A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 10:45 UTC
[CI][SpecDecode] Fix spec decode tests, use flash attention backend for spec decode CI tests. (#8975)
Lily Liu committed
Sep 30, 2024 at 17:51 UTC
bce324487a8e36140143ea37f4b27d273a0fd661
2 files changed
+4
-3
.buildkite/test-pipeline.yaml
-2
@@ -207,8 +207,6 @@ steps:
207
- vllm/spec_decode
208
- tests/spec_decode
209
commands:
210
- # See https://github.com/vllm-project/vllm/issues/5152
211
- - export VLLM_ATTENTION_BACKEND=XFORMERS
210
- pytest -v -s spec_decode/e2e/test_multistep_correctness.py
211
- pytest -v -s spec_decode --ignore=spec_decode/e2e/test_multistep_correctness.py
212
tests/spec_decode/test_multi_step_worker.py
+4
-1
@@ -673,7 +673,10 @@ def test_use_draft_model_runner_advance_step():
673
worker.model_runner._gpu_advance_step.side_effect = ValueError(
674
exception_secret)
675
676
- seq_group_metadata_list, _, _ = create_batch(batch_size, k)
676
+ seq_group_metadata_list, _, _ = create_batch(batch_size,
677
+ k,
678
+ block_size=block_size,
679
+ num_gpu_blocks=num_gpu_blocks)
680
681
# Fallback (should not call) when num_steps=1.
682
execute_model_req = ExecuteModelRequest(