File size: 18,378 Bytes
c34ff1f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 | """Integration tests for draft model forward passes with real weights.
Covers DFlash, Eagle3, PEagle, and MTP models with shared parametrized tests
for training, multi-batch, and vocab boundary scenarios, plus model-specific
parameter variation tests.
"""
from collections.abc import Callable
from dataclasses import dataclass, field
from functools import partial
from typing import Any
import pytest
import torch
from speculators.models.mtp import shift_batch_mtp
from speculators.models.mtp.core import compute_step_weights
from tests.conftest import requires_cuda, requires_transformers_version
from tests.integration.conftest import (
HIDDEN_SIZE,
TINY_QWEN3_5_KWARGS,
VOCAB_SIZE,
make_batch,
make_dflash_model,
make_eagle3_model,
make_mtp_model,
make_peagle_model,
make_sample,
)
MAX_LEN = 128
HIDDEN_MULTIPLIER = 3
LOSS_MASK_CASES = ["all", "none", "random", "alternating"]
SAMPLE_CONFIGS = [
pytest.param([128], id="single_sample"),
pytest.param([64, 64], id="two_equal"),
pytest.param([32, 96], id="two_unequal"),
pytest.param([8] * 20, id="twenty_tiny"),
]
MULTI_BATCH_CONFIGS: list[list[int]] = [
[16, 16, 8, 10, 15, 12],
[32, 32, 32],
[64, 64],
[32, 3, 17],
[128],
[16],
[],
]
# ---------------------------------------------------------------------------
# Model specs
# ---------------------------------------------------------------------------
_requires_qwen3_5 = requires_transformers_version("5.2.0")
@dataclass(frozen=True)
class ModelSpec:
name: str
factory: Callable[..., Any]
forward_kwargs: dict[str, Any] = field(default_factory=dict)
hidden_size: int = HIDDEN_SIZE
hidden_multiplier: int = HIDDEN_MULTIPLIER
batch_factory: Callable[..., Any] = make_batch
DFLASH_SPEC = ModelSpec(
name="dflash", factory=make_dflash_model, forward_kwargs={"max_anchors": 8}
)
EAGLE3_SPEC = ModelSpec(
name="eagle3", factory=make_eagle3_model, forward_kwargs={"ttt_steps": 2}
)
PEAGLE_SPEC = ModelSpec(
name="peagle",
factory=make_peagle_model,
forward_kwargs={
"num_depths": 4,
"down_sample_ratio": 0.7,
"down_sample_ratio_min": 0.2,
},
)
MTP_SPEC = ModelSpec(
name="mtp",
factory=make_mtp_model,
forward_kwargs={"step_weights": [0.51, 0.31, 0.18]},
hidden_size=TINY_QWEN3_5_KWARGS["hidden_size"],
hidden_multiplier=1,
batch_factory=partial(make_batch, num_target_layers=1, preprocess=shift_batch_mtp),
)
ALL_SPECS = [
pytest.param(DFLASH_SPEC, id="dflash"),
pytest.param(EAGLE3_SPEC, id="eagle3"),
pytest.param(PEAGLE_SPEC, id="peagle"),
pytest.param(MTP_SPEC, id="mtp", marks=_requires_qwen3_5),
]
VOCAB_SPECS = [
pytest.param(DFLASH_SPEC, id="dflash"),
pytest.param(EAGLE3_SPEC, id="eagle3"),
pytest.param(PEAGLE_SPEC, id="peagle"),
]
# ---------------------------------------------------------------------------
# Helpers
# ---------------------------------------------------------------------------
def _make_samples(
seq_lengths: list[int],
loss_mask_pattern: str = "all",
vocab_size: int = VOCAB_SIZE,
hidden_size: int = HIDDEN_SIZE,
hidden_multiplier: int = HIDDEN_MULTIPLIER,
boundary_token_ids: list[int] | None = None,
) -> list[dict[str, torch.Tensor]]:
return [
make_sample(
seq_len=sl,
hidden_size=hidden_size,
hidden_multiplier=hidden_multiplier,
vocab_size=vocab_size,
loss_mask_pattern=loss_mask_pattern,
include_verifier_states=True,
boundary_token_ids=boundary_token_ids,
)
for sl in seq_lengths
]
# ---------------------------------------------------------------------------
# Fixtures
# ---------------------------------------------------------------------------
@pytest.fixture(params=ALL_SPECS)
def model_and_spec(request):
spec: ModelSpec = request.param
model = spec.factory()
yield model, spec
del model
torch.cuda.empty_cache()
@pytest.fixture
def draft_vocab_model(request):
spec: ModelSpec = request.param
model = spec.factory(draft_vocab_size=32)
t2d = torch.zeros(VOCAB_SIZE, dtype=torch.bool)
t2d[:32] = True
d2t = torch.arange(32, dtype=torch.long)
model.load_vocab_mappings(t2d.to("cuda"), d2t.to("cuda"))
yield model, spec
del model
torch.cuda.empty_cache()
# ---------------------------------------------------------------------------
# Shared tests
# ---------------------------------------------------------------------------
@requires_cuda
class TestTraining:
"""Forward + backward pass across all models."""
