feat: GB200 GPT-oss disagg recipe (#4954)

Signed-off-by: jthomson04 <jothomson@nvidia.com>
Signed-off-by: jthomson04 <jwillthomson19@gmail.com>
Co-authored-by: dagil-nvidia <dagil@nvidia.com>

recipes/gpt-oss-120b/trtllm/disagg/decode.yaml

-# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
-# SPDX-License-Identifier: Apache-2.0
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-enable_attention_dp: true
-disable_overlap_scheduler: false
-moe_config:
-    backend: CUTLASS
-cuda_graph_config:
-    enable_padding: true
-cache_transceiver_config:
-  backend: UCX
-  max_tokens_in_buffer: 65536
-print_iter_log: false
-stream_interval: 10

recipes/gpt-oss-120b/trtllm/disagg/deploy.yaml

+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: llm-config
+data:
+  prefill.yaml: |
+    cache_transceiver_config:
+      backend: UCX
+      max_tokens_in_buffer: 9216
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 30
+    disable_overlap_scheduler: true
+    enable_attention_dp: false
+    kv_cache_config:
+      dtype: fp8
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.8
+    max_batch_size: 64
+    max_num_tokens: 20000
+    max_seq_len: 9000
+    moe_config:
+      backend: TRTLLM
+    moe_expert_parallel_size: 1
+    num_postprocess_workers: 4
+    pipeline_parallel_size: 1
+    print_iter_log: true
+    stream_interval: 20
+    tensor_parallel_size: 1
+    trust_remote_code: true
+  decode.yaml: |
+    allreduce_strategy: AUTO
+    attention_dp_config:
+      enable_balance: true
+    cache_transceiver_config:
+      backend: UCX
+      max_tokens_in_buffer: 9216
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 1280
+    disable_overlap_scheduler: false
+    enable_attention_dp: false
+    kv_cache_config:
+      dtype: fp8
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.85
+    max_batch_size: 1280
+    max_num_tokens: 20000
+    max_seq_len: 11000
+    moe_config:
+      backend: TRTLLM
+    moe_expert_parallel_size: 1
+    num_postprocess_workers: 4
+    pipeline_parallel_size: 1
+    print_iter_log: true
+    stream_interval: 20
+    tensor_parallel_size: 4
+    trust_remote_code: true
+
+---
+apiVersion: nvidia.com/v1alpha1
+kind: DynamoGraphDeployment
+metadata:
+  name: gpt-oss-disagg
+spec:
+  backendFramework: trtllm
+  pvcs:
+    - name: model-cache
+      create: false
+  services:
+    Frontend:
+      componentType: frontend
+      dynamoNamespace: gpt-oss-disagg
+      extraPodSpec:
+        affinity:
+          podAntiAffinity:
+            requiredDuringSchedulingIgnoredDuringExecution:
+            - labelSelector:
+                matchExpressions:
+                - key: nvidia.com/dynamo-graph-deployment-name
+                  operator: In
+                  values:
+                  - gpt-oss-disagg-frontend
+              topologyKey: kubernetes.io/hostname
+        mainContainer:
+          args:
+          - python3 -m dynamo.frontend --router-mode round-robin --http-port 8000
+          command:
+          - /bin/sh
+          - -c
+          image: nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:0.7.0
+      replicas: 1
+    TrtllmPrefillWorker:
+      componentType: main
+      dynamoNamespace: gpt-oss-disagg
+      envFromSecret: hf-token-secret
+      volumeMounts:
+        - name: model-cache
+          mountPoint: /opt/models
+      sharedMemory:
+        size: 80Gi
+      extraPodSpec:
+        affinity:
+          nodeAffinity:
+            requiredDuringSchedulingIgnoredDuringExecution:
+              nodeSelectorTerms:
+              - matchExpressions:
+                - key: nvidia.com/gpu.present
+                  operator: In
+                  values:
+                  - "true"
+        mainContainer:
+          args:
