DSv4 B300 TRT

.github/configs/nvidia-master.yaml

@@ -2682,6 +2682,31 @@ dsv4-fp4-b300-vllm:
       - { tp: 4, ep: 4, dp-attn: true, conc-start: 256, conc-end: 512 }
       - { tp: 8, ep: 8, dp-attn: true, conc-start: 2048, conc-end: 2048 }
 
+dsv4-fp4-b300-trt:
+  image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-4999884
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: b300
+  precision: fp4
+  framework: trt
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 4, conc-start: 1, conc-end: 64 }
+      - { tp: 4, ep: 4, dp-attn: true, conc-start: 64, conc-end: 256 }
+      - { tp: 8, ep: 8, dp-attn: true, conc-start: 64, conc-end: 512 }
+      - { tp: 8, conc-start: 1, conc-end: 32 }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 4, conc-start: 1, conc-end: 64 }
+      - { tp: 4, ep: 4, dp-attn: true, conc-start: 64, conc-end: 256 }
+      - { tp: 8, ep: 8, dp-attn: true, conc-start: 64, conc-end: 512 }
+      - { tp: 8, conc-start: 1, conc-end: 32 }
+
 dsv4-fp4-b300-vllm-mtp:
   image: vllm/vllm-openai:v0.20.0-cu130
   model: deepseek-ai/DeepSeek-V4-Pro

.github/workflows/benchmark-tmpl.yml

@@ -164,12 +164,13 @@ jobs:
           clean: true
           submodules: true
 
-      - name: Cleanup stale eval outputs (pre-run)
-        if: ${{ inputs.run-eval || inputs.eval-only }}
+      - name: Cleanup stale outputs (pre-run)
         run: |
           rm -f meta_env.json || true
           rm -f results*.json || true
           rm -f sample*.jsonl || true
+          rm -f server.log || true
+          rm -f gpu_metrics.csv || true
 
       - name: Launch job script
         env:

benchmarks/benchmark_lib.sh

@@ -165,11 +165,12 @@ wait_for_server_ready() {
 }
 
 # Run benchmark serving with standardized parameters
-# All parameters are required except --use-chat-template, --dsv4, and --trust-remote-code
+# All parameters are required except --endpoint, --use-chat-template, --dsv4, and --trust-remote-code
 # Parameters:
 #   --model: Model name
 #   --port: Server port
 #   --backend: Backend type - e.g., 'vllm' or 'openai'
+#   --endpoint: Optional API endpoint override
 #   --input-len: Random input sequence length
 #   --output-len: Random output sequence length
 #   --random-range-ratio: Random range ratio
@@ -194,6 +195,7 @@ run_benchmark_serving() {
     local model=""
     local port=""
     local backend=""
+    local endpoint=""
     local input_len=""
     local output_len=""
     local random_range_ratio=""
@@ -221,6 +223,10 @@ run_benchmark_serving() {
                 backend="$2"
                 shift 2
                 ;;
+            --endpoint)
+                endpoint="$2"
+                shift 2
+                ;;
             --input-len)
                 input_len="$2"
                 shift 2
@@ -356,6 +362,10 @@ run_benchmark_serving() {
         --result-dir "$result_dir"
         --result-filename "$result_filename.json"
     )
+
+    if [[ -n "$endpoint" ]]; then
+        benchmark_cmd+=(--endpoint "$endpoint")
+    fi
 
