Skip to content

[NVIDIA] GPTOSS GB200 DISAGG Configurations + Assign EP explicitly for AGG #387

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 6 commits into from
Jan 7, 2026
Merged

[NVIDIA] GPTOSS GB200 DISAGG Configurations + Assign EP explicitly for AGG #387

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
0feefdb
1. Update B200 AGG configs to EP. These were mistakenly set to EP=1 w…
Dec 31, 2025
7aae8cd
- Add GPTOSS DISAGG configurations
Jan 6, 2026
4752c9a
- Update PR # in perfchangelog
Jan 6, 2026
5c93796
Update with dynamo release branch and release container
Jan 6, 2026
e07fcc9
Updated perfchangelog description based on PR feedback
Jan 6, 2026
b011864
update postprocessing pattern matching
Jan 6, 2026
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
295 changes: 283 additions & 12 deletions .github/configs/nvidia-master.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -180,30 +180,30 @@ gptoss-fp4-b200-trt:
precision: fp4
framework: trt
multinode: false
# For all sequence lengths, if CONC >= 256, then EP=TP and DP_ATTN=true
seq-len-configs:
# DP Attn at higher concurrencies, TP attn at middle to lower. TP=1 turns out to be highest as artifact of concurrency limit=128
- isl: 1024
osl: 1024
search-space:
- { tp: 2, dp-attn: true, conc-start: 32, conc-end: 128 }
- { 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: 32 }
- { tp: 4, conc-start: 4, conc-end: 64 }
- { tp: 1, conc-start: 128, conc-end: 128 }
- { tp: 2, ep: 2, dp-attn: true, conc-start: 64, conc-end: 128 }
- { tp: 4, ep: 4, dp-attn: true, conc-start: 64, conc-end: 64 }
- { tp: 2, conc-start: 8, conc-end: 32 }
- { tp: 4, conc-start: 4, conc-end: 16 }
- { 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: 2, ep: 2, dp-attn: true, conc-start: 64, conc-end: 128 }
- { tp: 2, conc-start: 4, conc-end: 16 }
- { tp: 4, conc-start: 4, conc-end: 128 }
- { tp: 8, conc-start: 4, conc-end: 16 }
- { tp: 8, conc-start: 4, conc-end: 8 }
# DP Attn at higher concurrencies, TP attn at middle to lower. TP=1 turns out to be highest as artifact of concurrency limit=128
Comment thread
jgangani marked this conversation as resolved.
- 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: 1, conc-start: 128, conc-end: 128 }
- { tp: 2, ep: 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: 32 }
- { tp: 8, conc-start: 4, conc-end: 8 }
Expand Down Expand Up @@ -1047,3 +1047,274 @@ dsr1-fp4-gb200-dynamo-sglang:
dp-attn: true
additional-settings:
- "DECODE_NODES=8"

gptoss-fp4-gb200-dynamo-trt:
image: nvcr.io#nvidia/ai-dynamo/tensorrtllm-runtime:0.7.0.post2
model: openai/gpt-oss-120b
model-prefix: gptoss
runner: gb200
precision: fp4
framework: dynamo-trt
multinode: true
disagg: true
seq-len-configs:
- isl: 1024
osl: 1024
search-space:
#Right of pareto
#P: 1xTP1 D:1xTP4
- spec-decoding: "none"
conc-list: [ 1, 2, 4, 16, 32, 64, 128 ]
prefill:
num-worker: 1
tp: 1
ep: 1
dp-attn: false
additional-settings:
- "PREFILL_NODES=1"
- "PREFILL_MAX_NUM_TOKENS=20000"
- "PREFILL_MAX_BATCH_SIZE=32"
decode:
num-worker: 1
tp: 4
ep: 1
dp-attn: false
additional-settings:
- "DECODE_NODES=1"
- "DECODE_MAX_NUM_TOKENS=20000"
- "DECODE_MAX_BATCH_SIZE=256"
- "DECODE_GPU_MEM_FRACTION=0.9"

