Attention
|
|||||||
from vllm.attention.layers.mm_encoder_attention import MMEncoderAttention
|
from vllm.attention.layers.mm_encoder_attention import MMEncoderAttention
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.utils.mem_utils import get_max_shared_memory_bytes
|
from vllm.utils.mem_utils import get_max_shared_memory_bytes
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
|
FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
|
||||||
# This will change depending on the compute capability.
|
# This will change depending on the compute capability.
|
||||||
@@ -150,7 +151,7 @@ def test_paged_attention(
|
|||||||
|
|
||||||
global PARTITION_SIZE
|
global PARTITION_SIZE
|
||||||
|
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
scale = float(1.0 / (head_size**0.5))
|
scale = float(1.0 / (head_size**0.5))
|
||||||
num_query_heads, num_kv_heads = num_heads
|
num_query_heads, num_kv_heads = num_heads
|
||||||
|
|||||||
@@ -9,6 +9,7 @@ import torch
|
|||||||
from tests.kernels.utils import DEFAULT_OPCHECK_TEST_UTILS, opcheck
|
from tests.kernels.utils import DEFAULT_OPCHECK_TEST_UTILS, opcheck
|
||||||
from vllm import _custom_ops as ops
|
from vllm import _custom_ops as ops
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
COPYING_DIRECTION = [("cuda", "cpu"), ("cuda", "cuda"), ("cpu", "cuda")]
|
COPYING_DIRECTION = [("cuda", "cpu"), ("cuda", "cuda"), ("cpu", "cuda")]
|
||||||
DTYPES = [torch.bfloat16, torch.float]
|
DTYPES = [torch.bfloat16, torch.float]
|
||||||
@@ -64,7 +65,7 @@ def test_reshape_and_cache(
|
|||||||
) -> None:
|
) -> None:
|
||||||
if kv_cache_dtype == "fp8" and head_size % 16:
|
if kv_cache_dtype == "fp8" and head_size % 16:
|
||||||
pytest.skip()
|
pytest.skip()
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
torch.cuda.set_device(device)
|
torch.cuda.set_device(device)
|
||||||
# Create a random slot mapping.
|
# Create a random slot mapping.
|
||||||
@@ -185,7 +186,7 @@ def test_reshape_and_cache_flash(
|
|||||||
kv_cache_layout: str,
|
kv_cache_layout: str,
|
||||||
implementation: str,
|
implementation: str,
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
torch.cuda.set_device(device)
|
torch.cuda.set_device(device)
|
||||||
assert implementation in ["cuda", "triton"]
|
assert implementation in ["cuda", "triton"]
|
||||||
@@ -355,7 +356,7 @@ def test_swap_blocks(
|
|||||||
if kv_cache_dtype == "fp8" and head_size % 16:
|
if kv_cache_dtype == "fp8" and head_size % 16:
|
||||||
pytest.skip()
|
pytest.skip()
|
||||||
|
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
|
|
||||||
src_device = device if direction[0] == "cuda" else "cpu"
|
src_device = device if direction[0] == "cuda" else "cpu"
|
||||||
dst_device = device if direction[1] == "cuda" else "cpu"
|
dst_device = device if direction[1] == "cuda" else "cpu"
|
||||||
@@ -444,7 +445,7 @@ def test_fp8_e4m3_conversion(
|
|||||||
seed: int,
|
seed: int,
|
||||||
device: str,
|
device: str,
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
|
|
||||||
low = -224.0
|
low = -224.0
|
||||||
high = 224.0
|
high = 224.0
|
||||||
@@ -507,7 +508,7 @@ def test_concat_and_cache_mla(
|
|||||||
device: str,
|
device: str,
|
||||||
kv_cache_dtype: str,
|
kv_cache_dtype: str,
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
torch.cuda.set_device(device)
|
torch.cuda.set_device(device)
|
||||||
|
|
||||||
@@ -584,7 +585,7 @@ def test_concat_and_cache_ds_mla(
|
|||||||
if dtype.itemsize != 2:
|
if dtype.itemsize != 2:
|
||||||
pytest.skip("ds_mla only supports 16-bit input")
|
pytest.skip("ds_mla only supports 16-bit input")
|
||||||
kv_cache_dtype = "fp8_ds_mla"
|
kv_cache_dtype = "fp8_ds_mla"
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
torch.cuda.set_device(device)
|
torch.cuda.set_device(device)
|
||||||
|
|
||||||
@@ -695,7 +696,7 @@ def test_swap_blocks_mla(
|
|||||||
device: str,
|
device: str,
|
||||||
kv_cache_dtype: str,
|
kv_cache_dtype: str,
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
torch.cuda.set_device(device)
|
torch.cuda.set_device(device)
|
||||||
|
|
||||||
@@ -947,7 +948,7 @@ def test_concat_and_cache_mla_cpu(
|
|||||||
) -> None:
|
) -> None:
|
||||||
device = "cpu"
|
device = "cpu"
|
||||||
kv_cache_dtype = "auto"
|
kv_cache_dtype = "auto"
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
|
|
||||||
total_slots = num_blocks * block_size
|
total_slots = num_blocks * block_size
|
||||||
|
|||||||
@@ -6,6 +6,7 @@ import pytest
|
|||||||
import torch
|
import torch
|
||||||
|
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
from vllm.v1.attention.backends.flash_attn import cascade_attention, merge_attn_states
|
from vllm.v1.attention.backends.flash_attn import cascade_attention, merge_attn_states
|
||||||
|
|
||||||
try:
|
try:
|
||||||
@@ -39,7 +40,7 @@ def test_merge_kernel(
|
|||||||
dtype: torch.dtype,
|
dtype: torch.dtype,
|
||||||
):
|
):
|
||||||
torch.set_default_device("cuda")
|
torch.set_default_device("cuda")
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
num_query_heads = num_heads[0]
|
num_query_heads = num_heads[0]
|
||||||
num_kv_heads = num_heads[1]
|
num_kv_heads = num_heads[1]
|
||||||
assert num_query_heads % num_kv_heads == 0
|
assert num_query_heads % num_kv_heads == 0
|
||||||
@@ -103,7 +104,7 @@ def test_cascade(
|
|||||||
f'to: "{fa_version_unsupported_reason(fa_version)}"'
|
f'to: "{fa_version_unsupported_reason(fa_version)}"'
|
||||||
)
|
)
|
||||||
|
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
|
|
||||||
window_size = (-1, -1)
|
window_size = (-1, -1)
|
||||||
scale = head_size**-0.5
|
scale = head_size**-0.5
|
||||||
|
|||||||
@@ -8,6 +8,7 @@ import pytest
|
|||||||
import torch
|
import torch
|
||||||
|
|
||||||
from vllm.platforms import CpuArchEnum, current_platform
|
from vllm.platforms import CpuArchEnum, current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
from vllm.v1.attention.backends.cpu_attn import _get_attn_isa
|
from vllm.v1.attention.backends.cpu_attn import _get_attn_isa
|
||||||
|
|
||||||
if not current_platform.is_cpu():
|
if not current_platform.is_cpu():
|
||||||
@@ -190,7 +191,7 @@ def varlen_with_paged_kv(
|
|||||||
use_sink: bool,
|
use_sink: bool,
|
||||||
isa: str,
|
isa: str,
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
num_seqs = len(seq_lens)
|
num_seqs = len(seq_lens)
|
||||||
query_lens = [x[0] for x in seq_lens]
|
query_lens = [x[0] for x in seq_lens]
|
||||||
kv_lens = [x[1] for x in seq_lens]
|
kv_lens = [x[1] for x in seq_lens]
|
||||||
|
|||||||
@@ -6,6 +6,7 @@ import pytest
|
|||||||
import torch
|
import torch
|
||||||
|
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
try:
|
try:
|
||||||
from vllm.vllm_flash_attn import (
|
from vllm.vllm_flash_attn import (
|
||||||
@@ -129,7 +130,7 @@ def test_varlen_with_paged_kv(
|
|||||||
"Flash attention with quantized inputs is only "
|
"Flash attention with quantized inputs is only "
|
||||||
"supported on version 3 with bfloat16 base type"
|
"supported on version 3 with bfloat16 base type"
|
||||||
)
|
)
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
num_seqs = len(seq_lens)
|
num_seqs = len(seq_lens)
|
||||||
query_lens = [x[0] for x in seq_lens]
|
query_lens = [x[0] for x in seq_lens]
|
||||||
kv_lens = [x[1] for x in seq_lens]
|
kv_lens = [x[1] for x in seq_lens]
|
||||||
|
|||||||
@@ -5,6 +5,7 @@
|
|||||||
import pytest
|
import pytest
|
||||||
|
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
try:
|
try:
|
||||||
import flashinfer
|
import flashinfer
|
||||||
@@ -101,7 +102,7 @@ def test_flashinfer_decode_with_paged_kv(
|
|||||||
sliding_window: int | None,
|
sliding_window: int | None,
|
||||||
) -> None:
|
) -> None:
|
||||||
torch.set_default_device("cuda")
|
torch.set_default_device("cuda")
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
num_seqs = len(kv_lens)
|
num_seqs = len(kv_lens)
|
||||||
num_query_heads = num_heads[0]
|
num_query_heads = num_heads[0]
|
||||||
num_kv_heads = num_heads[1]
|
num_kv_heads = num_heads[1]
|
||||||
@@ -196,7 +197,7 @@ def test_flashinfer_prefill_with_paged_kv(
|
|||||||
sliding_window: int | None,
|
sliding_window: int | None,
|
||||||
) -> None:
|
) -> None:
|
||||||
torch.set_default_device("cuda")
|
torch.set_default_device("cuda")
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
num_seqs = len(seq_lens)
|
num_seqs = len(seq_lens)
|
||||||
query_lens = [x[0] for x in seq_lens]
|
query_lens = [x[0] for x in seq_lens]
|
||||||
kv_lens = [x[1] for x in seq_lens]
|
kv_lens = [x[1] for x in seq_lens]
|
||||||
@@ -299,7 +300,7 @@ def test_flashinfer_prefill_with_paged_fp8_kv(
|
|||||||
) -> None:
|
) -> None:
|
||||||
pytest.skip("TODO: fix the accuracy issue")
|
pytest.skip("TODO: fix the accuracy issue")
|
||||||
torch.set_default_device("cuda")
|
torch.set_default_device("cuda")
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
num_seqs = len(seq_lens)
|
num_seqs = len(seq_lens)
|
||||||
query_lens = [x[0] for x in seq_lens]
|
query_lens = [x[0] for x in seq_lens]
|
||||||
kv_lens = [x[1] for x in seq_lens]
|
kv_lens = [x[1] for x in seq_lens]
|
||||||
@@ -409,7 +410,7 @@ def test_flashinfer_decode_with_paged_fp8_kv(
|
|||||||
) -> None:
|
) -> None:
|
||||||
# test doesn't work for num_heads = (16,16)
|
# test doesn't work for num_heads = (16,16)
|
||||||
torch.set_default_device("cuda")
|
torch.set_default_device("cuda")
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
num_seqs = len(kv_lens)
|
num_seqs = len(kv_lens)
|
||||||
num_query_heads = num_heads[0]
|
num_query_heads = num_heads[0]
|
||||||
num_kv_heads = num_heads[1]
|
num_kv_heads = num_heads[1]
|
||||||
|
|||||||
@@ -10,6 +10,7 @@ from tests.kernels.quantization.nvfp4_utils import (
|
|||||||
)
|
)
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.utils.math_utils import round_up
|
from vllm.utils.math_utils import round_up
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
if not current_platform.is_device_capability_family(100):
|
if not current_platform.is_device_capability_family(100):
|
||||||
pytest.skip(
|
pytest.skip(
|
||||||
@@ -80,7 +81,7 @@ def test_flashinfer_trtllm_decode_with_baseline(
|
|||||||
has_sinks: bool,
|
has_sinks: bool,
|
||||||
) -> None:
|
) -> None:
|
||||||
torch.set_default_device("cuda")
|
torch.set_default_device("cuda")
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
|
|
||||||
q_quant_dtype, kv_quant_dtype, o_quant_dtype = quant_dtypes
|
q_quant_dtype, kv_quant_dtype, o_quant_dtype = quant_dtypes
|
||||||
q_quant_dtype = q_quant_dtype or dtype
|
q_quant_dtype = q_quant_dtype or dtype
|
||||||
@@ -279,7 +280,7 @@ def test_flashinfer_trtllm_prefill_with_baseline(
|
|||||||
has_sinks: bool,
|
has_sinks: bool,
|
||||||
) -> None:
|
) -> None:
|
||||||
torch.set_default_device("cuda")
|
torch.set_default_device("cuda")
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
|
|
||||||
q_quant_dtype, kv_quant_dtype, o_quant_dtype = quant_dtypes
|
q_quant_dtype, kv_quant_dtype, o_quant_dtype = quant_dtypes
|
||||||
q_quant_dtype = q_quant_dtype or dtype
|
q_quant_dtype = q_quant_dtype or dtype
|
||||||
|
|||||||
@@ -5,7 +5,7 @@ import pytest
|
|||||||
import torch
|
import torch
|
||||||
|
|
||||||
from vllm.model_executor.layers.lightning_attn import linear_decode_forward_triton
|
from vllm.model_executor.layers.lightning_attn import linear_decode_forward_triton
|
||||||
from vllm.platforms import current_platform
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
NUM_HEADS = [4, 8]
|
NUM_HEADS = [4, 8]
|
||||||
HEAD_SIZES = [64]
|
HEAD_SIZES = [64]
|
||||||
@@ -124,7 +124,7 @@ def test_linear_decode_forward_triton(
|
|||||||
torch.set_default_device("cuda")
|
torch.set_default_device("cuda")
|
||||||
torch.manual_seed(42)
|
torch.manual_seed(42)
|
||||||
torch.cuda.manual_seed_all(42)
|
torch.cuda.manual_seed_all(42)
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
base = 0.01
|
base = 0.01
|
||||||
q = base * torch.randn(batch_size, num_heads, 1, head_size, dtype=dtype)
|
q = base * torch.randn(batch_size, num_heads, 1, head_size, dtype=dtype)
|
||||||
k = base * torch.randn(batch_size, num_heads, 1, head_size, dtype=dtype)
|
k = base * torch.randn(batch_size, num_heads, 1, head_size, dtype=dtype)
|
||||||
@@ -167,7 +167,7 @@ def test_linear_decode_forward_triton_with_padding(
|
|||||||
torch.set_default_device("cuda")
|
torch.set_default_device("cuda")
|
||||||
torch.manual_seed(42)
|
torch.manual_seed(42)
|
||||||
torch.cuda.manual_seed_all(42)
|
torch.cuda.manual_seed_all(42)
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
|
|
||||||
batch_size = 4
|
batch_size = 4
|
||||||
base = 0.01
|
base = 0.01
|
||||||
@@ -231,7 +231,7 @@ def test_lightning_attention_reference(
|
|||||||
torch.set_default_device("cuda")
|
torch.set_default_device("cuda")
|
||||||
torch.manual_seed(42)
|
torch.manual_seed(42)
|
||||||
