File size: 10,144 Bytes
1faccd4 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 | # Copyright 2026 Amazon.com Inc and/or its affiliates
#
# 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.
import torch
from tensordict import TensorDict
from verl.workers.utils.padding import left_right_2_no_padding, no_padding_2_padding
def test_padding_conversion_with_log_probs():
"""Test that log probability tensors remain in padded format after conversion
This test verifies the fix for the bug where ratio values were ~451,728 instead of ~1.0.
The key insight is that old_log_probs should STAY in padded format and be sliced
in the loss computation to match log_prob from model output, rather than being
converted to nested format.
"""
batch_size = 4
max_seq_len = 128
max_response_len = 64
# Create test data with varying sequence lengths
input_ids = torch.randint(0, 1000, (batch_size, max_seq_len))
# Create attention masks with different valid lengths per sample
attention_mask = torch.zeros(batch_size, max_seq_len)
valid_lens = [100, 120, 90, 128] # Different lengths for each batch item
for i, vlen in enumerate(valid_lens):
attention_mask[i, :vlen] = 1
# Create response masks aligned with the end of each sequence
response_mask = torch.zeros(batch_size, max_response_len)
response_lens = [50, 60, 45, 64] # Different response lengths
for i, rlen in enumerate(response_lens):
response_mask[i, :rlen] = 1
# Create position IDs
position_ids = torch.arange(max_seq_len).unsqueeze(0).expand(batch_size, -1)
# Add log probability tensors in padded format
old_log_probs = torch.randn(batch_size, max_seq_len)
ref_log_prob = torch.randn(batch_size, max_seq_len)
advantages = torch.randn(batch_size, max_response_len)
rollout_log_probs = torch.randn(batch_size, max_seq_len)
data = TensorDict(
{
"input_ids": input_ids,
"attention_mask": attention_mask,
"response_mask": response_mask,
"position_ids": position_ids,
"old_log_probs": old_log_probs,
"ref_log_prob": ref_log_prob,
"advantages": advantages,
"rollout_log_probs": rollout_log_probs,
}
)
# Convert to no-padding format
data_converted = left_right_2_no_padding(data)
# Verify input_ids and position_ids are nested tensors
assert isinstance(data_converted["input_ids"], torch.Tensor)
assert data_converted["input_ids"].is_nested
assert data_converted["position_ids"].is_nested
# Verify log probs REMAIN in padded format (NOT converted to nested)
# They will be sliced in the loss computation to match log_prob format
assert isinstance(data_converted["old_log_probs"], torch.Tensor)
assert not data_converted["old_log_probs"].is_nested, "old_log_probs should remain in padded format"
assert not data_converted["ref_log_prob"].is_nested, "ref_log_prob should remain in padded format"
assert not data_converted["advantages"].is_nested, "advantages should remain in padded format"
assert not data_converted["rollout_log_probs"].is_nested, "rollout_log_probs should remain in padded format"
# Verify they maintain their original shapes
assert data_converted["old_log_probs"].shape == (batch_size, max_seq_len)
assert data_converted["ref_log_prob"].shape == (batch_size, max_seq_len)
assert data_converted["advantages"].shape == (batch_size, max_response_len)
assert data_converted["rollout_log_probs"].shape == (batch_size, max_seq_len)
# Verify that nested tensors (input_ids, position_ids) have correct number of elements per batch item
for i, vlen in enumerate(valid_lens):
assert data_converted["input_ids"][i].numel() == vlen, (
f"Batch {i}: input_ids should have {vlen} elements, got {data_converted['input_ids'][i].numel()}"
)
def test_padding_conversion_without_log_probs():
"""Test that padding conversion works correctly when log prob tensors are not present"""
batch_size = 4
max_seq_len = 128
max_response_len = 64
# Create minimal test data
input_ids = torch.randint(0, 1000, (batch_size, max_seq_len))
attention_mask = torch.ones(batch_size, max_seq_len)
response_mask = torch.ones(batch_size, max_response_len)
position_ids = torch.arange(max_seq_len).unsqueeze(0).expand(batch_size, -1)
data = TensorDict(
{
"input_ids": input_ids,
"attention_mask": attention_mask,
"response_mask": response_mask,
"position_ids": position_ids,
}
)
# Convert to no-padding format
data_converted = left_right_2_no_padding(data)
# Verify basic conversion works
assert data_converted["input_ids"].is_nested
