File size: 9,509 Bytes
2dd5f57 | 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 | """Tests for CLI arguments."""
import argparse
import pytest
from speculators import losses
from speculators.losses import eager
from speculators.models.dflash.core import DFlashDraftModel
from speculators.models.dspark.core import DSparkDraftModel
from speculators.models.eagle3.core import Eagle3DraftModel
from speculators.models.peagle.core import PEagleDraftModel
from speculators.train.config import TrainConfig
def _parse(monkeypatch, extra: list[str]) -> argparse.Namespace:
cfg = TrainConfig.resolve(["--verifier-name-or-path", "dummy", *extra])
return argparse.Namespace(**cfg.flatten())
# ---------------------------------------------------------------------------
# Ensure CLI args flow correctly through vars(args) into get_trainer_kwargs
# ---------------------------------------------------------------------------
def test_dflash_default_uses_kl(monkeypatch):
args = _parse(monkeypatch, [])
train_kw, val_kw = DFlashDraftModel.get_trainer_kwargs(**vars(args))
assert "kl_div" in train_kw["loss_config"]
assert train_kw["loss_config"]["kl_div"][0] is losses.kl_div_loss
assert "kl_div" in val_kw["loss_config"]
assert train_kw["gamma"] == 4.0
assert val_kw["gamma"] == 4.0
def test_dflash_explicit_ce(monkeypatch):
args = _parse(monkeypatch, ["--loss-fn", "ce"])
train_kw, val_kw = DFlashDraftModel.get_trainer_kwargs(**vars(args))
assert "ce" in train_kw["loss_config"]
assert train_kw["loss_config"]["ce"][0] is losses.ce_loss
assert "ce" in val_kw["loss_config"]
assert train_kw["gamma"] == 4.0
assert val_kw["gamma"] == 4.0
def test_dflash_explicit_decay_gamma(monkeypatch):
args = _parse(monkeypatch, ["--dflash-decay-gamma", "7.0"])
train_kw, val_kw = DFlashDraftModel.get_trainer_kwargs(**vars(args))
assert train_kw["gamma"] == 7.0
assert val_kw["gamma"] == 7.0
def test_dflash_decay_gamma_falls_back_when_omitted():
train_kw, val_kw = DFlashDraftModel.get_trainer_kwargs(loss_fn="kl_div")
assert train_kw["gamma"] == 4.0
assert val_kw["gamma"] == 4.0
def test_dflash_compound_loss(monkeypatch):
args = _parse(monkeypatch, ["--loss-fn", '{"ce": 0.1, "tv": 0.9}'])
train_kw, val_kw = DFlashDraftModel.get_trainer_kwargs(**vars(args))
assert "ce" in train_kw["loss_config"]
assert "tv" in train_kw["loss_config"]
assert train_kw["loss_config"]["ce"][1] == 0.1
assert train_kw["loss_config"]["tv"][1] == 0.9
assert "ce" in val_kw["loss_config"]
assert "tv" in val_kw["loss_config"]
def test_eagle3_default_uses_kl(monkeypatch):
args = _parse(monkeypatch, [])
train_kw, val_kw = Eagle3DraftModel.get_trainer_kwargs(**vars(args))
assert "kl_div" in train_kw["loss_config"]
assert train_kw["loss_config"]["kl_div"][0] is losses.kl_div_loss
assert "kl_div" in val_kw["loss_config"]
def test_eagle3_explicit_ce(monkeypatch):
args = _parse(monkeypatch, ["--loss-fn", "ce"])
train_kw, val_kw = Eagle3DraftModel.get_trainer_kwargs(**vars(args))
assert "ce" in train_kw["loss_config"]
assert train_kw["loss_config"]["ce"][0] is losses.ce_loss
assert "ce" in val_kw["loss_config"]
def test_peagle_default_uses_kl(monkeypatch):
args = _parse(monkeypatch, [])
train_kw, val_kw = PEagleDraftModel.get_trainer_kwargs(**vars(args))
assert "kl_div" in train_kw["loss_config"]
assert train_kw["loss_config"]["kl_div"][0] is losses.kl_div_loss
assert "kl_div" in val_kw["loss_config"]
def test_peagle_explicit_ce(monkeypatch):
args = _parse(monkeypatch, ["--loss-fn", "ce"])
train_kw, val_kw = PEagleDraftModel.get_trainer_kwargs(**vars(args))
assert "ce" in train_kw["loss_config"]
assert train_kw["loss_config"]["ce"][0] is losses.ce_loss
assert "ce" in val_kw["loss_config"]
def test_dspark_default_uses_kl(monkeypatch):
args = _parse(monkeypatch, [])
train_kw, val_kw = DSparkDraftModel.get_trainer_kwargs(**vars(args))
assert "kl_div" in train_kw["loss_config"]
assert train_kw["loss_config"]["kl_div"][0] is losses.kl_div_loss
assert train_kw["tv_loss_fn"] is losses.tv_loss
assert "kl_div" in val_kw["loss_config"]
assert train_kw["confidence_head_alpha"] == 1.0
assert val_kw["confidence_head_alpha"] == 1.0
def test_dspark_explicit_eager(monkeypatch):
args = _parse(monkeypatch, ["--loss-implementation", "eager"])
train_kw, val_kw = DSparkDraftModel.get_trainer_kwargs(**vars(args))
assert train_kw["loss_config"]["kl_div"][0] is eager.kl_div_loss
assert train_kw["tv_loss_fn"] is eager.tv_loss
assert val_kw["tv_loss_fn"] is eager.tv_loss
