[NVIDIA] B200 TRT Update GPT-OSS configs

.github/workflows/gptoss-tmpl.yml

@@ -100,8 +100,8 @@ jobs:
       osl: ${{ inputs.osl }}
       max-model-len: ${{ inputs.max-model-len }}
       random-range-ratio: ${{ inputs.random-range-ratio }}
-      runner: b200-trt
-      image: 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2.post2'
+      runner: b200-nvs
+      image: 'nvcr.io#nvidia/tensorrt-llm/release:1.2.0rc0.post1'
       model: 'openai/gpt-oss-120b'
       tp-list: '[1, 2, 4, 8]'
       framework: 'trt'

.github/workflows/runner-sweep-test.yml

@@ -30,6 +30,7 @@ on:
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2.post1'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2.post2'
+          - 'nvcr.io#nvidia/tensorrt-llm/release:1.2.0rc0.post1'
           - 'nvcr.io#nvidia/tensorrt-llm/release:gpt-oss-dev'
           - 'nvcr.io#nvidia/ai-dynamo/tensorrtllm-runtime:0.5.1-rc0.pre3'
           - 'rocm/7.0:rocm7.0_ubuntu_22.04_sgl-dev-v0.5.2-rocm7.0-mi30x-20250915'

.github/workflows/runner-test.yml

@@ -60,6 +60,7 @@ on:
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2.post1'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2.post2'
+          - 'nvcr.io#nvidia/tensorrt-llm/release:1.2.0rc0.post1'
           - 'nvcr.io#nvidia/tensorrt-llm/release:gpt-oss-dev'
           - 'rocm/7.0-preview:rocm7.0_preview_ubuntu_22.04_vllm_0.10.1_instinct_rc1'
           - 'rocm/7.0-preview:rocm7.0_preview_ubuntu_22.04_sgl-dev-v0.5.2rc2-mi30x_rc1'

benchmarks/gptoss_fp4_b200_trt_slurm.sh

@@ -30,42 +30,27 @@ EP_SIZE="1"
 MOE_BACKEND="TRTLLM"
 DP_ATTENTION=false
 
-# Lower concurrencies: Concurrency < 256
-# MoE backend=TRTLLM
-# Use TP Attention; Switch to MoE Expert parallel for conurrency >=16 (1k1k and 1k8k)
-TEP_REQUIRED=false
-if [[ "$TP" == "4" || "$TP" == "8" ]]; then 
-    if [[ "$ISL" == "1024" && "$OSL" == "1024" ]]; then
-        TEP_REQUIRED=true
-    elif [[ "$ISL" == "1024" && "$OSL" == "8192" ]]; then
-        TEP_REQUIRED=true
-    fi
-fi
-if [[ "$TEP_REQUIRED" == "true" && $CONC -ge 16 ]]; then
-    EP_SIZE="$TP"
-fi
-
 # Higher concurrencies: Concurrency >= 256
 #   MoE Backend = CUTLASS
 #   Use DP attention with expert parallel MoE
 if [[ $CONC -ge 256 ]]; then
     EP_SIZE="$TP"
     DP_ATTENTION=true
-    MOE_BACKEND="CUTLASS"
 fi
 
 echo "Final configuration: EP_SIZE='$EP_SIZE', MOE_BACKEND='$MOE_BACKEND', DP_ATTENTION='$DP_ATTENTION'"
 
 EXTRA_CONFIG_FILE="gptoss-fp4.yml"
 export TRTLLM_ENABLE_PDL=1
+export NCCL_GRAPH_REGISTER=0
 
 cat > $EXTRA_CONFIG_FILE << EOF
 cuda_graph_config:
     enable_padding: true
     max_batch_size: $CONC
 enable_attention_dp: $DP_ATTENTION
 kv_cache_config:
-    dtype: auto
+    dtype: fp8
     enable_block_reuse: false
     free_gpu_memory_fraction: 0.85
 print_iter_log: true
@@ -105,12 +90,6 @@ mpirun -n 1 --oversubscribe --allow-run-as-root \
 set +x
 while IFS= read -r line; do
     printf '%s\n' "$line"
-    if [[ "$line" =~ [Ee][Rr][Rr][Oo][Rr] ]]; then
-        sleep 5
-        tail -n100 $SERVER_LOG
-        echo "JOB $SLURM_JOB_ID ran on NODE $SLURMD_NODENAME"
-        exit 1
-    fi
     if [[ "$line" == *"Application startup complete"* ]]; then
         break
     fi
@@ -127,4 +106,4 @@ python3 bench_serving/benchmark_serving.py \
 --request-rate inf --ignore-eos \
 --save-result --percentile-metrics 'ttft,tpot,itl,e2el' \
 --result-dir /workspace/ \
---result-filename $RESULT_FILENAME.json
+--result-filename $RESULT_FILENAME.json