mediapipe-nightly 0.10.21.post20250114__cp39-cp39-manylinux_2_28_x86_64.whl

mediapipe/tasks/python/genai/converter/quantization_util.py

@@ -0,0 +1,516 @@
+# Copyright 2024 The MediaPipe Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Utilities for quantizing tensors.
+
+Note that this is a reduced fork version of the praxis libraries to provide a
+self-contained library for packaging.
+"""
+
+from typing import Any, List, Optional, Sequence, Tuple, Union
+
+import jax
+from jax import lax
+from jax import numpy as jnp
+import numpy as np
+
+
+JTensor = jax.Array
+_UINT4_ZP = 8  # Default zero point for unsigned 4-bit.
+
+
+def _get_scan_range() -> np.ndarray:
+  # Produce candidate scan values.
+  return np.linspace(1.0, 0.5, num=11)
+
+
+def _get_mean_error(bound, t, min_value, max_value, p_value):
+  scale = bound / max_value
+  candidate = jnp.divide(t, scale)
+  candidate = jnp.clip(jnp.round(candidate), min_value, max_value)
+  candidate = jnp.multiply(candidate, scale)
+  pmean_error = jnp.mean(jnp.abs(jnp.subtract(candidate, t)) ** p_value)
+  return pmean_error
+
+
+def _get_best_bound_per_tensor(
+    t: JTensor,
+    bound: JTensor,
+    min_value: float,
+    max_value: float,
+    p_value: float = 1.0,
+) -> JTensor:
+  """Scan around [0.5, 1] * hard max value to get bound value for whole tensor.
+
+  This does a scan to get bound value(s) that minimize mean absolute error (MAE)
+  between original tensor 't' and quantized tensor. It's (almost) equivalent to
+  maximizing entropy.
+
+  Args:
+    t: The input float tensor.
+    bound: The hard max value for tensor 't'. It has the same length as shape.
+    min_value: Minimal value for the quantization bound.
+    max_value: Maximal value for the quantization bound.
+    p_value: Exponent of the p-mean error metric. Default to 1.0 which is MAE.
+
+  Returns:
+    The best bound values for 't', that minimize p-mean error.
+  """
+
+  def _quant(scaling_factors):
+    return _get_mean_error(
+        bound * scaling_factors, t, min_value, max_value, p_value
+    )
+
+  scaling_factors = _get_scan_range()
+  diffs = jax.vmap(_quant)(scaling_factors)
+  best_scaling = scaling_factors[jnp.argmin(diffs)].astype(bound.dtype)
+  return bound * best_scaling
+
+
+def _quantrow(
+    vec: JTensor,
+    bound: JTensor,
+    min_value: float,
+    max_value: float,
+    p_value: float,
+    factors: np.ndarray,
+) -> JTensor:
+  """Get best rescaling factor from a list of factors applied a channel.
+
+  Args:
+    vec: The vector in a channel.
+    bound: The hard bound (max(abs(vec))) of the vector.
+    min_value: The target min value.
+    max_value: The target max value.
+    p_value: Exponent of the p-mean error metric.
+    factors: The values to be applied on top of bound.
+
+  Returns:
+    adjusted bound value out of the list of factors applied to bound.
+  """
+
+  def _quant(bounds):
+    return _get_mean_error(bounds, vec, min_value, max_value, p_value)
+
+  diffs = jax.vmap(_quant)(bound * factors)
+  best_scaling = factors[jnp.argmin(diffs)]
+  return bound * best_scaling
+
+
+def _get_best_bound_per_channel(
+    t: JTensor,
+    bound: JTensor,
+    min_value: float,
+    max_value: float,
+    p_value: float = 1.0,
+) -> JTensor:
+  """Scan around [0.5, 1] * hard max value to get bound value for each channel.
+
+  This does a scan to get bound value(s) that minimize mean absolute error (MAE)
+  between original tensor 't' and quantized tensor. It's (almost) equivalent to
+  maximizing entropy.
