Skip to content

GitLab

Commit faf97b39 authored by Catheriany's avatar Catheriany
Browse files

issue/228: gguf冗余移除，rms_norm测例

parent c2e87202
Hide whitespace changes
Inline Side-by-side
Showing with 68 additions and 172 deletions
src/infiniop-test/src/ops/rms_norm.cpp
View file @ faf97b39
......@@ -105,6 +105,10 @@ std::vector<std::string> Test::tensor_names() {
return {"x", "w", "ans", "y"};
}
std::vector<std::string> Test::output_names() {
return {"y"};
}
std::string Test::toString() const {
std::ostringstream oss;
oss << op_name() << std::endl;
......
test/infiniop-test/test_generate/testcases/rms_norm.py
View file @ faf97b39
from ast import List
import numpy as np
import gguf
from typing import Optional
from numpy.lib.stride_tricks import as_strided
from typing import List
from .. import InfiniopTestWriter, InfiniopTestCase, np_dtype_to_ggml, gguf_strides
from .. import InfiniopTestWriter, InfiniopTestCase, np_dtype_to_ggml, gguf_strides, contiguous_gguf_strides
def random_tensor(shape: tuple, dtype: np.dtype) -> np.ndarray:
return np.random.uniform(-1.0, 1.0, shape).astype(dtype) * 0.001
......@@ -29,46 +26,35 @@ def rms_norm(x: np.ndarray, w: np.ndarray, epsilon: float) -> np.ndarray:
class RMSNormTestCase(InfiniopTestCase):
def __init__(
self,
shape: tuple,
atype: np.dtype,
wtype: np.dtype,
x: np.ndarray,
w: np.ndarray,
y: np.ndarray,
shape: List[int] | None,
x_strides: List[int] | None,
y_strides: List[int] | None,
epsilon: float = 1e-5,
x_strides: Optional[tuple] = None,
y_strides: Optional[tuple] = None,
):
super().__init__("rms_norm")
self.x = x
self.w = w
self.y = y
self.shape = shape
self.atype = atype
self.w_shape = (shape[1],)
if x_strides is not None:
itemsize = np.dtype(atype).itemsize
byte_strides = tuple(s * itemsize for s in x_strides)
max_offset = sum((dim - 1) * abs(stride) for dim, stride in zip(shape, x_strides))
buffer_size = max_offset + 1
buffer = np.random.uniform(-1.0, 1.0, buffer_size).astype(atype) * 0.001
self.x = as_strided(buffer, shape=shape, strides=byte_strides)
else:
self.x = random_tensor(shape, atype)
self.w = random_tensor(self.w_shape, wtype)
self.epsilon = epsilon
self.x_strides=x_strides
self.y_strides=y_strides
if y_strides is not None:
itemsize_out = np.dtype(atype).itemsize
byte_strides_out = tuple(s * itemsize_out for s in y_strides)
max_offset_out = sum((dim - 1) * abs(stride) for dim, stride in zip(shape, y_strides))
buffer_size_out = max_offset_out + 1
buffer_out = np.zeros(buffer_size_out, dtype=atype)
self.y = as_strided(buffer_out, shape=shape, strides=byte_strides_out)
else:
self.y = np.zeros_like(self.x)
self.ans = rms_norm(self.x.astype(np.float64), self.w.astype(np.float64), self.epsilon)
def write_test(self, test_writer: "InfiniopTestWriter"):
super().write_test(test_writer)
test_writer.add_float32(test_writer.gguf_key("epsilon"), self.epsilon)
if self.shape is not None:
test_writer.add_array(test_writer.gguf_key("x.shape"), self.shape)
test_writer.add_array(test_writer.gguf_key("y.shape"), self.shape)
if self.x_strides is not None:
test_writer.add_array(test_writer.gguf_key("x.strides"), gguf_strides(*self.x_strides))
test_writer.add_array(
test_writer.gguf_key("y.strides"),
gguf_strides(*self.y_strides if self.y_strides is not None else contiguous_gguf_strides(self.shape))
)
test_writer.add_tensor(
test_writer.gguf_key("x"),
self.x,
......@@ -79,150 +65,56 @@ class RMSNormTestCase(InfiniopTestCase):
self.w,
