[Speculative decoding 7/9] Speculative decoding end-to-end correctness tests. (#3951)

This commit is contained in:
Cade Daniel
2024-04-23 08:02:36 +00:00
committed by GitHub
parent 050f285ff6
commit 62b8aebc6f
22 changed files with 1164 additions and 175 deletions
+6 -2
View File
@@ -91,12 +91,16 @@ def test_correct_output_format(which_tokens_accepted: str, seed: int,
bonus_token_ids,
)
# Bonus tokens are currently disabled. Verify they're set to -1.
# See https://github.com/vllm-project/vllm/issues/4212
expected_bonus_token_ids = bonus_token_ids.clone() * 0 - 1
if which_tokens_accepted == "all_tokens_accepted":
# Expect all tokens to be equal to draft tokens.
assert torch.equal(output_token_ids[:, :-1], draft_token_ids)
# Expect all bonus tokens to be included.
assert torch.equal(output_token_ids[:, -1:], bonus_token_ids)
assert torch.equal(output_token_ids[:, -1:], expected_bonus_token_ids)
elif which_tokens_accepted == "no_tokens_accepted":
# Expect first token to be equal to recovered tokens.
assert torch.equal(output_token_ids[:, 0], recovered_token_ids[:, 0])
@@ -106,7 +110,7 @@ def test_correct_output_format(which_tokens_accepted: str, seed: int,
torch.ones_like(output_token_ids[:, 1:]) * -1)
elif which_tokens_accepted == "some_tokens_accepted":
recovered_plus_bonus = torch.cat(
(recovered_token_ids, bonus_token_ids), dim=-1)
(recovered_token_ids, expected_bonus_token_ids), dim=-1)
# Assert first rejected token is a recovered token or bonus token.
assert torch.equal(
recovered_plus_bonus[torch.arange(0, batch_size),
+2 -1
View File
@@ -636,7 +636,8 @@ def test_sampler_top_k_top_p(seed: int, device: str):
def mock_sample(probs, *args, **kwargs):
nonlocal sample_probs
sample_probs = probs
return [[prob.topk(1, dim=-1).indices.tolist(), [0]] for prob in probs]
return ([[prob.topk(1, dim=-1).indices.tolist(), [0]]
for prob in probs], None)
with patch("vllm.model_executor.layers.sampler._sample", mock_sample):
sampler(logits=fake_logits, sampling_metadata=sampling_metadata)