File size: 8,583 Bytes
79a20c7 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 | from __future__ import annotations
import json
from pathlib import Path
import pytest
import torch
from dlb.io import atomic_json_write
def test_token_projector_clamps_every_call_without_mutating_input():
from dlb.adapters.conditional_runtime import token_project_fn
prefix = torch.tensor([[4, 5], [6, 7]])
project = token_project_fn(prefix)
state = torch.tensor([[99, 99, 8], [99, 99, 9]])
result = project(state)
assert result.tolist() == [[4, 5, 8], [6, 7, 9]]
assert state.tolist() == [[99, 99, 8], [99, 99, 9]]
state[:, :2] = 88
assert project(state).tolist() == [[4, 5, 8], [6, 7, 9]]
def test_vocab_projector_writes_clean_one_hot_prefix_and_preserves_suffix():
from dlb.adapters.conditional_runtime import vocab_project_fn
prefix = torch.tensor([[1, 3]])
state = torch.randn(1, 4, 5)
suffix = state[:, 2:, :].clone()
result = vocab_project_fn(prefix)(state)
assert torch.equal(result[:, :2, :], torch.nn.functional.one_hot(prefix, 5).to(state))
assert torch.equal(result[:, 2:, :], suffix)
assert torch.equal(state[:, 2:, :], suffix)
def test_rdlm_projector_encodes_prefix_tokens_as_base_n_digits():
from dlb.adapters.conditional_runtime import base_n_digits, rdlm_project_fn
prefix = torch.tensor([[5, 8]]) # base-3 little-endian: 5 -> [2, 1], 8 -> [2, 2]
state = torch.randn(1, 5, 4)
suffix = state[:, 4:, :].clone()
result = rdlm_project_fn(prefix, base=3, digits_per_token=2)(state)
expected_digits = base_n_digits(prefix, base=3, digits_per_token=2)
expected = torch.nn.functional.one_hot(expected_digits, 4).to(state)
assert expected_digits.tolist() == [[2, 1, 2, 2]]
assert torch.equal(result[:, :4, :], expected)
assert torch.equal(result[:, 4:, :], suffix)
assert torch.equal(state[:, 4:, :], suffix)
def test_embedding_projector_writes_clean_prefix_and_preserves_suffix():
from dlb.adapters.conditional_runtime import embedding_project_fn
clean = torch.tensor([[[1.0, 2.0], [3.0, 4.0]]])
state = torch.randn(1, 5, 2)
suffix = state[:, 2:, :].clone()
result = embedding_project_fn(clean)(state)
assert torch.equal(result[:, :2, :], clean)
assert torch.equal(result[:, 2:, :], suffix)
assert torch.equal(state[:, 2:, :], suffix)
def test_candi_prompt_mask_supports_upstream_subtraction():
from dlb.adapters.conditional_runtime import candi_prompt_conditioning
prefix = torch.tensor([[1, 2, 3]])
prompt_tokens, mask = candi_prompt_conditioning(prefix, sequence_length=5)
assert prompt_tokens.tolist() == [[1, 2, 3, 0, 0]]
assert mask.dtype == torch.float32
assert mask.tolist() == [[1.0, 1.0, 1.0, 0.0, 0.0]]
assert torch.equal(1 - mask[:, :3], torch.zeros_like(mask[:, :3]))
with pytest.raises(ValueError, match="shorter than prefix"):
candi_prompt_conditioning(prefix, sequence_length=2)
def test_candi_prompt_adapter_returns_bool_clean_mask_after_float_arithmetic():
from dlb.adapters.conditional_runtime import (
adapt_candi_generate_sample_prompt,
candi_prompt_conditioning,
)
prefix = torch.tensor([[1, 2, 3]])
prompt_tokens, mask = candi_prompt_conditioning(prefix, sequence_length=5)
state = torch.zeros(1, 5, 8)
def original(prompt_tokens, x, prompt_mask):
assert torch.equal(1 - prompt_mask[:, :3], torch.zeros_like(prompt_mask[:, :3]))
return x, prompt_mask
_, clean_mask = adapt_candi_generate_sample_prompt(original)(
prompt_tokens, state, mask
)
assert clean_mask.dtype == torch.bool
assert torch.equal(~clean_mask[:, :3], torch.zeros_like(clean_mask[:, :3]))
assert clean_mask.tolist() == [[True, True, True, False, False]]
def test_projectors_reject_incompatible_shapes():
from dlb.adapters.conditional_runtime import (
embedding_project_fn,
rdlm_project_fn,
token_project_fn,
