[Feature] Add Hopper DeepGEMM E8M0 for DeepSeekV3.1 scale_fmt (#23666)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: youkaichao <youkaichao@gmail.com>
Co-authored-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
Wentao Ye
2025-08-27 14:09:08 +00:00
committed by GitHub
co-authored by youkaichao
parent 513c1fe255
commit 3af47c3cc6
10 changed files with 68 additions and 53 deletions
+2 -3
View File
@@ -16,7 +16,7 @@ from vllm.model_executor.layers.fused_moe.fused_moe import (
fused_topk, modular_triton_fused_moe)
from vllm.platforms import current_platform
from vllm.utils import has_deep_gemm
from vllm.utils.deep_gemm import is_blackwell_deep_gemm_e8m0_used
from vllm.utils.deep_gemm import is_deep_gemm_e8m0_used
dg_available = has_deep_gemm()
@@ -226,8 +226,7 @@ def test_w8a8_block_fp8_fused_moe(M, N, K, E, topk, block_size, dtype, seed,
@pytest.mark.parametrize("topk", TOP_KS)
@pytest.mark.parametrize("seed", SEEDS)
@pytest.mark.skipif(not dg_available, reason="DeepGemm kernels not available.")
@pytest.mark.skipif(is_blackwell_deep_gemm_e8m0_used(),
reason="Not E8M0 scale MOE")
@pytest.mark.skipif(is_deep_gemm_e8m0_used(), reason="Not E8M0 scale MOE")
@torch.inference_mode()
def test_w8a8_block_fp8_deep_gemm_fused_moe(M, N, K, E, topk, seed,
monkeypatch):
@@ -20,8 +20,7 @@ from vllm.model_executor.layers.fused_moe.modular_kernel import (
FusedMoEModularKernel)
from vllm.platforms import current_platform
from vllm.utils import has_deep_ep, has_deep_gemm
from vllm.utils.deep_gemm import (is_blackwell_deep_gemm_e8m0_used,
is_deep_gemm_supported)
from vllm.utils.deep_gemm import is_deep_gemm_e8m0_used, is_deep_gemm_supported
from ...utils import multi_gpu_test
from .parallel_utils import ProcessGroupInfo, parallel_launch
@@ -374,7 +373,7 @@ NUM_EXPERTS = [32]
@multi_gpu_test(num_gpus=2)
@requires_deep_ep
@requires_deep_gemm
@pytest.mark.skipif(is_blackwell_deep_gemm_e8m0_used(),
@pytest.mark.skipif(is_deep_gemm_e8m0_used(),
reason="Skipping test for Blackwell DeepGEMM")
def test_ht_deepep_deepgemm_moe(mnk: tuple[int, int, int], num_experts: int,
topk: int, world_dp_size: tuple[int, int]):
@@ -432,7 +431,7 @@ USE_FP8_DISPATCH = [False]
@multi_gpu_test(num_gpus=2)
@requires_deep_ep
@requires_deep_gemm
@pytest.mark.skipif(is_blackwell_deep_gemm_e8m0_used(),
@pytest.mark.skipif(is_deep_gemm_e8m0_used(),
reason="Skipping test for Blackwell DeepGEMM")
def test_ll_deepep_deepgemm_moe(
mnk: tuple[int, int, int],