[1/N][Attention] Restructure attention: move files (#31916)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Matthew Bonanni
2026-01-09 13:10:24 -08:00
committed by GitHub
parent 1f8b7c536b
commit 2612ba9285
195 changed files with 426 additions and 396 deletions
@@ -9,10 +9,10 @@ import pytest
from tests.utils import wait_for_gpu_memory_to_clear
from tests.v1.attention.utils import full_cg_backend_configs as backend_configs
from vllm import LLM, SamplingParams
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.config import CompilationConfig
from vllm.platforms import current_platform
from vllm.utils.torch_utils import is_torch_equal_or_newer
from vllm.v1.attention.backends.registry import AttentionBackendEnum
@contextlib.contextmanager
+1 -1
View File
@@ -10,10 +10,10 @@ import torch
from tests.quantization.utils import is_quant_method_supported
from vllm import LLM, SamplingParams
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.config import CompilationConfig, CompilationMode, CUDAGraphMode, PassConfig
from vllm.platforms import current_platform
from vllm.utils.torch_utils import is_torch_equal_or_newer
from vllm.v1.attention.backends.registry import AttentionBackendEnum
from ...utils import create_new_process_for_each_test
+2 -2
View File
@@ -9,8 +9,6 @@ from tests.compile.backend import LazyInitPass, TestBackend
from tests.utils import flat_product
from tests.v1.attention.utils import BatchSpec, create_common_attn_metadata
from vllm._custom_ops import cutlass_scaled_fp4_mm, scaled_fp4_quant
from vllm.attention.backends.abstract import AttentionMetadata
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.attention.layer import Attention
from vllm.compilation.fusion_attn import ATTN_OP, AttnFusionPass
from vllm.compilation.fx_utils import find_op_nodes
@@ -37,6 +35,8 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
from vllm.model_executor.layers.quantization.utils.w8a8_utils import Fp8LinearOp
from vllm.platforms import current_platform
from vllm.utils.flashinfer import has_flashinfer
from vllm.v1.attention.backend import AttentionMetadata
from vllm.v1.attention.backends.registry import AttentionBackendEnum
from vllm.v1.kv_cache_interface import AttentionSpec
FP8_DTYPE = current_platform.fp8_dtype()
+1 -1
View File
@@ -5,7 +5,6 @@ import pytest
import torch
from tests.compile.backend import TestBackend
from vllm.attention.backends.abstract import AttentionType
from vllm.attention.layer import Attention
from vllm.compilation.matcher_utils import FLASHINFER_ROTARY_OP, RMS_OP, ROTARY_OP
from vllm.compilation.noop_elimination import NoOpEliminationPass
@@ -25,6 +24,7 @@ from vllm.config import (
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding
from vllm.platforms import current_platform
from vllm.v1.attention.backend import AttentionType
RSQRT_OP = torch.ops.aten.rsqrt.default
INDEX_SELECT_OP = torch.ops.aten.index.Tensor
+1 -1
View File
@@ -3,8 +3,8 @@
import pytest
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.config.multimodal import MultiModalConfig
from vllm.v1.attention.backends.registry import AttentionBackendEnum
def test_mm_encoder_attn_backend_str_conversion():
+1 -1
View File
@@ -299,7 +299,7 @@ def test_compilation_config():
def test_attention_config():
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.v1.attention.backends.registry import AttentionBackendEnum
parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
@@ -6,9 +6,9 @@ import pytest
import torch
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.platforms import current_platform
from vllm.utils.torch_utils import set_random_seed
from vllm.v1.attention.backends.fa_utils import is_flash_attn_varlen_func_available
NUM_HEADS = [(4, 4), (8, 2)]
HEAD_SIZES = [128, 256]
+2 -2
View File
@@ -10,7 +10,7 @@ from tests.kernels.allclose_default import get_default_atol, get_default_rtol