@pytest.mark.parametrize("seq_lengths", SAMPLE_CONFIGS)
@pytest.mark.parametrize("loss_mask_pattern", LOSS_MASK_CASES)
def test_forward_backward(self, model_and_spec, seq_lengths, loss_mask_pattern):
model, spec = model_and_spec
samples = _make_samples(
seq_lengths,
loss_mask_pattern=loss_mask_pattern,
hidden_size=spec.hidden_size,
hidden_multiplier=spec.hidden_multiplier,
)
batch = spec.batch_factory(
max_len=MAX_LEN, samples=samples, hidden_size=spec.hidden_size
)
draft_tokens, loss, metrics = model(**batch, **spec.forward_kwargs)
assert loss.isfinite(), f"Loss is not finite: {loss.item()}"
assert "loss_sum" in metrics
assert "loss_total" in metrics
loss.backward()
@requires_cuda
class TestMultiBatch:
"""Run multiple batches back-to-back to test statefulness and cache clearing."""
def test_varying_batches(self, model_and_spec):
model, spec = model_and_spec
torch.compiler.reset()
for seq_lengths in MULTI_BATCH_CONFIGS:
samples = _make_samples(
seq_lengths,
hidden_size=spec.hidden_size,
hidden_multiplier=spec.hidden_multiplier,
)
batch = spec.batch_factory(
max_len=MAX_LEN, samples=samples, hidden_size=spec.hidden_size
)
draft_tokens, loss, metrics = model(**batch, **spec.forward_kwargs)
assert loss.isfinite(), f"Loss not finite for seq_lengths={seq_lengths}"
loss.backward()
def test_varying_loss_masks_across_batches(self, model_and_spec):
model, spec = model_and_spec
torch.compiler.reset()
for pattern in LOSS_MASK_CASES:
samples = _make_samples(
[64, 64],
loss_mask_pattern=pattern,
hidden_size=spec.hidden_size,
hidden_multiplier=spec.hidden_multiplier,
)
batch = spec.batch_factory(
max_len=MAX_LEN, samples=samples, hidden_size=spec.hidden_size
)
draft_tokens, loss, metrics = model(**batch, **spec.forward_kwargs)
assert loss.isfinite(), f"Loss not finite for loss_mask_pattern={pattern}"
loss.backward()
@requires_cuda
class TestVocabBoundary:
"""Tests with draft vocab mapping."""
@pytest.mark.parametrize("draft_vocab_model", VOCAB_SPECS, indirect=True)
def test_boundary_tokens(self, draft_vocab_model):
model, spec = draft_vocab_model
samples = _make_samples([128], vocab_size=32, boundary_token_ids=[0, 31])
batch = make_batch(max_len=MAX_LEN, samples=samples, hidden_size=HIDDEN_SIZE)
draft_tokens, loss, metrics = model(**batch, **spec.forward_kwargs)
assert loss.isfinite()
loss.backward()
# ---------------------------------------------------------------------------
# Model-specific parameter tests
# ---------------------------------------------------------------------------
@requires_cuda
class TestDFlashParams:
@pytest.mark.parametrize("block_size", [2, 4, 8])
def test_varying_block_size(self, block_size):
model = make_dflash_model(block_size=block_size)
samples = _make_samples([128])
batch = make_batch(max_len=MAX_LEN, samples=samples, hidden_size=HIDDEN_SIZE)
draft_tokens, loss, metrics = model(**batch, max_anchors=4)
assert loss.isfinite()
loss.backward()
@pytest.mark.parametrize("max_anchors", [2, 8, 16])
def test_varying_max_anchors(self, max_anchors):
model = make_dflash_model()
samples = _make_samples([128])
batch = make_batch(max_len=MAX_LEN, samples=samples, hidden_size=HIDDEN_SIZE)
draft_tokens, loss, metrics = model(**batch, max_anchors=max_anchors)
assert loss.isfinite()
loss.backward()
@pytest.mark.parametrize("draft_attn_impl", ["sdpa", "eager"])
@pytest.mark.parametrize("seq_lengths", SAMPLE_CONFIGS)
def test_attention_backend(self, draft_attn_impl, seq_lengths):
model = make_dflash_model(draft_attn_impl=draft_attn_impl)
samples = _make_samples(seq_lengths)
batch = make_batch(max_len=MAX_LEN, samples=samples, hidden_size=HIDDEN_SIZE)
draft_tokens, loss, metrics = model(**batch, max_anchors=8)
assert loss.isfinite()
loss.backward()
ATTN_BACKENDS = ["simple_flex_attention", "sdpa", "eager"]
@pytest.mark.parametrize("seq_lengths", SAMPLE_CONFIGS)
def test_attention_backends_match(self, seq_lengths):
"""All attention backends produce equivalent outputs for the same input."""
