File size: 10,348 Bytes
6e802a4 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 | # test_loader_guard.py: The checkpoint loader refuses silent drops (LoRA, plan tensors, speaker table).
import sys
from pathlib import Path
import pytest
import torch
import torch.nn as nn
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
from train import (LoRALinear, _remap_plain_keys_for_lora, checkpoint_speaker_rows,
load_checkpoint)
class TinyPlain(nn.Module):
def __init__(self):
super().__init__()
self.in_proj = nn.Linear(4, 8, bias=False)
self.out_proj = nn.Linear(8, 4, bias=False)
self.embedding = nn.Embedding(10, 4)
class TinyLora(nn.Module):
def __init__(self, rank=2):
super().__init__()
self.in_proj = LoRALinear(nn.Linear(4, 8, bias=False), rank=rank,
alpha=rank, dropout=0.0)
self.out_proj = LoRALinear(nn.Linear(8, 4, bias=False), rank=rank,
alpha=rank, dropout=0.0)
self.embedding = nn.Embedding(10, 4)
def _save_ckpt(tmp_path, sd, name="ckpt.pt", step=123):
path = tmp_path / name
torch.save({"step": step, "model_state_dict": sd}, path)
return str(path)
def test_remap_plain_projection_keys():
plain_sd = TinyPlain().state_dict()
lora_sd = TinyLora().state_dict()
remapped, n = _remap_plain_keys_for_lora(plain_sd, lora_sd)
assert n == 2, f"expected both projections remapped, got {n}"
assert "in_proj.original.weight" in remapped
assert "out_proj.original.weight" in remapped
assert "in_proj.weight" not in remapped
assert "embedding.weight" in remapped
torch.testing.assert_close(remapped["in_proj.original.weight"],
plain_sd["in_proj.weight"])
def test_remap_leaves_matching_and_unknown_keys_alone():
lora_model_sd = TinyLora().state_dict()
sd = {
"embedding.weight": torch.zeros(10, 4),
"totally.unknown.weight": torch.zeros(2, 2),
}
remapped, n = _remap_plain_keys_for_lora(sd, lora_model_sd)
assert n == 0
assert set(remapped) == set(sd)
def test_fullft_ckpt_loads_exactly_into_lora_model(tmp_path):
src = TinyPlain()
with torch.no_grad():
src.in_proj.weight += 1.0
path = _save_ckpt(tmp_path, src.state_dict())
dst = TinyLora()
step = load_checkpoint(path, dst, device="cpu")
assert step == 123
torch.testing.assert_close(dst.in_proj.original.weight, src.in_proj.weight)
x = torch.randn(3, 4)
torch.testing.assert_close(dst.in_proj(x), src.in_proj(x))
def test_lora_ckpt_into_plain_model_still_errors(tmp_path):
path = _save_ckpt(tmp_path, TinyLora().state_dict())
with pytest.raises(RuntimeError, match="LoRA adapters"):
load_checkpoint(path, TinyPlain(), device="cpu")
def test_dropped_projection_weight_is_a_hard_error(tmp_path):
sd = TinyPlain().state_dict()
sd["mixer.in_proj.weight"] = torch.zeros(8, 4)
path = _save_ckpt(tmp_path, sd)
with pytest.raises(RuntimeError, match="silently dropped"):
load_checkpoint(path, TinyPlain(), device="cpu")
class TinyNeoX(nn.Module):
def __init__(self):
super().__init__()
self.attention = nn.Module()
self.attention.query_key_value = nn.Linear(4, 12, bias=False)
self.attention.dense = nn.Linear(4, 4, bias=False)
self.embedding = nn.Embedding(10, 4)
def test_neox_projection_unexpected_is_a_hard_error(tmp_path):
sd = TinyNeoX().state_dict()
sd["layers.0.attention.query_key_value.weight"] = torch.zeros(12, 4)
path = _save_ckpt(tmp_path, sd)
with pytest.raises(RuntimeError, match="silently dropped"):
load_checkpoint(path, TinyNeoX(), device="cpu")
def test_missing_projection_is_a_hard_error(tmp_path):
sd = TinyNeoX().state_dict()
del sd["attention.query_key_value.weight"]
path = _save_ckpt(tmp_path, sd)
with pytest.raises(RuntimeError, match="received NO value"):
load_checkpoint(path, TinyNeoX(), device="cpu")
def test_hybrid_graft_missing_attention_is_allowed(tmp_path):
class TinyHybrid(nn.Module):
def __init__(self):
super().__init__()
self.backbone = nn.Module()
self.backbone.backbone = nn.Module()
self.backbone.backbone.layers = nn.ModuleDict({
"0": nn.ModuleDict({"mixer": nn.ModuleDict(
{"in_proj": nn.Linear(4, 12, bias=False)})}),
})
self.hybrid_replaced_layers = [0]
model = TinyHybrid()
sd = {k: v for k, v in model.state_dict().items()
if not k.endswith("in_proj.weight")}
path = _save_ckpt(tmp_path, sd)
load_checkpoint(path, model, device="cpu")
def test_architecture_mismatched_projection_shape_is_a_hard_error(tmp_path):
sd = TinyPlain().state_dict()
sd["in_proj.weight"] = torch.zeros(24, 4)
path = _save_ckpt(tmp_path, sd)
with pytest.raises(RuntimeError, match="different SHAPE"):
load_checkpoint(path, TinyPlain(), device="cpu")
def test_shape_mismatched_speaker_table_still_only_warns(tmp_path):
sd = TinyPlain().state_dict()