from tests.kernels.utils import opcheck
from vllm import _custom_ops as ops
from vllm.attention.layer import Attention
from vllm.attention.layers.mm_encoder_attention import MMEncoderAttention
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
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
@@ -30,7 +30,7 @@ NUM_PREFILL_SEQS = [3] # Arbitrary values for testing
NUM_HEADS = [(40, 40), (64, 8)] # Arbitrary values for testing
# This should be sync with get_supported_head_sizes() in
# vllm.attention.ops.paged_attn.PagedAttention
# vllm.v1.attention.ops.paged_attn.PagedAttention
HEAD_SIZES = [32, 80, 128, 256]
BLOCK_SIZES = [16, 32]
@@ -6,13 +6,13 @@ from unittest.mock import patch
import pytest
import torch
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.attention.selector import _cached_get_attn_backend, get_attn_backend
from vllm.config import AttentionConfig, VllmConfig, set_current_vllm_config
from vllm.platforms import current_platform
from vllm.platforms.cpu import CpuPlatform
from vllm.platforms.cuda import CudaPlatform
from vllm.platforms.rocm import RocmPlatform
from vllm.v1.attention.backends.registry import AttentionBackendEnum
from vllm.v1.attention.selector import _cached_get_attn_backend, get_attn_backend
@pytest.fixture(autouse=True)
@@ -182,7 +182,7 @@ def test_backend_selection(
expected = name
assert backend.get_name() == expected
elif name == "FLASH_ATTN_MLA":
from vllm.attention.utils.fa_utils import (
from vllm.v1.attention.backends.fa_utils import (
flash_attn_supports_mla,
)
+1 -1
View File
@@ -270,7 +270,7 @@ def test_reshape_and_cache_flash(
v_scale,
)
elif implementation == "triton":
from vllm.attention.ops.triton_reshape_and_cache_flash import (
from vllm.v1.attention.ops.triton_reshape_and_cache_flash import (
triton_reshape_and_cache_flash,
)
+2 -2
View File
@@ -7,12 +7,12 @@ import random
import pytest
import torch
from vllm.attention.ops.flashmla import (
from vllm.triton_utils import triton
from vllm.v1.attention.ops.flashmla import (
flash_mla_with_kvcache,
get_mla_metadata,
is_flashmla_dense_supported,
)
from vllm.triton_utils import triton
def cal_diff(
@@ -5,7 +5,7 @@ import torch
def test_sparse_flashmla_metadata_smoke():
import vllm.attention.ops.flashmla as fm
import vllm.v1.attention.ops.flashmla as fm
ok, reason = fm.is_flashmla_sparse_supported()
if not ok:
@@ -34,7 +34,7 @@ def test_sparse_flashmla_metadata_smoke():
def test_sparse_flashmla_decode_smoke():
import vllm.attention.ops.flashmla as fm
import vllm.v1.attention.ops.flashmla as fm
ok, reason = fm.is_flashmla_sparse_supported()
if not ok:
@@ -97,7 +97,7 @@ def test_sparse_flashmla_decode_smoke():
def test_sparse_flashmla_prefill_smoke():
import vllm.attention.ops.flashmla as fm
import vllm.v1.attention.ops.flashmla as fm
ok, reason = fm.is_flashmla_sparse_supported()
if not ok:
@@ -5,10 +5,10 @@ import pytest
import torch
from vllm._custom_ops import merge_attn_states as merge_attn_states_cuda
from vllm.attention.ops.triton_merge_attn_states import (
from vllm.platforms import current_platform
from vllm.v1.attention.ops.triton_merge_attn_states import (
merge_attn_states as merge_attn_states_triton,
)
from vllm.platforms import current_platform
# Naive PyTorch Implements section 2.2 of https://www.arxiv.org/pdf/2501.01005
+3 -3
View File
@@ -12,14 +12,14 @@ from unittest.mock import patch
import pytest
import torch
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.attention.layers.mm_encoder_attention import MMEncoderAttention
from vllm.attention.selector import _cached_get_attn_backend
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
from vllm.platforms import current_platform
from vllm.platforms.cpu import CpuPlatform
from vllm.platforms.cuda import CudaPlatform
from vllm.platforms.rocm import RocmPlatform
from vllm.utils.torch_utils import set_random_seed
