Skip to content

[https://nvbugs/6018046][fix] Drop max_batch_size 32→8 for throughput_pp4_mtp (matching throughput_bs8_mtp), l #14201

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
tensorrt-cicd wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[https://nvbugs/6018046][fix] Drop max_batch_size 32→8 for throughput_pp4_mtp (matching throughput_bs8_mtp), l #14201

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
11 changes: 9 additions & 2 deletions tests/integration/defs/accuracy/test_llm_api_pytorch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2738,7 +2738,7 @@ class TestDeepSeekR1(LlmapiAccuracyTestHarness):
False,
True,
True,
32,
8,
"CUTLASS",
marks=pytest.mark.skip_less_mpi_world_size(4)),
],
Expand All @@ -2756,7 +2756,14 @@ def test_nvfp4_multi_gpus(self, tp_size, pp_size, ep_size, mtp_nextn, fp8kv,
if moe_backend == "TRTLLM" and sm_version in (120, 121):
pytest.skip(f"{moe_backend} backend does not support SM 120 or 121")

kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.70)
# PP4 + MTP + ep_size=1 keeps all 256 experts on every rank, so the
# weight footprint per GPU is much higher than the EP-sharded variants.
# The default 0.70 KV fraction leaves too little headroom for runtime
# NCCL / cuBLAS workspaces on Blackwell, which manifests as mid-run
# CUBLAS_STATUS_EXECUTION_FAILED in the router GEMM (NVBug 6018046).
kv_fraction = 0.5 if (pp_size > 1 and ep_size == 1
and mtp_nextn > 0) else 0.70
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=kv_fraction)
pytorch_config = dict(
disable_overlap_scheduler=not overlap_scheduler,
cuda_graph_config=CudaGraphConfig() if cuda_graph else None,
Expand Down
Loading