4b quantized embedding table operator (#3050)

Summary:

4b quantized embedding table operator

Differential Revision: D56123408

Author: manuelcandales

kernels/quantized/cpu/op_embedding4b.cpp

@@ -0,0 +1,346 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <executorch/runtime/kernel/kernel_includes.h>
+#include <algorithm>
+#include <cinttypes>
+#include <cmath>
+
+namespace torch {
+namespace executor {
+namespace native {
+
+using Tensor = exec_aten::Tensor;
+using Scalar = exec_aten::Scalar;
+using ScalarType = exec_aten::ScalarType;
+
+namespace {
+
+/**
+ * Asserts that the parameters are valid.
+ */
+void check_embedding_4bit_args(
+    const Tensor& weight,
+    const Tensor& weight_scales,
+    const optional<Tensor>& opt_weight_zero_points,
+    const int64_t weight_quant_min,
+    const int64_t weight_quant_max,
+    const Tensor& indices,
+    exec_aten::optional<ScalarType> out_dtype,
+    Tensor& out) {
+  ET_CHECK_MSG(
+      weight.dim() == 2, "weight must be 2D but got() %zd dims", weight.dim());
+
+  ET_CHECK_MSG(
+      weight_scales.dim() == 1 || weight_scales.dim() == 2,
+      "weight_scales must be 1D or 2D but got() %zd dims",
+      weight_scales.dim());
+
+  ET_CHECK_MSG(
+      weight_scales.size(0) == weight.size(0),
+      "Number of scales must be == weight.size(0)=%zd"
+      ", but got %zd",
+      weight_scales.size(0),
+      weight.size(0));
+
+  if (weight_scales.dim() == 2) {
+    auto num_groups = weight_scales.size(1);
+    ET_CHECK_MSG(
+        // each 8b uint8 column is 2 columns
+        (2 * weight.size(1)) % num_groups == 0,
+        "Number of groups must divide weight.size(1)=%zd"
+        ", but got # of groups = %zd",
+        weight.size(1),
+        num_groups);
+  }
+
+  ET_CHECK_MSG(
+      weight.scalar_type() == ScalarType::Byte,
+      "weight.scalar_type() %" PRId8 " is not supported:",
+      static_cast<int8_t>(weight.scalar_type()));
+
+  ET_CHECK_MSG(
+      out.scalar_type() == ScalarType::Float ||
+          out.scalar_type() == ScalarType::Half,
+      "out.scalar_type() %" PRId8 " is not supported:",
+      static_cast<int8_t>(out.scalar_type()));
+
+  ET_CHECK_MSG(
+      weight_scales.scalar_type() == ScalarType::Float ||
+          weight_scales.scalar_type() == ScalarType::Half,
+      "weight_scales.scalar_type() %" PRId8 " is not supported:",
+      static_cast<int8_t>(weight_scales.scalar_type()));
+
+  if (opt_weight_zero_points.has_value()) {
+    ET_CHECK_MSG(
+        opt_weight_zero_points.value().dim() == weight_scales.dim(),
+        "weight_zero_points's rank match that of weight_scales. "
+        "weight_zero_points rank: %" PRId8 ", weight_scales rank: %" PRId8,
+        static_cast<int8_t>(opt_weight_zero_points.value().dim()),
+        static_cast<int8_t>(weight_scales.dim()));
+
+    ET_CHECK_MSG(
+        opt_weight_zero_points.value().scalar_type() == out.scalar_type(),
+        "weight zero points scalar type %" PRId8
+        " does not match out.scalar_type()",
+        static_cast<int8_t>(opt_weight_zero_points.value().scalar_type()));
+
+    for (int32_t i = 0; i < weight_scales.dim(); ++i) {
+      ET_CHECK_MSG(
+          opt_weight_zero_points.value().size(i) == weight_scales.size(i),
+          "Dimension size misatch at dim %" PRId8
+          "Weight_zero_point size = %zd"
+          ", weight_scales size = %zd.",
+          i,
+          opt_weight_zero_points.value().size(i),
+          weight_scales.size(i));
+    }
+  }
+
+  ET_CHECK_MSG(
+      indices.scalar_type() == ScalarType::Long,
+      "indices.scalar_type() %" PRId8 " is not Long only Long is supported:",
+      static_cast<int8_t>(indices.scalar_type()));
+
+  ET_CHECK_MSG(
+      weight_quant_min <= weight_quant_max,
+      "weight quant min: %" PRId64
+      " is greater than weight quant max: %" PRId64,
+      weight_quant_min,
+      weight_quant_max);
+
+  if (out_dtype.has_value()) {
+    ET_CHECK_MSG(
+        out.scalar_type() == out_dtype.value(),
+        "output_dtype must match the dtype of the out tensor");
+  }
+}
+
+static inline int32_t weight_value(const unsigned char* w_data, int32_t index) {
+  int32_t odd = index & 1;
+  index >>= 1;
+  if (odd) {
