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

[torch.compile] Fuse RMSNorm with quant (#9138)

Browse Source 
Signed-off-by: luka <luka@neuralmagic.com>
Co-authored-by: youkaichao <youkaichao@126.com>
This commit is contained in:
Luka Govedič
2024-11-08 16:20:08 -05:00
committed by GitHub
parent e1b5a82179
commit 4f93dfe952
17 changed files with 1335 additions and 368 deletions
Download Patch File Download Diff File Expand all files Collapse all files

75
tests/kernels/test_layernorm.py
View File

@@ -1,13 +1,14 @@
import pytest
import torch
from tests.kernels.quant_utils import FP8_DTYPE
from tests.kernels.utils import opcheck
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.platforms import current_platform
DTYPES = [torch.half, torch.bfloat16, torch.float]
NUM_TOKENS = [7, 83, 4096] # Arbitrary values for testing
HIDDEN_SIZES = [768, 769, 770, 771, 5120, 5124, 5125, 5126, 8192,
HIDDEN_SIZES = [8, 768, 769, 770, 771, 5120, 5124, 5125, 5126, 8192,
8199] # Arbitrary values for testing
ADD_RESIDUAL = [False, True]
SEEDS = [0]
@@ -59,3 +60,75 @@ def test_rms_norm(
else:
opcheck(torch.ops._C.rms_norm,
(out, x, layer.weight.data, layer.variance_epsilon))
@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
@pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
@pytest.mark.parametrize("add_residual", ADD_RESIDUAL)
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("quant_scale", [1.0, 0.01, 10.0])
@pytest.mark.parametrize("seed", SEEDS)
@pytest.mark.parametrize("device", CUDA_DEVICES)
def test_fused_rms_norm_quant(
num_tokens: int,
hidden_size: int,
add_residual: bool,
dtype: torch.dtype,
quant_scale: float,
seed: int,
device: str,
) -> None:
current_platform.seed_everything(seed)
torch.set_default_device(device)
weight = torch.empty(hidden_size, dtype=dtype).normal_(mean=1.0, std=0.1)
scale = 1 / (2 * hidden_size)
x = torch.randn(num_tokens, hidden_size, dtype=dtype)
x *= scale
if add_residual:
residual = torch.randn_like(x) * scale
residual_fused = residual.clone()
else:
residual = residual_fused = None
out_norm = torch.empty_like(x)
out_quant = torch.empty_like(x, dtype=FP8_DTYPE)
out_quant_fused = torch.empty_like(out_quant)
quant_scale_t = torch.tensor(quant_scale, dtype=torch.float32)
if add_residual:
torch.ops._C.fused_add_rms_norm_static_fp8_quant(
out_quant_fused, x, residual_fused, weight, quant_scale_t, 1e-6)
# Unfused kernel is in-place so it goes second
# Also use a separate clone of x to avoid modifying the input
x_unfused = x.clone()
torch.ops._C.fused_add_rms_norm(x_unfused, residual, weight, 1e-6)
torch.ops._C.static_scaled_fp8_quant(out_quant, x_unfused,
quant_scale_t)
torch.cuda.synchronize()
torch.testing.assert_close(residual_fused,
residual,
atol=1e-2,
rtol=1e-2)
opcheck(
torch.ops._C.fused_add_rms_norm_static_fp8_quant,
(out_quant_fused, x, residual_fused, weight, quant_scale_t, 1e-6))
else:
torch.ops._C.rms_norm_static_fp8_quant(out_quant_fused, x, weight,
quant_scale_t, 1e-6)
torch.ops._C.rms_norm(out_norm, x, weight, 1e-6)
torch.ops._C.static_scaled_fp8_quant(out_quant, out_norm,
quant_scale_t)
opcheck(torch.ops._C.rms_norm_static_fp8_quant,
(out_quant_fused, x, weight, quant_scale_t, 1e-6))
torch.testing.assert_close(out_quant_fused.to(dtype=torch.float32),
out_quant.to(dtype=torch.float32),
atol=1e-3,
rtol=1e-3)