+          - |
+            python3 -m dynamo.trtllm \
+              --model-path "${MODEL_PATH}" \
+              --served-model-name "openai/gpt-oss-120b" \
+              --extra-engine-args "${ENGINE_ARGS}" \
+              --disaggregation-mode prefill
+          command:
+          - /bin/sh
+          - -c
+          image: nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:0.7.0
+          env:
+          - name: TRTLLM_ENABLE_PDL
+            value: "1"
+          - name: TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL
+            value: "True"
+          - name: OVERRIDE_QUANT_ALGO
+            value: "W4A8_MXFP4_MXFP8"
+          - name: NCCL_GRAPH_REGISTER
+            value: "0"
+          - name: OMPI_MCA_coll_ucc_enable
+            value: "0"
+          - name: SERVED_MODEL_NAME
+            value: "openai/gpt-oss-120b"
+          - name: ENGINE_ARGS
+            value: "/opt/dynamo/configs/prefill.yaml"
+          - name: MODEL_PATH
+            value: "/opt/models/hub/models--openai--gpt-oss-120b/snapshots/b5c939de8f754692c1647ca79fbf85e8c1e70f8a"
+          - name: HF_HOME
+            value: /opt/models
+          volumeMounts:
+          - mountPath: /opt/dynamo/configs
+            name: llm-config
+            readOnly: true
+          workingDir: /workspace/examples/backends/trtllm
+        volumes:
+        - configMap:
+            name: llm-config
+          name: llm-config
+      replicas: 1
+      resources:
+        limits:
+          gpu: "1"
+        requests:
+          gpu: "1"
+    TrtllmDecodeWorker:
+      componentType: main
+      dynamoNamespace: gpt-oss-disagg
+      envFromSecret: hf-token-secret
+      volumeMounts:
+        - name: model-cache
+          mountPoint: /opt/models
+      sharedMemory:
+        size: 80Gi
+      extraPodSpec:
+        affinity:
+          nodeAffinity:
+            requiredDuringSchedulingIgnoredDuringExecution:
+              nodeSelectorTerms:
+              - matchExpressions:
+                - key: nvidia.com/gpu.present
+                  operator: In
+                  values:
+                  - "true"
+        mainContainer:
+          args:
+          - |
+            python3 -m dynamo.trtllm \
+              --model-path "${MODEL_PATH}" \
+              --served-model-name "openai/gpt-oss-120b" \
+              --extra-engine-args "${ENGINE_ARGS}" \
+              --disaggregation-mode decode
+          command:
+          - /bin/sh
+          - -c
+          image: nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:0.7.0
+          env:
+          - name: TRTLLM_ENABLE_PDL
+            value: "1"
+          - name: TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL
+            value: "True"
+          - name: OVERRIDE_QUANT_ALGO
+            value: "W4A8_MXFP4_MXFP8"
+          - name: NCCL_GRAPH_REGISTER
+            value: "0"
+          - name: OMPI_MCA_coll_ucc_enable
+            value: "0"
+          - name: SERVED_MODEL_NAME
+            value: "openai/gpt-oss-120b"
+          - name: ENGINE_ARGS
+            value: "/opt/dynamo/configs/decode.yaml"
+          - name: MODEL_PATH
+            value: "/opt/models/hub/models--openai--gpt-oss-120b/snapshots/b5c939de8f754692c1647ca79fbf85e8c1e70f8a"
+          - name: HF_HOME
+            value: /opt/models
+          volumeMounts:
+          - mountPath: /opt/dynamo/configs
+            name: llm-config
+            readOnly: true
+          workingDir: /workspace/examples/backends/trtllm
+        volumes:
+        - configMap:
+            name: llm-config
+          name: llm-config
+      replicas: 1
+      resources:
+        limits:
+          gpu: "4"
+        requests:
+          gpu: "4"
\ No newline at end of file

