"""
pypto.scaled_mm 3D/4D ST测试用例配置
用于System Test自动化测试框架
"""
from dataclasses import dataclass
from typing import Tuple, Union
import pypto
import torch
@dataclass
class BaseScaledConfig:
a_shape: Tuple[int, ...]
b_shape: Tuple[int, ...]
m_tile_shape: Tuple[int, int]
k_tile_shape: Tuple[int, int]
n_tile_shape: Tuple[int, int]
view_shape: Union[Tuple[int, int, int], Tuple[int, int, int, int]]
in_dtype: pypto.DataType
out_dtype: pypto.DataType
a_trans: bool = False
b_trans: bool = False
scale_a_trans: bool = False
scale_b_trans: bool = False
a_format: str = "ND"
b_format: str = "ND"
c_format: str = "ND"
has_bias: bool = False
bias_batch: int = 0
bias_shape_type: str = "1_n"
enable_ksplit: bool = False
DTYPE_CONFIG = {
"DT_FP8E4M3": {"pto": pypto.DataType.DT_FP8E4M3, "torch": torch.float8_e4m3fn},
"DT_FP8E5M2": {"pto": pypto.DataType.DT_FP8E5M2, "torch": torch.float8_e5m2},
"DT_FP16": {"pto": pypto.DataType.DT_FP16, "torch": torch.float16},
"DT_FP32": {"pto": pypto.DataType.DT_FP32, "torch": torch.float32},
}
TOLERANCE_CONFIG = {
"DT_FP16": {"atol": 1e-3, "rtol": 1e-3},
"DT_FP32": {"atol": 1e-3, "rtol": 1e-3},
"DT_FP8E4M3": {"atol": 1e-3, "rtol": 1e-3},
"DT_FP8E5M2": {"atol": 1e-3, "rtol": 1e-3},
}
@classmethod
def from_test_case(cls, case: dict):
return cls(
a_shape=tuple(case["a_shape"]),
b_shape=tuple(case["b_shape"]),
m_tile_shape=tuple(case["m_tile_shape"]),
k_tile_shape=tuple(case["k_tile_shape"]),
n_tile_shape=tuple(case["n_tile_shape"]),
view_shape=tuple(case["view_shape"]),
in_dtype=cls.DTYPE_CONFIG[case["in_dtype"]]["pto"],
out_dtype=cls.DTYPE_CONFIG[case["out_dtype"]]["pto"],
a_trans=case.get("a_trans", False),
b_trans=case.get("b_trans", False),
scale_a_trans=case.get("scale_a_trans", False),
scale_b_trans=case.get("scale_b_trans", False),
a_format=case.get("a_format", "ND"),
b_format=case.get("b_format", "ND"),
c_format=case.get("c_format", "ND"),
has_bias=case.get("has_bias", False),
bias_batch=case.get("bias_batch", 0),
bias_shape_type=case.get("bias_shape_type", "1_n"),
enable_ksplit=case.get("enable_ksplit", False),
)
@classmethod
def get_torch_dtype(cls, dtype_str: str) -> torch.dtype:
return cls.DTYPE_CONFIG[dtype_str]["torch"]
@classmethod
def get_tolerance(cls, dtype_str: str) -> Tuple[float, float]:
info = cls.TOLERANCE_CONFIG[dtype_str]
return info["atol"], info["rtol"]
@classmethod
def pto_to_torch(cls, pto_dtype: pypto.DataType) -> torch.dtype:
for info in cls.DTYPE_CONFIG.values():
if info["pto"] == pto_dtype:
return info["torch"]
raise ValueError(f"Unsupported pypto.DataType: {pto_dtype}")
@dataclass
class ScaledBmmConfig(BaseScaledConfig):
@property
def batch_a(self) -> int:
return self.a_shape[0]
@property
def batch_b(self) -> int:
return self.b_shape[0]
@property
def batch(self) -> int:
return max(self.batch_a, self.batch_b)
def get_logical_dims_3d(self):
m = self.a_shape[2] if self.a_trans else self.a_shape[1]
k = self.a_shape[1] if self.a_trans else self.a_shape[2]
