fix: remove incomplete quantize kernel

huggingface · Mar 15, 2024 · 5737cfb · 5737cfb
1 parent f3389c0
commit 5737cfb
Show file tree

Hide file tree

Showing 9 changed files with 0 additions and 147 deletions.
diff --git a/bench/library/benchmark.py b/bench/library/benchmark.py
@@ -20,17 +20,6 @@ def bench_fn():
  return bench_fn
 
 
-def get_quantize_symmetric_bench(src_dtype, dst_dtype, per_axis, device):
- a = torch.rand([10240, 10240], dtype=src_dtype).to(device)
- scale = torch.ones((10240,)) * 0.5 if per_axis else torch.tensor(0.5)
- scale = scale.to(src_dtype).to(device)
-
- def bench_fn():
- return torch.ops.quanto.quantize_symmetric(a, scale, dst_dtype)
-
- return bench_fn
-
-
 def get_unpack_bench(bits, device):
  qmax = 2**bits
  a = torch.randint(0, qmax, [10240, 10240], dtype=torch.uint8).to(device)
@@ -92,9 +81,6 @@ def elapsed_time(self, other):
 
 GET_BENCH_FUNCTIONS = {
  "dqmm_w8a16": lambda device: get_dqmm_bench(torch.float16, device),
- "quantize_symmetric_fp32_int8_per_tensor": lambda device: get_quantize_symmetric_bench(
- torch.float32, torch.int8, False, device
- ),
  "unpack_2bit": lambda device: get_unpack_bench(2, device),
  "unpack_4bit": lambda device: get_unpack_bench(4, device),
 }

diff --git a/quanto/library/ext/cpp/__init__.py b/quanto/library/ext/cpp/__init__.py
@@ -33,11 +33,6 @@ def dqmm_cpp(input: torch.Tensor, other: torch.Tensor, other_scale: torch.Tensor
  return ext().dqmm(input, other, other_scale)
 
 
-@torch.library.impl("quanto_ext::quantize_symmetric", ["CPU"])
-def quantize_symmetric_cpp(t: torch.Tensor, scale: torch.Tensor, dtype: torch.Tensor.dtype):
- return ext().quantize_symmetric(t, scale, dtype)
-
-
 @torch.library.impl("quanto_ext::unpack", ["CPU", "CUDA"])
 def unpack_cpp(t: torch.Tensor, bits: int):
  return ext().unpack(t, bits)
diff --git a/quanto/library/ext/cpp/pybind_module.cpp b/quanto/library/ext/cpp/pybind_module.cpp
@@ -1,22 +1,14 @@
 #include <torch/extension.h>
 #include "mm.h"
-#include "quantize.h"
 #include "unpack.h"
 
 // !IMPORTANT! Some python objects such as dtype, device, are not mapped to C++ types,
 // and need to be explicitly converted using dedicated helpers before calling a C++ method.
 // As a consequence, when an operation takes such an object as parameter, instead
 // of creating a binding directly to the C++ method, you must create a binding to a
 // lambda method that converts the unmapped types and calls the C++ method.
-// See the binding of quantize_symmetric for instance.
 
 PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
  m.def("dqmm", &dqmm, "dqmm");
- m.def("quantize_symmetric",
- [](const torch::Tensor& t, const torch::Tensor& scale, py::object dtype) {
- return quantize_symmetric(t,
- scale,
- torch::python::detail::py_object_to_dtype(dtype));
- }, "quantize_symmetric");
  m.def("unpack", &unpack, "unpack");
 }
diff --git a/quanto/library/ext/cpp/quantize.cpp b/quanto/library/ext/cpp/quantize.cpp
diff --git a/quanto/library/ext/cpp/quantize.h b/quanto/library/ext/cpp/quantize.h
diff --git a/quanto/library/ops.py b/quanto/library/ops.py
@@ -54,5 +54,4 @@ def impl(*args, **kwargs):
 
 
 define("dqmm", "(Tensor input, Tensor other, Tensor other_scale) -> Tensor")
-define("quantize_symmetric", "(Tensor self, Tensor scale, ScalarType dtype) -> Tensor")
 define("unpack", "(Tensor self, int bits) -> Tensor")
diff --git a/quanto/library/python/__init__.py b/quanto/library/python/__init__.py
@@ -1,3 +1,2 @@
 from .mm import *
-from .quantize import *
 from .unpack import *
diff --git a/quanto/library/python/quantize.py b/quanto/library/python/quantize.py
diff --git a/test/library/test_quantize.py b/test/library/test_quantize.py