mirror of
https://git.datalinker.icu/vllm-project/vllm.git
synced 2025-12-10 05:15:42 +08:00
e489ad7a21
- **Add SPDX license headers to python source files**
- **Check for SPDX headers using pre-commit**
commit 9d7ef44c3cfb72ca4c32e1c677d99259d10d4745
Author: Russell Bryant <rbryant@redhat.com>
Date: Fri Jan 31 14:18:24 2025 -0500
Add SPDX license headers to python source files
This commit adds SPDX license headers to python source files as
recommended to
the project by the Linux Foundation. These headers provide a concise way
that is
both human and machine readable for communicating license information
for each
source file. It helps avoid any ambiguity about the license of the code
and can
also be easily used by tools to help manage license compliance.
The Linux Foundation runs license scans against the codebase to help
ensure
we are in compliance with the licenses of the code we use, including
dependencies. Having these headers in place helps that tool do its job.
More information can be found on the SPDX site:
- https://spdx.dev/learn/handling-license-info/
Signed-off-by: Russell Bryant <rbryant@redhat.com>
commit 5a1cf1cb3b80759131c73f6a9dddebccac039dea
Author: Russell Bryant <rbryant@redhat.com>
Date: Fri Jan 31 14:36:32 2025 -0500
Check for SPDX headers using pre-commit
Signed-off-by: Russell Bryant <rbryant@redhat.com>
---------
Signed-off-by: Russell Bryant <rbryant@redhat.com>
117 lines
4.5 KiB
Python
117 lines
4.5 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
|
|
import pytest
|
|
import torch
|
|
|
|
import vllm._custom_ops as ops
|
|
from tests.kernels.quant_utils import (FP8_DTYPE,
|
|
ref_dynamic_per_tensor_fp8_quant,
|
|
ref_dynamic_per_token_quant)
|
|
from tests.kernels.utils import opcheck
|
|
from vllm.platforms import current_platform
|
|
|
|
DTYPES = [torch.half, torch.bfloat16, torch.float]
|
|
HIDDEN_SIZES = [1, 2, 3, 4, 16, 67, 768, 2048, 5120, 5137, 8192,
|
|
8193] # Arbitrary values for testing
|
|
HIDDEN_SIZES += list(range(1024, 1033)) # vectorized conversion edge cases
|
|
NUM_TOKENS = [1, 7, 83, 4096] # Arbitrary values for testing
|
|
SCALE_UBS = [True, False]
|
|
SEEDS = [0]
|
|
|
|
|
|
def opcheck_fp8_quant(output,
|
|
input,
|
|
scale=None,
|
|
scale_ub=None,
|
|
use_per_token_if_dynamic=False):
|
|
if scale is not None:
|
|
opcheck(torch.ops._C.static_scaled_fp8_quant, (output, input, scale))
|
|
elif use_per_token_if_dynamic:
|
|
scale = torch.empty((input.shape[0], 1),
|
|
device=input.device,
|
|
dtype=torch.float32)
|
|
opcheck(torch.ops._C.dynamic_per_token_scaled_fp8_quant,
|
|
(output, input, scale, scale_ub))
|
|
else:
|
|
scale = torch.empty((input.numel() // input.shape[-1], 1),
|
|
device=input.device,
|
|
dtype=torch.float32)
|
|
opcheck(torch.ops._C.dynamic_scaled_fp8_quant, (output, input, scale))
|
|
|
|
|
|
@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
|
|
@pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
|
|
@pytest.mark.parametrize("dtype", DTYPES)
|
|
@pytest.mark.parametrize("scale_ub", SCALE_UBS)
|
|
@pytest.mark.parametrize("seed", SEEDS)
|
|
@torch.inference_mode()
|
|
def test_dynamic_per_token_fp8_quant(num_tokens: int, hidden_size: int,
|
|
dtype: torch.dtype, scale_ub: bool,
|
|
seed: int) -> None:
|
|
current_platform.seed_everything(seed)
|
|
|
|
x = torch.rand(num_tokens, hidden_size, dtype=dtype,
|
|
device="cuda") + 1e-6 # avoid nans
|
|
|
|
scale_ub = torch.mean(x).to(dtype=torch.float32, device='cuda') \
|
|
if scale_ub else None
|
|
ref_out, ref_scales = ref_dynamic_per_token_quant(x, FP8_DTYPE, scale_ub)
|
|
ops_out, ops_scales = ops.scaled_fp8_quant(x,
|
|
scale_ub=scale_ub,
|
|
use_per_token_if_dynamic=True)
|
|
|
|
torch.testing.assert_close(ref_scales, ops_scales)
|
|
torch.testing.assert_close(ref_out.to(dtype=torch.float32),
|
|
ops_out.to(dtype=torch.float32))
|
|
|
|
opcheck_fp8_quant(ops_out,
|
|
x,
|
|
None,
|
|
scale_ub,
|
|
use_per_token_if_dynamic=True)
|
|
|
|
|
|
@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
|
|
@pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
|
|
@pytest.mark.parametrize("dtype", DTYPES)
|
|
@pytest.mark.parametrize("seed", SEEDS)
|
|
@torch.inference_mode()
|
|
def test_dynamic_per_tensor_fp8_quant(num_tokens: int, hidden_size: int,
|
|
dtype: torch.dtype, seed: int) -> None:
|
|
current_platform.seed_everything(seed)
|
|
|
|
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda")
|
|
|
|
ref_out, ref_scale = ref_dynamic_per_tensor_fp8_quant(x)
|
|
ops_out, ops_scale = ops.scaled_fp8_quant(x)
|
|
|
|
torch.testing.assert_close(ref_scale, ops_scale)
|
|
torch.testing.assert_close(ref_out.to(dtype=torch.float32),
|
|
ops_out.to(dtype=torch.float32))
|
|
|
|
opcheck_fp8_quant(ops_out, x)
|
|
|
|
|
|
# Regression test for a case with large activations where an int32 index cannot
|
|
# represent the number of elements.
|
|
@torch.inference_mode()
|
|
@pytest.mark.parametrize("seed", SEEDS)
|
|
def test_fp8_quant_large(seed: int) -> None:
|
|
current_platform.seed_everything(seed)
|
|
|
|
num_tokens = 1024000 # Mistral-Nemo's max_position_embeddings
|
|
hidden_size = 1152 # Smallest hidden_size to reproduce the error
|
|
dtype = torch.bfloat16
|
|
|
|
x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda")
|
|
ref_out, scale = ref_dynamic_per_tensor_fp8_quant(x)
|
|
ops_out, _ = ops.scaled_fp8_quant(x, scale)
|
|
|
|
# Minimize memory footprint in this test by freeing x and upconverting
|
|
# the outputs in place. (torch.allclose does not support fp8)
|
|
del x
|
|
ref_out = ref_out.to(dtype=dtype)
|
|
ops_out = ops_out.to(dtype=dtype)
|
|
|
|
torch.testing.assert_close(ref_out, ops_out)
|