[CPU] Migrate unquantized MoE to the modular-kernel experts structure (#50133)

Signed-off-by: jiang1.li <jiang1.li@intel.com>
Signed-off-by: Fadi Arafeh <fadi.arafeh@arm.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Fadi Arafeh <fadi.arafeh@arm.com>
This commit is contained in:
Li, Jiang
2026-08-03 14:55:04 +08:00
committed by GitHub
co-authored by Claude Opus 5 Fadi Arafeh
parent b3f97dae24
commit 0a6446005d
20 changed files with 972 additions and 839 deletions
-79
View File
@@ -1482,85 +1482,6 @@ def test_moe_sum_pad_aware(topk: int, dtype: torch.dtype, topk_ids_dtype: torch.
opcheck(torch.ops._moe_C.moe_sum, (input, actual, topk_ids, expert_map))
@pytest.mark.usefixtures("default_vllm_config")
@pytest.mark.parametrize("m", [1, 33])
@pytest.mark.parametrize("n,k", [(128, 128)])
@pytest.mark.parametrize("e", [8])
@pytest.mark.parametrize("topk", [2])
@pytest.mark.parametrize("dtype", [torch.float32, torch.bfloat16])
@pytest.mark.parametrize("with_bias", [False, True])
@pytest.mark.parametrize("activation", [MoEActivation.SILU])
@pytest.mark.skipif(not current_platform.is_cpu(), reason="CPU only test")
def test_cpu_fused_moe_basic(
m: int,
n: int,
k: int,
e: int,
topk: int,
dtype: torch.dtype,
with_bias: bool,
activation: MoEActivation,
):
from vllm.model_executor.layers.fused_moe.cpu_fused_moe import CPUFusedMOE
device = "cpu"
set_random_seed(7)
a = torch.randn((m, k), device=device, dtype=dtype) / 10
w13 = torch.randn((e, 2 * n, k), device=device, dtype=dtype) / 10
w2 = torch.randn((e, k, n), device=device, dtype=dtype) / 10
router_logits = torch.randn((m, e), device=device, dtype=dtype)
b1 = b2 = None
if with_bias:
b1 = torch.randn((e, 2 * n), device=device, dtype=dtype) / 10
b2 = torch.randn((e, k), device=device, dtype=dtype) / 10
ref = (
torch_moe(a, w13, w2, router_logits, topk, b1, b2)
if with_bias
else torch_moe(a, w13, w2, router_logits, topk)
)
class _Dummy(torch.nn.Module):
def __init__(self, w13, w2, b1=None, b2=None):
super().__init__()
self.w13_weight = torch.nn.Parameter(w13, requires_grad=False)
self.w2_weight = torch.nn.Parameter(w2, requires_grad=False)
if b1 is not None:
self.w13_bias = torch.nn.Parameter(b1, requires_grad=False)
if b2 is not None:
self.w2_bias = torch.nn.Parameter(b2, requires_grad=False)
layer = _Dummy(w13, w2, b1, b2).to(dtype)
fused = CPUFusedMOE(layer)
out = fused(
layer=layer,
x=a,
use_grouped_topk=False,
top_k=topk,
router_logits=router_logits,
renormalize=False,
global_num_experts=e,
expert_map=None,
custom_routing_function=None,
scoring_func="softmax",
routed_scaling_factor=1.0,
e_score_correction_bias=None,
apply_router_weight_on_input=False,
activation=activation,
)
# Tolerances: fp32 tight; bf16 looser (esp. with bias)
if dtype == torch.float32:
atol = 1e-3
elif with_bias:
atol = 8e-2
else:
atol = 5e-2
torch.testing.assert_close(out, ref, atol=atol, rtol=0)
def _batched_fused_marlin_moe_cases() -> list[Any]:
cases = [
pytest.param(