torch.cuda.manual_seed_all(42)
|
torch.cuda.manual_seed_all(42)
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
|
|
||||||
base = 0.01
|
base = 0.01
|
||||||
q = base * torch.randn(batch_size, num_heads, seq_len, head_size, dtype=dtype)
|
q = base * torch.randn(batch_size, num_heads, seq_len, head_size, dtype=dtype)
|
||||||
|
|||||||
@@ -19,6 +19,7 @@ from vllm.platforms import current_platform
|
|||||||
from vllm.platforms.cpu import CpuPlatform
|
from vllm.platforms.cpu import CpuPlatform
|
||||||
from vllm.platforms.cuda import CudaPlatform
|
from vllm.platforms.cuda import CudaPlatform
|
||||||
from vllm.platforms.rocm import RocmPlatform
|
from vllm.platforms.rocm import RocmPlatform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
|
|
||||||
@pytest.fixture(autouse=True)
|
@pytest.fixture(autouse=True)
|
||||||
@@ -123,7 +124,7 @@ def test_mha_attn_forward(
|
|||||||
dtype: torch.dtype,
|
dtype: torch.dtype,
|
||||||
device: str,
|
device: str,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
torch.set_default_dtype(dtype)
|
torch.set_default_dtype(dtype)
|
||||||
|
|
||||||
@@ -168,7 +169,7 @@ def test_mha_attn_varlen_forward(
|
|||||||
dtype: torch.dtype,
|
dtype: torch.dtype,
|
||||||
device: str,
|
device: str,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
torch.set_default_dtype(dtype)
|
torch.set_default_dtype(dtype)
|
||||||
|
|
||||||
|
|||||||
@@ -13,7 +13,7 @@ import torch.nn.functional as F
|
|||||||
from vllm.attention.ops.chunked_prefill_paged_decode import chunked_prefill_paged_decode
|
from vllm.attention.ops.chunked_prefill_paged_decode import chunked_prefill_paged_decode
|
||||||
from vllm.attention.ops.prefix_prefill import context_attention_fwd
|
from vllm.attention.ops.prefix_prefill import context_attention_fwd
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
|
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE, set_random_seed
|
||||||
|
|
||||||
NUM_HEADS = [64]
|
NUM_HEADS = [64]
|
||||||
NUM_QUERIES_PER_KV = [1, 64]
|
NUM_QUERIES_PER_KV = [1, 64]
|
||||||
@@ -125,7 +125,7 @@ def test_contexted_kv_attention(
|
|||||||
):
|
):
|
||||||
pytest.skip("ROCm custom paged attention does not support fp8_e5m2 KV cache")
|
pytest.skip("ROCm custom paged attention does not support fp8_e5m2 KV cache")
|
||||||
|
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
|
|
||||||
# Need this, otherwise when we capture the graph the process
|
# Need this, otherwise when we capture the graph the process
|
||||||
@@ -346,7 +346,7 @@ def test_contexted_kv_attention_alibi(
|
|||||||
):
|
):
|
||||||
pytest.skip("ROCm custom paged attention does not support fp8_e5m2 KV cache")
|
pytest.skip("ROCm custom paged attention does not support fp8_e5m2 KV cache")
|
||||||
|
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
|
|
||||||
# Need this, otherwise when we capture the graph the process
|
# Need this, otherwise when we capture the graph the process
|
||||||
|
|||||||
@@ -8,6 +8,7 @@ import torch
|
|||||||
from vllm.attention.ops.triton_unified_attention import unified_attention
|
from vllm.attention.ops.triton_unified_attention import unified_attention
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.utils.math_utils import next_power_of_2
|
from vllm.utils.math_utils import next_power_of_2
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
NUM_HEADS = [(4, 4), (8, 2)]
|
NUM_HEADS = [(4, 4), (8, 2)]
|
||||||
HEAD_SIZES = [128, 256]
|
HEAD_SIZES = [128, 256]
|
||||||
@@ -113,7 +114,7 @@ def test_triton_unified_attn(
|
|||||||
) -> None:
|
) -> None:
|
||||||
torch.set_default_device("cuda")
|
torch.set_default_device("cuda")
|
||||||
|
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
num_seqs = len(seq_lens)
|
num_seqs = len(seq_lens)
|
||||||
query_lens = [x[0] for x in seq_lens]
|
query_lens = [x[0] for x in seq_lens]
|
||||||
kv_lens = [x[1] for x in seq_lens]
|
kv_lens = [x[1] for x in seq_lens]
|
||||||
|
|||||||
@@ -18,7 +18,7 @@ from vllm.model_executor.layers.activation import (
|
|||||||
SiluAndMul,
|
SiluAndMul,
|
||||||
SwigluOAIAndMul,
|
SwigluOAIAndMul,
|
||||||
)
|
)
|
||||||
from vllm.platforms import current_platform
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
DTYPES = [torch.half, torch.bfloat16, torch.float]
|
DTYPES = [torch.half, torch.bfloat16, torch.float]
|
||||||
NUM_TOKENS = [7, 83, 2048] # Arbitrary values for testing
|
NUM_TOKENS = [7, 83, 2048] # Arbitrary values for testing
|
||||||
@@ -52,7 +52,7 @@ def test_act_and_mul(
|
|||||||
seed: int,
|
seed: int,
|
||||||
device: str,
|
device: str,
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
x = torch.randn(num_tokens, 2 * d, dtype=dtype)
|
x = torch.randn(num_tokens, 2 * d, dtype=dtype)
|
||||||
if activation == "silu_and_mul":
|
if activation == "silu_and_mul":
|
||||||
@@ -129,7 +129,7 @@ def test_activation(
|
|||||||
seed: int,
|
seed: int,
|
||||||
device: str,
|
device: str,
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
x = torch.randn(num_tokens, d, dtype=dtype)
|
x = torch.randn(num_tokens, d, dtype=dtype)
|
||||||
layer = activation[0]()
|
layer = activation[0]()
|
||||||
|
|||||||
@@ -8,6 +8,7 @@ from tests.kernels.utils import opcheck
|
|||||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||||
from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding
|
from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
DTYPES = [torch.bfloat16, torch.float16]
|
DTYPES = [torch.bfloat16, torch.float16]
|
||||||
IS_NEOX = [True, False]
|
IS_NEOX = [True, False]
|
||||||
@@ -64,7 +65,7 @@ def test_fused_qk_norm_rope_matches_reference(
|
|||||||
rotary_ratio: float,
|
rotary_ratio: float,
|
||||||
):
|
):
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
num_heads, num_kv_heads, head_dim = 16, 4, 128
|
num_heads, num_kv_heads, head_dim = 16, 4, 128
|
||||||
num_tokens = 4
|
num_tokens = 4
|
||||||
|
|
||||||
|
|||||||
@@ -7,7 +7,7 @@ import torch
|
|||||||
from tests.kernels.quant_utils import FP8_DTYPE
|
from tests.kernels.quant_utils import FP8_DTYPE
|
||||||
from tests.kernels.utils import opcheck
|
from tests.kernels.utils import opcheck
|
||||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||||
from vllm.platforms import current_platform
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
DTYPES = [torch.half, torch.bfloat16, torch.float]
|
DTYPES = [torch.half, torch.bfloat16, torch.float]
|
||||||
NUM_TOKENS = [7, 83, 4096] # Arbitrary values for testing
|
NUM_TOKENS = [7, 83, 4096] # Arbitrary values for testing
|
||||||
@@ -34,7 +34,7 @@ def test_rms_norm(
|
|||||||
device: str,
|
device: str,
|
||||||
strided_input: bool,
|
strided_input: bool,
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
layer = RMSNorm(hidden_size).to(dtype=dtype)
|
layer = RMSNorm(hidden_size).to(dtype=dtype)
|
||||||
layer.weight.data.normal_(mean=1.0, std=0.1)
|
layer.weight.data.normal_(mean=1.0, std=0.1)
|
||||||
@@ -88,7 +88,7 @@ def test_fused_rms_norm_quant(
|
|||||||
device: str,
|
device: str,
|
||||||
strided_input: bool,
|
strided_input: bool,
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
|
|
||||||
weight = torch.empty(hidden_size, dtype=dtype).normal_(mean=1.0, std=0.1)
|
weight = torch.empty(hidden_size, dtype=dtype).normal_(mean=1.0, std=0.1)
|
||||||
|
|||||||
@@ -10,6 +10,7 @@ from transformers import __version__ as TRANSFORMERS_VERSION
|
|||||||
from vllm.model_executor.layers.rotary_embedding import get_rope
|
from vllm.model_executor.layers.rotary_embedding import get_rope
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.transformers_utils.config import get_config
|
from vllm.transformers_utils.config import get_config
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
||||||
|
|
||||||
@@ -24,7 +25,7 @@ def generate_test_data(
|
|||||||
device: torch.device,
|
device: torch.device,
|
||||||
):
|
):
|
||||||
"""Generate test data for given configuration."""
|
"""Generate test data for given configuration."""
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
# Create 2D positions (3, num_tokens) for multimodal case
|
# Create 2D positions (3, num_tokens) for multimodal case
|
||||||
positions = torch.randint(
|
positions = torch.randint(
|
||||||
0, max_position_embeddings // 4, (3, num_tokens), device=device
|
0, max_position_embeddings // 4, (3, num_tokens), device=device
|
||||||
|
|||||||
@@ -9,7 +9,7 @@ import torch
|
|||||||
|
|
||||||
from tests.kernels.allclose_default import get_default_atol, get_default_rtol
|
from tests.kernels.allclose_default import get_default_atol, get_default_rtol
|
||||||
from vllm.model_executor.layers.rotary_embedding import get_rope
|
from vllm.model_executor.layers.rotary_embedding import get_rope
|
||||||
from vllm.platforms import current_platform
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
IS_NEOX_STYLE = [True, False]
|
IS_NEOX_STYLE = [True, False]
|
||||||
DTYPES = [torch.bfloat16, torch.float]
|
DTYPES = [torch.bfloat16, torch.float]
|
||||||
@@ -79,7 +79,7 @@ def test_rotary_embedding(
|
|||||||
if rotary_dim is None:
|
if rotary_dim is None:
|
||||||
rotary_dim = head_size
|
rotary_dim = head_size
|
||||||
|
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
if rotary_dim is None:
|
if rotary_dim is None:
|
||||||
rotary_dim = head_size
|
rotary_dim = head_size
|
||||||
|
|||||||
@@ -12,7 +12,7 @@ from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
|
|||||||
causal_conv1d_fn,
|
causal_conv1d_fn,
|
||||||
causal_conv1d_update,
|
causal_conv1d_update,
|
||||||
)
|
)
|
||||||
from vllm.platforms import current_platform
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
|
|
||||||
def causal_conv1d_ref(
|
def causal_conv1d_ref(
|
||||||
@@ -154,7 +154,7 @@ def test_causal_conv1d_update(dim, width, seqlen, has_bias, silu_activation, ity
|
|||||||
if itype == torch.bfloat16:
|
if itype == torch.bfloat16:
|
||||||
rtol, atol = 1e-2, 5e-2
|
rtol, atol = 1e-2, 5e-2
|
||||||
# set seed
|
# set seed
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
batch = 2
|
batch = 2
|
||||||
x = torch.randn(batch, dim, seqlen, device=device, dtype=itype)
|
x = torch.randn(batch, dim, seqlen, device=device, dtype=itype)
|
||||||
x_ref = x.clone()
|
x_ref = x.clone()
|
||||||
@@ -201,7 +201,7 @@ def test_causal_conv1d_update_with_batch_gather(
|
|||||||
rtol, atol = 1e-2, 5e-2
|
rtol, atol = 1e-2, 5e-2
|
||||||
|
|
||||||
# set seed
|
# set seed
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
|
|
||||||
padding = 5 if with_padding else 0
|
padding = 5 if with_padding else 0
|
||||||
padded_batch_size = batch_size + padding
|
padded_batch_size = batch_size + padding
|
||||||
@@ -278,7 +278,7 @@ def test_causal_conv1d_varlen(
|
|||||||
if itype == torch.bfloat16:
|
if itype == torch.bfloat16:
|
||||||
rtol, atol = 1e-2, 5e-2
|
rtol, atol = 1e-2, 5e-2
|
||||||
# set seed
|
# set seed
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
seqlens = []
|
seqlens = []
|
||||||
batch_size = batch
|
batch_size = batch
|
||||||
padding = 3 if with_padding else 0
|
padding = 3 if with_padding else 0
|
||||||
|
|||||||
@@ -12,8 +12,8 @@ from vllm.distributed.parallel_state import (
|
|||||||
initialize_model_parallel,
|
initialize_model_parallel,
|
||||||
)
|
)
|
||||||
from vllm.model_executor.layers.mamba.mamba_mixer2 import Mixer2RMSNormGated
|
from vllm.model_executor.layers.mamba.mamba_mixer2 import Mixer2RMSNormGated
|
||||||
from vllm.platforms import current_platform
|
|
||||||
from vllm.utils.system_utils import update_environment_variables
|
from vllm.utils.system_utils import update_environment_variables
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
|
|
||||||
@multi_gpu_test(num_gpus=2)
|
@multi_gpu_test(num_gpus=2)
|
||||||
@@ -68,7 +68,7 @@ def mixer2_gated_norm_tensor_parallel(
|
|||||||
dtype: torch.dtype,
|
dtype: torch.dtype,
|
||||||
device: str,
|
device: str,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
|
|
||||||
device = torch.device(f"cuda:{local_rank}")
|
device = torch.device(f"cuda:{local_rank}")
|
||||||
torch.cuda.set_device(device)
|
torch.cuda.set_device(device)
|
||||||
|
|||||||
@@ -13,7 +13,7 @@ from vllm.model_executor.layers.mamba.ops.mamba_ssm import (
|
|||||||
selective_scan_fn,
|
selective_scan_fn,
|
||||||
selective_state_update,
|
selective_state_update,
|
||||||
)
|
)
|
||||||
from vllm.platforms import current_platform
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
|
|
||||||
def selective_state_update_ref(
|
def selective_state_update_ref(
|
||||||
@@ -271,7 +271,7 @@ def test_selective_scan(
|
|||||||
rtolw = max(rtolw, rtol)
|
rtolw = max(rtolw, rtol)
|
||||||
atolw = max(atolw, atol)
|
atolw = max(atolw, atol)
|
||||||
# set seed
|
# set seed
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
batch_size = 1
|
batch_size = 1
|
||||||
dim = 4
|
dim = 4
|
||||||
dstate = 8
|
dstate = 8
|
||||||