assert data_converted["position_ids"].is_nested
assert "old_log_probs" not in data_converted
assert "ref_log_prob" not in data_converted
def test_padding_roundtrip():
"""Test that converting from padding to nested and back preserves values in the response region"""
batch_size = 2
max_seq_len = 64
max_response_len = 32
prompt_len = max_seq_len - max_response_len # 32
# Create simple test data with known values
input_ids = torch.arange(1, max_seq_len + 1).unsqueeze(0).expand(batch_size, -1).clone()
attention_mask = torch.ones(batch_size, max_seq_len)
response_mask = torch.ones(batch_size, max_response_len)
position_ids = torch.arange(max_seq_len).unsqueeze(0).expand(batch_size, -1)
# Create nested prompts and responses (required by no_padding_2_padding)
prompt_list = [input_ids[i, :prompt_len] for i in range(batch_size)]
response_list = [input_ids[i, prompt_len:] for i in range(batch_size)]
prompts_nested = torch.nested.as_nested_tensor(prompt_list, layout=torch.jagged)
responses_nested = torch.nested.as_nested_tensor(response_list, layout=torch.jagged)
data = TensorDict(
{
"input_ids": input_ids,
"prompts": prompts_nested,
"responses": responses_nested,
"attention_mask": attention_mask,
"response_mask": response_mask,
"position_ids": position_ids,
}
)
# Convert to nested format
data_nested = left_right_2_no_padding(data)
# Verify input_ids is nested
assert data_nested["input_ids"].is_nested
# Convert back to padding format
recovered = no_padding_2_padding(data_nested["input_ids"], data_nested)
# Verify the shape is correct (response region only)
assert recovered.shape == (batch_size, max_response_len)
# Verify values are correct (left-shifted by 1 for log_probs alignment)
# Response tokens are 33,34,...,64 -> left-shifted: 32,33,...,63
expected = torch.arange(prompt_len, max_seq_len, dtype=torch.long).unsqueeze(0).expand(batch_size, -1)
torch.testing.assert_close(recovered, expected)
def test_no_padding_2_padding_varying_lengths():
"""Test no_padding_2_padding with varied prompt/response lengths."""
batch_size = 4
max_seq_len = 100
max_response_len = 50
prompt_lens = [10, 30, 5, 40]
response_lens = [40, 20, 45, 10]
input_ids = torch.zeros(batch_size, max_seq_len, dtype=torch.long)
for i in range(batch_size):
total_len = prompt_lens[i] + response_lens[i]
input_ids[i, :total_len] = torch.arange(1, total_len + 1)
attention_mask = torch.zeros(batch_size, max_seq_len)
for i in range(batch_size):
attention_mask[i, : prompt_lens[i] + response_lens[i]] = 1
response_mask = torch.zeros(batch_size, max_response_len)
for i in range(batch_size):
response_mask[i, : response_lens[i]] = 1
position_ids = torch.arange(max_seq_len).unsqueeze(0).expand(batch_size, -1).clone()
prompt_list = [input_ids[i, : prompt_lens[i]] for i in range(batch_size)]
response_list = [input_ids[i, prompt_lens[i] : prompt_lens[i] + response_lens[i]] for i in range(batch_size)]
prompts_nested = torch.nested.as_nested_tensor(prompt_list, layout=torch.jagged)
responses_nested = torch.nested.as_nested_tensor(response_list, layout=torch.jagged)
data = TensorDict(
{
"input_ids": input_ids,
"attention_mask": attention_mask,
"response_mask": response_mask,
"position_ids": position_ids,
"prompts": prompts_nested,
"responses": responses_nested,
}
)
data_nested = left_right_2_no_padding(data)
input_ids_nested = data_nested["input_ids"]
log_probs_values = input_ids_nested.values().float()
log_probs_nested = torch.nested.nested_tensor_from_jagged(log_probs_values, offsets=input_ids_nested.offsets())
result_slice_response = no_padding_2_padding(log_probs_nested, data_nested)
# Verify no_padding_2_padding produces correct values (left-shifted by 1)
for i in range(batch_size):
resp_len = response_lens[i]
expected_start = prompt_lens[i]
expected_values = torch.arange(expected_start, expected_start + resp_len, dtype=torch.float)
torch.testing.assert_close(
result_slice_response[i, :resp_len],
expected_values,
rtol=1e-5,
atol=1e-6,
msg=f"Batch {i} (prompt_len={prompt_lens[i]}, resp_len={resp_len}): values incorrect",
)
print("All varied length tests passed")
if __name__ == "__main__":
test_padding_conversion_with_log_probs()
test_padding_conversion_without_log_probs()
test_padding_roundtrip()
test_no_padding_2_padding_varying_lengths()
print("All padding conversion tests passed!")
|