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 15, 06:00 UTC
Adding "AMD: Multi-step Tests" to amdproduction. (#19508)
Signed-off-by: Yida Wu <yidawu@alumni.cmu.edu> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Concurrensee committed
Jun 13, 2025 at 19:08 UTC
d65668b4e88008adccb91c6598c93c2d53487484
2 files changed
+11
-1
.buildkite/test-pipeline.yaml
+1
-1
@@ -675,7 +675,7 @@ steps:
675
- pytest -v -s plugins/lora_resolvers # unit tests for in-tree lora resolver plugins
676
677
- label: Multi-step Tests (4 GPUs) # 36min
678
- mirror_hardwares: [amdexperimental]
678
+ mirror_hardwares: [amdexperimental, amdproduction]
679
working_dir: "/vllm-workspace/tests"
680
num_gpus: 4
681
source_file_dependencies:
tests/multi_step/test_correctness_llm.py
+10
@@ -8,6 +8,7 @@ from typing import Optional
8
9
import pytest
10
11
+from vllm.platforms import current_platform
12
from vllm.utils import STR_BACKEND_ENV_VAR
13
14
from ..models.utils import check_logprobs_close, check_outputs_equal
@@ -71,6 +72,12 @@ def test_multi_step_llm(
72
num_logprobs: corresponds to the `logprobs` argument to the OpenAI
73
completions endpoint; `None` -> 1 logprob returned.
74
"""
75
+ if current_platform.is_rocm() and \
76
+ (attention_backend == "FLASHINFER" or enable_chunked_prefill):
77
+ pytest.skip(
78
+ "Multi-Step with FLASHINFER or Chunked-Prefill is not supported"
79
+ "on ROCm")
80
+
81
with monkeypatch.context() as m:
82
m.setenv(STR_BACKEND_ENV_VAR, attention_backend)
83
@@ -221,6 +228,9 @@ def test_multi_step_llm_w_prompt_logprobs(
228
@pytest.mark.parametrize("num_prompts", NUM_PROMPTS)
229
@pytest.mark.parametrize("num_logprobs", [None, 5])
230
@pytest.mark.parametrize("attention_backend", ["FLASH_ATTN"])
231
+@pytest.mark.skipif(
232
+ current_platform.is_rocm(),
233
+ reason="Multi-Step + Chunked-Prefill not supported on ROCm")
234
def test_multi_step_llm_chunked_prefill_prefix_cache(
235
vllm_runner,
236
example_prompts,