add gptoss trt docker

Author: Ankur-singh

benchmarks/gptoss_fp4_b200_trt_docker.sh

@@ -0,0 +1,78 @@
+#!/usr/bin/env bash
+
+# === Required Env Vars === 
+# HF_TOKEN
+# HF_HUB_CACHE
+# IMAGE
+# MODEL
+# ISL
+# OSL
+# MAX_MODEL_LEN
+# RANDOM_RANGE_RATIO
+# TP
+# CONC
+# RESULT_FILENAME
+# PORT
+
+# GPTOSS TRTLLM Deployment Guide:
+# https://github.com/NVIDIA/TensorRT-LLM/blob/main/docs/source/deployment-guide/quick-start-recipe-for-gpt-oss-on-trtllm.md
+
+# ========= Determine DP_ATTENTION, EP_SIZE and MOE_BACKEND based on ISL, OSL, CONC =========
+EP_SIZE="1"
+MOE_BACKEND="TRTLLM"
+DP_ATTENTION=false
+
+# 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
+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: fp8
+    enable_block_reuse: false
+    free_gpu_memory_fraction: 0.85
+print_iter_log: true
+stream_interval: 20
+num_postprocess_workers: 4
+moe_config:
+    backend: $MOE_BACKEND
+EOF
+
+if [[ "$DP_ATTENTION" == "true" ]]; then
+    cat << EOF >> $EXTRA_CONFIG_FILE
+attention_dp_config:
+    enable_balance: true
+EOF
+fi
+
+echo "Generated config file contents:"
+cat $EXTRA_CONFIG_FILE
+
+set -x
+
+MAX_NUM_TOKENS=20000
+
+# Launch TRT-LLM server
+mpirun -n 1 --oversubscribe --allow-run-as-root \
+    trtllm-serve $MODEL --port=$PORT \
+    --trust_remote_code \
+    --backend=pytorch \
+    --max_batch_size 512 \
+    --max_seq_len=$MAX_MODEL_LEN \
+    --max_num_tokens=$MAX_NUM_TOKENS \
+    --tp_size=$TP --ep_size=$EP_SIZE \
+    --extra_llm_api_options=$EXTRA_CONFIG_FILE