@@ -401,7 +401,7 @@ def test_selective_state_update(dim, dstate, has_z, itype):
|
|||||||
if torch.version.hip:
|
if torch.version.hip:
|
||||||
atol *= 2
|
atol *= 2
|
||||||
# set seed
|
# set seed
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
batch_size = 1
|
batch_size = 1
|
||||||
state = torch.randn(batch_size, dim, dstate, dtype=itype, device=device)
|
state = torch.randn(batch_size, dim, dstate, dtype=itype, device=device)
|
||||||
x = torch.randn(batch_size, dim, device=device, dtype=itype)
|
x = torch.randn(batch_size, dim, device=device, dtype=itype)
|
||||||
@@ -438,7 +438,7 @@ def test_selective_state_update_varlen(dim, dstate, has_z, itype, max_seq_len):
|
|||||||
if torch.version.hip:
|
if torch.version.hip:
|
||||||
atol *= 2
|
atol *= 2
|
||||||
# set seed
|
# set seed
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
batch_size = 4
|
batch_size = 4
|
||||||
token_counts = torch.randint(1, max_seq_len + 1, (batch_size,), device=device)
|
token_counts = torch.randint(1, max_seq_len + 1, (batch_size,), device=device)
|
||||||
total_tokens = int(token_counts.sum().item())
|
total_tokens = int(token_counts.sum().item())
|
||||||
@@ -857,7 +857,7 @@ def test_selective_state_update_with_num_accepted_tokens(
|
|||||||
if torch.version.hip:
|
if torch.version.hip:
|
||||||
atol *= 2
|
atol *= 2
|
||||||
|
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
batch_size = 4
|
batch_size = 4
|
||||||
|
|
||||||
tokens_per_seq = torch.randint(1, max_seq_len + 1, (batch_size,), device=device)
|
tokens_per_seq = torch.randint(1, max_seq_len + 1, (batch_size,), device=device)
|
||||||
@@ -983,7 +983,7 @@ def test_selective_state_update_varlen_with_num_accepted(
|
|||||||
if torch.version.hip:
|
if torch.version.hip:
|
||||||
atol *= 2
|
atol *= 2
|
||||||
|
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
batch_size = 4
|
batch_size = 4
|
||||||
|
|
||||||
tokens_per_seq = torch.randint(1, max_seq_len + 1, (batch_size,), device=device)
|
tokens_per_seq = torch.randint(1, max_seq_len + 1, (batch_size,), device=device)
|
||||||
|
|||||||
@@ -9,7 +9,7 @@ from einops import rearrange, repeat
|
|||||||
from vllm.model_executor.layers.mamba.ops.ssd_combined import (
|
from vllm.model_executor.layers.mamba.ops.ssd_combined import (
|
||||||
mamba_chunk_scan_combined_varlen,
|
mamba_chunk_scan_combined_varlen,
|
||||||
)
|
)
|
||||||
from vllm.platforms import current_platform
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
from vllm.v1.attention.backends.mamba2_attn import compute_varlen_chunk_metadata
|
from vllm.v1.attention.backends.mamba2_attn import compute_varlen_chunk_metadata
|
||||||
|
|
||||||
# Added by the IBM Team, 2024
|
# Added by the IBM Team, 2024
|
||||||
@@ -82,7 +82,7 @@ def ssd_minimal_discrete(X, A, B, C, block_len, initial_states=None):
|
|||||||
|
|
||||||
|
|
||||||
def generate_random_inputs(batch_size, seqlen, n_heads, d_head, itype, device="cuda"):
|
def generate_random_inputs(batch_size, seqlen, n_heads, d_head, itype, device="cuda"):
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
A = -torch.exp(torch.rand(n_heads, dtype=itype, device=device))
|
A = -torch.exp(torch.rand(n_heads, dtype=itype, device=device))
|
||||||
dt = F.softplus(
|
dt = F.softplus(
|
||||||
torch.randn(batch_size, seqlen, n_heads, dtype=itype, device=device) - 4
|
torch.randn(batch_size, seqlen, n_heads, dtype=itype, device=device) - 4
|
||||||
|
|||||||
@@ -10,7 +10,7 @@ from tqdm import tqdm
|
|||||||
|
|
||||||
from vllm.config import VllmConfig, set_current_vllm_config
|
from vllm.config import VllmConfig, set_current_vllm_config
|
||||||
from vllm.model_executor.layers.fused_moe.config import FUSED_MOE_UNQUANTIZED_CONFIG
|
from vllm.model_executor.layers.fused_moe.config import FUSED_MOE_UNQUANTIZED_CONFIG
|
||||||
from vllm.platforms import current_platform
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
from .common import (
|
from .common import (
|
||||||
Config,
|
Config,
|
||||||
@@ -40,7 +40,7 @@ def rank_worker(
|
|||||||
config: Config,
|
config: Config,
|
||||||
weights: WeightTensors,
|
weights: WeightTensors,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(pgi.rank)
|
set_random_seed(pgi.rank)
|
||||||
|
|
||||||
# sanity check
|
# sanity check
|
||||||
from vllm import envs
|
from vllm import envs
|
||||||
|
|||||||
@@ -9,7 +9,7 @@ from typing import Any
|
|||||||
import torch
|
import torch
|
||||||
|
|
||||||
from vllm.config import VllmConfig
|
from vllm.config import VllmConfig
|
||||||
from vllm.platforms import current_platform
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
from .common import Config, RankTensors, WeightTensors, make_modular_kernel
|
from .common import Config, RankTensors, WeightTensors, make_modular_kernel
|
||||||
from .parallel_utils import ProcessGroupInfo, parallel_launch_with_config
|
from .parallel_utils import ProcessGroupInfo, parallel_launch_with_config
|
||||||
@@ -82,7 +82,7 @@ def rank_worker(
|
|||||||
config: Config,
|
config: Config,
|
||||||
weights: WeightTensors,
|
weights: WeightTensors,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(pgi.rank)
|
set_random_seed(pgi.rank)
|
||||||
|
|
||||||
# sanity check
|
# sanity check
|
||||||
from vllm import envs
|
from vllm import envs
|
||||||
|
|||||||
@@ -21,6 +21,7 @@ from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
|
|||||||
from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
|
from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.triton_utils import tl
|
from vllm.triton_utils import tl
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
MNK_FACTORS = [
|
MNK_FACTORS = [
|
||||||
(1, 128, 128),
|
(1, 128, 128),
|
||||||
@@ -115,7 +116,7 @@ def test_batched_mm(
|
|||||||
):
|
):
|
||||||
"""Note: float8_e4m3fn is not supported on CUDA architecture < 89,
|
"""Note: float8_e4m3fn is not supported on CUDA architecture < 89,
|
||||||
and those tests will be skipped on unsupported hardware."""
|
and those tests will be skipped on unsupported hardware."""
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
|
|
||||||
use_fp8_w8a8 = dtype == torch.float8_e4m3fn
|
use_fp8_w8a8 = dtype == torch.float8_e4m3fn
|
||||||
|
|
||||||
@@ -252,7 +253,7 @@ def test_fused_moe_batched_experts(
|
|||||||
):
|
):
|
||||||
"""Note: float8_e4m3fn is not supported on CUDA architecture < 89,
|
"""Note: float8_e4m3fn is not supported on CUDA architecture < 89,
|
||||||
and those tests will be skipped on unsupported hardware."""
|
and those tests will be skipped on unsupported hardware."""
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
|
|
||||||
use_fp8_w8a8 = dtype == torch.float8_e4m3fn
|
use_fp8_w8a8 = dtype == torch.float8_e4m3fn
|
||||||
|
|
||||||
|
|||||||
@@ -8,6 +8,7 @@ from tests.kernels.allclose_default import get_default_atol, get_default_rtol
|
|||||||
from vllm._custom_ops import cpu_fused_moe, cpu_prepack_moe_weight
|
from vllm._custom_ops import cpu_fused_moe, cpu_prepack_moe_weight
|
||||||
from vllm.model_executor.layers.activation import SiluAndMul, SwigluOAIAndMul
|
from vllm.model_executor.layers.activation import SiluAndMul, SwigluOAIAndMul
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
if not current_platform.is_cpu():
|
if not current_platform.is_cpu():
|
||||||
pytest.skip("skipping CPU-only tests", allow_module_level=True)
|
pytest.skip("skipping CPU-only tests", allow_module_level=True)
|
||||||
@@ -114,7 +115,7 @@ def test_cpu_fused_moe(
|
|||||||
act: str,
|
act: str,
|
||||||
isa: str,
|
isa: str,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
|
|
||||||
topk_num = max(expert_num // 2, 1)
|
topk_num = max(expert_num // 2, 1)
|
||||||
up_dim = 2 * intermediate_size
|
up_dim = 2 * intermediate_size
|
||||||
|
|||||||
@@ -20,6 +20,7 @@ from vllm.model_executor.layers.fused_moe.cutlass_moe import (
|
|||||||
from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
|
from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
|
||||||
from vllm.model_executor.layers.fused_moe.utils import moe_kernel_quantize_input
|
from vllm.model_executor.layers.fused_moe.utils import moe_kernel_quantize_input
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
NUM_EXPERTS = [40, 64]
|
NUM_EXPERTS = [40, 64]
|
||||||
TOP_KS = [6, 8]
|
TOP_KS = [6, 8]
|
||||||
@@ -277,7 +278,7 @@ def test_cutlass_moe_8_bit_no_graph(
|
|||||||
workspace_init,
|
workspace_init,
|
||||||
ep_size: int | None = None,
|
ep_size: int | None = None,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
|
monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
|
||||||
with set_current_vllm_config(vllm_config):
|
with set_current_vllm_config(vllm_config):
|
||||||
mt = MOETensors8Bit.make_moe_tensors_8bit(m, k, n, e, per_act_token, per_out_ch)
|
mt = MOETensors8Bit.make_moe_tensors_8bit(m, k, n, e, per_act_token, per_out_ch)
|
||||||
@@ -332,7 +333,7 @@ def test_cutlass_moe_8_bit_cuda_graph(
|
|||||||
monkeypatch,
|
monkeypatch,
|
||||||
workspace_init,
|
workspace_init,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
|
monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
|
||||||
with set_current_vllm_config(vllm_config):
|
with set_current_vllm_config(vllm_config):
|
||||||
dtype = torch.half
|
dtype = torch.half
|
||||||
@@ -469,7 +470,7 @@ def test_run_cutlass_moe_fp8(
|
|||||||
ep_size: int,
|
ep_size: int,
|
||||||
workspace_init,
|
workspace_init,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
with set_current_vllm_config(vllm_config):
|
with set_current_vllm_config(vllm_config):
|
||||||
mt = MOETensors8Bit.make_moe_tensors_8bit(
|
mt = MOETensors8Bit.make_moe_tensors_8bit(
|
||||||
m, k, n, e, per_act_token, per_out_channel
|
m, k, n, e, per_act_token, per_out_channel
|
||||||
|
|||||||
@@ -22,13 +22,13 @@ from vllm.model_executor.layers.fused_moe.config import (
|
|||||||
)
|
)
|
||||||
from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts
|
from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts
|
||||||
from vllm.model_executor.layers.fused_moe.modular_kernel import FusedMoEModularKernel
|
from vllm.model_executor.layers.fused_moe.modular_kernel import FusedMoEModularKernel
|
||||||
from vllm.platforms import current_platform
|
|
||||||
from vllm.utils.deep_gemm import (
|
from vllm.utils.deep_gemm import (
|
||||||
get_mk_alignment_for_contiguous_layout,
|
get_mk_alignment_for_contiguous_layout,
|
||||||
is_deep_gemm_e8m0_used,
|
is_deep_gemm_e8m0_used,
|
||||||
is_deep_gemm_supported,
|
is_deep_gemm_supported,
|
||||||
)
|
)
|
||||||
from vllm.utils.import_utils import has_deep_ep, has_deep_gemm
|
from vllm.utils.import_utils import has_deep_ep, has_deep_gemm
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
from vllm.v1.worker.workspace import init_workspace_manager
|
from vllm.v1.worker.workspace import init_workspace_manager
|
||||||
|
|
||||||
from ...utils import multi_gpu_test
|
from ...utils import multi_gpu_test
|
||||||
@@ -367,7 +367,7 @@ def _test_deepep_deepgemm_moe(
|
|||||||
device = torch.device(f"cuda:{pgi.local_rank}")
|
device = torch.device(f"cuda:{pgi.local_rank}")
|
||||||
init_workspace_manager(device)
|
init_workspace_manager(device)
|
||||||
|
|
||||||
current_platform.seed_everything(pgi.rank)
|
set_random_seed(pgi.rank)
|
||||||
|
|
||||||
w1 = w1.to(device=torch.cuda.current_device())
|
w1 = w1.to(device=torch.cuda.current_device())
|
||||||
w2 = w2.to(device=torch.cuda.current_device())
|
w2 = w2.to(device=torch.cuda.current_device())
|
||||||
@@ -456,7 +456,7 @@ def test_ht_deepep_deepgemm_moe(
|
|||||||
"""
|
"""
|
||||||
|
|
||||||
m, n, k = mnk
|
m, n, k = mnk
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
|
|
||||||
if topk > num_experts:
|
if topk > num_experts:
|
||||||
pytest.skip(f"Skipping test: topk={topk} > E={num_experts}")
|
pytest.skip(f"Skipping test: topk={topk} > E={num_experts}")
|
||||||
@@ -531,7 +531,7 @@ def test_ll_deepep_deepgemm_moe(
|
|||||||
assert not is_deep_gemm_e8m0_used()
|
assert not is_deep_gemm_e8m0_used()
|
||||||
|
|
||||||
m, n, k = mnk
|
m, n, k = mnk
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
|
|
||||||
if topk > num_experts:
|
if topk > num_experts:
|
||||||
pytest.skip(f"Skipping test: topk={topk} > E={num_experts}")
|
pytest.skip(f"Skipping test: topk={topk} > E={num_experts}")
|
||||||
|
|||||||
@@ -20,8 +20,8 @@ from vllm.model_executor.layers.fused_moe.modular_kernel import FusedMoEModularK
|
|||||||
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
|
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
|
||||||
per_token_group_quant_fp8,
|
per_token_group_quant_fp8,
|
||||||
)
|
)
|
||||||
from vllm.platforms import current_platform
|
|
||||||
from vllm.utils.import_utils import has_deep_ep
|
from vllm.utils.import_utils import has_deep_ep
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
from vllm.v1.worker.workspace import init_workspace_manager
