Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 16 additions & 0 deletions tests/integration/defs/disaggregated/test_disaggregated.py
Original file line number Diff line number Diff line change
Expand Up @@ -84,6 +84,21 @@ def get_ucx_tls():
return "^ib,gdr_copy"


def apply_pre_hopper_nccl_workaround(env: dict) -> None:
"""Keep NCCL off PCIe peer-to-peer on pre-Hopper CI nodes.

The 8xL40S CI nodes run with the AMD IOMMU in DMA translation mode, so
PCIe P2P writes across the host bridge fault (``AMD-Vi IO_PAGE_FAULT``)
and never land. NCCL still picks its P2P/CUMEM transport there, and the
first PP send/recv or TP collective in the warmup forward hangs the
worker before it can register with the disagg server
(https://nvbugs/6728119). Restricting P2P to PCIe-switch peers makes NCCL
fall back to SHM on those nodes; NVLink platforms are unaffected.
"""
if get_sm_version() < 90:
env.setdefault("NCCL_P2P_LEVEL", "PXB")


def cleanup_output_files():
"""Clean up output files from previous runs."""
for file in ['output.json', 'output_streaming.json']:
Expand Down Expand Up @@ -1059,6 +1074,7 @@ def run_disaggregated_test(example_dir,

run_env = env.copy() if env else os.environ.copy()
run_env["UCX_TLS"] = get_ucx_tls()
apply_pre_hopper_nccl_workaround(run_env)
Comment thread
chuangz0 marked this conversation as resolved.
ctx_run_env = run_env.copy()
if ctx_env:
ctx_run_env.update(ctx_env)
Expand Down
4 changes: 0 additions & 4 deletions tests/integration/test_lists/waives.txt
Original file line number Diff line number Diff line change
Expand Up @@ -156,11 +156,7 @@ full:H20/accuracy/test_llm_api_pytorch_multimodal.py::TestExaone4_5_33B::test_au
full:H20/accuracy/test_llm_api_pytorch_multimodal.py::TestExaone4_5_33B::test_auto_dtype[full_budget] SKIP (https://nvbugs/6597570)
full:H20/accuracy/test_llm_api_pytorch_multimodal.py::TestNanoV3Omni::test_auto_dtype[bf16] SKIP (https://nvbugs/6618649)
full:L40S/accuracy/test_llm_api_pytorch_encode.py::TestEncoderEncode::test_encoder_encode_cuda_graph_matches_eager_logits[bert-yelp] SKIP (https://nvbugs/6802114)
full:L40S/disaggregated/test_disaggregated.py::test_disaggregated_ctxpp2_genpp2 SKIP (https://nvbugs/6728119)
full:L40S/disaggregated/test_disaggregated.py::test_disaggregated_ctxpp2_gentp2 SKIP (https://nvbugs/6728119)
full:L40S/disaggregated/test_disaggregated.py::test_disaggregated_ctxtp2_genpp2 SKIP (https://nvbugs/6728119)
full:L40S/disaggregated/test_disaggregated.py::test_disaggregated_deepseek_v3_lite_fp8_attention_dp_overlap[DeepSeek-V3-Lite-fp8] SKIP (https://nvbugs/6581064)
full:L40S/disaggregated/test_disaggregated.py::test_disaggregated_multi_gpu SKIP (https://nvbugs/6728119)
full:RTXPro6000D/accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[dep4_latency_moe_cutlass-torch_compile=True] SKIP (https://nvbugs/5929339)
full:RTX_PRO_6000_Blackwell_Server_Edition/accuracy/test_disaggregated_serving.py::TestDeepSeekV3Lite::test_auto_dtype[mtp_nextn=0-overlap_scheduler=False] SKIP (https://nvbugs/6616033)
full:RTX_PRO_6000_Blackwell_Server_Edition/accuracy/test_disaggregated_serving.py::TestDeepSeekV3Lite::test_auto_dtype[mtp_nextn=0-overlap_scheduler=True] SKIP (https://nvbugs/6616033)
Expand Down
Loading