raw_dtype=np_dtype_to_ggml(self.w.dtype),
)
test_writer.add_tensor(
test_writer.gguf_key("ans"),
self.ans,
raw_dtype=np_dtype_to_ggml(np.float64),
)
test_writer.add_tensor(
test_writer.gguf_key("y"),
self.y,
raw_dtype=np_dtype_to_ggml(self.y.dtype),
)
ans = rms_norm(self.x.astype(np.float64), self.w.astype(np.float64), self.epsilon)
test_writer.add_tensor(
test_writer.gguf_key("ans"),
ans,
raw_dtype=np_dtype_to_ggml(np.float64),
)
if __name__ == "__main__":
test_writer = InfiniopTestWriter("rms_norm.gguf")
test_cases = [
RMSNormTestCase(
shape=(2, 256),
atype=np.float32,
wtype=np.float32,
),
RMSNormTestCase(
shape=(4, 512),
atype=np.float32,
wtype=np.float32,
),
RMSNormTestCase(
shape=(8, 1024),
atype=np.float32,
wtype=np.float32,
),
RMSNormTestCase(
shape=(1, 768),
atype=np.float32,
wtype=np.float32,
),
RMSNormTestCase(
shape=(8, 256),
atype=np.float32,
wtype=np.float32,
),
RMSNormTestCase(
shape=(500, 4096),
atype=np.float32,
wtype=np.float32,
),
RMSNormTestCase(
shape=(2, 256),
atype=np.float16,
wtype=np.float16,
),
RMSNormTestCase(
shape=(4, 512),
atype=np.float16,
wtype=np.float16,
),
RMSNormTestCase(
shape=(500, 4096),
atype=np.float16,
wtype=np.float16,
),
RMSNormTestCase(
shape=(4, 512),
atype=np.float16,
wtype=np.float32,
),
RMSNormTestCase(
shape=(500, 4096),
atype=np.float16,
wtype=np.float32,
),
RMSNormTestCase(
shape=(4, 512),
atype=np.float32,
wtype=np.float32,
x_strides=(1024, 1),
),
RMSNormTestCase(
shape=(4, 512),
atype=np.float32,
wtype=np.float32,
x_strides=(512, 2),
),
RMSNormTestCase(
shape=(500, 4096),
atype=np.float32,
wtype=np.float32,
x_strides=(9192, 1),
),
RMSNormTestCase(
shape=(500, 4096),
atype=np.float32,
wtype=np.float32,
x_strides=(4096, 2),
),
RMSNormTestCase(
shape=(4, 512),
atype=np.float16,
wtype=np.float16,
x_strides=(1024, 1),
),
RMSNormTestCase(
shape=(500, 4096),
atype=np.float16,
wtype=np.float16,
x_strides=(9192, 1),
),
RMSNormTestCase(
shape=(4, 512),
atype=np.float16,
wtype=np.float32,
x_strides=(1024, 1),
),
RMSNormTestCase(
shape=(500, 4096),
atype=np.float16,
wtype=np.float32,
x_strides=(9192, 1),
),
RMSNormTestCase(
shape=(4, 512),
atype=np.float32,
wtype=np.float32,
y_strides=(1024, 1),
),
RMSNormTestCase(
shape=(500, 4096),
atype=np.float32,
wtype=np.float32,
y_strides=(8192, 2),
),
RMSNormTestCase(
shape=(4, 512),
atype=np.float32,
wtype=np.float32,
x_strides=(1024, 1),
y_strides=(512, 2),
),
RMSNormTestCase(
shape=(4, 512),
atype=np.float16,
wtype=np.float16,
y_strides=(2048, 1),
),
test_cases = []
_TEST_CASES_ = [
# shape, x_strides, y_strides
((2, 256), None, None),
((4, 512), None, None),
((8, 1024), None, None),
((1, 768), None, None),
((8, 256), None, None),
((500, 4096), None, None),
((4, 512), (1024, 1), None),
((4, 512), (512, 1), None),
((500, 4096), (9192, 1), None),
((500, 4096), (4096, 1), None),
((4, 512), None, (1024, 1)),
((500, 4096), None, (8192, 1)),
((4, 512), (1024, 1), (512, 1)),
((4, 512), None, (2048, 1)),
]
_TENSOR_DTYPES_ = [np.float32, np.float16]
for dtype in _TENSOR_DTYPES_:
for shape, x_strides, y_strides in _TEST_CASES_:
w = np.random.rand(shape[-1]).astype(dtype)
x = np.random.rand(*shape).astype(dtype)
y = np.empty(tuple(0 for _ in shape), dtype=dtype)
epsilon = 1e-5
test_case = RMSNormTestCase(
x=x,
w=w,
y=y,
shape=shape,
x_strides=x_strides,
y_strides=y_strides,
epsilon=epsilon
)
test_cases.append(test_case)
test_writer.add_tests(test_cases)
test_writer.save()
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment