biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Misc] Removed force_fp8_e4m3fnuz from FP8LinearOp (#23725)

Browse Source 
Signed-off-by: Julien Lin <jullin@nvidia.com>
Signed-off-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
This commit is contained in:
nvjullin
2025-09-04 21:25:40 +08:00
committed by GitHub
parent c9f7081f9c
commit 37241077d5
5 changed files with 45 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
tests/compile/test_silu_mul_quant_fusion.py
View File

@@ -17,9 +17,10 @@ from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.quantization.utils.quant_utils import (
GroupShape, kFp8StaticTensorSym, kNvfp4Quant)
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
Fp8LinearOp)
Fp8LinearOp, cutlass_fp8_supported)
from vllm.platforms import current_platform
from ..utils import override_cutlass_fp8_supported
from .backend import TestBackend
FP8_DTYPE = current_platform.fp8_dtype()
@@ -32,7 +33,7 @@ def is_nvfp4_supported():
class TestSiluMulFp8QuantModel(torch.nn.Module):
def __init__(self, hidden_size: int, force_fp8_e4m3fnuz: bool, **kwargs):
def __init__(self, hidden_size: int, cuda_force_torch: bool, **kwargs):
super().__init__()
self.silu_and_mul = SiluAndMul()
self.wscale = torch.rand(1, dtype=torch.float32)
@@ -40,11 +41,11 @@ class TestSiluMulFp8QuantModel(torch.nn.Module):
self.w = torch.rand(hidden_size, hidden_size).to(dtype=FP8_DTYPE).t()
self.fp8_linear = Fp8LinearOp(
force_fp8_e4m3fnuz=force_fp8_e4m3fnuz,
act_quant_static=True,
act_quant_group_shape=GroupShape.PER_TENSOR,
)
with override_cutlass_fp8_supported(not cuda_force_torch):
self.fp8_linear = Fp8LinearOp(
act_quant_static=True,
act_quant_group_shape=GroupShape.PER_TENSOR,
)
def forward(self, x):
y = self.silu_and_mul(x)
@@ -96,12 +97,15 @@ class TestSiluMulNvfp4QuantModel(torch.nn.Module):
@pytest.mark.parametrize(
"model_class", [TestSiluMulFp8QuantModel, TestSiluMulNvfp4QuantModel]
if is_nvfp4_supported() else [TestSiluMulFp8QuantModel])
@pytest.mark.parametrize("force_fp8_e4m3fnuz", [True, False])
# cuda_force_torch used to test torch code path on platforms that
# cutlass_fp8_supported() == True.
@pytest.mark.parametrize("cuda_force_torch",
[True, False] if cutlass_fp8_supported() else [True])
@pytest.mark.skipif(envs.VLLM_TARGET_DEVICE not in ["cuda", "rocm"],
reason="Only test on CUDA and ROCm")
def test_fusion_silu_and_mul_quant(num_tokens, hidden_size, model_class,
force_fp8_e4m3fnuz):
if model_class == TestSiluMulNvfp4QuantModel and force_fp8_e4m3fnuz:
cuda_force_torch):
if model_class == TestSiluMulNvfp4QuantModel and cuda_force_torch:
pytest.skip("Duplicate tests for NVFP4")
torch.set_default_device("cuda")
@@ -114,8 +118,7 @@ def test_fusion_silu_and_mul_quant(num_tokens, hidden_size, model_class,
fusion_pass = ActivationQuantFusionPass(config)
backend = TestBackend(NoOpEliminationPass(config), fusion_pass)
model = model_class(hidden_size=hidden_size,
force_fp8_e4m3fnuz=force_fp8_e4m3fnuz)
model = model_class(hidden_size, cuda_force_torch)
# First dimension dynamic
x = torch.rand(num_tokens, hidden_size * 2)