[MoE] Move various experts classes to fused_moe/experts/ (#41979)
Signed-off-by: Jackmin801 <ongjackm@gmail.com> Signed-off-by: Robert Shaw <robertgshaw2@gmail.com> Signed-off-by: Jackmin801 <56836461+Jackmin801@users.noreply.github.com> Signed-off-by: Bill Nell <bnell@redhat.com> Co-authored-by: Jackmin801 <ongjackm@gmail.com> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Robert Shaw <robertgshaw2@gmail.com> Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com> Co-authored-by: Jackmin801 <56836461+Jackmin801@users.noreply.github.com>
This commit is contained in:
co-authored by
Jackmin801
Claude
Robert Shaw
Robert Shaw
Jackmin801
parent
21943d4c25
commit
1b57eb41f2
@@ -237,7 +237,7 @@ if has_mori():
|
||||
)
|
||||
|
||||
if has_flashinfer_cutlass_fused_moe() and current_platform.has_device_capability(100):
|
||||
from vllm.model_executor.layers.fused_moe.flashinfer_cutlass_moe import (
|
||||
from vllm.model_executor.layers.fused_moe.experts.flashinfer_cutlass_moe import (
|
||||
FlashInferExperts,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.prepare_finalize.flashinfer_nvlink_two_sided import ( # noqa: E501
|
||||
@@ -298,7 +298,7 @@ if has_flashinfer_cutlass_fused_moe() and current_platform.has_device_capability
|
||||
)
|
||||
|
||||
if has_aiter():
|
||||
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
|
||||
from vllm.model_executor.layers.fused_moe.experts.rocm_aiter_moe import (
|
||||
AiterExperts,
|
||||
)
|
||||
|
||||
|
||||
@@ -18,12 +18,12 @@ from vllm.model_executor.layers.fused_moe.config import (
|
||||
RoutingMethodType,
|
||||
fp8_w8a8_moe_quant_config,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.experts.flashinfer_cutlass_moe import (
|
||||
FlashInferExperts,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.experts.trtllm_fp8_moe import (
|
||||
TrtLlmFp8ExpertsMonolithic,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.flashinfer_cutlass_moe import (
|
||||
FlashInferExperts,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts
|
||||
from vllm.model_executor.layers.quantization.utils.flashinfer_utils import (
|
||||
rotate_weights_for_fi_trtllm_fp8_per_tensor_moe,
|
||||
|
||||
@@ -22,7 +22,7 @@ from vllm.model_executor.layers.fused_moe.config import (
|
||||
FusedMoEParallelConfig,
|
||||
RoutingMethodType,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.flashinfer_cutlass_moe import (
|
||||
from vllm.model_executor.layers.fused_moe.experts.flashinfer_cutlass_moe import (
|
||||
FlashInferExperts,
|
||||
is_valid_flashinfer_cutlass_fused_moe,
|
||||
)
|
||||
|
||||
@@ -5,7 +5,7 @@
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
from vllm.model_executor.layers.fused_moe.fused_marlin_moe import (
|
||||
from vllm.model_executor.layers.fused_moe.experts.marlin_moe import (
|
||||
fused_marlin_moe,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.router.grouped_topk_router import (
|
||||
|
||||
@@ -32,7 +32,7 @@ from vllm.model_executor.layers.fused_moe.config import (
|
||||
int4_w4a16_moe_quant_config,
|
||||
int8_w8a16_moe_quant_config,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.fused_marlin_moe import (
|
||||
from vllm.model_executor.layers.fused_moe.experts.marlin_moe import (
|
||||
batched_fused_marlin_moe,
|
||||
fused_marlin_moe,
|
||||
)
|
||||
|
||||
@@ -20,7 +20,7 @@ if not current_platform.is_rocm():
|
||||
pytest.skip("This test can only run on ROCm.", allow_module_level=True)
|
||||
|
||||
# this import statement is needed to ensure the ops are registered
|
||||
import vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe # noqa: F401
|
||||
import vllm.model_executor.layers.fused_moe.experts.rocm_aiter_moe # noqa: F401
|
||||
|
||||
# need to import once to ensure the ops are registered
|
||||
# Check if aiter package is installed
|
||||
|
||||
@@ -15,7 +15,7 @@ from vllm.model_executor.layers.fused_moe.activation import MoEActivation
|
||||
from vllm.model_executor.layers.fused_moe.config import (
|
||||
FUSED_MOE_UNQUANTIZED_CONFIG,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.fused_moe import TritonExperts
|
||||
from vllm.model_executor.layers.fused_moe.experts.triton_moe import TritonExperts
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
# Test parameters
|
||||
@@ -151,7 +151,7 @@ def test_triton_experts_no_mul_activation(
|
||||
@torch.inference_mode()
|
||||
def test_workspace_shapes_no_mul_vs_gated():
|
||||
"""Test that workspace shapes differ correctly between gated and non-gated."""
|
||||
from vllm.model_executor.layers.fused_moe.fused_moe import TritonExperts
|
||||
from vllm.model_executor.layers.fused_moe.experts.triton_moe import TritonExperts
|
||||
|
||||
M, N, K, topk = 64, 256, 128, 2
|
||||
|
||||
@@ -192,7 +192,7 @@ def test_workspace_shapes_no_mul_vs_gated():
|
||||
@torch.inference_mode()
|
||||
def test_adjust_n_for_activation():
|
||||
"""Test the adjust_N_for_activation method."""
|
||||
from vllm.model_executor.layers.fused_moe.fused_moe import TritonExperts
|
||||
from vllm.model_executor.layers.fused_moe.experts.triton_moe import TritonExperts
|
||||
|
||||
experts = TritonExperts(
|
||||
moe_config=make_dummy_moe_config(),
|
||||
|
||||
@@ -158,7 +158,7 @@ def test_select_cuda_flashinfer_trtllm_backend(mock_is_supported_trtllm, monkeyp
|
||||
return_value=(False, None),
|
||||
)
|
||||
@patch(
|
||||
"vllm.model_executor.layers.fused_moe.flashinfer_cutlass_moe.FlashInferExperts.is_supported_config",
|
||||
"vllm.model_executor.layers.fused_moe.experts.flashinfer_cutlass_moe.FlashInferExperts.is_supported_config",
|
||||
return_value=(True, None),
|
||||
)
|
||||
@pytest.mark.skipif(
|
||||
|
||||
@@ -17,12 +17,14 @@ from vllm.model_executor.layers.fused_moe.config import (
|
||||
FusedMoEQuantConfig,
|
||||
RoutingMethodType,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.experts.triton_moe import (
|
||||
TritonExperts,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
|
||||
BatchedTritonExperts,
|
||||
NaiveBatchedExperts,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.fused_moe import (
|
||||
TritonExperts,
|
||||
fused_experts,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.modular_kernel import FusedMoEKernel
|
||||
|
||||
Reference in New Issue
Block a user