     # Add --use-chat-template if requested
     if [[ "$use_chat_template" == true ]]; then

benchmarks/single_node/dsv4_fp4_b300_trt.sh

@@ -0,0 +1,163 @@
+#!/usr/bin/env bash
+
+# DeepSeek-V4-Pro single-node TRTLLM recipe for B300. The configured image
+# already contains NVIDIA/TensorRT-LLM@feat/deepseek_v4; do not build TRTLLM at
+# runtime from this benchmark path.
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME \
+    DP_ATTENTION \
+    EP_SIZE
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+echo "TP: $TP, CONC: $CONC, ISL: $ISL, OSL: $OSL, EP_SIZE: $EP_SIZE, DP_ATTENTION: $DP_ATTENTION"
+
+export TRTLLM_DSV4_USE_MPIRUN="${TRTLLM_DSV4_USE_MPIRUN:-1}"
+export TRTLLM_DSV4_SANITIZE_SLURM_MPI_ENV="${TRTLLM_DSV4_SANITIZE_SLURM_MPI_ENV:-1}"
+
+sanitize_slurm_mpi_env_for_trtllm() {
+    if [[ "${TRTLLM_DSV4_SANITIZE_SLURM_MPI_ENV:-0}" != "1" ]]; then
+        return 0
+    fi
+
+    echo "Sanitizing Slurm/PMI environment for TensorRT-LLM launch"
+    while IFS='=' read -r name _; do
+        case "$name" in
+            SLURM_*|PMIX*|PMI*|OMPI_*|ORTE_*)
+                unset "$name"
+                ;;
+        esac
+    done < <(env)
+}
+
+sanitize_slurm_mpi_env_for_trtllm
+
+export NCCL_NVLS_ENABLE="${NCCL_NVLS_ENABLE:-0}"
+echo "NCCL_NVLS_ENABLE: $NCCL_NVLS_ENABLE"
+
+if [[ "$MODEL" != /* ]]; then
+    hf download "$MODEL"
+fi
+
+nvidia-smi
+
+SERVER_LOG="$PWD/server.log"
+PORT=${PORT:-8888}
+EXTRA_CONFIG_FILE="dsv4-fp4-trt.yml"
+
+MOE_BACKEND="TRTLLM"
+MAX_BATCH_SIZE=$(( CONC > 16 ? CONC : 16 ))
+CUDA_GRAPH_MAX_BATCH_SIZE="$MAX_BATCH_SIZE"
+KV_CACHE_FREE_MEM_FRACTION="${KV_CACHE_FREE_MEM_FRACTION:-0.50}"
+
+ATTENTION_DP_CONFIG=""
+if [[ "$DP_ATTENTION" == "true" ]]; then
+    ATTENTION_DP_CONFIG="
+attention_dp_config:
+    batching_wait_iters: 0
+    enable_balance: true
+    timeout_iters: 60"
+fi
+
+cat > "$EXTRA_CONFIG_FILE" << EOF
+cuda_graph_config:
+    enable_padding: true
+    max_batch_size: $CUDA_GRAPH_MAX_BATCH_SIZE
+enable_attention_dp: $DP_ATTENTION$ATTENTION_DP_CONFIG
+print_iter_log: true
+kv_cache_config:
+    tokens_per_block: 128
+    dtype: fp8
+    free_gpu_memory_fraction: $KV_CACHE_FREE_MEM_FRACTION
+    enable_block_reuse: false
+stream_interval: 10
+num_postprocess_workers: 4
+moe_config:
+    backend: $MOE_BACKEND
+EOF
+
+echo "Generated config file contents:"
+cat "$EXTRA_CONFIG_FILE"
+
+MAX_MODEL_LEN=$(( MAX_MODEL_LEN > 8192 ? MAX_MODEL_LEN : 8192 ))
+MAX_NUM_TOKENS=$(( ISL + OSL + 256 ))
+MAX_NUM_TOKENS=$(( MAX_NUM_TOKENS > 8192 ? MAX_NUM_TOKENS : 8192 ))
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+    MAX_NUM_TOKENS="$EVAL_MAX_MODEL_LEN"
+fi
+
+if [[ "${RUN_EVAL:-false}" == "true" || "${EVAL_ONLY:-false}" == "true" ]]; then
+    # DeepSeek-V4-Pro has hidden size 7168. The current TRTLLM fused-HC MHC
+    # path corrupts eval generations for this shape; keep eval servers on the
+    # unfused path until the fused kernel is guarded or supports 7168.
+    export TRTLLM_MHC_ENABLE_FUSED_HC="${TRTLLM_MHC_ENABLE_FUSED_HC:-0}"
+    echo "TRTLLM_MHC_ENABLE_FUSED_HC: $TRTLLM_MHC_ENABLE_FUSED_HC"
+fi
+
+start_gpu_monitor --output "$PWD/gpu_metrics.csv"
+
+set -x
+SERVE_CMD=(
+    trtllm-serve "$MODEL" \
+    --host 0.0.0.0 \
+    --port "$PORT" \
+    --trust_remote_code \
+    --backend pytorch \
+    --max_batch_size "$MAX_BATCH_SIZE" \
+    --max_seq_len "$MAX_MODEL_LEN" \
+    --max_num_tokens "$MAX_NUM_TOKENS" \
+    --tp_size "$TP" \
+    --ep_size "$EP_SIZE" \
+    --custom_tokenizer deepseek_v4 \
+    --config "$EXTRA_CONFIG_FILE"
+)
+
+if [[ "${TRTLLM_DSV4_USE_MPIRUN:-1}" == "0" ]]; then
+    "${SERVE_CMD[@]}" > "$SERVER_LOG" 2>&1 &
+else
+    mpirun -n 1 --oversubscribe --allow-run-as-root \
+        "${SERVE_CMD[@]}" \
+        > "$SERVER_LOG" 2>&1 &
+fi
+
+SERVER_PID=$!
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+run_benchmark_serving \
+    --model "$MODEL" \
+    --port "$PORT" \
+    --backend openai-chat \
+    --endpoint /v1/chat/completions \
+    --input-len "$ISL" \
+    --output-len "$OSL" \
+    --random-range-ratio "$RANDOM_RANGE_RATIO" \
+    --num-prompts "$(( CONC * 10 ))" \
+    --max-concurrency "$CONC" \
+    --result-filename "$RESULT_FILENAME" \
+    --result-dir "$PWD/" \
+    --trust-remote-code \
+    --server-pid "$SERVER_PID"
+
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+stop_gpu_monitor
+set +x