|
from vllm.v1.worker.workspace import init_workspace_manager
|
||||||
|
|
||||||
from ...utils import multi_gpu_test
|
from ...utils import multi_gpu_test
|
||||||
@@ -446,7 +446,7 @@ def test_deep_ep_moe(
|
|||||||
low_latency_mode = False
|
low_latency_mode = False
|
||||||
use_fp8_dispatch = False
|
use_fp8_dispatch = False
|
||||||
|
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
world_size, dp_size = world_dp_size
|
world_size, dp_size = world_dp_size
|
||||||
config = TestConfig(dtype=dtype, topk=topk, m=m, k=k, n=n, num_experts=num_experts)
|
config = TestConfig(dtype=dtype, topk=topk, m=m, k=k, n=n, num_experts=num_experts)
|
||||||
|
|
||||||
@@ -507,7 +507,7 @@ def test_low_latency_deep_ep_moe(
|
|||||||
f"hidden sizes {DeepEPLLPrepareAndFinalize.SUPPORTED_HIDDEN_SIZES}"
|
f"hidden sizes {DeepEPLLPrepareAndFinalize.SUPPORTED_HIDDEN_SIZES}"
|
||||||
)
|
)
|
||||||
|
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
world_size, dp_size = world_dp_size
|
world_size, dp_size = world_dp_size
|
||||||
config = TestConfig(dtype=dtype, topk=topk, m=m, k=k, n=n, num_experts=num_experts)
|
config = TestConfig(dtype=dtype, topk=topk, m=m, k=k, n=n, num_experts=num_experts)
|
||||||
|
|
||||||
|
|||||||
@@ -22,6 +22,7 @@ from vllm.model_executor.layers.quantization.utils.flashinfer_utils import (
|
|||||||
from vllm.model_executor.layers.quantization.utils.fp8_utils import input_to_float8
|
from vllm.model_executor.layers.quantization.utils.fp8_utils import input_to_float8
|
||||||
from vllm.model_executor.models.llama4 import Llama4MoE
|
from vllm.model_executor.models.llama4 import Llama4MoE
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
try:
|
try:
|
||||||
from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
|
from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
|
||||||
@@ -158,7 +159,7 @@ def test_flashinfer_per_tensor_moe_fp8_no_graph(
|
|||||||
):
|
):
|
||||||
if not current_platform.has_device_capability(100):
|
if not current_platform.has_device_capability(100):
|
||||||
pytest.skip("Test is only supported for sm >= 100")
|
pytest.skip("Test is only supported for sm >= 100")
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
|
monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
|
||||||
with set_current_vllm_config(vllm_config):
|
with set_current_vllm_config(vllm_config):
|
||||||
td = TestData.make_moe_tensors_8bit(m, k, n, e, reorder=True)
|
td = TestData.make_moe_tensors_8bit(m, k, n, e, reorder=True)
|
||||||
@@ -222,7 +223,7 @@ def test_flashinfer_cutlass_moe_fp8_no_graph(
|
|||||||
monkeypatch,
|
monkeypatch,
|
||||||
workspace_init,
|
workspace_init,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
|
monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "8192")
|
||||||
with set_current_vllm_config(vllm_config):
|
with set_current_vllm_config(vllm_config):
|
||||||
td = TestData.make_moe_tensors_8bit(
|
td = TestData.make_moe_tensors_8bit(
|
||||||
|
|||||||
@@ -23,6 +23,7 @@ from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
|
|||||||
from vllm.model_executor.layers.fused_moe.modular_kernel import FusedMoEModularKernel
|
from vllm.model_executor.layers.fused_moe.modular_kernel import FusedMoEModularKernel
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
|
from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
if not has_flashinfer_cutlass_fused_moe() or not current_platform.has_device_capability(
|
if not has_flashinfer_cutlass_fused_moe() or not current_platform.has_device_capability(
|
||||||
100
|
100
|
||||||
@@ -60,7 +61,7 @@ def test_flashinfer_fp4_moe_no_graph(
|
|||||||
activation: str,
|
activation: str,
|
||||||
workspace_init,
|
workspace_init,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
with set_current_vllm_config(
|
with set_current_vllm_config(
|
||||||
VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
|
VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
|
||||||
):
|
):
|
||||||
|
|||||||
@@ -19,6 +19,7 @@ from vllm.model_executor.layers.fused_moe.fused_moe import (
|
|||||||
fused_grouped_topk,
|
fused_grouped_topk,
|
||||||
)
|
)
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
|
|
||||||
@pytest.mark.skipif(
|
@pytest.mark.skipif(
|
||||||
@@ -52,7 +53,7 @@ def test_grouped_topk(
|
|||||||
)
|
)
|
||||||
get_cached_compilation_config.cache_clear()
|
get_cached_compilation_config.cache_clear()
|
||||||
|
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
hidden_states = torch.randn((n_token, n_hidden), dtype=dtype, device="cuda")
|
hidden_states = torch.randn((n_token, n_hidden), dtype=dtype, device="cuda")
|
||||||
gating_output = torch.randn((n_token, n_expert), dtype=dtype, device="cuda")
|
gating_output = torch.randn((n_token, n_expert), dtype=dtype, device="cuda")
|
||||||
e_score_correction_bias = torch.randn(
|
e_score_correction_bias = torch.randn(
|
||||||
|
|||||||
@@ -15,7 +15,7 @@ from vllm.config import VllmConfig, set_current_vllm_config
|
|||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
|
from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
|
||||||
from vllm.utils.import_utils import has_deep_ep, has_deep_gemm, has_pplx
|
from vllm.utils.import_utils import has_deep_ep, has_deep_gemm, has_pplx
|
||||||
from vllm.utils.torch_utils import cuda_device_count_stateless
|
from vllm.utils.torch_utils import cuda_device_count_stateless, set_random_seed
|
||||||
from vllm.v1.worker.workspace import init_workspace_manager
|
from vllm.v1.worker.workspace import init_workspace_manager
|
||||||
|
|
||||||
from .modular_kernel_tools.common import (
|
from .modular_kernel_tools.common import (
|
||||||
@@ -82,7 +82,7 @@ def rank_worker(
|
|||||||
device = torch.device(f"cuda:{pgi.local_rank}")
|
device = torch.device(f"cuda:{pgi.local_rank}")
|
||||||
init_workspace_manager(device)
|
init_workspace_manager(device)
|
||||||
|
|
||||||
current_platform.seed_everything(pgi.rank)
|
set_random_seed(pgi.rank)
|
||||||
|
|
||||||
# sanity check
|
# sanity check
|
||||||
from vllm import envs
|
from vllm import envs
|
||||||
|
|||||||
@@ -34,6 +34,7 @@ from vllm.model_executor.layers.fused_moe.prepare_finalize import (
|
|||||||
)
|
)
|
||||||
from vllm.model_executor.layers.utils import shuffle_weight
|
from vllm.model_executor.layers.utils import shuffle_weight
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
MNK = [
|
MNK = [
|
||||||
(1, 512, 384),
|
(1, 512, 384),
|
||||||
@@ -211,7 +212,7 @@ def test_oai_triton_moe(
|
|||||||
unfused: bool,
|
unfused: bool,
|
||||||
workspace_init,
|
workspace_init,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
(
|
(
|
||||||
w1,
|
w1,
|
||||||
w2,
|
w2,
|
||||||
|
|||||||
@@ -60,6 +60,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import quantize_w
|
|||||||
from vllm.model_executor.models.mixtral import MixtralMoE
|
from vllm.model_executor.models.mixtral import MixtralMoE
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.scalar_type import ScalarType, scalar_types
|
from vllm.scalar_type import ScalarType, scalar_types
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
from vllm.v1.worker.workspace import init_workspace_manager
|
from vllm.v1.worker.workspace import init_workspace_manager
|
||||||
|
|
||||||
NUM_EXPERTS = [8, 64, 192]
|
NUM_EXPERTS = [8, 64, 192]
|
||||||
@@ -234,7 +235,7 @@ def test_fused_moe(
|
|||||||
monkeypatch,
|
monkeypatch,
|
||||||
workspace_init,
|
workspace_init,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
|
|
||||||
monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", str(chunk_size))
|
monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", str(chunk_size))
|
||||||
|
|
||||||
|
|||||||
@@ -14,12 +14,13 @@ from vllm.model_executor.layers.fused_moe.moe_align_block_size import (
|
|||||||
)
|
)
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.utils.math_utils import round_up
|
from vllm.utils.math_utils import round_up
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
NUM_TOKENS = [1, 3, 256, 2256, 4096]
|
NUM_TOKENS = [1, 3, 256, 2256, 4096]
|
||||||
NUM_EXPERTS = [32, 160, 256, 257]
|
NUM_EXPERTS = [32, 160, 256, 257]
|
||||||
TOP_KS = [1, 2, 16, 32]
|
TOP_KS = [1, 2, 16, 32]
|
||||||
BLOCK_SIZES = [32, 128]
|
BLOCK_SIZES = [32, 128]
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
|
|
||||||
|
|
||||||
def _group_tokens_by_expert(
|
def _group_tokens_by_expert(
|
||||||
|
|||||||
@@ -17,11 +17,12 @@ from vllm.model_executor.layers.fused_moe.moe_permute_unpermute import (
|
|||||||
moe_unpermute,
|
moe_unpermute,
|
||||||
)
|
)
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
NUM_EXPERTS = [16, 64, 256]
|
NUM_EXPERTS = [16, 64, 256]
|
||||||
TOP_KS = [2, 6, 8]
|
TOP_KS = [2, 6, 8]
|
||||||
EP_SIZE = [1, 4, 16]
|
EP_SIZE = [1, 4, 16]
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
|
|
||||||
if current_platform.is_rocm():
|
if current_platform.is_rocm():
|
||||||
pytest.skip(
|
pytest.skip(
|
||||||
@@ -226,7 +227,7 @@ def test_moe_permute_unpermute(
|
|||||||
n_local_expert, expert_map, _ = determine_expert_map(ep_size, ep_rank, n_expert)
|
n_local_expert, expert_map, _ = determine_expert_map(ep_size, ep_rank, n_expert)
|
||||||
expert_map = expert_map.cuda()
|
expert_map = expert_map.cuda()
|
||||||
start_expert = n_local_expert * ep_rank
|
start_expert = n_local_expert * ep_rank
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
hidden_states = torch.randn((n_token, n_hidden), device="cuda").to(dtype)
|
hidden_states = torch.randn((n_token, n_hidden), device="cuda").to(dtype)
|
||||||
gating_output = torch.randn((n_token, n_expert), device="cuda").to(dtype)
|
gating_output = torch.randn((n_token, n_expert), device="cuda").to(dtype)
|
||||||
topk_weights, topk_ids, token_expert_indices = fused_topk(
|
topk_weights, topk_ids, token_expert_indices = fused_topk(
|
||||||
|
|||||||
@@ -16,6 +16,7 @@ from vllm.model_executor.layers.fused_moe.config import nvfp4_moe_quant_config
|
|||||||
from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp4
|
from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp4
|
||||||
from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
|
from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
if not current_platform.has_device_capability(100):
|
if not current_platform.has_device_capability(100):
|
||||||
pytest.skip(
|
pytest.skip(
|
||||||
@@ -42,7 +43,7 @@ MNK_FACTORS = [
|
|||||||
def test_cutlass_fp4_moe_no_graph(
|
def test_cutlass_fp4_moe_no_graph(
|
||||||
m: int, n: int, k: int, e: int, topk: int, dtype: torch.dtype, workspace_init
|
m: int, n: int, k: int, e: int, topk: int, dtype: torch.dtype, workspace_init
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
with set_current_vllm_config(
|
with set_current_vllm_config(
|
||||||
VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
|
VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
|
||||||
):
|
):
|
||||||
|
|||||||
@@ -14,6 +14,7 @@ from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
|
|||||||
from vllm.model_executor.layers.fused_moe.modular_kernel import FusedMoEModularKernel
|
from vllm.model_executor.layers.fused_moe.modular_kernel import FusedMoEModularKernel
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.utils.math_utils import cdiv
|
from vllm.utils.math_utils import cdiv
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
from ...utils import multi_gpu_test
|
from ...utils import multi_gpu_test
|
||||||
from .parallel_utils import ProcessGroupInfo, parallel_launch
|
from .parallel_utils import ProcessGroupInfo, parallel_launch
|
||||||
@@ -290,7 +291,7 @@ def test_cutlass_moe_pplx(
|
|||||||
world_dp_size: tuple[int, int],
|
world_dp_size: tuple[int, int],
|
||||||
use_internode: bool,
|
use_internode: bool,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
|
|
||||||
with set_current_vllm_config(vllm_config):
|
with set_current_vllm_config(vllm_config):
|
||||||
dtype = torch.half
|
dtype = torch.half
|
||||||
|
|||||||
@@ -44,8 +44,8 @@ from vllm.model_executor.layers.fused_moe.modular_kernel import FusedMoEModularK
|
|||||||
from vllm.model_executor.layers.fused_moe.topk_weight_and_reduce import (
|
from vllm.model_executor.layers.fused_moe.topk_weight_and_reduce import (
|
||||||
TopKWeightAndReduceDelegate,
|
TopKWeightAndReduceDelegate,
|
||||||
)
|
)
|
||||||
from vllm.platforms import current_platform
|
|
||||||
from vllm.utils.math_utils import round_up
|
from vllm.utils.math_utils import round_up
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
from vllm.v1.worker.workspace import init_workspace_manager
|
from vllm.v1.worker.workspace import init_workspace_manager
|
||||||
|
|
||||||
from ...utils import multi_gpu_test
|
from ...utils import multi_gpu_test
|
||||||
@@ -184,7 +184,7 @@ def test_fused_moe_batched_experts(
|
|||||||
dtype: torch.dtype,
|
dtype: torch.dtype,
|
||||||
workspace_init,