# P: 1xTP1 D:4xTP2
- spec-decoding: "none"
conc-list: [ 16 ]
prefill:
num-worker: 1
tp: 1
ep: 1
dp-attn: false
additional-settings:
- "PREFILL_NODES=1"
- "PREFILL_MAX_NUM_TOKENS=20000"
- "PREFILL_MAX_BATCH_SIZE=32"
Comment thread
jgangani marked this conversation as resolved.
decode:
num-worker: 4
tp: 2
ep: 1
dp-attn: false
additional-settings:
- "DECODE_NODES=2"
- "DECODE_MAX_NUM_TOKENS=20000"
- "DECODE_MAX_BATCH_SIZE=32"
- "DECODE_GPU_MEM_FRACTION=0.9"

# P: 1xTP1 D:1xDEP2
- spec-decoding: "none"
conc-list: [ 256, 512, 1024, 2048, 2560 ]
prefill:
num-worker: 1
tp: 1
ep: 1
dp-attn: false
additional-settings:
- "PREFILL_NODES=1"
- "PREFILL_MAX_NUM_TOKENS=20000"
- "PREFILL_MAX_BATCH_SIZE=32"
decode:
num-worker: 1
tp: 2
ep: 2
dp-attn: true
additional-settings:
- "DECODE_NODES=1"
- "DECODE_MAX_NUM_TOKENS=20000"
- "DECODE_MAX_BATCH_SIZE=1536"
- "DECODE_GPU_MEM_FRACTION=0.9"

# P: 1xTP1 D:2xDEP2
- spec-decoding: "none"
conc-list: [ 512, 1024, 2048, 2560 ]
prefill:
num-worker: 1
tp: 1
ep: 1
dp-attn: false
additional-settings:
- "PREFILL_NODES=1"
- "PREFILL_MAX_NUM_TOKENS=20000"
- "PREFILL_MAX_BATCH_SIZE=32"
decode:
num-worker: 2
tp: 2
ep: 2
dp-attn: true
additional-settings:
- "DECODE_NODES=1"
- "DECODE_MAX_NUM_TOKENS=20000"
- "DECODE_MAX_BATCH_SIZE=1536"
- "DECODE_GPU_MEM_FRACTION=0.9"

# P: 1xTP1 D:1xDEP4
- spec-decoding: "none"
conc-list: [ 256, 1024, 1536 ]
prefill:
num-worker: 1
tp: 1
ep: 1
dp-attn: false
additional-settings:
- "PREFILL_NODES=1"
- "PREFILL_MAX_NUM_TOKENS=20000"
- "PREFILL_MAX_BATCH_SIZE=32"
decode:
num-worker: 1
tp: 4
ep: 4
dp-attn: true
additional-settings:
- "DECODE_NODES=1"
- "DECODE_MAX_NUM_TOKENS=20000"
- "DECODE_MAX_BATCH_SIZE=512"
- "DECODE_GPU_MEM_FRACTION=0.9"

# P: 1xTP1 D:3xDEP4
- spec-decoding: "none"
conc-list: [ 3072 ]
prefill:
num-worker: 1
tp: 1
ep: 1
dp-attn: false
additional-settings:
- "PREFILL_NODES=1"
- "PREFILL_MAX_NUM_TOKENS=20000"
- "PREFILL_MAX_BATCH_SIZE=32"
decode:
num-worker: 3
tp: 4
ep: 4
dp-attn: true
additional-settings:
- "DECODE_NODES=1"
- "DECODE_MAX_NUM_TOKENS=20000"
- "DECODE_MAX_BATCH_SIZE=1024"
- "DECODE_GPU_MEM_FRACTION=0.9"

- isl: 8192
osl: 1024
search-space:
# Right side of pareto
- spec-decoding: "none"
conc-list: [1]
prefill:
num-worker: 1
tp: 1
ep: 1
dp-attn: false
additional-settings:
- "PREFILL_NODES=1"
- "PREFILL_MAX_NUM_TOKENS=20000"
- "PREFILL_MAX_BATCH_SIZE=32"
decode:
num-worker: 1
tp: 8
ep: 1
dp-attn: false
additional-settings:
- "DECODE_NODES=2"
- "DECODE_MAX_NUM_TOKENS=20000"
- "DECODE_MAX_BATCH_SIZE=4"
- "DECODE_GPU_MEM_FRACTION=0.9"