perf-changelog.yaml

@@ -1765,7 +1765,7 @@
     - "TP=2 and TP=4, concurrency 4-256 for 1k1k and 8k1k sequence lengths"
     - "Add --gpu-memory-utilization 0.9 to server launch"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1133
-  
+
 - config-keys:
     - dsv4-fp8-h200-vllm
   description:
@@ -1886,7 +1886,7 @@
     - "Image pinned to lmsysorg/sglang:deepseek-v4-b300@sha256:26e116bd211e300dbb76924d56c5cbe6cc3ee5ee2fe314859cb8774f5bc070f3"
     - "DP-attention path enables SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN=1 for better SWA eviction behavior"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1185
-  
+
 - config-keys:
     - dsv4-fp4-b200-sglang
   description:
@@ -2122,3 +2122,11 @@
     - "run_benchmark_serving uses --dsv4 (chat-formatted prompts) per the AGENTS.md MTP rule, since EAGLE acceptance regresses on raw random tokens"
     - "Search space mirrors the non-MTP H200 SGLang entry: TP=8 EP=1, conc 1 and 4-64 for both 1k1k and 8k1k, with spec-decoding: mtp"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1265
+
+- config-keys:
+    - dsv4-fp4-b300-trt
+  description:
+    - "Add B300 TensorRT-LLM DeepSeek-V4-Pro eval coverage using the feat/deepseek_v4 image"
+    - "Disable TRTLLM fused MHC hyper-connection for eval servers via TRTLLM_MHC_ENABLE_FUSED_HC=0 because the current fused kernel corrupts DeepSeek-V4-Pro hidden size 7168 generations"
+    - "Keep this as eval-only PR validation until the TensorRT-LLM fused MHC kernel is guarded or supports hidden size 7168"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1233

runners/launch_b300-nv.sh

@@ -276,6 +276,7 @@ else
         LEGACY_FW_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
         BENCH_SCRIPT="${BENCH_BASE}${LEGACY_FW_SUFFIX}${SPEC_SUFFIX}.sh"
     fi
+
     LOCK_FILE="${SQUASH_FILE}.lock"
 
     # TODO(Cam): the deepseek-v4 sglang images (lmsysorg/sglang:deepseek-v4-blackwell
@@ -310,6 +311,7 @@ else
     JOB_ID=$(squeue --name="$RUNNER_NAME" -u "$USER" -h -o %A | head -n1)
 
     srun --jobid=$JOB_ID \
+        --mpi=none \
         --container-image=$SQUASH_FILE \
         --container-mounts=$GITHUB_WORKSPACE:$CONTAINER_MOUNT_DIR,$HF_HUB_CACHE_MOUNT:$HF_HUB_CACHE_MOUNT \
         --no-container-mount-home \

utils/bench_serving/backend_request_func.py

@@ -341,8 +341,9 @@ async def async_request_openai_chat_completions(
 
     async with aiohttp.ClientSession(trust_env=True,
                                      timeout=AIOHTTP_TIMEOUT) as session:
-        content = [{"type": "text", "text": request_func_input.prompt}]
+        content = request_func_input.prompt
         if request_func_input.multi_modal_content:
+            content = [{"type": "text", "text": request_func_input.prompt}]
             content.append(request_func_input.multi_modal_content)
         payload = {
             "model": request_func_input.model_name \