[ROCM][CI] Attention test speedup (#52763)
Signed-off-by: Stefan Koncarevic <stefan.koncarevic@amd.com>
This commit is contained in:
@@ -14,6 +14,8 @@ from vllm.platforms import current_platform
|
||||
from vllm.utils.mem_utils import get_max_shared_memory_bytes
|
||||
from vllm.utils.torch_utils import set_random_seed
|
||||
|
||||
pytestmark = pytest.mark.skip_global_cleanup
|
||||
|
||||
FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
|
||||
# This will change depending on the compute capability.
|
||||
# - 512 as a buffer
|
||||
|
||||
@@ -12,6 +12,8 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import scaled_deq
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.utils.torch_utils import nvfp4_split_data_scale, set_random_seed
|
||||
|
||||
pytestmark = pytest.mark.skip_global_cleanup
|
||||
|
||||
COPYING_DIRECTION = [("cuda", "cpu"), ("cuda", "cuda"), ("cpu", "cuda")]
|
||||
DTYPES = [torch.bfloat16, torch.float]
|
||||
NUM_TOKENS = [42] # Arbitrary values for testing
|
||||
|
||||
@@ -11,6 +11,8 @@ from vllm.platforms import current_platform
|
||||
from vllm.triton_utils import triton
|
||||
from vllm.utils.platform_utils import num_compute_units
|
||||
|
||||
pytestmark = pytest.mark.skip_global_cleanup
|
||||
|
||||
|
||||
def cal_diff(
|
||||
x: torch.Tensor,
|
||||
|
||||
@@ -16,6 +16,8 @@ from vllm.v1.attention.ops.flashmla import (
|
||||
is_flashmla_dense_supported,
|
||||
)
|
||||
|
||||
pytestmark = pytest.mark.skip_global_cleanup
|
||||
|
||||
|
||||
def cal_diff(
|
||||
x: torch.Tensor, y: torch.Tensor, name: str, use_fp8: bool = False
|
||||
|
||||
@@ -15,6 +15,14 @@ from vllm.v1.attention.ops.triton_merge_attn_states import (
|
||||
merge_attn_states as merge_attn_states_triton,
|
||||
)
|
||||
|
||||
pytestmark = [
|
||||
pytest.mark.skip_global_cleanup,
|
||||
pytest.mark.skipif(
|
||||
not current_platform.is_cuda_alike(),
|
||||
reason="merge_attn_states kernels require CUDA or ROCm.",
|
||||
),
|
||||
]
|
||||
|
||||
|
||||
# Naive PyTorch Implements section 2.2 of https://www.arxiv.org/pdf/2501.01005
|
||||
# can be used to combine partial attention results (in the split-KV case)
|
||||
@@ -153,12 +161,6 @@ def test_merge_attn_states(
|
||||
input_dtype: torch.dtype,
|
||||
use_fp8: bool,
|
||||
):
|
||||
if not current_platform.is_cuda():
|
||||
pytest.skip(
|
||||
"Currently only support compare triton merge_attn_states "
|
||||
"with custom cuda merge_attn_states kernel"
|
||||
)
|
||||
|
||||
NUM_TOKENS = num_tokens
|
||||
NUM_HEADS = num_query_heads
|
||||
HEAD_SIZE = head_size
|
||||
|
||||
@@ -17,6 +17,8 @@ from vllm.v1.attention.ops.chunked_prefill_paged_decode import (
|
||||
)
|
||||
from vllm.v1.attention.ops.prefix_prefill import context_attention_fwd
|
||||
|
||||
pytestmark = pytest.mark.skip_global_cleanup
|
||||
|
||||
NUM_HEADS = [64]
|
||||
NUM_QUERIES_PER_KV = [1, 64]
|
||||
HEAD_SIZES = [24, 128]
|
||||
|
||||
@@ -8,6 +8,8 @@ from vllm.platforms import current_platform
|
||||
from vllm.utils.math_utils import cdiv
|
||||
from vllm.v1.attention.ops.triton_decode_attention import decode_attention_fwd
|
||||
|
||||
pytestmark = pytest.mark.skip_global_cleanup
|
||||
|
||||
DEVICE_TYPE = current_platform.device_type
|
||||
|
||||
|
||||
|
||||
@@ -11,6 +11,8 @@ from vllm.utils.torch_utils import set_random_seed
|
||||
from vllm.v1.attention.ops.triton_unified_attention import unified_attention
|
||||
from vllm.v1.kv_cache_interface import KVQuantMode
|
||||
|
||||
pytestmark = pytest.mark.skip_global_cleanup
|
||||
|
||||
DEVICE_TYPE = current_platform.device_type
|
||||
|
||||
NUM_HEADS = [(4, 4), (8, 2), (5, 1)]
|
||||
|
||||
@@ -21,6 +21,8 @@ from vllm.v1.attention.ops.triton_unified_attention_diffkv import (
|
||||
unified_attention_diffkv,
|
||||
)
|
||||
|
||||
pytestmark = pytest.mark.skip_global_cleanup
|
||||
|
||||
DEVICE_TYPE = current_platform.device_type
|
||||
|
||||
# (num_query_heads, num_kv_heads): MHA, GQA, and the num_kv_heads==1
|
||||
|
||||
Reference in New Issue
Block a user