iHore/vLLM
8
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

[CI] Move smaller H200 workloads to 18GB MIG slices (#56695)

Browse Source 
Signed-off-by: Kevin Luu <51931015+khluu@users.noreply.github.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Kevin H. Luu
2026-09-13 21:30:11 -07:00
committed by GitHub
co-authored by OpenAI Codex
parent a2685f2cda
commit 214248c7d8
4 changed files with 40 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.buildkite/test_areas/cuda.yaml
+2 -2
View File
@@ -30,8 +30,8 @@ steps:
- vllm/
- tests/cuda
- label: ":nvidia: (H200 MIG 35GB) CUDAGraph"
device: h200_35gb
- label: ":nvidia: (H200 MIG 18GB) CUDAGraph"
device: h200_18gb
key: cudagraph
timeout_in_minutes: 30
source_file_dependencies:
.buildkite/test_areas/engine.yaml
+17 -3
View File
@@ -74,18 +74,20 @@ steps:
depends_on:
- image-build-amd
- label: ":nvidia: (H200 MIG 35GB) E2E Core"
device: h200_35gb
- label: ":nvidia: (H200 MIG 18GB) E2E Core"
device: h200_18gb
key: e2e-core-1-gpu
timeout_in_minutes: 50
source_file_dependencies:
- vllm/v1/
- tests/v1/e2e/general/
commands:
- pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
- pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py --ignore v1/e2e/general/test_kv_sharing_fast_prefill.py -k "not test_mamba_prefix_cache_mrv2"
mirror:
amd:
label: ":amd: (MI250) E2E Core"
commands:
- pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
device: mi250_1
timeout_in_minutes: 50
depends_on:
@@ -95,6 +97,18 @@ steps:
- tests/v1/e2e/general/
- vllm/platforms/rocm.py
# Keep the large Gemma-3n and Mamba MRV2 warmup allocations on 35GB slices.
- label: ":nvidia: (H200 MIG 35GB) E2E Core Large Memory"
device: h200_35gb
key: e2e-core-large-memory
timeout_in_minutes: 50
source_file_dependencies:
- vllm/v1/
- tests/v1/e2e/general/
commands:
- pytest -v -s v1/e2e/general/test_kv_sharing_fast_prefill.py
- pytest -v -s v1/e2e/general/test_mamba_prefix_cache.py::test_mamba_prefix_cache_mrv2 v1/e2e/general/test_mamba_prefix_cache.py::test_mamba_prefix_cache_mrv2_async
- label: ":nvidia: (L4) V1 E2E"
key: v1-e2e-2-gpus
timeout_in_minutes: 25 # TODO: Fix timeout after we have more confidence in the test stability
.buildkite/test_areas/misc.yaml
+2 -2
View File
@@ -103,8 +103,8 @@ steps:
depends_on:
- image-build-amd
- label: ":nvidia: (H200 MIG 35GB) V1 Executor + Worker"
device: h200_35gb
- label: ":nvidia: (H200 MIG 18GB) V1 Executor + Worker"
device: h200_18gb
key: v1-executor-worker
timeout_in_minutes: 45
source_file_dependencies: *v1-core-kv-metrics-deps
.buildkite/test_areas/samplers.yaml
+19 -4
View File
@@ -2,8 +2,8 @@ group: Samplers
depends_on:
- image-build
steps:
- label: ":nvidia: (H200 MIG 35GB) Samplers"
device: h200_35gb
- label: ":nvidia: (H200 MIG 18GB) Samplers"
device: h200_18gb
key: samplers-test
timeout_in_minutes: 60
source_file_dependencies:
@@ -15,8 +15,8 @@ steps:
commands:
# VLLM_USE_FLASHINFER_SAMPLER defaults to 1 now, so we need to pin both
# values explicitly to still cover the PyTorch-native (Triton) path.
- VLLM_USE_FLASHINFER_SAMPLER=0 pytest -v -s samplers
- VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
- VLLM_USE_FLASHINFER_SAMPLER=0 pytest -v -s samplers -k "not test_beam_search_passes_multimodal_data"
- VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers -k "not test_beam_search_passes_multimodal_data"
mirror:
amd:
label: ":amd: (MI250) Samplers"
@@ -35,3 +35,18 @@ steps:
- vllm/platforms/rocm.py
commands:
- pytest -v -s samplers
# Qwen2-Audio-7B leaves no room for KV cache on a 16 GiB slice.
- label: ":nvidia: (H200 MIG 35GB) Samplers Multimodal Beam Search"
device: h200_35gb
key: samplers-multimodal-beam-search
timeout_in_minutes: 30
source_file_dependencies:
- vllm/model_executor/layers
- vllm/sampling_metadata.py
- tests/samplers
- tests/conftest.py
- vllm/entrypoints/generate/beam_search
commands:
- VLLM_USE_FLASHINFER_SAMPLER=0 pytest -v -s samplers/test_beam_search.py::test_beam_search_passes_multimodal_data
- VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers/test_beam_search.py::test_beam_search_passes_multimodal_data