[Spec Decode] Reject placeholder (-1) draft tokens in rejection sampler (#46533)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Nick Hill
2026-06-24 03:32:32 +00:00
committed by GitHub
parent ce9f64020b
commit 7ee4d22009
4 changed files with 75 additions and 3 deletions
+33
View File
@@ -1150,3 +1150,36 @@ def test_synthetic_all_rejected(all_greedy: bool):
for row in result:
assert row[0] != PLACEHOLDER_TOKEN_ID
assert (row[1:] == PLACEHOLDER_TOKEN_ID).all()
def test_placeholder_draft_token_rejected_random(rejection_sampler):
"""A placeholder draft id (-1) must be rejected in non-greedy sampling
without indexing the probability tensors by the invalid id.
"""
vocab_size = 100
spec_tokens = [[1, vocab_size - 1, PLACEHOLDER_TOKEN_ID]]
output_tokens = [[1, vocab_size - 1, 7, 9]]
temperature = torch.ones(1, dtype=torch.float32, device=DEVICE_TYPE)
metadata = create_sampling_metadata(
all_greedy=False,
temperature=temperature,
generators={0: torch.Generator(device=DEVICE_TYPE).manual_seed(0)},
)
logits = create_logits_tensor(output_tokens, vocab_size=vocab_size)
bonus_token_tensor = torch.tensor([output_tokens[0][-1]], device=logits.device)
spec_decode_metadata = create_spec_decode_metadata(spec_tokens, logits)
mock_sampler_output(rejection_sampler, bonus_token_tensor)
output = rejection_sampler(
spec_decode_metadata,
draft_probs=None,
logits=logits,
sampling_metadata=metadata,
)
sampled = output.sampled_token_ids
assert sampled[0, 0].item() == 1
assert sampled[0, 1].item() == vocab_size - 1
recovered = sampled[0, 2].item()
assert 0 <= recovered < vocab_size
@@ -280,3 +280,32 @@ def test_synthetic_rejection_sample(
f"Step {i}: observed rate {observed_rate:.4f} deviates from "
f"expected rate {expected_rate:.4f} by more than {deviation_tol}."
)
def test_placeholder_draft_token_rejected():
"""A placeholder draft id (-1) must be rejected without reading the logit
tensors out of bounds, for any sampling method.
"""
torch.manual_seed(0)
device = "cuda"
num_trials = 64
K = 1
temperature = 0.6
target_logits_1d = torch.randn(VOCAB_SIZE, device=device) / temperature
draft_logits_1d = torch.randn(VOCAB_SIZE, device=device) / temperature
inputs = _build_rejection_sample_inputs(
target_logits_1d,
draft_logits_1d,
K,
temperature=temperature,
num_trials=num_trials,
)
inputs["draft_sampled"].view(num_trials, K + 1)[:, 1:] = -1
sampled, num_sampled = rejection_sample(**inputs, num_speculative_steps=K)
assert torch.equal(num_sampled, torch.ones_like(num_sampled))
recovered = sampled[:, 0]
assert (recovered >= 0).all() and (recovered < VOCAB_SIZE).all()