from vllm.v1.attention.backends.registry import AttentionBackendEnum
from vllm.v1.attention.selector import _cached_get_attn_backend
@pytest.fixture(autouse=True)
@@ -4,7 +4,7 @@
import torch
from torch.testing import assert_close
from vllm.attention.ops.common import pack_seq_triton, unpack_seq_triton
from vllm.v1.attention.ops.common import pack_seq_triton, unpack_seq_triton
def test_pack_seq_basic_fp8():
@@ -10,10 +10,12 @@ import pytest
import torch
import torch.nn.functional as F
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.platforms import current_platform
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE, set_random_seed
from vllm.v1.attention.ops.chunked_prefill_paged_decode import (
chunked_prefill_paged_decode,
)
from vllm.v1.attention.ops.prefix_prefill import context_attention_fwd
NUM_HEADS = [64]
NUM_QUERIES_PER_KV = [1, 64]
@@ -4,10 +4,10 @@
import pytest
import torch
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.attention.selector import _cached_get_attn_backend, get_attn_backend
from vllm.config import AttentionConfig, VllmConfig, set_current_vllm_config
from vllm.platforms.rocm import RocmPlatform
from vllm.v1.attention.backends.registry import AttentionBackendEnum
from vllm.v1.attention.selector import _cached_get_attn_backend, get_attn_backend
@pytest.fixture(autouse=True)
@@ -19,7 +19,7 @@ def clear_cache():
@pytest.mark.skip(reason="Skipped for now. Should be revisited.")
def test_selector(monkeypatch: pytest.MonkeyPatch):
# Set the current platform to ROCm using monkeypatch
monkeypatch.setattr("vllm.attention.selector.current_platform", RocmPlatform())
monkeypatch.setattr("vllm.v1.attention.selector.current_platform", RocmPlatform())
# Test standard ROCm attention
attention_config = AttentionConfig(backend=AttentionBackendEnum.ROCM_ATTN)
@@ -4,8 +4,8 @@
import pytest
import torch
from vllm.attention.ops.triton_decode_attention import decode_attention_fwd
from vllm.utils.math_utils import cdiv
from vllm.v1.attention.ops.triton_decode_attention import decode_attention_fwd
@pytest.mark.parametrize("B", [3, 5])
@@ -5,7 +5,7 @@ import pytest
import torch
import torch.nn.functional as F
from vllm.attention.ops.triton_prefill_attention import context_attention_fwd
from vllm.v1.attention.ops.triton_prefill_attention import context_attention_fwd
def ref_masked_attention(
@@ -5,10 +5,10 @@
import pytest
import torch
from vllm.attention.ops.triton_unified_attention import unified_attention
from vllm.platforms import current_platform
from vllm.utils.math_utils import next_power_of_2
from vllm.utils.torch_utils import set_random_seed
from vllm.v1.attention.ops.triton_unified_attention import unified_attention
NUM_HEADS = [(4, 4), (8, 2)]
HEAD_SIZES = [128, 256]
+1 -1
View File
@@ -13,11 +13,11 @@ import torch
from torch._prims_common import TensorLikeType
from tests.kernels.quant_utils import native_w8a8_block_matmul
from vllm.attention.backends.abstract import AttentionType
from vllm.model_executor.custom_op import CustomOp
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.fused_moe.utils import moe_kernel_quantize_input
from vllm.utils.torch_utils import make_tensor_with_pad
from vllm.v1.attention.backend import AttentionType
# For now, disable "test_aot_dispatch_dynamic" since there are some
# bugs related to this test in PyTorch 2.4.
@@ -14,10 +14,10 @@ import pytest
from transformers import AutoProcessor
from vllm import LLM, EngineArgs, SamplingParams
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.multimodal.utils import encode_image_url
from vllm.multimodal.video import sample_frames_from_video
from vllm.platforms import current_platform
from vllm.v1.attention.backends.registry import AttentionBackendEnum
from ....utils import create_new_process_for_each_test
from ...utils import dummy_hf_overrides
+1 -1
View File
@@ -9,7 +9,7 @@ Note: these tests will only pass on L4 GPU.