recipes/gpt-oss-120b/trtllm/disagg/perf.yaml

+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+apiVersion: batch/v1
+kind: Job
+metadata:
+  name: gpt-oss-120b-disagg-bench
+spec:
+  backoffLimit: 1
+  completions: 1
+  parallelism: 1
+  template:
+    metadata:
+      labels:
+        app: gpt-oss-120b-disagg-bench
+    spec:
+      affinity:
+        podAntiAffinity:
+          requiredDuringSchedulingIgnoredDuringExecution:
+            - labelSelector:
+                matchExpressions:
+                  - key: nvidia.com/dynamo-graph-deployment-name
+                    operator: In
+                    values:
+                      - gpt-oss-disagg
+              topologyKey: kubernetes.io/hostname
+      containers:
+      - command:
+        - /bin/sh
+        - -c
+        - |
+          apt-get update && apt-get install -y curl jq procps git && apt-get clean
+          pip install git+https://github.com/ai-dynamo/aiperf.git@4d3fa29403c8f75da22a14f1f7b3aeb27db9288f;
+          echo "aiperf installation completed";
+          sysctl -w net.ipv4.ip_local_port_range="1024 65000"
+          cat /proc/sys/net/ipv4/ip_local_port_range
+          export COLUMNS=200
+          EPOCH=$(date +%s)
+          ## utility functions -- can be moved to a bash script / configmap
+          wait_for_model_ready() {
+            echo "Waiting for model '$TARGET_MODEL' at $ENDPOINT/v1/models (checking every 5s)..."
+            while ! curl -s "http://$ENDPOINT/v1/models" | jq -e --arg model "$TARGET_MODEL" '.data[]? | select(.id == $model)' >/dev/null 2>&1; do
+                echo "[$(date '+%H:%M:%S')] Model not ready yet, sleeping 5s before checking again http://$ENDPOINT/v1/models"
+                sleep 5
+            done
+            echo "✅ Model '$TARGET_MODEL' is now available!"
+            echo "Model '$TARGET_MODEL' is now available!"
+            curl -s "http://$ENDPOINT/v1/models" | jq .
+          }
+          run_perf() {
+            local concurrency=$1
+            local isl=$2
+            local osl=$3
+            key=concurrency_${concurrency}
+            export ARTIFACT_DIR="${ROOT_ARTIFACT_DIR}/${EPOCH}_${JOB_NAME}/${key}"
+            mkdir -p "$ARTIFACT_DIR"
+            echo "ARTIFACT_DIR: $ARTIFACT_DIR"
+            aiperf profile --artifact-dir $ARTIFACT_DIR \
+                --model $TARGET_MODEL \
+                --tokenizer /model-cache/hub/models--openai--gpt-oss-120b/snapshots/b5c939de8f754692c1647ca79fbf85e8c1e70f8a  \
+                --endpoint-type chat  \
+                --endpoint /v1/chat/completions \
+                --streaming \
+                --url http://$ENDPOINT \
+                --synthetic-input-tokens-mean $isl \
+                --synthetic-input-tokens-stddev 0 \
+                --output-tokens-mean $osl \
+                --output-tokens-stddev 0 \
+                --extra-inputs "max_tokens:$osl" \
+                --extra-inputs "min_tokens:$osl" \
+                --extra-inputs "ignore_eos:true" \
+                --extra-inputs "{\"nvext\":{\"ignore_eos\":true}}" \
+                --extra-inputs "repetition_penalty:1.0" \
+                --extra-inputs "temperature: 0.0" \
+                --concurrency $concurrency \
+                --request-count $((10*concurrency)) \
+                --warmup-request-count $concurrency \
+                --num-dataset-entries 12800 \
+                --random-seed 100 \
+                --workers-max 252 \
+                -H 'Authorization: Bearer NOT USED' \
+                -H 'Accept: text/event-stream'\
+                --record-processors 32 \
+                --ui simple
+            echo "ARTIFACT_DIR: $ARTIFACT_DIR"
+            ls -la $ARTIFACT_DIR
+          }
+          #### Actual execution ####
+          wait_for_model_ready
+          mkdir -p "${ROOT_ARTIFACT_DIR}/${EPOCH}_${JOB_NAME}"
+          # Calculate total concurrency based on per-GPU concurrency and GPU count
+          TOTAL_CONCURRENCY=$((CONCURRENCY_PER_GPU * DEPLOYMENT_GPU_COUNT))
+          echo "Calculated total concurrency: $TOTAL_CONCURRENCY (${CONCURRENCY_PER_GPU} per GPU × ${DEPLOYMENT_GPU_COUNT} GPUs)"
+          # Write input_config.json
+          cat > "${ROOT_ARTIFACT_DIR}/${EPOCH}_${JOB_NAME}/input_config.json" <<EOF
+          {
+            "gpu_count": $DEPLOYMENT_GPU_COUNT,
+            "concurrency_per_gpu": $CONCURRENCY_PER_GPU,
+            "total_concurrency": $TOTAL_CONCURRENCY,
+            "mode": "$DEPLOYMENT_MODE",
+            "isl": $ISL,
+            "osl": $OSL,
+            "endpoint": "$ENDPOINT",
+            "model endpoint": "$TARGET_MODEL"
+          }
+          EOF
+
+          # Run perf with calculated total concurrency
+          run_perf $TOTAL_CONCURRENCY $ISL $OSL
+          echo "done with concurrency $TOTAL_CONCURRENCY"
+        env:
+        - name: TARGET_MODEL
+          value: openai/gpt-oss-120b
+        - name: ENDPOINT
+          value: gpt-oss-disagg-frontend:8000
+        - name: CONCURRENCY_PER_GPU
+          value: "256"
+        - name: DEPLOYMENT_GPU_COUNT
+          value: "6"
+        - name: ISL
+          value: "8192"
+        - name: OSL
+          value: "1024"
+        - name: DEPLOYMENT_MODE
+          value: disagg
+        - name: AIPERF_HTTP_CONNECTION_LIMIT
+          value: "252"
+        - name: JOB_NAME
+          valueFrom:
+            fieldRef:
+              apiVersion: v1
+              fieldPath: metadata.labels['job-name']
+        - name: ROOT_ARTIFACT_DIR
+          value: /model-cache/perf
+        - name: HF_HOME
+          value: /model-cache
+        - name: PYTHONUNBUFFERED
+          value: "1"
+        image: python:3.12-slim
+        imagePullPolicy: IfNotPresent
+        name: perf
+        securityContext:
+          privileged: true
+        volumeMounts:
+        - name: model-cache
+          mountPath: /model-cache
+        workingDir: /workspace
+      imagePullSecrets:
+      - name: nvcrimagepullsecret
+      restartPolicy: Never
+      volumes:
+      - name: model-cache
+        persistentVolumeClaim:
+          claimName: model-cache

recipes/gpt-oss-120b/trtllm/disagg/prefill.yaml

-# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
-# SPDX-License-Identifier: Apache-2.0
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-enable_attention_dp: false
-disable_overlap_scheduler: true
-moe_config:
-    backend: CUTLASS
-enable_chunked_prefill: true
-cuda_graph_config:
-    max_batch_size: 32
-    enable_padding: true
-cache_transceiver_config:
-  backend: UCX
-  max_tokens_in_buffer: 65536
-print_iter_log: false
-stream_interval: 10