|
workspace_init,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
|
|
||||||
a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
|
a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
|
||||||
w1 = torch.randn((e, 2 * n, k), device="cuda", dtype=dtype) / 10
|
w1 = torch.randn((e, 2 * n, k), device="cuda", dtype=dtype) / 10
|
||||||
@@ -491,7 +491,7 @@ def test_pplx_prepare_finalize_slow(
|
|||||||
if per_act_token_quant and block_shape is not None:
|
if per_act_token_quant and block_shape is not None:
|
||||||
pytest.skip("Skip illegal quantization combination")
|
pytest.skip("Skip illegal quantization combination")
|
||||||
|
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
m, n, k = mnk
|
m, n, k = mnk
|
||||||
world_size, dp_size = world_dp_size
|
world_size, dp_size = world_dp_size
|
||||||
device = "cuda"
|
device = "cuda"
|
||||||
@@ -809,7 +809,7 @@ def test_pplx_moe_slow(
|
|||||||
block_shape: list[int] | None,
|
block_shape: list[int] | None,
|
||||||
use_internode: bool,
|
use_internode: bool,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
m, n, k = mnk
|
m, n, k = mnk
|
||||||
world_size, dp_size = world_dp_size
|
world_size, dp_size = world_dp_size
|
||||||
|
|
||||||
@@ -888,7 +888,7 @@ def _pplx_test_loop(
|
|||||||
new_vllm_config.parallel_config.enable_expert_parallel = True
|
new_vllm_config.parallel_config.enable_expert_parallel = True
|
||||||
_set_vllm_config(new_vllm_config, pgi.world_size, pgi.rank, pgi.local_rank)
|
_set_vllm_config(new_vllm_config, pgi.world_size, pgi.rank, pgi.local_rank)
|
||||||
|
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
combos = itertools.product(
|
combos = itertools.product(
|
||||||
PPLX_COMBOS, NUM_EXPERTS, TOP_KS, DTYPES, [False, True], [None, [128, 128]]
|
PPLX_COMBOS, NUM_EXPERTS, TOP_KS, DTYPES, [False, True], [None, [128, 128]]
|
||||||
)
|
)
|
||||||
@@ -982,7 +982,7 @@ def test_pplx_prepare_finalize(
|
|||||||
world_dp_size: tuple[int, int],
|
world_dp_size: tuple[int, int],
|
||||||
use_internode: bool,
|
use_internode: bool,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
world_size, dp_size = world_dp_size
|
world_size, dp_size = world_dp_size
|
||||||
parallel_launch(
|
parallel_launch(
|
||||||
world_size * dp_size,
|
world_size * dp_size,
|
||||||
@@ -1005,7 +1005,7 @@ def test_pplx_moe(
|
|||||||
use_internode: bool,
|
use_internode: bool,
|
||||||
use_shared_experts: bool,
|
use_shared_experts: bool,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(7)
|
set_random_seed(7)
|
||||||
world_size, dp_size = world_dp_size
|
world_size, dp_size = world_dp_size
|
||||||
parallel_launch(
|
parallel_launch(
|
||||||
world_size,
|
world_size,
|
||||||
|
|||||||
@@ -13,6 +13,7 @@ from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
|
|||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.utils.deep_gemm import DeepGemmQuantScaleFMT, has_deep_gemm
|
from vllm.utils.deep_gemm import DeepGemmQuantScaleFMT, has_deep_gemm
|
||||||
from vllm.utils.math_utils import cdiv, round_up
|
from vllm.utils.math_utils import cdiv, round_up
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
if current_platform.is_fp8_fnuz():
|
if current_platform.is_fp8_fnuz():
|
||||||
pytest.skip(
|
pytest.skip(
|
||||||
@@ -201,7 +202,7 @@ def token_random(E, T, H2, tokens_per_expert):
|
|||||||
@torch.inference_mode()
|
@torch.inference_mode()
|
||||||
def test_silu_mul_fp8_quant_deep_gemm(E: int, T: int, H: int, fp8_type: torch.dtype):
|
def test_silu_mul_fp8_quant_deep_gemm(E: int, T: int, H: int, fp8_type: torch.dtype):
|
||||||
group_size = 128
|
group_size = 128
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
|
|
||||||
tokens_per_expert = torch.randint(
|
tokens_per_expert = torch.randint(
|
||||||
low=0,
|
low=0,
|
||||||
|
|||||||
@@ -11,6 +11,7 @@ from vllm.model_executor.layers.quantization.utils.fp8_utils import (
|
|||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.triton_utils import triton
|
from vllm.triton_utils import triton
|
||||||
from vllm.utils.deep_gemm import is_deep_gemm_e8m0_used
|
from vllm.utils.deep_gemm import is_deep_gemm_e8m0_used
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
FLOAT8_DTYPE = torch.float8_e4m3fn
|
FLOAT8_DTYPE = torch.float8_e4m3fn
|
||||||
GROUP_SIZE = 128
|
GROUP_SIZE = 128
|
||||||
@@ -72,7 +73,7 @@ def reference(x: torch.Tensor, use_ue8m0: bool) -> tuple[torch.Tensor, torch.Ten
|
|||||||
reason="ROCm does not support DeepGemm.",
|
reason="ROCm does not support DeepGemm.",
|
||||||
)
|
)
|
||||||
def test_silu_mul_fp8_quant_deep_gemm(T: int, N: int):
|
def test_silu_mul_fp8_quant_deep_gemm(T: int, N: int):
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
|
|
||||||
input = torch.rand((T, N), dtype=torch.bfloat16, device="cuda")
|
input = torch.rand((T, N), dtype=torch.bfloat16, device="cuda")
|
||||||
|
|
||||||
|
|||||||
@@ -13,7 +13,7 @@ from vllm.model_executor.layers.quantization.awq_triton import (
|
|||||||
awq_dequantize_triton,
|
awq_dequantize_triton,
|
||||||
awq_gemm_triton,
|
awq_gemm_triton,
|
||||||
)
|
)
|
||||||
from vllm.platforms import current_platform
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
device = "cuda"
|
device = "cuda"
|
||||||
|
|
||||||
@@ -86,7 +86,7 @@ def test_dequantize(qweight_rows, qweight_cols, group_size):
|
|||||||
zeros_cols = qweight_cols
|
zeros_cols = qweight_cols
|
||||||
zeros_dtype = torch.int32
|
zeros_dtype = torch.int32
|
||||||
|
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
|
|
||||||
qweight = torch.randint(
|
qweight = torch.randint(
|
||||||
0,
|
0,
|
||||||
@@ -141,7 +141,7 @@ def test_gemm(N, K, M, splitK, group_size):
|
|||||||
qzeros_rows = scales_rows
|
qzeros_rows = scales_rows
|
||||||
qzeros_cols = qweight_cols
|
qzeros_cols = qweight_cols
|
||||||
|
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
|
|
||||||
input = torch.rand((input_rows, input_cols), dtype=input_dtype, device=device)
|
input = torch.rand((input_rows, input_cols), dtype=input_dtype, device=device)
|
||||||
qweight = torch.randint(
|
qweight = torch.randint(
|
||||||
|
|||||||
@@ -17,6 +17,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
|
|||||||
)
|
)
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.scalar_type import ScalarType, scalar_types
|
from vllm.scalar_type import ScalarType, scalar_types
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
IS_SUPPORTED_BY_GPU = (
|
IS_SUPPORTED_BY_GPU = (
|
||||||
current_platform.is_cuda() and current_platform.get_device_capability()[0] >= 9
|
current_platform.is_cuda() and current_platform.get_device_capability()[0] >= 9
|
||||||
@@ -248,7 +249,7 @@ def compute_moe_reference_output(setup: MoETestSetup) -> torch.Tensor:
|
|||||||
@pytest.mark.parametrize("random_zero", [True, False])
|
@pytest.mark.parametrize("random_zero", [True, False])
|
||||||
def test_cutlass_w4a8_moe_mm_end_to_end(shape, random_zero):
|
def test_cutlass_w4a8_moe_mm_end_to_end(shape, random_zero):
|
||||||
num_experts, N, K = shape
|
num_experts, N, K = shape
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
setup = make_moe_test_setup(
|
setup = make_moe_test_setup(
|
||||||
num_experts=num_experts, K=K, N=N, max_blocks=64, random_zero=random_zero
|
num_experts=num_experts, K=K, N=N, max_blocks=64, random_zero=random_zero
|
||||||
)
|
)
|
||||||
@@ -308,7 +309,7 @@ class W4A8MoELayer(torch.nn.Module):
|
|||||||
reason="W4A8 Grouped GEMM is not supported on this GPU type.",
|
reason="W4A8 Grouped GEMM is not supported on this GPU type.",
|
||||||
)
|
)
|
||||||
def test_cutlass_w4a8_moe_mm_cuda_graph():
|
def test_cutlass_w4a8_moe_mm_cuda_graph():
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
# Fixed config for CUDA graph test (single parameter point).
|
# Fixed config for CUDA graph test (single parameter point).
|
||||||
num_experts = 8
|
num_experts = 8
|
||||||
K = 512
|
K = 512
|
||||||
|
|||||||
@@ -12,6 +12,7 @@ from nvfp4_utils import (
|
|||||||
from vllm import _custom_ops as ops
|
from vllm import _custom_ops as ops
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.utils.flashinfer import flashinfer_scaled_fp4_mm
|
from vllm.utils.flashinfer import flashinfer_scaled_fp4_mm
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
if not current_platform.has_device_capability(100):
|
if not current_platform.has_device_capability(100):
|
||||||
pytest.skip(
|
pytest.skip(
|
||||||
@@ -72,7 +73,7 @@ def test_flashinfer_nvfp4_gemm(
|
|||||||
if backend == "trtllm" and dtype == torch.float16:
|
if backend == "trtllm" and dtype == torch.float16:
|
||||||
pytest.skip("Only torch.bfloat16 is supported for TRTLLM FP4 GEMM operations")
|
pytest.skip("Only torch.bfloat16 is supported for TRTLLM FP4 GEMM operations")
|
||||||
|
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
m, n, packed_k = shape
|
m, n, packed_k = shape
|
||||||
k = packed_k * 2
|
k = packed_k * 2
|
||||||
block_size = 16
|
block_size = 16
|
||||||
|
|||||||
@@ -6,6 +6,7 @@ import torch
|
|||||||
from vllm import _custom_ops as ops
|
from vllm import _custom_ops as ops
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.utils.flashinfer import flashinfer_scaled_fp8_mm
|
from vllm.utils.flashinfer import flashinfer_scaled_fp8_mm
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
if not current_platform.has_device_capability(100):
|
if not current_platform.has_device_capability(100):
|
||||||
pytest.skip(
|
pytest.skip(
|
||||||
@@ -38,7 +39,7 @@ def test_flashinfer_fp8_gemm(
|
|||||||
device: str,
|
device: str,
|
||||||
autotune: bool,
|
autotune: bool,
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
m, n, k = shape
|
m, n, k = shape
|
||||||
a = torch.randn((m, k), dtype=dtype, device=device)
|
a = torch.randn((m, k), dtype=dtype, device=device)
|
||||||
b = torch.randn((n, k), dtype=dtype, device=device) / k
|
b = torch.randn((n, k), dtype=dtype, device=device) / k
|
||||||
|
|||||||
@@ -11,7 +11,7 @@ from tests.kernels.quant_utils import (
|
|||||||
ref_dynamic_per_token_quant,
|
ref_dynamic_per_token_quant,
|
||||||
)
|
)
|
||||||
from tests.kernels.utils import opcheck
|
from tests.kernels.utils import opcheck
|
||||||
from vllm.platforms import current_platform
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
DTYPES = [torch.bfloat16, torch.float]
|
DTYPES = [torch.bfloat16, torch.float]
|
||||||
HIDDEN_SIZES = [17, 1024, 1025, 1026, 5137, 8193]
|
HIDDEN_SIZES = [17, 1024, 1025, 1026, 5137, 8193]
|
||||||
@@ -51,7 +51,7 @@ def opcheck_fp8_quant(
|
|||||||
def test_dynamic_per_token_fp8_quant(
|
def test_dynamic_per_token_fp8_quant(
|
||||||
num_tokens: int, hidden_size: int, dtype: torch.dtype, scale_ub: bool, seed: int
|
num_tokens: int, hidden_size: int, dtype: torch.dtype, scale_ub: bool, seed: int
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
|
|
||||||
x = (
|
x = (
|
||||||
torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda") + 1e-6
|
torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda") + 1e-6
|
||||||
@@ -81,7 +81,7 @@ def test_dynamic_per_token_fp8_quant(
|
|||||||
def test_dynamic_per_tensor_fp8_quant(
|
def test_dynamic_per_tensor_fp8_quant(
|
||||||
num_tokens: int, hidden_size: int, dtype: torch.dtype, seed: int
|
num_tokens: int, hidden_size: int, dtype: torch.dtype, seed: int
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
|
|
||||||
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda")
|
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda")
|
||||||
|
|
||||||
@@ -101,7 +101,7 @@ def test_dynamic_per_tensor_fp8_quant(
|
|||||||
@torch.inference_mode()
|
@torch.inference_mode()
|
||||||
@pytest.mark.parametrize("seed", SEEDS)
|
@pytest.mark.parametrize("seed", SEEDS)
|
||||||
def test_fp8_quant_large(seed: int) -> None:
|
def test_fp8_quant_large(seed: int) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
|
|
||||||
num_tokens = 1024000 # Mistral-Nemo's max_position_embeddings
|
num_tokens = 1024000 # Mistral-Nemo's max_position_embeddings
|
||||||
hidden_size = 1152 # Smallest hidden_size to reproduce the error
|
hidden_size = 1152 # Smallest hidden_size to reproduce the error
|
||||||
|
|||||||
@@ -7,7 +7,7 @@ import torch
|
|||||||
|
|
||||||
from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
|
from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
|
||||||
from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
|
from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
|
||||||
from vllm.platforms import current_platform
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
|
|
||||||
@pytest.mark.parametrize(
|
@pytest.mark.parametrize(
|
||||||
@@ -30,7 +30,7 @@ def test_quantfp8_group_functionality(
|
|||||||
Tests both CUDA and native implementations, column-major scales,
|
Tests both CUDA and native implementations, column-major scales,
|
||||||
and verifies consistency between implementations.