sd["embedding.weight"] = torch.zeros(20, 4)
path = _save_ckpt(tmp_path, sd, step=5)
assert load_checkpoint(path, TinyPlain(), device="cpu") == 5
def test_depth_arch_swap_is_allowed(tmp_path):
class GruDepth(nn.Module):
def __init__(self):
super().__init__()
self.depth_module = nn.Module()
self.depth_module.gru = nn.Linear(4, 4, bias=False)
self.depth_module.h_proj = nn.Linear(4, 4, bias=False)
class TfDepth(nn.Module):
def __init__(self):
super().__init__()
self.depth_module = nn.Module()
self.depth_module.depth_tf = nn.Module()
self.depth_module.depth_tf.self_attn = nn.Module()
self.depth_module.depth_tf.self_attn.out_proj = nn.Linear(4, 4, bias=False)
self.depth_module.h_proj = nn.Linear(4, 4, bias=False)
path = _save_ckpt(tmp_path, GruDepth().state_dict(), step=9)
assert load_checkpoint(path, TfDepth(), device="cpu") == 9
def test_gru_to_mamba2_depth_swap_is_allowed(tmp_path):
class GruDepth(nn.Module):
def __init__(self):
super().__init__()
self.depth_module = nn.Module()
self.depth_module.gru = nn.Linear(4, 4, bias=False)
class SsmDepth(nn.Module):
def __init__(self):
super().__init__()
self.depth_module = nn.Module()
self.depth_module.depth_ssm = nn.Module()
self.depth_module.depth_ssm.in_proj = nn.Linear(4, 4, bias=False)
self.depth_module.depth_ssm.out_proj = nn.Linear(4, 4, bias=False)
path = _save_ckpt(tmp_path, GruDepth().state_dict(), step=11)
assert load_checkpoint(path, SsmDepth(), device="cpu") == 11
def test_same_arch_missing_depth_projection_still_raises(tmp_path):
class TfDepth(nn.Module):
def __init__(self):
super().__init__()
self.depth_module = nn.Module()
self.depth_module.depth_tf = nn.Module()
self.depth_module.depth_tf.self_attn = nn.Module()
self.depth_module.depth_tf.self_attn.out_proj = nn.Linear(4, 4, bias=False)
sd = TfDepth().state_dict()
del sd["depth_module.depth_tf.self_attn.out_proj.weight"]
path = _save_ckpt(tmp_path, sd)
with pytest.raises(RuntimeError, match="received NO value"):
load_checkpoint(path, TfDepth(), device="cpu")
def test_matched_load_roundtrip_unaffected(tmp_path):
src = TinyLora()
path = _save_ckpt(tmp_path, src.state_dict(), step=7)
dst = TinyLora()
step = load_checkpoint(path, dst, device="cpu")
assert step == 7
torch.testing.assert_close(dst.in_proj.lora_A, src.in_proj.lora_A)
class TinyDelay(nn.Module):
def __init__(self):
super().__init__()
self.in_proj = nn.Linear(4, 8, bias=False)
self.delay_heads = nn.ModuleList([nn.Linear(8, 16, bias=False)])
class TinyDepth(nn.Module):
def __init__(self):
super().__init__()
self.in_proj = nn.Linear(4, 8, bias=False)
self.depth_gru = nn.GRU(8, 8, batch_first=True)
def test_delay_ckpt_into_depth_model_warm_start(tmp_path):
src = TinyDelay()
path = _save_ckpt(tmp_path, src.state_dict(), step=100000)
dst = TinyDepth()
before = {k: v.clone() for k, v in dst.state_dict().items()
if k.startswith("depth_gru")}
load_checkpoint(path, dst, device="cpu")
torch.testing.assert_close(dst.in_proj.weight, src.in_proj.weight)
for k, v in before.items():
torch.testing.assert_close(dst.state_dict()[k], v)
class TinySpeaker(nn.Module):
def __init__(self, num_speakers):
super().__init__()
self.speaker_encoder = nn.Module()
self.speaker_encoder.embedding = nn.Embedding(num_speakers, 4)
self.in_proj = nn.Linear(4, 8, bias=False)
def test_speaker_table_mismatch_is_a_hard_error(tmp_path):
path = _save_ckpt(tmp_path, TinySpeaker(16).state_dict())
with pytest.raises(RuntimeError, match=r"Speaker table.*--num_speakers 16"):
load_checkpoint(path, TinySpeaker(8), device="cpu")
def test_speaker_table_reinit_only_when_asked(tmp_path):
src = TinySpeaker(16)
with torch.no_grad():
src.in_proj.weight += 1.0
path = _save_ckpt(tmp_path, src.state_dict())
dst = TinySpeaker(8)
load_checkpoint(path, dst, device="cpu", allow_speaker_table_reinit=True)
torch.testing.assert_close(dst.in_proj.weight, src.in_proj.weight)
assert dst.speaker_encoder.embedding.weight.shape[0] == 8
def test_matching_speaker_table_loads_exactly(tmp_path):
src = TinySpeaker(16)
path = _save_ckpt(tmp_path, src.state_dict())
dst = TinySpeaker(16)
load_checkpoint(path, dst, device="cpu")
torch.testing.assert_close(dst.speaker_encoder.embedding.weight,
src.speaker_encoder.embedding.weight)
def test_checkpoint_speaker_rows_reads_the_table_size(tmp_path):
assert checkpoint_speaker_rows(
_save_ckpt(tmp_path, TinySpeaker(16).state_dict())) == 16
assert checkpoint_speaker_rows(
_save_ckpt(tmp_path, TinyPlain().state_dict(), name="plain.pt")) is None
assert checkpoint_speaker_rows(str(tmp_path / "absent.pt")) is None
|