+
+  Args:
+    t: The input float tensor.
+    bound: The hard max value for tensor 't'. It has the same length as shape.
+    min_value: Minimal value for the quantization bound.
+    max_value: Maximal value for the quantization bound.
+    p_value: Exponent of the p-mean error metric. Default to 1.0 which is MAE.
+
+  Returns:
+    The best bound values for 't', that minimize p-mean error.
+  """
+  assert len(t.shape) == 2
+  assert len(bound.shape) == 2
+  assert t.shape[1] == bound.shape[1]
+  assert bound.shape[0] == 1
+  scans = _get_scan_range()
+
+  def _quant(tensor, bound, min_value, max_value, p_value, factors):
+    ret = np.zeros(bound.shape)
+    for i in range(len(tensor)):
+      best = _quantrow(
+          tensor[i], bound[i], min_value, max_value, p_value, factors
+      )
+      ret[i] = best
+    return ret
+
+  t = t.transpose()
+  t_split = list(t)
+  res = _quant(t_split, bound[0, :], min_value, max_value, p_value, scans)
+  res = res.reshape(bound.shape)
+  return res
+
+
+def get_best_bound(
+    t: JTensor,
+    bound: JTensor,
+    min_value: float,
+    max_value: float,
+    p_value: float = 1.0,
+    per_channel: bool = False,
+) -> JTensor:
+  """Scan multiple factors on max value to get best bound value.
+
+  This does a scan to get bound value(s) that minimize mean absolute error (MAE)
+  between original tensor 't' and quantized tensor. It's (almost) equivalent to
+  maximizing entropy.
+
+  Args:
+    t: The input float tensor.
+    bound: The hard max value for tensor 't'. It has the same length as shape.
+    min_value: Minimal value for the quantization bound.
+    max_value: Maximal value for the quantization bound.
+    p_value: Exponent of the p-mean error metric. Default to 1.0 which is MAE.
+    per_channel: if get best bound for entire tensor or per channel.
+
+  Returns:
+    The best bound values for 't', that minimize p-mean error.
+  """
+  if per_channel:
+    return _get_best_bound_per_channel(t, bound, min_value, max_value, p_value)
+  else:
+    return _get_best_bound_per_tensor(t, bound, min_value, max_value, p_value)
+
+
+def get_min_max(
+    bits: int = 8,
+    unsigned: bool = False,
+    use_fp: bool = False,
+) -> Tuple[float, float]:
+  """Gets the min/max range for a given number of bits.
+
+  Args:
+    bits: Target number of bits for quantization.
+    unsigned: If True compute min and max for unsigned number, else for signed.
+    use_fp: in floating point.
+
+  Returns:
+    min/max values for the provide number of bits.
+  """
+  if use_fp:
+    # TODO: support other fp types.
+    return -448.0, 448.0
+  # Calculation instead of jax.iinfo is used to support bits beside 4 and 8.
+  if unsigned:
+    # For unsigned 8 bits precision it is [0, 255]
+    return 0, 2**bits - 1
+  else:
+    # For signed 8 bits precision it is [-128, 127]
+    return -1 * 2 ** (bits - 1), 2 ** (bits - 1) - 1
+
+
+def pass_through(x: JTensor, fn: Any) -> JTensor:
+  # Create an exactly-zero expression with Sterbenz lemma that has an
+  # exactly-one gradient.
+  return x - jax.lax.stop_gradient(x) + jax.lax.stop_gradient(fn(x))
+
+
+def reduce_precision(
+    t: JTensor,
+    contract_dims: Optional[Sequence[int]],
+    need_gradient: bool = False,
+    bits: int = 8,
+    optimization_on_bound: bool = False,
+    p_value: float = 1.0,
+    percentile: float = 1.0,
+    use_symmetric: bool = True,
+    use_fp: bool = False,
+    add_scale_eps: bool = False,
+    per_channel: bool = False,
+    random_rounding: bool = False,
+    key: Optional[jax.Array] = None,
+) -> Tuple[JTensor, JTensor, Optional[JTensor]]:
+  """Reduce the precision of a tensor.