import pytest
from tests.quantization.utils import is_quant_method_supported
from vllm.attention.utils.fa_utils import flash_attn_supports_fp8
from vllm.v1.attention.backends.fa_utils import flash_attn_supports_fp8
from vllm.platforms import current_platform
from ..utils import check_logprobs_close
+2 -2
View File
@@ -1,10 +1,10 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from vllm.attention.backends.abstract import (
from vllm.v1.attention.backend import (
AttentionBackend,
AttentionImpl,
)
from vllm.attention.backends.registry import (
from vllm.v1.attention.backends.registry import (
AttentionBackendEnum,
MambaAttentionBackendEnum,
register_backend,
@@ -15,8 +15,6 @@ from tests.v1.attention.utils import (
create_vllm_config,
try_get_attention_backend,
)
from vllm.attention.backends.abstract import AttentionType
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.config import ModelConfig
from vllm.platforms import current_platform
from vllm.utils.math_utils import cdiv
@@ -25,6 +23,8 @@ from vllm.utils.torch_utils import (
is_torch_equal_or_newer,
set_random_seed,
)
from vllm.v1.attention.backend import AttentionType
from vllm.v1.attention.backends.registry import AttentionBackendEnum
from vllm.v1.attention.backends.utils import (
CommonAttentionMetadata,
set_kv_cache_layout,
+3 -3
View File
@@ -18,15 +18,15 @@ from tests.v1.attention.utils import (
try_get_attention_backend,
)
from vllm import _custom_ops as ops
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.attention.ops.flashmla import is_flashmla_dense_supported
from vllm.attention.utils.fa_utils import flash_attn_supports_mla
from vllm.config.vllm import set_current_vllm_config
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.utils.math_utils import cdiv
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
from vllm.v1.attention.backends.fa_utils import flash_attn_supports_mla
from vllm.v1.attention.backends.mla.common import QueryLenSupport
from vllm.v1.attention.backends.registry import AttentionBackendEnum
from vllm.v1.attention.backends.utils import CommonAttentionMetadata
from vllm.v1.attention.ops.flashmla import is_flashmla_dense_supported
from vllm.v1.kv_cache_interface import FullAttentionSpec
BACKENDS_TO_TEST = [
@@ -7,9 +7,9 @@ from unittest.mock import MagicMock, patch
import pytest
import torch
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.attention.selector import AttentionSelectorConfig
from vllm.platforms import current_platform
from vllm.v1.attention.backends.registry import AttentionBackendEnum
from vllm.v1.attention.selector import AttentionSelectorConfig
# ROCm-specific attention backend selection tests
pytestmark = pytest.mark.skipif(
@@ -21,7 +21,6 @@ from tests.v1.attention.utils import (
create_vllm_config,
)
from vllm import _custom_ops as ops
from vllm.attention.ops import flashmla
from vllm.config import set_current_vllm_config
from vllm.model_executor.layers.linear import ColumnParallelLinear
from vllm.platforms import current_platform
@@ -31,6 +30,7 @@ from vllm.v1.attention.backends.mla.flashmla_sparse import (
triton_convert_req_index_to_global_index,
)
from vllm.v1.attention.backends.utils import split_prefill_chunks
from vllm.v1.attention.ops import flashmla
SPARSE_BACKEND_BATCH_SPECS = {
name: BATCH_SPECS[name]
+2 -2
View File
@@ -7,8 +7,6 @@ from dataclasses import dataclass
import pytest
import torch
from vllm.attention.backends.abstract import AttentionImpl
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.config import (
CacheConfig,
CompilationConfig,
@@ -20,6 +18,8 @@ from vllm.config import (
VllmConfig,