def test_dspark_compound_loss(monkeypatch):
args = _parse(monkeypatch, ["--loss-fn", '{"ce": 0.1, "tv": 0.9}'])
train_kw, val_kw = DSparkDraftModel.get_trainer_kwargs(**vars(args))
assert "ce" in train_kw["loss_config"]
assert train_kw["loss_config"]["ce"][0] is losses.ce_loss
assert train_kw["loss_config"]["ce"][1] == 0.1
assert "tv" in train_kw["loss_config"]
assert train_kw["loss_config"]["tv"][0] is losses.tv_loss
assert train_kw["loss_config"]["tv"][1] == 0.9
assert "ce" in val_kw["loss_config"]
assert "tv" in val_kw["loss_config"]
def test_dspark_confidence_head_alpha(monkeypatch):
args = _parse(monkeypatch, ["--confidence-head-alpha", "0.5"])
train_kw, val_kw = DSparkDraftModel.get_trainer_kwargs(**vars(args))
assert train_kw["confidence_head_alpha"] == 0.5
assert val_kw["confidence_head_alpha"] == 0.5
# ---------------------------------------------------------------------------
# Per-speculator-type defaults for draft_arch, norm_before_fc, norm_output
# ---------------------------------------------------------------------------
def test_eagle3_defaults_to_llama_arch(monkeypatch):
args = _parse(monkeypatch, [])
assert args.draft_arch == "llama"
def test_eagle3_defaults_norm_before_fc_true(monkeypatch):
args = _parse(monkeypatch, [])
assert args.norm_before_fc is True
def test_eagle3_defaults_norm_output_true(monkeypatch):
args = _parse(monkeypatch, [])
assert args.norm_output is True
def test_dflash_defaults_to_qwen3_arch(monkeypatch):
args = _parse(monkeypatch, ["--speculator-type", "dflash"])
assert args.draft_arch == "qwen3"
def test_dflash_defaults_norm_before_fc_false(monkeypatch):
args = _parse(monkeypatch, ["--speculator-type", "dflash"])
assert args.norm_before_fc is False
def test_dflash_defaults_norm_output_false(monkeypatch):
args = _parse(monkeypatch, ["--speculator-type", "dflash"])
assert args.norm_output is False
# ---------------------------------------------------------------------------
# Per-speculator-type defaults for num_layers, per_position_loss_weight, loss_fn
# (best-practices recipe from https://github.com/vllm-project/speculators/issues/979)
# ---------------------------------------------------------------------------
def test_dflash_defaults_num_layers_to_5(monkeypatch):
args = _parse(monkeypatch, ["--speculator-type", "dflash"])
assert args.num_layers == 5
def test_dflash_defaults_per_position_loss_weight_to_dpace(monkeypatch):
args = _parse(monkeypatch, ["--speculator-type", "dflash"])
assert args.per_position_loss_weight == "dpace"
def test_dflash_defaults_loss_fn_to_ce(monkeypatch):
args = _parse(monkeypatch, ["--speculator-type", "dflash"])
assert args.loss_fn == "ce"
def test_dflash_defaults_block_size_to_16(monkeypatch):
args = _parse(monkeypatch, ["--speculator-type", "dflash"])
assert args.block_size == 16
def test_dspark_defaults_block_size_to_8(monkeypatch):
# block_size is shared with dspark, which never had block_size=16 validated.
args = _parse(monkeypatch, ["--speculator-type", "dspark"])
assert args.block_size == 8
def test_dflash_explicit_flags_override_new_defaults(monkeypatch):
args = _parse(
monkeypatch,
[
"--speculator-type",
"dflash",
"--num-layers",
"3",
"--per-position-loss-weight",
"fixed-exp-decay",
"--loss-fn",
"kl_div",
"--block-size",
"8",
],
)
assert args.num_layers == 3
assert args.per_position_loss_weight == "fixed-exp-decay"
assert args.loss_fn == "kl_div"
assert args.block_size == 8
def test_eagle3_num_layers_and_loss_defaults_unchanged(monkeypatch):
args = _parse(monkeypatch, [])
assert args.num_layers == 1
assert args.per_position_loss_weight == "fixed-exp-decay"
assert args.loss_fn == "kl_div"
assert args.block_size == 8
def test_no_norm_before_fc_flag(monkeypatch):
args = _parse(monkeypatch, ["--no-norm-before-fc"])
assert args.norm_before_fc is False
def test_no_norm_output_flag(monkeypatch):
args = _parse(monkeypatch, ["--no-norm-output"])
assert args.norm_output is False
# ---------------------------------------------------------------------------
# --max-steps
# ---------------------------------------------------------------------------
def test_max_steps_default_is_none(monkeypatch):
args = _parse(monkeypatch, [])
assert args.max_steps is None
def test_max_steps_explicit(monkeypatch):
args = _parse(monkeypatch, ["--max-steps", "15"])
assert args.max_steps == 15
def test_max_steps_rejects_non_positive(monkeypatch):
with pytest.raises(SystemExit):
_parse(monkeypatch, ["--max-steps", "0"])
|