|
and verifies consistency between implementations.
|
||||||
"""
|
"""
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
|
|
||||||
x = torch.randn((batch_size, hidden_dim), dtype=torch.bfloat16, device="cuda") * 8
|
x = torch.randn((batch_size, hidden_dim), dtype=torch.bfloat16, device="cuda") * 8
|
||||||
expected_num_groups = (hidden_dim + group_size - 1) // group_size
|
expected_num_groups = (hidden_dim + group_size - 1) // group_size
|
||||||
@@ -83,7 +83,7 @@ def test_quantfp8_group_functionality(
|
|||||||
@pytest.mark.parametrize("use_ue8m0", [True, False])
|
@pytest.mark.parametrize("use_ue8m0", [True, False])
|
||||||
@torch.inference_mode()
|
@torch.inference_mode()
|
||||||
def test_quantfp8_group_multidimensional(seed: int, use_ue8m0: bool) -> None:
|
def test_quantfp8_group_multidimensional(seed: int, use_ue8m0: bool) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
|
|
||||||
group_size = 64
|
group_size = 64
|
||||||
|
|
||||||
@@ -136,7 +136,7 @@ def test_quantfp8_group_multidimensional(seed: int, use_ue8m0: bool) -> None:
|
|||||||
@pytest.mark.parametrize("seed", [42])
|
@pytest.mark.parametrize("seed", [42])
|
||||||
@torch.inference_mode()
|
@torch.inference_mode()
|
||||||
def test_quantfp8_group_edge_cases(seed: int) -> None:
|
def test_quantfp8_group_edge_cases(seed: int) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
|
|
||||||
batch_size = 16
|
batch_size = 16
|
||||||
group_size = 64
|
group_size = 64
|
||||||
|
|||||||
@@ -11,7 +11,7 @@ from huggingface_hub import snapshot_download
|
|||||||
import vllm._custom_ops as ops
|
import vllm._custom_ops as ops
|
||||||
from vllm.model_executor.layers.fused_moe import fused_experts
|
from vllm.model_executor.layers.fused_moe import fused_experts
|
||||||
from vllm.model_executor.layers.quantization.gguf import _fused_moe_gguf
|
from vllm.model_executor.layers.quantization.gguf import _fused_moe_gguf
|
||||||
from vllm.platforms import current_platform
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
GGUF_SAMPLE = snapshot_download("Isotr0py/test-gguf-sample")
|
GGUF_SAMPLE = snapshot_download("Isotr0py/test-gguf-sample")
|
||||||
GGUF_SAMPLE_MOE = snapshot_download("SzymonOzog/test-gguf-moe-sample")
|
GGUF_SAMPLE_MOE = snapshot_download("SzymonOzog/test-gguf-moe-sample")
|
||||||
@@ -91,7 +91,7 @@ def test_dequantize(
|
|||||||
@pytest.mark.parametrize("quant_type", QUANT_TYPES)
|
@pytest.mark.parametrize("quant_type", QUANT_TYPES)
|
||||||
@torch.inference_mode()
|
@torch.inference_mode()
|
||||||
def test_mmvq(hidden_size: int, dtype: torch.dtype, quant_type: GGMLQuantizationType):
|
def test_mmvq(hidden_size: int, dtype: torch.dtype, quant_type: GGMLQuantizationType):
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
|
|
||||||
tensors = get_gguf_sample_tensors(hidden_size, quant_type)
|
tensors = get_gguf_sample_tensors(hidden_size, quant_type)
|
||||||
x = torch.rand((1, hidden_size), dtype=dtype, device="cuda")
|
x = torch.rand((1, hidden_size), dtype=dtype, device="cuda")
|
||||||
@@ -134,7 +134,7 @@ def test_mmq(
|
|||||||
dtype: torch.dtype,
|
dtype: torch.dtype,
|
||||||
quant_type: GGMLQuantizationType,
|
quant_type: GGMLQuantizationType,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
|
|
||||||
tensors = get_gguf_sample_tensors(hidden_size, quant_type)
|
tensors = get_gguf_sample_tensors(hidden_size, quant_type)
|
||||||
x = torch.rand((num_tokens, hidden_size), dtype=dtype, device="cuda")
|
x = torch.rand((num_tokens, hidden_size), dtype=dtype, device="cuda")
|
||||||
@@ -169,7 +169,7 @@ def test_moe(
|
|||||||
quant_type: GGMLQuantizationType,
|
quant_type: GGMLQuantizationType,
|
||||||
top_k: int,
|
top_k: int,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
H, E = 1024, 256
|
H, E = 1024, 256
|
||||||
|
|
||||||
x = torch.rand((num_tokens, H), dtype=dtype, device="cuda")
|
x = torch.rand((num_tokens, H), dtype=dtype, device="cuda")
|
||||||
|
|||||||
@@ -7,7 +7,7 @@ import torch
|
|||||||
from tests.kernels.quant_utils import ref_dynamic_per_token_quant
|
from tests.kernels.quant_utils import ref_dynamic_per_token_quant
|
||||||
from tests.kernels.utils import opcheck
|
from tests.kernels.utils import opcheck
|
||||||
from vllm._custom_ops import scaled_int8_quant
|
from vllm._custom_ops import scaled_int8_quant
|
||||||
from vllm.platforms import current_platform
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
DTYPES = [torch.bfloat16, torch.float]
|
DTYPES = [torch.bfloat16, torch.float]
|
||||||
HIDDEN_SIZES = [17, 1024, 1025, 1026, 5137, 8193]
|
HIDDEN_SIZES = [17, 1024, 1025, 1026, 5137, 8193]
|
||||||
@@ -46,7 +46,7 @@ def opcheck_int8_quant_dynamic(output, input, symmetric=True):
|
|||||||
def test_dynamic_scaled_int8_quant(
|
def test_dynamic_scaled_int8_quant(
|
||||||
num_tokens: int, hidden_size: int, dtype: torch.dtype, seed: int
|
num_tokens: int, hidden_size: int, dtype: torch.dtype, seed: int
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
|
|
||||||
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda") * 1000
|
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda") * 1000
|
||||||
|
|
||||||
@@ -70,7 +70,7 @@ def test_dynamic_scaled_int8_quant(
|
|||||||
def test_dynamic_scaled_int8_azp_quant(
|
def test_dynamic_scaled_int8_azp_quant(
|
||||||
num_tokens: int, hidden_size: int, dtype: torch.dtype, seed: int
|
num_tokens: int, hidden_size: int, dtype: torch.dtype, seed: int
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
int8_traits = torch.iinfo(torch.int8)
|
int8_traits = torch.iinfo(torch.int8)
|
||||||
|
|
||||||
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda") * 1000 - 300
|
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda") * 1000 - 300
|
||||||
@@ -111,7 +111,7 @@ def test_dynamic_scaled_int8_azp_quant(
|
|||||||
def test_static_scaled_int8_quant(
|
def test_static_scaled_int8_quant(
|
||||||
num_tokens: int, hidden_size: int, dtype: torch.dtype, seed: int, scale: float
|
num_tokens: int, hidden_size: int, dtype: torch.dtype, seed: int, scale: float
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
int8_traits = torch.iinfo(torch.int8)
|
int8_traits = torch.iinfo(torch.int8)
|
||||||
|
|
||||||
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda") * 1000
|
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda") * 1000
|
||||||
@@ -144,7 +144,7 @@ def test_static_scaled_int8_azp_quant(
|
|||||||
scale: float,
|
scale: float,
|
||||||
azp: int,
|
azp: int,
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
int8_traits = torch.iinfo(torch.int8)
|
int8_traits = torch.iinfo(torch.int8)
|
||||||
|
|
||||||
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda") * 1000 - 300
|
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda") * 1000 - 300
|
||||||
|
|||||||
@@ -24,6 +24,7 @@ from compressed_tensors.transform.utils.hadamard import deterministic_hadamard_m
|
|||||||
from vllm._custom_ops import fusedQuantizeMx, matmul_mxf4_bf16_tn
|
from vllm._custom_ops import fusedQuantizeMx, matmul_mxf4_bf16_tn
|
||||||
from vllm.model_executor.layers.quantization.qutlass_utils import to_blocked
|
from vllm.model_executor.layers.quantization.qutlass_utils import to_blocked
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
if not torch.cuda.is_available():
|
if not torch.cuda.is_available():
|
||||||
pytest.skip("CUDA required for these tests.", allow_module_level=True)
|
pytest.skip("CUDA required for these tests.", allow_module_level=True)
|
||||||
@@ -205,7 +206,7 @@ LLAMA_MODELS = {
|
|||||||
|
|
||||||
@pytest.fixture(autouse=True)
|
@pytest.fixture(autouse=True)
|
||||||
def _seed_each_test():
|
def _seed_each_test():
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
np.random.seed(0)
|
np.random.seed(0)
|
||||||
torch.random.manual_seed(0)
|
torch.random.manual_seed(0)
|
||||||
|
|
||||||
|
|||||||
@@ -6,6 +6,7 @@ import torch
|
|||||||
from vllm import _custom_ops as ops
|
from vllm import _custom_ops as ops
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.scalar_type import scalar_types
|
from vllm.scalar_type import scalar_types
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
if not current_platform.has_device_capability(100):
|
if not current_platform.has_device_capability(100):
|
||||||
pytest.skip(
|
pytest.skip(
|
||||||
@@ -134,7 +135,7 @@ def test_quantize_to_fp4(
|
|||||||
seed: int,
|
seed: int,
|
||||||
device: str,
|
device: str,
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
|
|
||||||
m, n = shape
|
m, n = shape
|
||||||
@@ -156,7 +157,7 @@ def test_quantize_to_fp4(
|
|||||||
@torch.inference_mode()
|
@torch.inference_mode()
|
||||||
def test_quantize_to_fp4_padded(pad_shape: tuple[int, int]) -> None:
|
def test_quantize_to_fp4_padded(pad_shape: tuple[int, int]) -> None:
|
||||||
dtype = torch.float16
|
dtype = torch.float16
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
torch.set_default_device("cuda:0")
|
torch.set_default_device("cuda:0")
|
||||||
|
|
||||||
m, n = pad_shape
|
m, n = pad_shape
|
||||||
|
|||||||
@@ -25,6 +25,7 @@ from vllm import _custom_ops as ops # use existing nvfp4 gemm in vllm
|
|||||||
from vllm._custom_ops import fusedQuantizeNv
|
from vllm._custom_ops import fusedQuantizeNv
|
||||||
from vllm.model_executor.layers.quantization.qutlass_utils import to_blocked
|
from vllm.model_executor.layers.quantization.qutlass_utils import to_blocked
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
if not torch.cuda.is_available():
|
if not torch.cuda.is_available():
|
||||||
pytest.skip("CUDA required for these tests.", allow_module_level=True)
|
pytest.skip("CUDA required for these tests.", allow_module_level=True)
|
||||||
@@ -193,7 +194,7 @@ LLAMA_MODELS = {
|
|||||||
|
|
||||||
@pytest.fixture(autouse=True)
|
@pytest.fixture(autouse=True)
|
||||||
def _seed_each_test():
|
def _seed_each_test():
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
np.random.seed(0)
|
np.random.seed(0)
|
||||||
torch.random.manual_seed(0)
|
torch.random.manual_seed(0)
|
||||||
|
|
||||||
|
|||||||
@@ -6,6 +6,7 @@ from nvfp4_utils import FLOAT4_E2M1_MAX, FLOAT8_E4M3_MAX, dequantize_nvfp4_to_dt
|
|||||||
|
|
||||||
from vllm import _custom_ops as ops
|
from vllm import _custom_ops as ops
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
if not current_platform.has_device_capability(100):
|
if not current_platform.has_device_capability(100):
|
||||||
pytest.skip(
|
pytest.skip(
|
||||||
@@ -59,7 +60,7 @@ def test_nvfp4_gemm(
|
|||||||
seed: int,
|
seed: int,
|
||||||
device: str,
|
device: str,
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
m, n, packed_k = shape
|
m, n, packed_k = shape
|
||||||
k = packed_k * 2
|
k = packed_k * 2
|
||||||
block_size = 16
|
block_size = 16
|
||||||
|
|||||||
@@ -11,6 +11,7 @@ from tests.kernels.quantization.nvfp4_utils import (
|
|||||||
from vllm._custom_ops import scaled_fp4_quant
|
from vllm._custom_ops import scaled_fp4_quant
|
||||||
from vllm.model_executor.layers.activation import SiluAndMul
|
from vllm.model_executor.layers.activation import SiluAndMul
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
if not current_platform.has_device_capability(100):
|
if not current_platform.has_device_capability(100):
|
||||||
pytest.skip(
|
pytest.skip(
|
||||||
@@ -33,7 +34,7 @@ def test_silu_mul_nvfp4_quant(
|
|||||||
dtype: torch.dtype,
|
dtype: torch.dtype,
|
||||||
shape: tuple[int, int],
|
shape: tuple[int, int],
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
device = "cuda:0"
|
device = "cuda:0"
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
|
|
||||||
|
|||||||
@@ -11,6 +11,7 @@ import pytest
|
|||||||
import torch
|
import torch
|
||||||
|
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
device = "cuda"
|
device = "cuda"
|
||||||
|
|
||||||
@@ -85,7 +86,7 @@ def test_scaled_mm(
|
|||||||
):
|
):
|
||||||
is_floating_point_type = lambda t: torch.tensor([1, 1], dtype=t).is_floating_point()
|
is_floating_point_type = lambda t: torch.tensor([1, 1], dtype=t).is_floating_point()
|
||||||
|
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
|
|
||||||
# NOTE: There are cases, where if the matrix is large enough, an output
|
# NOTE: There are cases, where if the matrix is large enough, an output
|
||||||
# like 65504.4 can be produced, and can easily turn into inf when
|
# like 65504.4 can be produced, and can easily turn into inf when
|
||||||
|
|||||||
@@ -9,6 +9,7 @@ from vllm._custom_ops import (
|
|||||||
apply_repetition_penalties_torch,
|
apply_repetition_penalties_torch,
|
||||||
)
|
)
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
NUM_SEQS = [1, 2, 3, 4, 8, 13, 17, 32, 37, 256, 1023, 1024, 1025]
|
NUM_SEQS = [1, 2, 3, 4, 8, 13, 17, 32, 37, 256, 1023, 1024, 1025]
|
||||||
# [stress, stress, stress, Qwen, llama 4]
|
# [stress, stress, stress, Qwen, llama 4]
|
||||||
@@ -38,7 +39,7 @@ def test_apply_repetition_penalties(
|
|||||||
Test the apply_repetition_penalties custom op
|
Test the apply_repetition_penalties custom op
|
||||||
against a reference implementation.
