biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[SpecDecode] Simplified alternative padded-speculation acceptance rate fix (#29845)

Browse Source 
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
This commit is contained in:
Lucas Wilkinson
2025-12-22 16:06:10 -05:00
committed by GitHub
parent 9586354053
commit de71747655
8 changed files with 62 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
tests/v1/spec_decode/test_eagle.py
View File

@@ -306,10 +306,16 @@ def test_prepare_inputs_padded():
proposer = _create_proposer("eagle", num_speculative_tokens)
output_metadata, token_indices_to_sample = proposer.prepare_inputs_padded(
common_attn_metadata, spec_decode_metadata, valid_sampled_tokens_count
output_metadata, token_indices_to_sample, num_rejected_tokens_gpu = (
proposer.prepare_inputs_padded(
common_attn_metadata, spec_decode_metadata, valid_sampled_tokens_count
)
)
# Verify num_rejected_tokens_gpu is calculated correctly
expected_num_rejected = torch.tensor([1, 0, 2], dtype=torch.int32, device=device)
assert torch.equal(num_rejected_tokens_gpu, expected_num_rejected)
assert output_metadata.max_query_len == 3
assert torch.equal(output_metadata.query_start_loc, expected_query_start_loc)
assert torch.equal(token_indices_to_sample, expected_token_indices_to_sample)