+    return (int32_t)(w_data[index] & 0x0F) - 8;
+  } else {
+    return (int32_t)((w_data[index] >> 4) & 0x0F) - 8;
+  }
+}
+
+/**
+ * Retrieves the embeddings specified by indices, dequantizes them, and stores
+ * them in out. Weight will always be uint8
+ */
+template <typename CTYPE_PARAMS, typename CTYPE_OUT>
+void embedding_4bit_per_channel(
+    const Tensor& weight,
+    const Tensor& weight_scales,
+    const optional<Tensor>& opt_weight_zero_points,
+    const Tensor& indices,
+    Tensor& out) {
+  auto embedding_dim = weight.size(1) * 2;
+
+  int32_t num_groups_per_channel = 1;
+  if (weight_scales.dim() == 2) {
+    num_groups_per_channel = weight_scales.size(1);
+  }
+  int32_t group_size = embedding_dim / num_groups_per_channel;
+
+  CTYPE_OUT* out_data = out.mutable_data_ptr<CTYPE_OUT>();
+  const int64_t* indices_ptr = indices.const_data_ptr<int64_t>();
+
+  const CTYPE_PARAMS* scales = weight_scales.const_data_ptr<CTYPE_PARAMS>();
+  const CTYPE_PARAMS* zero_points = nullptr;
+  if (opt_weight_zero_points.has_value()) {
+    zero_points = opt_weight_zero_points.value().const_data_ptr<CTYPE_PARAMS>();
+  }
+
+  for (int i = 0; i < indices.numel(); i++) {
+    int64_t index = indices_ptr[i];
+    // If using groupwise embedding
+    int32_t qparams_index = index * num_groups_per_channel;
+    CTYPE_PARAMS zp = 0.0;
+    const CTYPE_PARAMS* scale_ptr = scales + qparams_index;
+    const CTYPE_PARAMS* zero_points_ptr = nullptr;
+    if (opt_weight_zero_points.has_value()) {
+      zero_points_ptr = zero_points + qparams_index;
+    }
+
+    const uint8_t* w_data = weight.data_ptr<uint8_t>() + weight.size(1) * index;
+
+    for (int j = 0; j < embedding_dim; ++j) {
+      int32_t group_id = j / group_size;
+      const CTYPE_PARAMS scale = scale_ptr[group_id];
+      if (opt_weight_zero_points.has_value()) {
+        zp = zero_points_ptr[group_id];
+      }
+      out_data[j] = static_cast<CTYPE_OUT>(
+          (static_cast<float>(weight_value(w_data, j)) -
+           static_cast<float>(zp)) *
+          static_cast<float>(scale));
+    }
+    out_data += embedding_dim;
+  }
+}
+
+void resize_out_tensor(
+    const Tensor& weight,
+    const Tensor& indices,
+    Tensor& out) {
+  exec_aten::SizesType expected_output_size[kTensorDimensionLimit];
+  for (size_t i = 0; i < indices.dim(); i++) {
+    expected_output_size[i] = indices.size(i);
+  }
+  const size_t embedding_dim = weight.size(1);
+  expected_output_size[out.dim() - 1] = embedding_dim;
+
+  exec_aten::ArrayRef<exec_aten::SizesType> output_size{
+      expected_output_size, static_cast<size_t>(out.dim())};
+
+  torch::executor::Error err = resize_tensor(out, output_size);
+  ET_CHECK_MSG(
+      err == torch::executor::Error::Ok,
+      "Failed to resize out Tensor in quantized_embedding_4bit_out");
+}
+
+} // namespace
+
+/**
+ * Retrieves the embeddings specified by indices, dequantizes them, and stores
+ * them in out. The weight is quantized per channel, with a scale and zero_point
+ * for each embedding.
+ *
+ * Corresponds as the out variant to torch.ops.quantized.embedding_4bit
+ *
+ * NOTE: quant_min, quant_max, and Dtype are not used in computation, but rather
+ * metadata that is passed around which can be useful for pattern matching. See
+ * https://github.com/pytorch/pytorch/pull/87093#discussion_r1000841181 for more
+ * info.
+ */
+Tensor& quantized_embedding_4bit_out(
+    // TODO Evaluate whether this name is appropriate for an operator that takes
+    // non quant input and returns fp output
+    const Tensor& weight,
+    const Tensor& weight_scales,
+    const optional<Tensor>& opt_weight_zero_points,
+    const int64_t weight_quant_min,
+    const int64_t weight_quant_max,
+    const Tensor& indices,
+    Tensor& out) {
+  ScalarType w_type = weight.scalar_type();
+  ScalarType out_type = out.scalar_type();
+
+  // TODO (jakeszwe): improve these to account for the size of out in relation
+  // to weight and indices accounting for a possible batch dimension
+  check_embedding_4bit_args(
+      weight,
+      weight_scales,
+      opt_weight_zero_points,
+      weight_quant_min,
+      weight_quant_max,
+      indices,
+      out_type,
+      out);
+