|
against a reference implementation.
|
||||||
"""
|
"""
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
torch.set_default_device("cuda:0")
|
torch.set_default_device("cuda:0")
|
||||||
|
|
||||||
# Create test data
|
# Create test data
|
||||||
@@ -95,7 +96,7 @@ def test_apply_repetition_penalties_zero_seqs() -> None:
|
|||||||
dtype = torch.float32
|
dtype = torch.float32
|
||||||
seed = 0
|
seed = 0
|
||||||
|
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
torch.set_default_device("cuda:0")
|
torch.set_default_device("cuda:0")
|
||||||
|
|
||||||
# Create test data
|
# Create test data
|
||||||
|
|||||||
@@ -10,7 +10,7 @@ from vllm.model_executor.layers.fla.ops.layernorm_guard import (
|
|||||||
layernorm_fn,
|
layernorm_fn,
|
||||||
rms_norm_ref,
|
rms_norm_ref,
|
||||||
)
|
)
|
||||||
from vllm.platforms import current_platform
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
|
|
||||||
def layer_norm_ref(
|
def layer_norm_ref(
|
||||||
@@ -114,7 +114,7 @@ def test_layer_norm_fwd_basic(
|
|||||||
is_rms_norm: bool,
|
is_rms_norm: bool,
|
||||||
) -> None:
|
) -> None:
|
||||||
"""Test basic layer norm forward pass without z (gate) tensor."""
|
"""Test basic layer norm forward pass without z (gate) tensor."""
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
device = torch.device("cuda:0")
|
device = torch.device("cuda:0")
|
||||||
|
|
||||||
# Create inputs
|
# Create inputs
|
||||||
@@ -156,7 +156,7 @@ def test_layer_norm_fwd_with_gate(
|
|||||||
is_rms_norm: bool,
|
is_rms_norm: bool,
|
||||||
) -> None:
|
) -> None:
|
||||||
"""Test layer norm forward pass with z (gate) tensor."""
|
"""Test layer norm forward pass with z (gate) tensor."""
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
device = torch.device("cuda:0")
|
device = torch.device("cuda:0")
|
||||||
|
|
||||||
# Create inputs
|
# Create inputs
|
||||||
@@ -213,7 +213,7 @@ def test_layer_norm_fwd_with_groups(
|
|||||||
f"hidden_size {hidden_size} not divisible by group_size {group_size}"
|
f"hidden_size {hidden_size} not divisible by group_size {group_size}"
|
||||||
)
|
)
|
||||||
|
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
device = torch.device("cuda:0")
|
device = torch.device("cuda:0")
|
||||||
|
|
||||||
# Create inputs
|
# Create inputs
|
||||||
@@ -253,7 +253,7 @@ def test_layer_norm_rows_per_block(
|
|||||||
dtype: torch.dtype,
|
dtype: torch.dtype,
|
||||||
) -> None:
|
) -> None:
|
||||||
"""Test that rows_per_block logic works correctly for various M sizes."""
|
"""Test that rows_per_block logic works correctly for various M sizes."""
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
device = torch.device("cuda:0")
|
device = torch.device("cuda:0")
|
||||||
hidden_size = 1024
|
hidden_size = 1024
|
||||||
|
|
||||||
@@ -278,7 +278,7 @@ def test_layer_norm_rows_per_block(
|
|||||||
def test_strided_input(dtype: torch.dtype) -> None:
|
def test_strided_input(dtype: torch.dtype) -> None:
|
||||||
"""Test that the kernel handles non-contiguous (strided)
|
"""Test that the kernel handles non-contiguous (strided)
|
||||||
inputs correctly."""
|
inputs correctly."""
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
device = torch.device("cuda:0")
|
device = torch.device("cuda:0")
|
||||||
num_tokens = 128
|
num_tokens = 128
|
||||||
hidden_size = 1024
|
hidden_size = 1024
|
||||||
@@ -318,7 +318,7 @@ def test_output_buffer_provided(
|
|||||||
dtype: torch.dtype,
|
dtype: torch.dtype,
|
||||||
) -> None:
|
) -> None:
|
||||||
"""Test that the kernel works when an output buffer is provided."""
|
"""Test that the kernel works when an output buffer is provided."""
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
device = torch.device("cuda:0")
|
device = torch.device("cuda:0")
|
||||||
|
|
||||||
# Create inputs
|
# Create inputs
|
||||||
@@ -359,7 +359,7 @@ def test_multidimensional_input(
|
|||||||
dtype: torch.dtype,
|
dtype: torch.dtype,
|
||||||
) -> None:
|
) -> None:
|
||||||
"""Test that the autograd function handles multidimensional inputs."""
|
"""Test that the autograd function handles multidimensional inputs."""
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
device = torch.device("cuda:0")
|
device = torch.device("cuda:0")
|
||||||
hidden_size = shape[-1]
|
hidden_size = shape[-1]
|
||||||
|
|
||||||
|
|||||||
@@ -18,8 +18,8 @@ from vllm.distributed.parallel_state import (
|
|||||||
get_tensor_model_parallel_world_size,
|
get_tensor_model_parallel_world_size,
|
||||||
)
|
)
|
||||||
from vllm.lora.ops.triton_ops import fused_moe_lora
|
from vllm.lora.ops.triton_ops import fused_moe_lora
|
||||||
from vllm.platforms import current_platform
|
|
||||||
from vllm.utils.network_utils import get_open_port
|
from vllm.utils.network_utils import get_open_port
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
|
|
||||||
@pytest.fixture(autouse=True)
|
@pytest.fixture(autouse=True)
|
||||||
@@ -265,7 +265,7 @@ def test_fused_moe_lora_kernel(
|
|||||||
seed,
|
seed,
|
||||||
):
|
):
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
# the number of randomly generated sentences.
|
# the number of randomly generated sentences.
|
||||||
num_sequences = 10
|
num_sequences = 10
|
||||||
# generate data
|
# generate data
|
||||||
@@ -358,7 +358,7 @@ def test_fused_moe_lora_kernel_fully_sharded(
|
|||||||
seed,
|
seed,
|
||||||
column_parallel,
|
column_parallel,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
# the number of randomly generated sentences.
|
# the number of randomly generated sentences.
|
||||||
num_sequences = 10
|
num_sequences = 10
|
||||||
# generate data
|
# generate data
|
||||||
@@ -415,7 +415,7 @@ def use_fused_moe_lora_kernel_tensor_parallel(
|
|||||||
def _get_shard_slice(shard_size):
|
def _get_shard_slice(shard_size):
|
||||||
return slice(local_rank * shard_size, (local_rank + 1) * shard_size)
|
return slice(local_rank * shard_size, (local_rank + 1) * shard_size)
|
||||||
|
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
|
|
||||||
device = torch.device(f"cuda:{local_rank}")
|
device = torch.device(f"cuda:{local_rank}")
|
||||||
torch.cuda.set_device(device)
|
torch.cuda.set_device(device)
|
||||||
|
|||||||
@@ -43,8 +43,8 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
|
|||||||
VocabParallelEmbedding,
|
VocabParallelEmbedding,
|
||||||
get_masked_input_and_mask,
|
get_masked_input_and_mask,
|
||||||
)
|
)
|
||||||
from vllm.model_executor.utils import set_random_seed
|
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
from .utils import DummyLoRAManager
|
from .utils import DummyLoRAManager
|
||||||
|
|
||||||
|
|||||||
@@ -9,7 +9,7 @@ import vllm.lora.ops.torch_ops as torch_ops
|
|||||||
import vllm.lora.ops.triton_ops as triton_ops
|
import vllm.lora.ops.triton_ops as triton_ops
|
||||||
from vllm.lora.ops.triton_ops import LoRAKernelMeta
|
from vllm.lora.ops.triton_ops import LoRAKernelMeta
|
||||||
from vllm.lora.ops.triton_ops.utils import _LORA_A_PTR_DICT, _LORA_B_PTR_DICT
|
from vllm.lora.ops.triton_ops.utils import _LORA_A_PTR_DICT, _LORA_B_PTR_DICT
|
||||||
from vllm.platforms import current_platform
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
from .utils import PunicaTensors, assert_close, generate_data_for_nslices
|
from .utils import PunicaTensors, assert_close, generate_data_for_nslices
|
||||||
|
|
||||||
@@ -395,7 +395,7 @@ def test_kernels(
|
|||||||
Tests LoRA kernels.
|
Tests LoRA kernels.
|
||||||
"""
|
"""
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
|
|
||||||
if op_type == "shrink":
|
if op_type == "shrink":
|
||||||
check_lora_shrink_kernel(
|
check_lora_shrink_kernel(
|
||||||
@@ -447,7 +447,7 @@ def test_kernels_hidden_size(
|
|||||||
Tests SGMV and LoRA kernels.
|
Tests SGMV and LoRA kernels.
|
||||||
"""
|
"""
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
|
|
||||||
if op_type == "shrink":
|
if op_type == "shrink":
|
||||||
check_lora_shrink_kernel(
|
check_lora_shrink_kernel(
|
||||||
|
|||||||
@@ -21,6 +21,7 @@ from vllm.model_executor.models.vision import (
|
|||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.utils.network_utils import get_open_port
|
from vllm.utils.network_utils import get_open_port
|
||||||
from vllm.utils.system_utils import update_environment_variables
|
from vllm.utils.system_utils import update_environment_variables
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
pytestmark = pytest.mark.cpu_test
|
pytestmark = pytest.mark.cpu_test
|
||||||
|
|
||||||
@@ -98,7 +99,7 @@ def run_dp_sharded_vision_model_vs_direct(
|
|||||||
"""
|
"""
|
||||||
|
|
||||||
# Set random seed for reproducibility
|
# Set random seed for reproducibility
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
|
|
||||||
device = f"{current_platform.device_name}:{local_rank}"
|
device = f"{current_platform.device_name}:{local_rank}"
|
||||||
current_platform.set_device(device)
|
current_platform.set_device(device)
|
||||||
@@ -284,7 +285,7 @@ def run_dp_sharded_mrope_vision_model_vs_direct(
|
|||||||
calling the model directly.
|
calling the model directly.
|
||||||
"""
|
"""
|
||||||
# Set random seed for reproducibility
|
# Set random seed for reproducibility
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
device = f"{current_platform.device_name}:{local_rank}"
|
device = f"{current_platform.device_name}:{local_rank}"
|
||||||
current_platform.set_device(device)
|
current_platform.set_device(device)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
@@ -408,7 +409,7 @@ def run_dp_sharded_mrope_vision_model_uneven_load_worker(
|
|||||||
):
|
):
|
||||||
"""Test run_dp_sharded_mrope_vision_model with uneven load distribution."""
|
"""Test run_dp_sharded_mrope_vision_model with uneven load distribution."""
|
||||||
# Set up distributed environment
|
# Set up distributed environment
|
||||||
current_platform.seed_everything(123)
|
set_random_seed(123)
|
||||||
device = f"{current_platform.device_name}:{local_rank}"
|
device = f"{current_platform.device_name}:{local_rank}"
|
||||||
current_platform.set_device(device)
|
current_platform.set_device(device)
|
||||||
torch.set_default_device(device)
|
torch.set_default_device(device)
|
||||||
|
|||||||
@@ -19,7 +19,11 @@ from vllm.attention.backends.registry import AttentionBackendEnum
|
|||||||
from vllm.config import ModelConfig
|
from vllm.config import ModelConfig
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.utils.math_utils import cdiv
|
from vllm.utils.math_utils import cdiv
|
||||||
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE, is_torch_equal_or_newer
|
from vllm.utils.torch_utils import (
|
||||||
|
STR_DTYPE_TO_TORCH_DTYPE,
|
||||||
|
is_torch_equal_or_newer,
|
||||||
|
set_random_seed,
|
||||||
|
)
|
||||||
from vllm.v1.attention.backends.utils import (
|
from vllm.v1.attention.backends.utils import (
|
||||||
CommonAttentionMetadata,
|
CommonAttentionMetadata,
|
||||||
set_kv_cache_layout,
|
set_kv_cache_layout,
|
||||||
@@ -320,7 +324,7 @@ def _test_backend_correctness(
|
|||||||
multiple GPUs. This tests that backends work correctly with different
|
multiple GPUs. This tests that backends work correctly with different
|
||||||
head counts.
|
head counts.