- spec-decoding: "none"
conc-list: [2, 4, 8, 16, 32, 64]
prefill:
num-worker: 1
tp: 1
ep: 1
dp-attn: false
additional-settings:
- "PREFILL_NODES=1"
- "PREFILL_MAX_NUM_TOKENS=20000"
- "PREFILL_MAX_BATCH_SIZE=32"
decode:
num-worker: 1
tp: 4
ep: 1
dp-attn: false
additional-settings:
- "DECODE_NODES=1"
- "DECODE_MAX_NUM_TOKENS=20000"
- "DECODE_MAX_BATCH_SIZE=128"
- "DECODE_GPU_MEM_FRACTION=0.9"

# Middle of pareto
# P: 2xTP1 D:1xTP4
- spec-decoding: "none"
conc-list: [128, 512]
prefill:
num-worker: 2
tp: 1
ep: 1
dp-attn: false
additional-settings:
- "PREFILL_NODES=1"
- "PREFILL_MAX_NUM_TOKENS=20000"
- "PREFILL_MAX_BATCH_SIZE=32"
decode:
num-worker: 1
tp: 4
ep: 1
dp-attn: false
additional-settings:
- "DECODE_NODES=1"
- "DECODE_MAX_NUM_TOKENS=20000"
- "DECODE_MAX_BATCH_SIZE=1024"
- "DECODE_GPU_MEM_FRACTION=0.9"

# P: 2xTP1 D:1xTP2
- spec-decoding: "none"
conc-list: [256, 384]
prefill:
num-worker: 2
tp: 1
ep: 1
dp-attn: false
additional-settings:
- "PREFILL_NODES=1"
- "PREFILL_MAX_NUM_TOKENS=20000"
- "PREFILL_MAX_BATCH_SIZE=32"
decode:
num-worker: 1
tp: 2
ep: 1
dp-attn: false
additional-settings:
- "DECODE_NODES=1"
- "DECODE_MAX_NUM_TOKENS=20000"
- "DECODE_MAX_BATCH_SIZE=512"
- "DECODE_GPU_MEM_FRACTION=0.9"

# P: 2xTP1 D:1xDEP2
- spec-decoding: "none"
conc-list: [128, 512]
prefill:
num-worker: 2
tp: 1
ep: 1
dp-attn: false
additional-settings:
- "PREFILL_NODES=1"
- "PREFILL_MAX_NUM_TOKENS=20000"
- "PREFILL_MAX_BATCH_SIZE=32"
decode:
num-worker: 1
tp: 2
ep: 2
dp-attn: true
additional-settings:
- "DECODE_NODES=1"
- "DECODE_MAX_NUM_TOKENS=20000"
- "DECODE_MAX_BATCH_SIZE=512"
- "DECODE_GPU_MEM_FRACTION=0.9"

13 changes: 10 additions & 3 deletions benchmarks/gptoss_fp4_b200_trt_slurm.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,9 +49,16 @@ moe_config:
EOF

if [[ "$DP_ATTENTION" == "true" ]]; then
export TRTLLM_MOE_ALLTOALL_BACKEND="mnnvlthroughput"
export TRTLLM_FORCE_ALLTOALL_METHOD="MNNVL"
export TRTLLM_MOE_A2A_WORKSPACE_MB="2048"
# DISABLE All2All for MoE TP
if [[ "$EP_SIZE" -eq 1 ]]; then
# DTP Alltoall Environment variables for EP_SIZE == 1
export TRTLLM_FORCE_ALLTOALL_METHOD="NotEnabled"
elif [[ "$EP_SIZE" -gt 1 ]]; then
# DEP
export TRTLLM_MOE_ALLTOALL_BACKEND="mnnvlthroughput"
export TRTLLM_FORCE_ALLTOALL_METHOD="MNNVL"
export TRTLLM_MOE_A2A_WORKSPACE_MB="2048"
fi
cat << EOF >> $EXTRA_CONFIG_FILE
attention_dp_config:
enable_balance: true
Expand Down
Loading
Loading