Skip to content

[NVIDIA] Update GPTOSS B200 TRTLLM #266

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

Merged
cquil11 merged 3 commits into from
Dec 2, 2025
Merged

[NVIDIA] Update GPTOSS B200 TRTLLM #266

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
b1fbf48
Update GPTOSS B200 AGG
Dec 2, 2025
1abbbb6
set dp attention env vars
Dec 2, 2025
eeeb6cd
Add DP attn comment
Dec 2, 2025
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
14 changes: 9 additions & 5 deletions .github/configs/nvidia-master.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -167,34 +167,38 @@ dsr1-fp8-h200-trt:
- { tp: 8, ep: 8, dp-attn: true, conc-start: 64, conc-end: 64 }

gptoss-fp4-b200-trt:
image: nvcr.io#nvidia/tensorrt-llm/release:1.2.0rc0.post1
image: nvcr.io#nvidia/tensorrt-llm/release:1.2.0rc2
model: openai/gpt-oss-120b
model-prefix: gptoss
runner: b200-trt
precision: fp4
framework: trt
# For all sequence lengths, if CONC >= 256, then EP=TP and DP_ATTN=true
# Enable DP_ATTENTION for conc >= 32
seq-len-configs:
- isl: 1024
osl: 1024
search-space:
- { tp: 2, dp-attn: true, conc-start: 32, conc-end: 128 }
Comment thread
cquil11 marked this conversation as resolved.
- { tp: 4, dp-attn: true, conc-start: 32, conc-end: 64 }
- { tp: 1, conc-start: 64, conc-end: 128 }
- { tp: 2, conc-start: 4, conc-end: 128 }
- { tp: 4, conc-start: 4, conc-end: 128 }
- { tp: 2, conc-start: 4, conc-end: 32 }
- { tp: 4, conc-start: 4, conc-end: 64 }
- { tp: 8, conc-start: 4, conc-end: 8 }
- isl: 1024
osl: 8192
search-space:
- { tp: 1, conc-start: 64, conc-end: 128 }
- { tp: 2, dp-attn: true, conc-start: 64, conc-end: 128 }
- { tp: 2, conc-start: 4, conc-end: 128 }
- { tp: 4, conc-start: 4, conc-end: 128 }
- { tp: 8, conc-start: 4, conc-end: 16 }
- isl: 8192
osl: 1024
search-space:
- { tp: 1, conc-start: 64, conc-end: 128 }
- { tp: 2, dp-attn: true, conc-start: 64, conc-end: 128 }
- { tp: 2, conc-start: 4, conc-end: 128 }
- { tp: 4, conc-start: 4, conc-end: 128 }
- { tp: 4, conc-start: 4, conc-end: 32 }
- { tp: 8, conc-start: 4, conc-end: 8 }

gptoss-fp4-b200-vllm:
Expand Down
4 changes: 3 additions & 1 deletion benchmarks/gptoss_fp4_b200_trt_slurm.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,7 +31,6 @@ echo "MOE_BACKEND set to '$MOE_BACKEND'"

EXTRA_CONFIG_FILE="gptoss-fp4.yml"
export TRTLLM_ENABLE_PDL=1
export NCCL_GRAPH_REGISTER=0

cat > $EXTRA_CONFIG_FILE << EOF
cuda_graph_config:
Expand All @@ -50,6 +49,9 @@ moe_config:
EOF

if [[ "$DP_ATTENTION" == "true" ]]; then
export TRTLLM_MOE_ALLTOALL_BACKEND="mnnvlthroughput"
export TRTLLM_FORCE_ALLTOALL_METHOD="MNNVL"
export TRTLLM_MOE_A2A_WORKSPACE_MB="2048"
cat << EOF >> $EXTRA_CONFIG_FILE
attention_dp_config:
enable_balance: true
Expand Down
Loading