samples = _make_samples(seq_lengths)
results = {}
for backend in self.ATTN_BACKENDS:
torch.manual_seed(0)
model = make_dflash_model(draft_attn_impl=backend)
batch = make_batch(
max_len=MAX_LEN, samples=samples, hidden_size=HIDDEN_SIZE
)
_, loss, _ = model(**batch, max_anchors=8)
results[backend] = loss.detach().cpu()
del model
torch.cuda.empty_cache()
ref_backend = self.ATTN_BACKENDS[0]
for backend in self.ATTN_BACKENDS[1:]:
torch.testing.assert_close(
results[backend],
results[ref_backend],
atol=1e-3,
rtol=1e-3,
msg=f"{backend} loss diverges from {ref_backend}",
)
@requires_cuda
class TestEagle3Params:
@pytest.mark.parametrize("ttt_steps", [1, 3, 5])
def test_varying_ttt_steps(self, ttt_steps):
model = make_eagle3_model()
samples = _make_samples([128])
batch = make_batch(max_len=MAX_LEN, samples=samples, hidden_size=HIDDEN_SIZE)
draft_tokens, loss, metrics = model(**batch, ttt_steps=ttt_steps)
assert len(draft_tokens) == ttt_steps
for dt in draft_tokens:
assert dt.shape == (1, MAX_LEN)
assert dt.dtype == torch.long
assert loss.isfinite()
loss.backward()
@pytest.mark.parametrize("draft_attn_impl", ["sdpa", "eager"])
@pytest.mark.parametrize("seq_lengths", SAMPLE_CONFIGS)
def test_attention_backend(self, draft_attn_impl, seq_lengths):
model = make_eagle3_model(draft_attn_impl=draft_attn_impl)
samples = _make_samples(seq_lengths)
batch = make_batch(max_len=MAX_LEN, samples=samples, hidden_size=HIDDEN_SIZE)
draft_tokens, loss, metrics = model(**batch, ttt_steps=2)
assert loss.isfinite()
loss.backward()
ATTN_BACKENDS = ["simple_flex_attention", "sdpa", "eager"]
@pytest.mark.parametrize("seq_lengths", SAMPLE_CONFIGS)
def test_attention_backends_match(self, seq_lengths):
"""All attention backends produce equivalent outputs for the same input."""
samples = _make_samples(seq_lengths)
results = {}
for backend in self.ATTN_BACKENDS:
torch.manual_seed(0)
model = make_eagle3_model(draft_attn_impl=backend)
batch = make_batch(
max_len=MAX_LEN, samples=samples, hidden_size=HIDDEN_SIZE
)
_, loss, _ = model(**batch, ttt_steps=2)
results[backend] = loss.detach().cpu()
del model
torch.cuda.empty_cache()
ref_backend = self.ATTN_BACKENDS[0]
for backend in self.ATTN_BACKENDS[1:]:
torch.testing.assert_close(
results[backend],
results[ref_backend],
atol=1e-3,
rtol=1e-3,
msg=f"{backend} loss diverges from {ref_backend}",
)
@requires_cuda
class TestNormOutputParams:
"""Tests for Eagle 3.1: norm_before_fc + norm_output."""