vocab_project_fn,
)
with pytest.raises(ValueError, match="rank two"):
token_project_fn(torch.tensor([1, 2]))(torch.zeros(1, 3, dtype=torch.long))
with pytest.raises(ValueError, match="vocabulary"):
vocab_project_fn(torch.tensor([[6]]))(torch.zeros(1, 2, 5))
with pytest.raises(ValueError, match="RDLM"):
rdlm_project_fn(torch.tensor([[4]]), base=5, digits_per_token=1)(
torch.zeros(1, 2, 3)
)
with pytest.raises(ValueError, match="embedding"):
embedding_project_fn(torch.zeros(2, 4, 3))(torch.zeros(1, 4, 3))
class FakeOwner:
def update(self, value):
return value
def test_scoped_patch_restores_even_after_sampler_error():
from dlb.adapters.conditional_runtime import patched_attribute
owner = FakeOwner()
original = owner.update
with pytest.raises(RuntimeError):
with patched_attribute(owner, "update", lambda *_: (_ for _ in ()).throw(RuntimeError())):
owner.update(None)
assert owner.update.__func__ is original.__func__
def _write_manifest_tree(root: Path) -> Path:
prompts = root / "data" / "conditional" / "lm1b-c64" / "prompts.jsonl"
prompts.parent.mkdir(parents=True)
rows = []
for prompt_id in range(4):
rows.append(
{
"prompt_id": prompt_id,
"source_index": 100 + prompt_id,
"prefix_token_ids": [prompt_id] * 64,
"reference_token_ids": [prompt_id + 10] * 64,
"source_sequence_sha256": f"{prompt_id:064x}",
}
)
prompts.write_text("\n".join(json.dumps(row, sort_keys=True) for row in rows) + "\n")
manifest = root / "data" / "manifests" / "conditional-lm1b-c64.json"
manifest.parent.mkdir(parents=True)
atomic_json_write(
manifest,
{
"schema_version": 1,
"protocol": "c64_zs_v1",
"dataset": "lm1b",
"source_split": "validation",
"source_processed_path": "data/processed/lm1b-bert-128/validation",
"source_manifest_path": "data/manifests/lm1b.json",
"source_manifest_sha256": "a" * 64,
"tokenizer_id": "bert-base-uncased",
"tokenizer_revision": "b" * 40,
"vocabulary_size": 30522,
"selection_algorithm": "sha256_seed_index_digest_ascending_v1",
"selection_seed": 42,
"source_row_count": 4,
"prompt_count": 4,
"prefix_length": 64,
"evaluation_continuation_length": 64,
"model_length": 128,
"prompt_file": "data/conditional/lm1b-c64/prompts.jsonl",
"prompt_file_sha256": __import__("hashlib").sha256(prompts.read_bytes()).hexdigest(),
},
)
return manifest
def test_load_conditioning_batch_follows_completion_boundaries(tmp_path: Path):
from dlb.adapters.conditional_runtime import load_conditioning_batch
from dlb.io import sha256_file
manifest = _write_manifest_tree(tmp_path)
digest = sha256_file(manifest)
batch = load_conditioning_batch(
manifest,
digest,
completion_id=0,
prompt_start=1,
batch_size=2,
device="cpu",
vocab_size=30522,
)
assert batch.prompt_ids == (1, 2)
assert batch.source_indices == (101, 102)
assert batch.prefix_token_ids.tolist() == [[1] * 64, [2] * 64]
assert batch.reference_token_ids.tolist() == [[11] * 64, [12] * 64]
with pytest.raises(ValueError, match="schedule boundary"):
load_conditioning_batch(
manifest,
digest,
completion_id=1,
prompt_start=0,
batch_size=257,
device="cpu",
vocab_size=30522,
)
def test_load_conditioning_batch_accepts_runtime_vocab_with_extra_mask_dimension(tmp_path: Path):
from dlb.adapters.conditional_runtime import load_conditioning_batch
from dlb.io import sha256_file
manifest = _write_manifest_tree(tmp_path)
digest = sha256_file(manifest)
batch = load_conditioning_batch(
manifest,
digest,
completion_id=0,
prompt_start=0,
batch_size=1,
device="cpu",
vocab_size=30523,
)
assert batch.prefix_token_ids.tolist() == [[0] * 64]
with pytest.raises(ValueError, match="smaller than prompt manifest"):
load_conditioning_batch(
manifest,
digest,
completion_id=0,
prompt_start=0,
batch_size=1,
device="cpu",
vocab_size=30521,
)
|