[CI] Refresh hybrid Model Runner V2 coverage (#51410)
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Jiangyun Zhu committed
Aug 8, 2026 at 13:10 UTC
44351f81d58861edc873c7678c500a4f40834450
5 files changed
+8
-16
.buildkite/intel_jobs/model_runner_v2_intel.yaml
+1
@@ -108,6 +108,7 @@ steps:
108
- vllm/v1/worker/gpu_worker.py
109
- tests/v1/spec_decode/test_max_len.py
110
- tests/v1/spec_decode/test_rejection_sampler_utils.py
111
+ - tests/v1/spec_decode/test_synthetic_rejection_sampler_utils.py
112
commands:
113
- >-
114
bash .buildkite/scripts/hardware_ci/run-intel-test.sh
.buildkite/test-amd.yaml
+1
-1
@@ -1851,7 +1851,7 @@ steps:
1851
commands:
1852
- set -x
1853
- export VLLM_USE_V2_MODEL_RUNNER=1
1854
- - pytest -v -s distributed/test_pipeline_parallel.py -k "not ray and not Jamba"
1854
+ - pytest -v -s distributed/test_pipeline_parallel.py -k "not ray"
1855
- pytest -v -s distributed/test_pp_cudagraph.py -k "not ray"
1856
1857
- label: Model Runner V2 Spec Decode # TBD
.buildkite/test_areas/model_runner_v2.yaml
+5
-1
@@ -96,7 +96,7 @@ steps:
96
commands:
97
- set -x
98
- export VLLM_USE_V2_MODEL_RUNNER=1
99
- - pytest -v -s distributed/test_pipeline_parallel.py -k "not ray and not Jamba"
99
+ - pytest -v -s distributed/test_pipeline_parallel.py -k "not ray"
100
- pytest -v -s distributed/test_pp_cudagraph.py -k "not ray"
101
- pytest -v -s v1/distributed/test_pp_dp_v2.py
102
@@ -110,6 +110,7 @@ steps:
110
- vllm/v1/worker/gpu_worker.py
111
- tests/v1/spec_decode/test_max_len.py
112
- tests/v1/spec_decode/test_rejection_sampler_utils.py
113
+ - tests/v1/spec_decode/test_synthetic_rejection_sampler_utils.py
114
- tests/v1/e2e/spec_decode/
115
commands:
116
- set -x
@@ -119,3 +120,6 @@ steps:
120
- pytest -v -s v1/spec_decode/test_synthetic_rejection_sampler_utils.py
121
- pytest -v -s v1/e2e/spec_decode/eagle/
122
- pytest -v -s v1/e2e/spec_decode/speculators/
123
+ - >-
124
+ pytest -v -s
125
+ 'v1/e2e/spec_decode/mtp/test_mtp.py::test_mtp_correctness[qwen3_5-hybrid]'
.buildkite/test_areas/quantization.yaml
+1
-7
@@ -13,13 +13,7 @@ steps:
13
- vllm/model_executor/layers/quantization
14
- tests/quantization
15
commands:
16
- # temporary install here since we need nightly, will move to requirements/test/cuda.in
17
- # after torchao 0.12 release, and pin a working version of torchao nightly here
18
-
19
- # since torchao nightly is only compatible with torch nightly currently
20
- # https://github.com/pytorch/ao/issues/2919, we'll have to skip new torchao tests for now
21
- # we can only upgrade after this is resolved
22
- # TODO(jerryzh168): resolve the above comment
16
+ # Pin torchao to the version compatible with the CI PyTorch and CUDA stack.
17
- uv pip install --system torchao==0.17.0 --index-url https://download.pytorch.org/whl/cu130
18
- uv pip install --system conch-triton-kernels
19
# The SM90-only checkpoint currently contains a removed weight_chan_scale
tests/v1/e2e/spec_decode/mtp/test_mtp.py
-7
@@ -1,7 +1,6 @@
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
-import os
4
from typing import Any
5
6
import pytest
@@ -72,12 +71,6 @@ def test_mtp_correctness(
71
draft_model = None
72
_skip_if_insufficient_gpus_for_tp(tp_size)
73
75
- if "Qwen3.5" in model_name and os.environ.get("VLLM_USE_V2_MODEL_RUNNER"):
76
- pytest.skip(
77
- "Model Runner V2 does not yet support hybrid models "
78
- "(Qwen3.5 mixes Mamba-style GDN with attention layers)."
79
- )
80
-
74
attn_backend = "TRITON_ATTN" if current_platform.is_rocm() else "auto"
75
76
# Skip multimodal profiling for models that don't need it in this test.