[Chore] Remove unused batched RoPE op & kernel (#24789)
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon
@@ -1,7 +1,7 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
from itertools import accumulate, product
|
||||
from itertools import product
|
||||
from typing import Callable, Optional
|
||||
|
||||
import pytest
|
||||
@@ -111,151 +111,6 @@ def test_rotary_embedding(
|
||||
"expected returned key to be None"
|
||||
|
||||
|
||||
@pytest.mark.parametrize("is_neox_style", IS_NEOX_STYLE)
|
||||
@pytest.mark.parametrize("tensor_shape_fn", TENSORS_SHAPES_FN)
|
||||
@pytest.mark.parametrize("batch_size", BATCH_SIZES)
|
||||
@pytest.mark.parametrize("seq_len", SEQ_LENS)
|
||||
@pytest.mark.parametrize("num_heads", NUM_HEADS)
|
||||
@pytest.mark.parametrize("head_size", HEAD_SIZES)
|
||||
@pytest.mark.parametrize("rotary_dim", ROTARY_DIMS)
|
||||
@pytest.mark.parametrize("dtype", DTYPES)
|
||||
@pytest.mark.parametrize("seed", SEEDS)
|
||||
@pytest.mark.parametrize("device", CUDA_DEVICES)
|
||||
@pytest.mark.parametrize("use_key", USE_KEY)
|
||||
@torch.inference_mode()
|
||||
def test_batched_rotary_embedding(
|
||||
is_neox_style: bool,
|
||||
tensor_shape_fn: Callable[[int, int, int, int], tuple[int]],
|
||||
batch_size: int,
|
||||
seq_len: int,
|
||||
num_heads: int,
|
||||
head_size: int,
|
||||
rotary_dim: Optional[int],
|
||||
dtype: torch.dtype,
|
||||
seed: int,
|
||||
device: str,
|
||||
use_key: bool,
|
||||
max_position: int = 8192,
|
||||
base: float = 10000,
|
||||
) -> None:
|
||||
current_platform.seed_everything(seed)
|
||||
torch.set_default_device(device)
|
||||
if rotary_dim is None:
|
||||
rotary_dim = head_size
|
||||
rope = get_rope(head_size, rotary_dim, max_position, base, is_neox_style, {
|
||||
"rope_type": "linear",
|
||||
"factor": (1, )
|
||||
})
|
||||
rope = rope.to(dtype=dtype, device=torch.get_default_device())
|
||||
|
||||
positions = torch.randint(0, max_position, (batch_size, seq_len))
|
||||
query_shape = tensor_shape_fn(batch_size, seq_len, num_heads, head_size)
|
||||
query = torch.randn(query_shape, dtype=dtype)
|
||||
key = torch.randn_like(query) if use_key else None
|
||||
|
||||
# slice tensor if required, noop otherwise
|
||||
query = query[..., :head_size]
|
||||
key = key[..., :head_size] if use_key else None
|
||||
|
||||
# NOTE(woosuk): The reference implementation should be executed first
|
||||
# because the custom kernel is in-place.
|
||||
ref_query, ref_key = rope.forward_native(positions, query, key)
|
||||
out_query, out_key = rope.forward(positions,
|
||||
query,
|
||||
key,
|
||||
offsets=torch.zeros(batch_size * seq_len,
|
||||
dtype=torch.long,
|
||||
device=device))
|
||||
# Compare the results.
|
||||
torch.testing.assert_close(out_query,
|
||||
ref_query,
|
||||
atol=get_default_atol(out_query),
|
||||
rtol=get_default_rtol(out_query))
|
||||
if use_key:
|
||||
torch.testing.assert_close(out_key,
|
||||
ref_key,
|
||||
atol=get_default_atol(out_key),
|
||||
rtol=get_default_rtol(out_key))
|
||||
else:
|
||||
assert ref_key is None and out_key is None, \
|
||||
"expected returned key to be None"
|
||||
|
||||
|
||||
@pytest.mark.parametrize("is_neox_style", IS_NEOX_STYLE)
|
||||
@pytest.mark.parametrize("batch_size", BATCH_SIZES)
|
||||
@pytest.mark.parametrize("seq_len", SEQ_LENS)
|
||||
@pytest.mark.parametrize("num_heads", NUM_HEADS)
|
||||
@pytest.mark.parametrize("head_size", HEAD_SIZES)
|
||||
@pytest.mark.parametrize("rotary_dim", ROTARY_DIMS)
|
||||
@pytest.mark.parametrize("dtype", DTYPES)
|
||||
@pytest.mark.parametrize("seed", SEEDS)
|
||||
@pytest.mark.parametrize("device", CUDA_DEVICES)
|
||||
@pytest.mark.parametrize("use_key", USE_KEY)
|
||||
@torch.inference_mode()
|
||||
def test_batched_rotary_embedding_multi_lora(
|
||||
is_neox_style: bool,
|
||||
batch_size: int,
|
||||
seq_len: int,
|
||||
num_heads: int,
|
||||
head_size: int,
|
||||
rotary_dim: Optional[int],
|
||||
dtype: torch.dtype,
|
||||
seed: int,
|
||||
device: str,
|
||||
use_key: bool,
|
||||
max_position: int = 8192,
|
||||
base: float = 10000,
|
||||
) -> None:
|
||||
current_platform.seed_everything(seed)
|
||||
torch.set_default_device(device)
|
||||
if rotary_dim is None:
|
||||
rotary_dim = head_size
|
||||
scaling_factors: list[int] = [1, 2, 4]
|
||||
rope = get_rope(head_size, rotary_dim, max_position, base, is_neox_style, {
|
||||
"rope_type": "linear",
|
||||
"factor": tuple(scaling_factors)
|
||||
})
|
||||
rope = rope.to(dtype=dtype, device=torch.get_default_device())
|
||||
|
||||
positions = torch.randint(0, max_position, (batch_size, seq_len))
|
||||
query = torch.randn(batch_size,
|
||||
seq_len,
|
||||
num_heads * head_size,
|
||||
dtype=dtype)
|
||||
key = torch.randn_like(query) if use_key else None
|
||||
|
||||
offset_map = torch.tensor(
|
||||
list(
|
||||
accumulate([0] + [
|
||||
max_position * scaling_factor * 2
|
||||
for scaling_factor in scaling_factors[:-1]
|
||||
])))
|
||||
query_types = torch.randint(0,
|
||||
len(scaling_factors), (batch_size, seq_len),
|
||||
device=device)
|
||||
query_offsets = offset_map[query_types]
|
||||
|
||||
# NOTE(woosuk): The reference implementation should be executed first
|
||||
# because the custom kernel is in-place.
|
||||
ref_query, ref_key = rope.forward_native(positions, query, key,
|
||||
query_offsets)
|
||||
out_query, out_key = rope.forward(positions, query, key,
|
||||
query_offsets.flatten())
|
||||
# Compare the results.
|
||||
torch.testing.assert_close(out_query,
|
||||
ref_query,
|
||||
atol=get_default_atol(out_query),
|
||||
rtol=get_default_rtol(out_query))
|
||||
if use_key:
|
||||
torch.testing.assert_close(out_key,
|
||||
ref_key,
|
||||
atol=get_default_atol(out_key),
|
||||
rtol=get_default_rtol(out_key))
|
||||
else:
|
||||
assert ref_key is None and out_key is None, \
|
||||
"expected returned key to be None"
|
||||
|
||||
|
||||
@torch.inference_mode()
|
||||
def test_rope_module_cache():
|
||||
MAX_POSITIONS = [123, 1234]
|
||||
|
||||
Reference in New Issue
Block a user