Add benchmark script for GPTOSS FP4 B200 TRT-LLM (#256)

* Add benchmark script for GPTOSS FP4 B200 TRT-LLM

* make changes to perf changelog

---------

Co-authored-by: Cameron Quilici <cjquilici@gmail.com>

Author: 2 people

benchmarks/gptoss_fp4_b200_trt_docker.sh

@@ -0,0 +1,90 @@
+#!/usr/bin/env bash
+
+# === Required Env Vars === 
+# MODEL
+# PORT
+# TP
+# EP_SIZE
+# DP_ATTENTION
+# CONC
+# ISL
+# OSL
+# MAX_MODEL_LEN
+# RANDOM_RANGE_RATIO
+# NUM_PROMPTS
+# RESULT_FILENAME
+
+SERVER_LOG=$(mktemp /tmp/server-XXXXXX.log)
+
+# GPTOSS TRTLLM Deployment Guide:
+# https://github.com/NVIDIA/TensorRT-LLM/blob/main/docs/source/deployment-guide/quick-start-recipe-for-gpt-oss-on-trtllm.md
+
+MOE_BACKEND="TRTLLM"
+echo "MOE_BACKEND set to '$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:
+    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 > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+# Source benchmark utilities
+source "$(dirname "$0")/benchmark_lib.sh"
+
+# Wait for server to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+pip install -q datasets pandas
+
+run_benchmark_serving \
+    --model "$MODEL" \
+    --port "$PORT" \
+    --backend openai \
+    --input-len "$ISL" \
+    --output-len "$OSL" \
+    --random-range-ratio "$RANDOM_RANGE_RATIO" \
+    --num-prompts "$NUM_PROMPTS" \
+    --max-concurrency "$CONC" \
+    --result-filename "$RESULT_FILENAME" \
+    --result-dir /workspace/

perf-changelog.yaml

@@ -86,3 +86,8 @@
   description: |
     - Updating MI355x Deepseek-R1 FP4 SGLang Image to upstream v0.5.6.post1 
     PR: https://github.com/InferenceMAX/InferenceMAX/pull/330
+- config-keys:
+    - gptoss-fp4-b200-trt
+  description: |
+    - Add benchmark script for GPTOSS FP4 B200 TRT-LLM
+    PR: https://github.com/InferenceMAX/InferenceMAX/pull/256