)
from vllm.config.model import ModelDType
from vllm.v1.attention.backend import AttentionImpl
from vllm.v1.attention.backends.registry import AttentionBackendEnum
from vllm.v1.attention.backends.utils import (
AttentionMetadataBuilder,
CommonAttentionMetadata,
+1 -1
View File
@@ -6,9 +6,9 @@ import random
import pytest
import torch
from vllm.attention.utils.fa_utils import flash_attn_supports_mla
from vllm.platforms import current_platform
from vllm.utils.flashinfer import has_flashinfer
from vllm.v1.attention.backends.fa_utils import flash_attn_supports_mla
skip_unsupported = pytest.mark.skipif(
not (current_platform.is_cuda() and current_platform.has_device_capability(80)),
@@ -14,12 +14,12 @@ from unittest.mock import patch
import pytest
from vllm.attention.backends.abstract import AttentionMetadata
from vllm.distributed.kv_transfer.kv_connector.factory import KVConnectorFactory
from vllm.distributed.kv_transfer.kv_connector.v1 import (
KVConnectorBase_V1,
KVConnectorRole,
)
from vllm.v1.attention.backend import AttentionMetadata
from vllm.v1.core.sched.output import SchedulerOutput
from .utils import create_scheduler, create_vllm_config
+1 -1
View File
@@ -13,7 +13,6 @@ from tests.v1.attention.utils import (
create_standard_kv_cache_spec,
try_get_attention_backend,
)
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.config import (
AttentionConfig,
CacheConfig,
@@ -27,6 +26,7 @@ from vllm.config import (
from vllm.config.load import LoadConfig
from vllm.model_executor.models.llama import LlamaForCausalLM
from vllm.platforms import current_platform
from vllm.v1.attention.backends.registry import AttentionBackendEnum
from vllm.v1.spec_decode.eagle import EagleProposer
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
+1 -1
View File
@@ -12,7 +12,6 @@ from tests.v1.attention.utils import (
create_standard_kv_cache_spec,
try_get_attention_backend,
)
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.config import (
CacheConfig,
DeviceConfig,
@@ -25,6 +24,7 @@ from vllm.config import (
from vllm.config.load import LoadConfig
from vllm.model_executor.models.llama import LlamaForCausalLM
from vllm.platforms import current_platform
from vllm.v1.attention.backends.registry import AttentionBackendEnum
from vllm.v1.spec_decode.eagle import EagleProposer
mimo_7b_dir = "XiaomiMiMo/MiMo-7B-Base"
+2 -2
View File
@@ -11,9 +11,9 @@ from tests.v1.attention.utils import (
create_vllm_config,
try_get_attention_backend,
)
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.attention.utils.fa_utils import is_flash_attn_varlen_func_available
from vllm.config import ParallelConfig, SpeculativeConfig
from vllm.v1.attention.backends.fa_utils import is_flash_attn_varlen_func_available
from vllm.v1.attention.backends.registry import AttentionBackendEnum
from vllm.v1.attention.backends.utils import CommonAttentionMetadata
if not is_flash_attn_varlen_func_available():
+2 -2
View File
@@ -5,8 +5,6 @@ import numpy as np
import pytest
import torch
from vllm.attention.backends.abstract import MultipleOf
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.attention.layer import Attention
from vllm.config import (
AttentionConfig,
@@ -27,6 +25,8 @@ from vllm.sampling_params import SamplingParams
from vllm.utils.mem_constants import GiB_bytes
from vllm.utils.system_utils import update_environment_variables
from vllm.utils.torch_utils import set_random_seed
from vllm.v1.attention.backend import MultipleOf
from vllm.v1.attention.backends.registry import AttentionBackendEnum
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.kv_cache_interface import (