+
+  Generic for all tensors.
+
+  Args:
+    t: Input tensor.
+    contract_dims: Specifies contracting dimensions of the input tensor.
+    need_gradient: If gradient is needed out of this function.
+    bits: Target number of bits.
+    optimization_on_bound: If MAE bound optimizer is used.
+    p_value: Exponent of the p-mean error metric. Default to 1.0 which is MAE.
+    percentile: Percentile Factor to apply on the min/max range. Setting this to
+      other than 1.0 disables optimization_on_bound.
+    use_symmetric: If the input tensor is quantized symmetrically.
+    use_fp: Use floating point.
+    add_scale_eps: Add eps value or replace zero value by 1 to avoid division by
+      zero.
+    per_channel: use per-channel clipping optimization.
+    random_rounding: round with uniform random.
+    key: rng key for rounding.
+
+  Returns:
+    A tuple of quantized tensor, quantization scale
+      and quantization zero point (optional).
+  """
+  min_value, max_value = get_min_max(bits, use_fp=use_fp)
+
+  if use_symmetric:
+    bound = jnp.max(jnp.abs(t), axis=contract_dims, keepdims=True)
+    scale_bound = max_value
+  else:
+    t_max = jnp.max(t, axis=contract_dims, keepdims=True)
+    t_min = jnp.min(t, axis=contract_dims, keepdims=True)
+    bound = t_max - t_min
+    scale_bound = max_value - min_value
+
+  if percentile < 1.0:
+    bound = jnp.multiply(bound, percentile)
+  elif optimization_on_bound:
+    bound = get_best_bound(
+        t, bound, min_value, max_value, p_value, per_channel=per_channel
+    )
+
+  scale = bound / scale_bound
+
+  if add_scale_eps:
+    # Add epsilon to avoid divide-by-zero.
+    scale = scale + jnp.finfo(t.dtype).eps
+  else:
+    scale = jnp.where(scale == 0.0, 1.0, scale)
+
+  if use_symmetric:
+    zp = None
+    t = jnp.divide(t, scale)
+  else:
+    zp = min_value - t_min / scale
+    t = jnp.divide(t, scale) + zp
+    zp = jnp.multiply(scale, zp)
+
+  if use_fp:
+    # No need to round.
+    t = jnp.clip(t, min_value, max_value).astype(jnp.float8_e4m3fn)
+    # TODO: refactor to remove this logic.
+    t = jax.lax.bitcast_convert_type(t, new_dtype=jnp.int8)
+  else:
+    if need_gradient:
+      t = pass_through(t, jnp.round)
+      t = jnp.clip(t, min_value, max_value)
+    else:
+      if random_rounding:
+        t = t + jax.random.uniform(
+            key=key, shape=t.shape, minval=-0.5, maxval=0.5
+        )
+      t = jnp.round(t)
+      container_dtype = (
+          jnp.int8 if bits <= 8 else jnp.int16 if bits <= 16 else jnp.int32
+      )
+      t = jnp.clip(t, min_value, max_value).astype(container_dtype)
+
+  return t, scale, zp
+
+
+def quantize_tensor(
+    var: np.ndarray,
+    axis: List[int],
+    factor: float = 1.0,
+    sym: bool = True,
+    number_bits: int = 8,
+    use_fp: bool = False,
+    add_scale_eps: bool = False,
+    optimization_on_bound: bool = False,
+    p_value: float = 1.0,
+    per_channel: bool = False,
+    block_size: int = 0,
+) -> Union[
+    Tuple[np.ndarray, np.ndarray], Tuple[np.ndarray, np.ndarray, np.ndarray]
+]:
+  """Quantize a tensor.
+
+  Args:
+    var: The variable to be quantized.
+    axis: The axis along which variable will be quantized.
+    factor: The clipping factor.
+    sym: Symmetric or asymmetric quantize the variable.
+    number_bits: Number of bits for quantized value.
+    use_fp: do fp with number of bits (i.e. fp8)
+    add_scale_eps: add epsilon to scale to avoid division by zero, else it will
+      replace zero scale by 1.
+    optimization_on_bound: If p-mean bound optimizer is used.
+    p_value: Exponent of the p-mean error metric. Default to 1.0 which is MAE.
+    per_channel: use per-channel clipping optimization.
+    block_size:  block size for sub-channel quantization. Defaults to 0, which
+      means off.
+
+  Returns:
+    Quantized tensors, along with scales and zero point.
+  """
+  # TODO: support jnp.float8_e5m2
+  assert number_bits == 8 or number_bits == 4
+  jnp_var = jnp.asarray(var)
+  # When using sub-channel, the contracting dim is split into a sub-channel
+  # dim followed by the block dim. Therefore the contracting dim
+  # (quantize_axis) should increment by one, and the corresponding pack_dim
+  # should also increment by one.
+  if block_size > 0:
+    shape = list(jnp_var.shape)
+    assert len(axis) == 1, 'Only support 1D sub-channel quantization'
+    sub_channels, rem = divmod(shape[axis[0]], block_size)
+    assert rem == 0
+    shape.insert(axis[0], sub_channels)
+    axis[0] += 1
+    shape[axis[0]] = block_size
+    jnp_var = jnp.reshape(jnp_var, shape)
+
+  qvar, scale, zp = reduce_precision(
+      jnp_var,
+      contract_dims=axis,
+      need_gradient=False,
+      bits=number_bits,
+      optimization_on_bound=optimization_on_bound,
+      percentile=factor,
+      use_symmetric=sym,
+      use_fp=use_fp,
+      add_scale_eps=add_scale_eps,
+      p_value=p_value,
+      per_channel=per_channel,
+  )
+  if sym:
+    return np.array(qvar), np.array(jnp.squeeze(scale, axis=axis))  # pytype: disable=wrong-arg-types  # jnp-type
+  else:
+    return (
+        np.array(qvar),
+        # CAVEAT: the following squeezes should squeeze along the quantization
+        # axis only.
+        np.array(jnp.squeeze(scale)),
+        np.array(jnp.squeeze(zp)),
+    )
+
+
+def pack_4bit(
+    x: np.ndarray, pack_dim: int, packed_dtype: jnp.dtype = jnp.int32
+) -> np.ndarray:
+  """Pack int8 or uint8 tensor where its values are actually int4 or uint4, to int32 or int8 nibble format along pack_dim.
+
+  Args:
+    x: Original int8 or uint8 tensor to pack.
+    pack_dim: Dimension to pack along. x.shape[pack_dim] must be divisible by 8,
+      when packed_dtype is int32 and divisible by 2 when target_type is int8.
+      Also pack_dim must be < x.ndim - 1.
+    packed_dtype: Target type to pack to, int32 or int8.
+
+  Returns:
+    int32 or int8 packed tensor where the pack_dim size is dividend by 8
+    from the original tensor x.
+  """
+  x = jnp.asarray(x)
+  if packed_dtype == jnp.int8 and x.dtype == jnp.uint8:
+    # It doesn't make sense to pack uint8 numbers into int4 as we'll
+    # the range overlap between uint8 and int4 is [0..7].
+    raise ValueError(
+        'only int8 input dtype is supported when packing into int8. '
+        f'Given {x.dtype}'
+    )
+
+  if x.dtype != jnp.int8 and x.dtype != jnp.uint8:
+    raise ValueError(
+        f'input dtype must be either int8 or uint8. Given {x.dtype}'
+    )
+  if pack_dim >= x.ndim - 1:
+    raise ValueError(
+        f'pack_dim must be < input ndim - 1. input shape {x.shape} and pack_dim'
+        f' {pack_dim}'
+    )
+  if packed_dtype != jnp.int32 and packed_dtype != jnp.int8:
+    raise ValueError(
+        f'packed_dtype must be either int32 or int8. Given {packed_dtype}'
+    )
+  if packed_dtype == jnp.int32 and x.shape[pack_dim] % 8 != 0:
+    raise ValueError(
+        'input shape[pack_dim] must be divisible by 8 when target_type '
+        f'is int32. Given shape {x.shape}'
+    )
+  if packed_dtype == jnp.int8 and x.shape[pack_dim] % 2 != 0:
+    raise ValueError(
+        'input shape[pack_dim] must be divisible by 2 when target_type '
+        f'is int8. Given shape {x.shape}'
+    )
+
+  int4s_per_packed_type = 8 if packed_dtype == jnp.int32 else 2
+
+  rep_shape = list(x.shape)
+  rep_shape.insert(pack_dim + 1, int4s_per_packed_type)
+  rep_shape[pack_dim] //= int4s_per_packed_type
+
+  shifts = lax.broadcasted_iota(packed_dtype, rep_shape, pack_dim + 1)
+  shifts <<= 2
+
+  # Promote x to packed_dtype
+  x = x & jnp.array(0x0F, packed_dtype)
+  x = lax.reshape(x, rep_shape)
+  x = x << shifts
+  x = lax.reduce(x, jnp.array(0x0, packed_dtype), lax.add, [pack_dim + 1])
+  return np.asarray(x)
+
+
+def update_to_uint4(
+    qx: np.ndarray, scale: np.ndarray, zp: Optional[np.ndarray] = None
+):
+  """Updates the quantized weights from int4 to uint4.
+
+  This is a conversion function designed for XNNPack as it expects the 4-bit
+  quantized weight to be represented differently from the original Pax setting.
+  Specifically, the differences are:
+    1) The dynamic range of weight values: int4 (Pax) vs. uint4 (XNNPack).
+    2) The dynamic range of zero-point: float (Pax) vs. uint4 (XNNPack).
+    3) The number of zero-point: per-channel (Pax) vs. per-tensor (XNNPack).
+
+  Args:
+    qx: np.array of shape [..., channel], which is the quantized weight values
+      from Pax in the shape of. The values are in the dynamic range of int4 but
+      are hosted as int8 type. Note that if the first dimension is 3, it means
+      the qkv matrices are concatenated together and should be treated
+      differently.
+    scale: np.array of shape [1(3), channel] as np.float type, which are the
+      scaling factors for dequantization per channel.
+    zp: (optional) np.array of shape [1 (or 3), channel] as np.float type, which
+      are the zero points for dequantization per channel.
+
+  Returns:
+    A tuple (qx, scale, zp):
+      qx: The updated np.array of shape [..., channel] as np.int8 type with
+        updated dynamic range as uint4 (with 8 as the default zero points).
+      scale: Same as the input scale.
+      zp: (optional) np.array of shape [1 (or 3)] as np.int8 type with the
+        updated zero point values in the dynamic range as uint4.
+  """
+  if qx.dtype != np.int8 or ('float' not in str(scale.dtype)):
+    raise ValueError(
+        'Unexpected dtype qx:' + str(qx.dtype) + ' scale:' + str(scale.dtype)
+    )
+
+  scale = scale.astype(np.float32)
+
+  def get_new_zp(old_zp):
+    new_zp = old_zp / (scale + np.finfo(np.float32).eps)
+    per_tensor_zp = np.mean(new_zp)
+    per_tensor_zp = per_tensor_zp.astype(np.int8) + _UINT4_ZP
+    return per_tensor_zp
+
+  if zp is not None:
+    if qx.shape[0] == 3:
+      per_tensor_zp = np.stack([get_new_zp(szp) for szp in zp], axis=0)
+    else:
+      per_tensor_zp = get_new_zp(zp)
+  else:
+    per_tensor_zp = (
+        _UINT4_ZP * np.ones(shape=(3)) if qx.shape[0] == 3 else _UINT4_ZP
+    )
+
+  qx = qx + _UINT4_ZP
+  return qx, scale, np.array(per_tensor_zp, dtype=np.int32)