n = self.b_shape[1] if self.b_trans else self.b_shape[2]
return self.batch, m, k, n
def get_logical_dims_4d(self):
b0 = max(self.a_shape[0], self.b_shape[0])
b1 = max(self.a_shape[1], self.b_shape[1])
m = self.a_shape[3] if self.a_trans else self.a_shape[2]
k = self.a_shape[2] if self.a_trans else self.a_shape[3]
n = self.b_shape[2] if self.b_trans else self.b_shape[3]
return b0, b1, m, k, n
SCALED_BMM_TESTS = [
{
"id": "BMM3D_BASIC",
"name": "scaled_bmm_3d_no_bias",
"a_shape": [4, 135, 192],
"b_shape": [1, 192, 351],
"m_tile_shape": [64, 64],
"k_tile_shape": [64, 256],
"n_tile_shape": [256, 256],
"view_shape": [2, 192, 215],
"in_dtype": "DT_FP8E4M3",
"out_dtype": "DT_FP16",
"a_trans": False,
"b_trans": False,
"scale_a_trans": True,
"scale_b_trans": False,
"a_format": "ND",
"b_format": "ND",
"c_format": "ND",
"has_bias": False,
"bias_batch": 0,
"enable_ksplit": False,
"products": ["950"],
},
{
"id": "BMM3D_BIAS_1N",
"name": "scaled_bmm_3d_bias_1n",
"a_shape": [3, 320, 299],
"b_shape": [1, 411, 320],
"m_tile_shape": [64, 64],
"k_tile_shape": [128, 256],
"n_tile_shape": [256, 256],
"view_shape": [1, 192, 32],
"in_dtype": "DT_FP8E5M2",
"out_dtype": "DT_FP32",
"a_trans": True,
"b_trans": True,
"scale_a_trans": True,
"scale_b_trans": False,
"a_format": "ND",
"b_format": "ND",
"c_format": "ND",
"has_bias": True,
"bias_batch": 1,
"bias_shape_type": "1_n",
"enable_ksplit": False,
"products": ["950"],
},
{
"id": "BMM3D_BIAS_B1N",
"name": "scaled_bmm_3d_bias_b1n",
"a_shape": [2, 256, 192],
"b_shape": [2, 224, 192],
"m_tile_shape": [64, 64],
"k_tile_shape": [128, 256],
"n_tile_shape": [256, 256],
"view_shape": [2, 192, 32],
"in_dtype": "DT_FP8E4M3",
"out_dtype": "DT_FP16",
"a_trans": False,
"b_trans": True,
"scale_a_trans": True,
"scale_b_trans": False,
"a_format": "NZ",
"b_format": "NZ",
"c_format": "ND",
"has_bias": True,
"bias_batch": 2,
"bias_shape_type": "b_1_n",
"enable_ksplit": False,
"products": ["950"],
},
{
"id": "BMM4D_BASIC",
"name": "scaled_bmm_4d_no_bias",
"a_shape": [2, 3, 384, 320],
"b_shape": [2, 3, 178, 320],
"m_tile_shape": [128, 256],
"k_tile_shape": [64, 256],
"n_tile_shape": [64, 192],
"view_shape": [1, 3, 192, 32],
"in_dtype": "DT_FP8E4M3",
"out_dtype": "DT_FP16",
"a_trans": False,
"b_trans": True,
"scale_a_trans": True,
"scale_b_trans": True,
"a_format": "NZ",
"b_format": "ND",
"c_format": "ND",
"has_bias": False,
"bias_batch": 0,
"enable_ksplit": False,
"products": ["950"],
},
{
"id": "BMM4D_BIAS_1N",
"name": "scaled_bmm_4d_bias_1n",
"a_shape": [2, 3, 154, 256],
"b_shape": [1, 3, 65, 256],
"m_tile_shape": [64, 64],
"k_tile_shape": [64, 256],
"n_tile_shape": [256, 256],
"view_shape": [1, 3, 192, 32],
"in_dtype": "DT_FP8E5M2",
"out_dtype": "DT_FP16",
"a_trans": False,
"b_trans": True,
"scale_a_trans": False,
"scale_b_trans": True,
"a_format": "ND",
"b_format": "ND",
"c_format": "ND",
"has_bias": True,
"bias_batch": 1,
"bias_shape_type": "1_n",
"enable_ksplit": False,
"products": ["950"],
},
]