+  constexpr auto name = "quantized_decomposed::embedding_4bit.out";
+  ET_SWITCH_TWO_TYPES(Float, Half, out_type, ctx, name, CTYPE_OUT, [&]() {
+    embedding_4bit_per_channel<CTYPE_OUT, CTYPE_OUT>(
+        weight, weight_scales, opt_weight_zero_points, indices, out);
+  });
+
+  return out;
+}
+
+Tensor& quantized_embedding_4bit_out(
+    RuntimeContext& context,
+    const Tensor& weight,
+    const Tensor& weight_scales,
+    const optional<Tensor>& opt_weight_zero_points,
+    int64_t weight_quant_min,
+    int64_t weight_quant_max,
+    const Tensor& indices,
+    Tensor& out) {
+  // TODO(larryliu): Add a context arg to the real op function and remove this
+  // wrapper
+  (void)context;
+  resize_out_tensor(weight, indices, out);
+  return quantized_embedding_4bit_out(
+      weight,
+      weight_scales,
+      opt_weight_zero_points,
+      weight_quant_min,
+      weight_quant_max,
+      indices,
+      out);
+}
+
+Tensor& quantized_embedding_4bit_dtype_out(
+    // TODO Evaluate whether this name is appropriate for an operator that takes
+    // non quant input and returns fp output
+    const Tensor& weight,
+    const Tensor& weight_scales,
+    const optional<Tensor>& opt_weight_zero_points,
+    const int64_t weight_quant_min,
+    const int64_t weight_quant_max,
+    const Tensor& indices,
+    exec_aten::optional<ScalarType> out_dtype,
+    Tensor& out) {
+  // TODO (jakeszwe): improve these to account for the size of out in relation
+  // to weight and indices accounting for a possible batch dimension
+  check_embedding_4bit_args(
+      weight,
+      weight_scales,
+      opt_weight_zero_points,
+      weight_quant_min,
+      weight_quant_max,
+      indices,
+      out_dtype,
+      out);
+
+  ScalarType weight_type = weight.scalar_type();
+  ScalarType params_type = weight_scales.scalar_type();
+  ScalarType out_type = out.scalar_type();
+
+  constexpr auto name = "quantized_decomposed::embedding_4bit.dtype_out";
+  ET_SWITCH_TWO_TYPES(Float, Half, params_type, ctx, name, CTYPE_P, [&]() {
+    ET_SWITCH_TWO_TYPES(Float, Half, out_type, ctx, name, CTYPE_OUT, [&]() {
+      embedding_4bit_per_channel<CTYPE_P, CTYPE_OUT>(
+          weight, weight_scales, opt_weight_zero_points, indices, out);
+    });
+  });
+
+  return out;
+}
+
+Tensor& quantized_embedding_4bit_dtype_out(
+    RuntimeContext& context,
+    const Tensor& weight,
+    const Tensor& weight_scales,
+    const optional<Tensor>& opt_weight_zero_points,
+    int64_t weight_quant_min,
+    int64_t weight_quant_max,
+    const Tensor& indices,
+    exec_aten::optional<ScalarType> out_dtype,
+    Tensor& out) {
+  // TODO(larryliu): Add a context arg to the real op function and remove this
+  // wrapper
+  (void)context;
+  resize_out_tensor(weight, indices, out);
+  return quantized_embedding_4bit_dtype_out(
+      weight,
+      weight_scales,
+      opt_weight_zero_points,
+      weight_quant_min,
+      weight_quant_max,
+      indices,
+      out_dtype,
+      out);
+}
+
+} // namespace native
+} // namespace executor
+} // namespace torch

kernels/quantized/cpu/targets.bzl

@@ -23,6 +23,9 @@ _QUANT_OPS = (
     op_target(
         name = "op_embedding",
     ),
+    op_target(
+        name = "op_embedding4b",
+    ),
     op_target(
         name = "op_mixed_mm",
         deps = [

kernels/quantized/quantized.yaml

@@ -46,6 +46,18 @@
     - arg_meta: null
       kernel_name: torch::executor::quantized_embedding_byte_dtype_out
 
+- func: quantized_decomposed::embedding_4bit.out(Tensor weight, Tensor weight_scales, Tensor? weight_zero_points, int weight_quant_min, int weight_quant_max, Tensor indices, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::quantized_embedding_4bit_out
+
+- func: quantized_decomposed::embedding_4bit.dtype_out(Tensor weight, Tensor weight_scales, Tensor? weight_zero_points, int weight_quant_min, int weight_quant_max, Tensor indices, ScalarType? dtype=None, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::quantized_embedding_4bit_dtype_out
+
 - func: quantized_decomposed::mixed_mm.out(Tensor input, Tensor weight, Tensor weight_scales, Tensor? weight_zero_points, *, Tensor(a!) out) -> Tensor(a!)
   variants: function
   kernels: