File size: 6,822 Bytes
53c10a4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 | # Copyright 2024 Bytedance Ltd. and/or its affiliates
# Copyright 2026 The OraRL 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.
"""Shared tensor and duck-typed batch utilities.
Parts of the masking and grouping behavior are derived from the Apache-2.0
EasyR1/verl implementation.
"""
from __future__ import annotations
from collections import defaultdict
from collections.abc import Hashable, Sequence
from typing import Any
import numpy as np
import torch
def group_rows(
group_ids: Sequence[Any] | np.ndarray | torch.Tensor,
size: int,
) -> dict[Any, list[int]]:
"""Return row indices grouped in first-seen order."""
if isinstance(group_ids, torch.Tensor):
if group_ids.ndim != 1:
raise ValueError("group_ids must be one-dimensional.")
values = group_ids.detach().cpu().tolist()
else:
array = np.asarray(group_ids, dtype=object)
if array.ndim != 1:
raise ValueError("group_ids must be one-dimensional.")
values = array.tolist()
if len(values) != size:
raise ValueError(f"group_ids has {len(values)} rows, expected {size}.")
grouped: dict[Any, list[int]] = defaultdict(list)
for row, value in enumerate(values):
if isinstance(value, np.generic):
value = value.item()
if not isinstance(value, Hashable):
raise TypeError(f"group id at row {row} is not hashable.")
grouped[value].append(row)
return dict(grouped)
def boolean_mask(
values: Sequence[bool] | np.ndarray | torch.Tensor,
size: int,
*,
device: torch.device,
name: str = "mask",
) -> torch.Tensor:
"""Materialize a one-dimensional boolean mask on ``device``."""
if isinstance(values, torch.Tensor):
result = values.detach().to(device=device, dtype=torch.bool)
else:
array = np.asarray(values, dtype=bool)
result = torch.as_tensor(array, dtype=torch.bool, device=device)
if result.ndim != 1 or result.numel() != size:
raise ValueError(f"{name} must be one-dimensional with {size} rows.")
return result
def validate_floating_tensor(name: str, value: torch.Tensor) -> None:
if not isinstance(value, torch.Tensor):
raise TypeError(f"{name} must be a torch.Tensor.")
if not value.is_floating_point():
raise TypeError(f"{name} must have a floating dtype.")
if not bool(torch.isfinite(value).all()):
raise ValueError(f"{name} contains non-finite values.")
def validate_response_mask(response_mask: torch.Tensor) -> None:
"""Accept boolean, integer, or floating response masks."""
if not isinstance(response_mask, torch.Tensor):
raise TypeError("response_mask must be a torch.Tensor.")
if response_mask.is_complex():
raise TypeError("response_mask must have a real-valued dtype.")
if not bool(torch.isfinite(response_mask).all()):
raise ValueError("response_mask contains non-finite values.")
def sequence_rewards(
rewards: torch.Tensor,
response_mask: torch.Tensor | None,
) -> tuple[torch.Tensor, bool]:
"""Collapse token rewards, returning ``(scores, token_input)``."""
validate_floating_tensor("rewards", rewards)
if rewards.ndim == 1:
if response_mask is not None:
raise ValueError("response_mask is only valid with token-level rewards.")
return rewards.detach(), False
if rewards.ndim != 2:
raise ValueError("rewards must be one- or two-dimensional.")
if response_mask is None or response_mask.shape != rewards.shape:
raise ValueError("token-level rewards require a matching response_mask.")
validate_response_mask(response_mask)
mask = response_mask.detach().to(dtype=rewards.dtype)
return (rewards.detach() * mask).sum(dim=-1), True
def sequence_advantages(
advantages: torch.Tensor,
response_mask: torch.Tensor,
) -> torch.Tensor:
"""Collapse token advantages with a response-length-neutral masked mean."""
validate_floating_tensor("advantages", advantages)
validate_response_mask(response_mask)
if advantages.ndim != 2 or advantages.shape != response_mask.shape:
raise ValueError("advantages and response_mask must be matching matrices.")
mask = response_mask.detach().to(dtype=advantages.dtype)
lengths = mask.sum(dim=-1)
if bool((lengths <= 0).any()):
raise ValueError("every row must contain at least one valid response token.")
return (advantages.detach() * mask).sum(dim=-1) / lengths
def broadcast_sequence_values(
values: torch.Tensor,
response_mask: torch.Tensor,
*,
dtype: torch.dtype | None = None,
) -> torch.Tensor:
"""Broadcast one scalar per row over valid response tokens."""
if values.ndim != 1 or values.shape[0] != response_mask.shape[0]:
raise ValueError("values must contain one scalar per response-mask row.")
target_dtype = response_mask.dtype if dtype is None else dtype
return values.to(dtype=target_dtype).unsqueeze(-1) * response_mask.to(dtype=target_dtype)
def sample_std(values: torch.Tensor) -> torch.Tensor:
"""Sample standard deviation, with zero for fewer than two rows."""
if values.numel() < 2:
return values.new_zeros(())
return values.std(unbiased=True)
def batch_reward_tokens(data: Any, preferred_key: str | None = None) -> torch.Tensor:
"""Read token rewards from a DataProto-like object."""
if preferred_key is not None:
if preferred_key not in data.batch:
raise ValueError(f"batch is missing reward key {preferred_key!r}.")
return data.batch[preferred_key]
for key in ("token_level_scores", "token_level_rewards"):
if key in data.batch:
return data.batch[key]
raise ValueError("batch requires token_level_scores or token_level_rewards.")
def copy_and_refresh_meta(data: Any) -> None:
"""Detach selected metadata and refresh token counts when available."""
if not hasattr(data, "meta_info"):
return
meta = getattr(data, "meta_info")
data.meta_info = dict(meta) if meta is not None else {}
if hasattr(data, "batch") and "attention_mask" in data.batch:
data.meta_info["global_token_num"] = (
data.batch["attention_mask"].sum(dim=-1).detach().cpu().tolist()
)
|