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, 21:00 UTC
[ROCm] Fix the Kernels, Core, and Prefix Caching AMD CI groups (#13970)
Signed-off-by: Sage Moore <sage@neuralmagic.com>
Sage Moore committed
Feb 27, 2025 at 12:31 UTC
38acae6e97897d78ddcb7b3d3608c2c076e1d623
3 files changed
+23
-1
.buildkite/run-amd-test.sh
+3
-1
@@ -92,7 +92,9 @@ if [[ $commands == *" kernels "* ]]; then
92
--ignore=kernels/test_moe.py \
93
--ignore=kernels/test_prefix_prefill.py \
94
--ignore=kernels/test_rand.py \
95
- --ignore=kernels/test_sampler.py"
95
+ --ignore=kernels/test_sampler.py \
96
+ --ignore=kernels/test_cascade_flash_attn.py \
97
+ --ignore=kernels/test_mamba_mixer2.py"
98
fi
99
100
#ignore certain Entrypoints tests
tests/core/block/e2e/test_correctness_sliding_window.py
+10
@@ -7,6 +7,7 @@ import pytest
7
8
from tests.kernels.utils import override_backend_env_variable
9
from vllm import LLM, SamplingParams
10
+from vllm.platforms import current_platform
11
12
from .conftest import get_text_from_llm_generator
13
@@ -42,6 +43,11 @@ def test_sliding_window_retrival(baseline_llm_generator, test_llm_generator,
43
44
Additionally, we compare the results of the v1 and v2 managers.
45
"""
46
+ if backend == "FLASHINFER" and current_platform.is_rocm():
47
+ pytest.skip("Flashinfer does not support ROCm/HIP.")
48
+ if backend == "XFORMERS" and current_platform.is_rocm():
49
+ pytest.skip("Xformers does not support ROCm/HIP.")
50
+
51
override_backend_env_variable(monkeypatch, backend)
52
53
sampling_params = SamplingParams(
@@ -101,6 +107,10 @@ def test_sliding_window_chunked_prefill(test_llm_generator, batch_size, seed,
107
The results with and without chunked prefill are not the same due to
108
numerical instabilities.
109
"""
110
+ if backend == "FLASHINFER" and current_platform.is_rocm():
111
+ pytest.skip("Flashinfer does not support ROCm/HIP.")
112
+ if backend == "XFORMERS" and current_platform.is_rocm():
113
+ pytest.skip("Xformers does not support ROCm/HIP.")
114
override_backend_env_variable(monkeypatch, backend)
115
116
sampling_params = SamplingParams(
tests/prefix_caching/test_prefix_caching.py
+10
@@ -12,6 +12,7 @@ from tests.kernels.utils import override_backend_env_variable
12
from vllm import SamplingParams, TokensPrompt
13
from vllm.core.scheduler import Scheduler
14
from vllm.engine.llm_engine import LLMEngine
15
+from vllm.platforms import current_platform
16
17
from ..models.utils import check_outputs_equal
18
@@ -53,6 +54,10 @@ def test_mixed_requests(
54
and the others don't. The cached position determines where
55
the sequence is at among the batch of prefills.
56
"""
57
+ if backend == "FLASHINFER" and current_platform.is_rocm():
58
+ pytest.skip("Flashinfer does not support ROCm/HIP.")
59
+ if backend == "XFORMERS" and current_platform.is_rocm():
60
+ pytest.skip("Xformers does not support ROCm/HIP.")
61
override_backend_env_variable(monkeypatch, backend)
62
63
with hf_runner(model, dtype=dtype) as hf_model:
@@ -103,6 +108,11 @@ def test_unstable_prompt_sequence(
108
backend: str,
109
monkeypatch,
110
) -> None:
111
+
112
+ if backend == "FLASHINFER" and current_platform.is_rocm():
113
+ pytest.skip("Flashinfer does not support ROCm/HIP.")
114
+ if backend == "XFORMERS" and current_platform.is_rocm():
115
+ pytest.skip("Xformers does not support ROCm/HIP.")
116
override_backend_env_variable(monkeypatch, backend)
117
118
with vllm_runner(