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

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

Browse Source
This commit is contained in:
Cade Daniel
2024-04-23 01:02:36 -07:00
committed by GitHub
parent 050f285ff6
commit 62b8aebc6f
22 changed files with 1164 additions and 175 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
tests/spec_decode/test_metrics.py
View File

@@ -119,7 +119,7 @@ def test_initial_metrics_has_correct_values(has_data: bool):
num_draft_tokens = 0
k = 5
num_possible_tokens = AsyncMetricsCollector.get_max_num_accepted_tokens(
max_num_emitted_tokens = AsyncMetricsCollector.get_max_num_emitted_tokens(
num_draft_tokens, k)
rej_sampler = MagicMock()
@@ -153,7 +153,7 @@ def test_initial_metrics_has_correct_values(has_data: bool):
assert (metrics.draft_acceptance_rate == num_accepted_tokens /
num_draft_tokens)
assert (metrics.system_efficiency == num_emitted_tokens /
num_possible_tokens)
max_num_emitted_tokens)
else:
assert math.isnan(metrics.draft_acceptance_rate)
assert math.isnan(metrics.system_efficiency)