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

[Hardware][ROCM] using current_platform.is_rocm (#9642)

Browse Source 
Signed-off-by: wangshuai09 <391746016@qq.com>
This commit is contained in:
wangshuai09
2024-10-28 12:07:00 +08:00
committed by GitHub
parent 34a9941620
commit 4e2d95e372
32 changed files with 165 additions and 151 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
tests/kernels/quant_utils.py
View File

@@ -2,12 +2,13 @@ from typing import Optional, Tuple, Union
import torch
from vllm.utils import is_hip
from vllm.platforms import current_platform
# Using the default value (240.0) from pytorch will cause accuracy
# issue on dynamic quantization models. Here use 224.0 for rocm.
ROCM_FP8_MAX = 224.0
FP8_DTYPE = torch.float8_e4m3fnuz if is_hip() else torch.float8_e4m3fn
FP8_DTYPE = torch.float8_e4m3fnuz if current_platform.is_rocm() \
else torch.float8_e4m3fn
def as_float32_tensor(x: Union[float, torch.tensor]) -> torch.tensor:
@@ -24,8 +25,10 @@ def ref_dynamic_per_token_quant(x: torch.tensor,
qtype_traits = torch.iinfo(quant_dtype) if quant_dtype == torch.int8 \
else torch.finfo(quant_dtype)
qtype_traits_max = ROCM_FP8_MAX if is_hip() else qtype_traits.max
qtype_traits_min = -ROCM_FP8_MAX if is_hip() else qtype_traits.min
qtype_traits_max = ROCM_FP8_MAX if current_platform.is_rocm() \
else qtype_traits.max
qtype_traits_min = -ROCM_FP8_MAX if current_platform.is_rocm() \
else qtype_traits.min
qtype_max = as_float32_tensor(qtype_traits_max)
s_1 = as_float32_tensor(1.0)
s_512 = as_float32_tensor(512.0)
@@ -66,8 +69,10 @@ def ref_dynamic_per_tensor_fp8_quant(x: torch.tensor) \
-> Tuple[torch.tensor, torch.tensor]:
fp8_traits = torch.finfo(FP8_DTYPE)
fp8_traits_max = ROCM_FP8_MAX if is_hip() else fp8_traits.max
fp8_traits_min = -ROCM_FP8_MAX if is_hip() else fp8_traits.min
fp8_traits_max = ROCM_FP8_MAX if current_platform.is_rocm() \
else fp8_traits.max
fp8_traits_min = -ROCM_FP8_MAX if current_platform.is_rocm() \
else fp8_traits.min
fp8_max = as_float32_tensor(fp8_traits_max)
one = as_float32_tensor(1.0)