|
||||||
"""
|
"""
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
|
|
||||||
hf_config_override = None
|
hf_config_override = None
|
||||||
if tensor_parallel_size > 1:
|
if tensor_parallel_size > 1:
|
||||||
|
|||||||
@@ -7,6 +7,7 @@ import pytest
|
|||||||
import torch
|
import torch
|
||||||
|
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
from vllm.v1.attention.backends.flash_attn import FlashAttentionBackend
|
from vllm.v1.attention.backends.flash_attn import FlashAttentionBackend
|
||||||
from vllm.v1.kv_offload.mediums import CPULoadStoreSpec, GPULoadStoreSpec
|
from vllm.v1.kv_offload.mediums import CPULoadStoreSpec, GPULoadStoreSpec
|
||||||
from vllm.v1.kv_offload.worker.cpu_gpu import CpuGpuOffloadingHandlers
|
from vllm.v1.kv_offload.worker.cpu_gpu import CpuGpuOffloadingHandlers
|
||||||
@@ -62,7 +63,7 @@ def test_transfer(
|
|||||||
seed: int,
|
seed: int,
|
||||||
device: str,
|
device: str,
|
||||||
) -> None:
|
) -> None:
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
|
|
||||||
# create per-layer GPU KV caches based on available attn_backends
|
# create per-layer GPU KV caches based on available attn_backends
|
||||||
attn_backends_list = BACKENDS_TO_TEST
|
attn_backends_list = BACKENDS_TO_TEST
|
||||||
|
|||||||
@@ -15,6 +15,7 @@ import torch_xla.core.xla_model
|
|||||||
from vllm.attention.layers.mm_encoder_attention import MMEncoderAttention
|
from vllm.attention.layers.mm_encoder_attention import MMEncoderAttention
|
||||||
from vllm.attention.selector import _cached_get_attn_backend
|
from vllm.attention.selector import _cached_get_attn_backend
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
|
|
||||||
|
|
||||||
@pytest.fixture(autouse=True)
|
@pytest.fixture(autouse=True)
|
||||||
@@ -63,7 +64,7 @@ def test_mha_attn_forward(
|
|||||||
head_size: int,
|
head_size: int,
|
||||||
device: str,
|
device: str,
|
||||||
):
|
):
|
||||||
current_platform.seed_everything(0)
|
set_random_seed(0)
|
||||||
# These are expected to be f32
|
# These are expected to be f32
|
||||||
q = torch.randn(batch_size, seq_len, num_heads * head_size, device=device)
|
q = torch.randn(batch_size, seq_len, num_heads * head_size, device=device)
|
||||||
k = torch.randn(batch_size, seq_len, num_kv_heads * head_size, device=device)
|
k = torch.randn(batch_size, seq_len, num_kv_heads * head_size, device=device)
|
||||||
|
|||||||
@@ -26,6 +26,7 @@ from vllm.platforms import current_platform
|
|||||||
from vllm.sampling_params import SamplingParams
|
from vllm.sampling_params import SamplingParams
|
||||||
from vllm.utils.mem_constants import GiB_bytes
|
from vllm.utils.mem_constants import GiB_bytes
|
||||||
from vllm.utils.system_utils import update_environment_variables
|
from vllm.utils.system_utils import update_environment_variables
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
from vllm.v1.core.kv_cache_utils import estimate_max_model_len, get_kv_cache_configs
|
from vllm.v1.core.kv_cache_utils import estimate_max_model_len, get_kv_cache_configs
|
||||||
from vllm.v1.core.sched.output import CachedRequestData, NewRequestData, SchedulerOutput
|
from vllm.v1.core.sched.output import CachedRequestData, NewRequestData, SchedulerOutput
|
||||||
from vllm.v1.kv_cache_interface import (
|
from vllm.v1.kv_cache_interface import (
|
||||||
@@ -776,7 +777,7 @@ def test_hybrid_attention_mamba_tensor_shapes():
|
|||||||
will not corrupt an attention block and vice versa
|
will not corrupt an attention block and vice versa
|
||||||
"""
|
"""
|
||||||
|
|
||||||
current_platform.seed_everything(42)
|
set_random_seed(42)
|
||||||
|
|
||||||
update_environment_variables(
|
update_environment_variables(
|
||||||
{
|
{
|
||||||
|
|||||||
@@ -2,10 +2,8 @@
|
|||||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||||
|
|
||||||
from vllm.model_executor.parameter import BasevLLMParameter, PackedvLLMParameter
|
from vllm.model_executor.parameter import BasevLLMParameter, PackedvLLMParameter
|
||||||
from vllm.model_executor.utils import set_random_seed
|
|
||||||
|
|
||||||
__all__ = [
|
__all__ = [
|
||||||
"set_random_seed",
|
|
||||||
"BasevLLMParameter",
|
"BasevLLMParameter",
|
||||||
"PackedvLLMParameter",
|
"PackedvLLMParameter",
|
||||||
]
|
]
|
||||||
|
|||||||
@@ -10,12 +10,6 @@ import torch
|
|||||||
from vllm.utils.torch_utils import is_torch_equal_or_newer
|
from vllm.utils.torch_utils import is_torch_equal_or_newer
|
||||||
|
|
||||||
|
|
||||||
def set_random_seed(seed: int | None) -> None:
|
|
||||||
from vllm.platforms import current_platform
|
|
||||||
|
|
||||||
current_platform.seed_everything(seed)
|
|
||||||
|
|
||||||
|
|
||||||
def set_weight_attrs(
|
def set_weight_attrs(
|
||||||
weight: torch.Tensor,
|
weight: torch.Tensor,
|
||||||
weight_attrs: dict[str, Any] | None,
|
weight_attrs: dict[str, Any] | None,
|
||||||
|
|||||||
@@ -372,6 +372,10 @@ class Platform:
|
|||||||
|
|
||||||
Loosely based on: https://github.com/Lightning-AI/pytorch-lightning/blob/2.4.0/src/lightning/fabric/utilities/seed.py#L20
|
Loosely based on: https://github.com/Lightning-AI/pytorch-lightning/blob/2.4.0/src/lightning/fabric/utilities/seed.py#L20
|
||||||
"""
|
"""
|
||||||
|
logger.info_once(
|
||||||
|
"`seed_everything` is deprecated. It will be removed in v0.14.0 or later. "
|
||||||
|
"Please use `vllm.utils.torch_utils.set_random_seed` instead."
|
||||||
|
)
|
||||||
if seed is not None:
|
if seed is not None:
|
||||||
random.seed(seed)
|
random.seed(seed)
|
||||||
np.random.seed(seed)
|
np.random.seed(seed)
|
||||||
|
|||||||
@@ -3,6 +3,7 @@
|
|||||||
import contextlib
|
import contextlib
|
||||||
import importlib.metadata
|
import importlib.metadata
|
||||||
import os
|
import os
|
||||||
|
import random
|
||||||
import threading
|
import threading
|
||||||
from collections.abc import Callable, Collection
|
from collections.abc import Callable, Collection
|
||||||
from functools import lru_cache
|
from functools import lru_cache
|
||||||
@@ -278,6 +279,13 @@ def kv_cache_dtype_str_to_dtype(
|
|||||||
return STR_DTYPE_TO_TORCH_DTYPE[kv_cache_dtype]
|
return STR_DTYPE_TO_TORCH_DTYPE[kv_cache_dtype]
|
||||||
|
|
||||||
|
|
||||||
|
def set_random_seed(seed: int | None) -> None:
|
||||||
|
if seed is not None:
|
||||||
|
random.seed(seed)
|
||||||
|
np.random.seed(seed)
|
||||||
|
torch.manual_seed(seed)
|
||||||
|
|
||||||
|
|
||||||
def create_kv_caches_with_random_flash(
|
def create_kv_caches_with_random_flash(
|
||||||
num_blocks: int,
|
num_blocks: int,
|
||||||
block_size: int,
|
block_size: int,
|
||||||
@@ -290,9 +298,7 @@ def create_kv_caches_with_random_flash(
|
|||||||
device: str | None = "cuda",
|
device: str | None = "cuda",
|
||||||
cache_layout: str | None = "NHD",
|
cache_layout: str | None = "NHD",
|
||||||
) -> tuple[list[torch.Tensor], list[torch.Tensor]]:
|
) -> tuple[list[torch.Tensor], list[torch.Tensor]]:
|
||||||
from vllm.platforms import current_platform
|
set_random_seed(seed)
|
||||||
|
|
||||||
current_platform.seed_everything(seed)
|
|
||||||
|
|
||||||
dtype = get_kv_cache_torch_dtype(cache_dtype, model_dtype)
|
dtype = get_kv_cache_torch_dtype(cache_dtype, model_dtype)
|
||||||
generic_kv_cache_shape = (num_blocks, 2, block_size, num_heads, head_size)
|
generic_kv_cache_shape = (num_blocks, 2, block_size, num_heads, head_size)
|
||||||
@@ -335,9 +341,8 @@ def create_kv_caches_with_random(
|
|||||||
raise ValueError(
|
raise ValueError(
|
||||||
f"Does not support key cache of type fp8 with head_size {head_size}"
|
f"Does not support key cache of type fp8 with head_size {head_size}"
|
||||||
)
|
)
|
||||||
from vllm.platforms import current_platform
|
|
||||||
|
|
||||||
current_platform.seed_everything(seed)
|
set_random_seed(seed)
|
||||||
|
|
||||||
dtype = get_kv_cache_torch_dtype(cache_dtype, model_dtype)
|
dtype = get_kv_cache_torch_dtype(cache_dtype, model_dtype)
|
||||||
|
|
||||||
|
|||||||
@@ -10,10 +10,10 @@ import torch
|
|||||||
from vllm import envs
|
from vllm import envs
|
||||||
from vllm.config import VllmConfig
|
from vllm.config import VllmConfig
|
||||||
from vllm.logger import init_logger
|
from vllm.logger import init_logger
|
||||||
from vllm.model_executor.utils import set_random_seed
|
|
||||||
from vllm.platforms import CpuArchEnum, current_platform
|
from vllm.platforms import CpuArchEnum, current_platform
|
||||||
from vllm.platforms.cpu import CpuPlatform, LogicalCPUInfo
|
from vllm.platforms.cpu import CpuPlatform, LogicalCPUInfo
|
||||||
from vllm.profiler.wrapper import TorchProfilerWrapper
|
from vllm.profiler.wrapper import TorchProfilerWrapper
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
from vllm.v1.worker.cpu_model_runner import CPUModelRunner
|
from vllm.v1.worker.cpu_model_runner import CPUModelRunner
|
||||||
from vllm.v1.worker.gpu_worker import Worker, init_worker_distributed_environment
|
from vllm.v1.worker.gpu_worker import Worker, init_worker_distributed_environment
|
||||||
|
|
||||||
|
|||||||
@@ -34,7 +34,6 @@ from vllm.distributed.parallel_state import (
|
|||||||
)
|
)
|
||||||
from vllm.logger import init_logger
|
from vllm.logger import init_logger
|
||||||
from vllm.lora.request import LoRARequest
|
from vllm.lora.request import LoRARequest
|
||||||
from vllm.model_executor import set_random_seed
|
|
||||||
from vllm.model_executor.models.interfaces import is_mixture_of_experts
|
from vllm.model_executor.models.interfaces import is_mixture_of_experts
|
||||||
from vllm.model_executor.warmup.kernel_warmup import kernel_warmup
|
from vllm.model_executor.warmup.kernel_warmup import kernel_warmup
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
@@ -43,6 +42,7 @@ from vllm.sequence import IntermediateTensors
|
|||||||
from vllm.tasks import SupportedTask
|
from vllm.tasks import SupportedTask
|
||||||
from vllm.utils.mem_constants import GiB_bytes
|
from vllm.utils.mem_constants import GiB_bytes
|
||||||
from vllm.utils.mem_utils import MemorySnapshot, memory_profiling
|
from vllm.utils.mem_utils import MemorySnapshot, memory_profiling
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
|
from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
|
||||||
from vllm.v1.engine import ReconfigureDistributedRequest, ReconfigureRankType
|
from vllm.v1.engine import ReconfigureDistributedRequest, ReconfigureRankType
|
||||||
from vllm.v1.kv_cache_interface import KVCacheConfig, KVCacheSpec
|
from vllm.v1.kv_cache_interface import KVCacheConfig, KVCacheSpec
|
||||||
|
|||||||
@@ -20,12 +20,11 @@ from vllm.distributed.kv_transfer import (
|
|||||||
)
|
)
|
||||||
from vllm.logger import init_logger
|
from vllm.logger import init_logger
|
||||||
from vllm.lora.request import LoRARequest
|
from vllm.lora.request import LoRARequest
|
||||||
from vllm.model_executor import set_random_seed
|
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.platforms.tpu import USE_TPU_INFERENCE
|
from vllm.platforms.tpu import USE_TPU_INFERENCE
|
||||||
from vllm.tasks import SupportedTask
|
from vllm.tasks import SupportedTask
|
||||||
from vllm.utils.math_utils import cdiv
|
from vllm.utils.math_utils import cdiv
|
||||||
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
|
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE, set_random_seed
|
||||||
from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
|
from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
|
||||||
from vllm.v1.kv_cache_interface import AttentionSpec, KVCacheConfig, KVCacheSpec
|
from vllm.v1.kv_cache_interface import AttentionSpec, KVCacheConfig, KVCacheSpec
|
||||||
from vllm.v1.outputs import ModelRunnerOutput
|
from vllm.v1.outputs import ModelRunnerOutput
|
||||||
|
|||||||
@@ -9,9 +9,9 @@ import torch.distributed
|
|||||||
from vllm.config import VllmConfig
|
from vllm.config import VllmConfig
|
||||||
from vllm.distributed import get_world_group
|
from vllm.distributed import get_world_group
|
||||||
from vllm.logger import init_logger
|
from vllm.logger import init_logger
|
||||||
from vllm.model_executor import set_random_seed
|
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.profiler.wrapper import TorchProfilerWrapper
|
from vllm.profiler.wrapper import TorchProfilerWrapper
|
||||||
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
from vllm.v1.worker.gpu_worker import Worker, init_worker_distributed_environment
|
from vllm.v1.worker.gpu_worker import Worker, init_worker_distributed_environment
|
||||||
from vllm.v1.worker.xpu_model_runner import XPUModelRunner
|
from vllm.v1.worker.xpu_model_runner import XPUModelRunner
|
||||||
|
|
||||||
|
|||||||
Reference in New Issue
Block a user