def test_norm_output(self):
model = make_eagle3_model(norm_before_fc=True, norm_output=True)
assert model.input_norm is not None
samples = _make_samples([128])
batch = make_batch(max_len=MAX_LEN, samples=samples, hidden_size=HIDDEN_SIZE)
draft_tokens, loss, _metrics = model(**batch, ttt_steps=3)
assert len(draft_tokens) == 3
assert loss.isfinite()
loss.backward()
def test_norm_output_without_norm_before_fc(self):
model = make_eagle3_model(norm_output=True)
assert model.input_norm is None
samples = _make_samples([128])
batch = make_batch(max_len=MAX_LEN, samples=samples, hidden_size=HIDDEN_SIZE)
draft_tokens, loss, _metrics = model(**batch, ttt_steps=3)
assert len(draft_tokens) == 3
assert loss.isfinite()
loss.backward()
def test_fc_norm(self):
model = make_eagle3_model(fc_norm=True, norm_output=True)
assert model.fc_norm is not None
assert len(model.fc_norm) == 3
assert model.input_norm is None
samples = _make_samples([128])
batch = make_batch(max_len=MAX_LEN, samples=samples, hidden_size=HIDDEN_SIZE)
draft_tokens, loss, _metrics = model(**batch, ttt_steps=3)
assert len(draft_tokens) == 3
assert loss.isfinite()
loss.backward()
def test_peagle_fc_norm(self):
model = make_peagle_model(fc_norm=True)
assert model.fc_norm is not None
assert len(model.fc_norm) == 3
samples = _make_samples([128])
batch = make_batch(max_len=MAX_LEN, samples=samples, hidden_size=HIDDEN_SIZE)
_draft_tokens, loss, _metrics = model(**batch, num_depths=4)
assert loss.isfinite()
loss.backward()
def test_peagle_norm_before_fc(self):
model = make_peagle_model()
assert model.input_norm is None
model = make_peagle_model(norm_before_fc=True)
assert model.input_norm is not None
samples = _make_samples([128])
batch = make_batch(max_len=MAX_LEN, samples=samples, hidden_size=HIDDEN_SIZE)
_draft_tokens, loss, _metrics = model(**batch, num_depths=4)
assert loss.isfinite()
loss.backward()
@requires_cuda
class TestPEagleParams:
@pytest.mark.parametrize("num_depths", [2, 4, 8])
def test_varying_num_depths(self, num_depths):
model = make_peagle_model(num_depths=num_depths)
samples = _make_samples([128])
batch = make_batch(max_len=MAX_LEN, samples=samples, hidden_size=HIDDEN_SIZE)
draft_tokens, loss, metrics = model(**batch, num_depths=num_depths)
assert loss.isfinite()
loss.backward()
@pytest.mark.parametrize("down_sample_ratio", [0.3, 0.7, 1.0])
def test_varying_down_sample_ratio(self, down_sample_ratio):
model = make_peagle_model()
samples = _make_samples([128])
batch = make_batch(max_len=MAX_LEN, samples=samples, hidden_size=HIDDEN_SIZE)
draft_tokens, loss, metrics = model(
**batch, num_depths=4, down_sample_ratio=down_sample_ratio
)
assert loss.isfinite()
loss.backward()
@pytest.mark.parametrize("draft_attn_impl", ["sdpa", "eager"])
@pytest.mark.parametrize("seq_lengths", SAMPLE_CONFIGS)
def test_attention_backend(self, draft_attn_impl, seq_lengths):
model = make_peagle_model(draft_attn_impl=draft_attn_impl)
samples = _make_samples(seq_lengths)
batch = make_batch(max_len=MAX_LEN, samples=samples, hidden_size=HIDDEN_SIZE)
draft_tokens, loss, metrics = model(**batch, num_depths=4)
assert loss.isfinite()
loss.backward()
ATTN_BACKENDS = ["simple_flex_attention", "sdpa", "eager"]
@pytest.mark.parametrize("seq_lengths", SAMPLE_CONFIGS)
def test_attention_backends_match(self, seq_lengths):
"""All attention backends produce equivalent outputs for the same input."""
samples = _make_samples(seq_lengths)
results = {}
for backend in self.ATTN_BACKENDS:
torch.manual_seed(0)
model = make_peagle_model(draft_attn_impl=backend)
batch = make_batch(
max_len=MAX_LEN, samples=samples, hidden_size=HIDDEN_SIZE
)
_, loss, _ = model(**batch, num_depths=4)
results[backend] = loss.detach().cpu()
del model
torch.cuda.empty_cache()
ref_backend = self.ATTN_BACKENDS[0]
for backend in self.ATTN_BACKENDS[1:]:
torch.testing.assert_close(
results[backend],
results[ref_backend],
atol=1e-3,
rtol=1e-3,
msg=f"{backend} loss diverges from {ref_backend}",
)
@requires_cuda
@requires_transformers_version("5.2.0")
class TestMTPParams:
@pytest.mark.parametrize("num_speculative_steps", [1, 2, 5])
def test_varying_num_speculative_steps(self, num_speculative_steps):
model = make_mtp_model(
num_speculative_steps=num_speculative_steps, torch_compile=False
)
step_weights = compute_step_weights(num_steps=num_speculative_steps)
samples = _make_samples(
[128],
hidden_size=TINY_QWEN3_5_KWARGS["hidden_size"],
hidden_multiplier=1,
vocab_size=TINY_QWEN3_5_KWARGS["vocab_size"],
)
batch = make_batch(
max_len=MAX_LEN,
samples=samples,
hidden_size=TINY_QWEN3_5_KWARGS["hidden_size"],
num_target_layers=1,
preprocess=shift_batch_mtp,
)
logits_list, loss, metrics = model(**batch, step_weights=step_weights)
assert loss.isfinite()
loss.backward()
|