biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[compile] Enable sequence parallelism matching w/o custom ops enabled (#27126)

Browse Source 
Signed-off-by: angelayi <yiangela7@gmail.com>
Signed-off-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Signed-off-by: ProExpertProg <lgovedic@redhat.com>
Co-authored-by: Luka Govedič <lgovedic@redhat.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: Luka Govedič <luka.govedic@gmail.com>
(cherry picked from commit f36292dbee)
This commit is contained in:
Angela Yi
2025-11-15 03:46:12 -08:00
committed by Kevin H. Luu
parent 5f6666fb5a
commit f67299f66d
6 changed files with 487 additions and 452 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
.buildkite/test-pipeline.yaml
View File

@@ -472,10 +472,11 @@ steps:
- vllm/
- tests/compile
commands:
# fp8 kv scales not supported on sm89, tested on Blackwell instead
- pytest -v -s compile/test_full_graph.py -k 'not test_fp8_kv_scale_compile'
# Limit to no custom ops to reduce running time
# Wrap with quotes to escape yaml and avoid starting -k string with a -
- "pytest -v -s compile/test_fusions_e2e.py -k 'TRITON and -quant_fp8'"
- "pytest -v -s compile/test_fusions_e2e.py -k 'TRITON and not +quant_fp8 and not Llama-4'"
- label: Cudagraph test
timeout_in_minutes: 20
@@ -913,7 +914,7 @@ steps:
- pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
- pytest -v -s tests/kernels/moe/test_flashinfer.py
- label: Blackwell Fusion Tests # 30 min
- label: Blackwell Fusion & Compile Tests # 30 min
timeout_in_minutes: 40
working_dir: "/vllm-workspace/"
gpu: b200
@@ -933,8 +934,10 @@ steps:
# this runner has 2 GPUs available even though num_gpus=2 is not set
- pytest -v -s tests/compile/test_fusion_all_reduce.py
# Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
# Wrap with quotes to escape yaml
- "pytest -v -s tests/compile/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and Llama-3.1 and -quant_fp8 and -rms_norm'"
# Wrap with quotes to escape yaml
- "pytest -v -s tests/compile/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and not +quant_fp8 and not +rms_norm'"
# test_fp8_kv_scale_compile requires FlashAttention (not supported on default L4/L40)
- pytest -v -s tests/compile/test_full_graph.py::test_fp8_kv_scale_compile
- label: Blackwell Fusion E2E Tests # 30 min
timeout_in_minutes: 40
@@ -957,8 +960,6 @@ steps:
- nvidia-smi
# Run all e2e fusion tests
- pytest -v -s tests/compile/test_fusions_e2e.py
# test_fp8_kv_scale_compile requires FlashAttention (not supported on default L4/L40)
- pytest -v -s tests/compile/test_full_graph.py::test_fp8_kv_scale_compile
- label: Blackwell GPT-OSS Eval
timeout_in_minutes: 60
@@ -1254,7 +1255,8 @@ steps:
- pytest -v -s tests/compile/test_async_tp.py
- pytest -v -s tests/compile/test_sequence_parallelism.py
- pytest -v -s tests/compile/test_fusion_all_reduce.py
- pytest -v -s tests/compile/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm
- "pytest -v -s tests/compile/test_fusions_e2e.py -k 'not Llama-4'"
- pytest -v -s tests/distributed/test_sequence_parallel.py
- pytest -v -s tests/distributed/test_context_parallel.py
- CUDA_VISIBLE_DEVICES=1,2 VLLM_ALL2ALL_BACKEND=deepep_high_throughput VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model Qwen/Qwen1.5-MoE-A2.7B --tp-size=1 --dp-size=2 --max-model-len 2048
- pytest -v -s tests/v1/distributed/test_dbo.py