[None] [chore] cherry pick changes on slurm scripts from release/1.1.0rc2 (#7750)

Signed-off-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com>

examples/disaggregated/slurm/benchmark/disaggr_torch.slurm

@@ -83,8 +83,6 @@ echo "ntasks_per_node: ${ntasks_per_node}"
 echo "==========================================="
 
 
-nsys_on=""
-# nsys_on=${full_logdir} # Uncomment this line to enable Nsys profiling
 numa_bind=true # Only allocate memory from nodes, this only works on GB200
 ctx_max_seq_len=$((isl + 10))
 gen_max_seq_len=$((isl + osl + 10))
@@ -96,6 +94,9 @@ logdir=${workdir}/slurm-${SLURM_JOB_ID}/benchmark-${isl}-${osl}
 mkdir -p ${logdir}
 full_logdir=${logdir}/ctx${num_ctx_servers}_gen${num_gen_servers}_dep${gen_tp_size}_batch${gen_batch_size}_eplb${eplb_num_slots}_mtp${mtp_size}
 
+nsys_on=""
+# nsys_on=${full_logdir} # Uncomment this line to enable Nsys profiling
+
 echo "concurrency: ${concurrency}"
 
 ctx_gpus=$((num_ctx_servers * ctx_tp_size * ctx_pp_size))

examples/disaggregated/slurm/benchmark/gen_worker_config.py

@@ -48,6 +48,11 @@ def gen_config_file(work_dir: str,
         server_port: Server port
     """
     ctx_config = {
+        'build_config': {
+            'max_batch_size': ctx_batch_size,
+            'max_num_tokens': ctx_max_num_tokens,
+            'max_seq_len': ctx_max_seq_len,
+        },
         'max_batch_size': ctx_batch_size,
         'max_num_tokens': ctx_max_num_tokens,
         'max_seq_len': ctx_max_seq_len,
@@ -79,6 +84,11 @@ def gen_config_file(work_dir: str,
         gen_moe_backend = "TRTLLM"
 
     gen_config = {
+        'build_config': {
+            'max_batch_size': gen_batch_size,
+            'max_num_tokens': gen_max_num_tokens,
+            'max_seq_len': gen_max_seq_len,
+        },
         'tensor_parallel_size': gen_tp_size,
         'moe_expert_parallel_size': gen_tp_size,
         'enable_attention_dp': True if gen_enable_attention_dp else False,

examples/disaggregated/slurm/benchmark/start_worker.sh

@@ -65,14 +65,14 @@ else
     nsys_file=${nsys_folder}/nsys_worker_proc_${instance_id}_${SLURM_PROCID}
     export TLLM_PROFILE_RECORD_GC=1
     export TLLM_NVTX_DEBUG=1
-    if [ "${role}" = "GEN" ]; then
+    if [ "${role}" = "GEN" ] && [ "$SLURM_PROCID" = "0" ]; then
         export TLLM_PROFILE_START_STOP=200-250
         nsys_prefix="nsys profile -e \"NSYS_MPI_STORE_TEAMS_PER_RANK=1\" -o ${nsys_file} -f true -t cuda,nvtx,python-gil -c cudaProfilerApi --cuda-graph-trace node --capture-range-end=stop --gpu-metrics-devices=none"
         echo "nsys_prefix: ${nsys_prefix}"
     elif [ "${role}" = "CTX" ]; then
         echo "nsys is not enabled on ctx_gpus"
     fi
-    trtllm-llmapi-launch ${numa_bind_cmd} ${nsys_prefix} \
+    ${nsys_prefix} trtllm-llmapi-launch ${numa_bind_cmd} \
         trtllm-serve ${model_path} \
             --host $(hostname) --port ${port} \
             --extra_llm_api_options ${config_file}

tensorrt_llm/serve/scripts/benchmark_dataset.py

@@ -494,11 +494,14 @@ class RandomDataset(BenchmarkDataset):
 
             # Filter out sequences that are too long or too short
             requests = []
-            for prompt, initial_prompt_len, cached_token_ids in zip(
-                    dataset, prompt_lengths, prompt_token_ids):
-                i = len(requests)
-                if i == num_requests:
-                    break
+            dataset_len = len(dataset)
+
+            for i in range(num_requests):
+                # Use modulo to cycle through the dataset when num_requests > dataset_len
+                dataset_idx = i % dataset_len
+                prompt = dataset[dataset_idx]
+                initial_prompt_len = prompt_lengths[dataset_idx]
+                cached_token_ids = prompt_token_ids[dataset_idx]
 
                 # Skip empty prompt
                 if initial_prompt_len == 0:
@@ -534,9 +537,6 @@ class RandomDataset(BenchmarkDataset):
                         prompt_len=total_input_len,
                         expected_output_len=int(output_lens[i]),
                     ))
-            assert len(requests) == num_requests, (
-                f"Only {len(requests)} requests sampled from sharegpt dataset, {num_requests} requests are needed"
-            )
         else:
             for i in range(num_requests):
                 inner_seq = ((offsets[i] + i + np.arange(input_lens[i])) %
@@ -1131,6 +1131,7 @@ class VisionArenaDataset(HuggingFaceDataset):
         if parser_fn is None:
             raise ValueError(f"Unsupported dataset path: {self.dataset_path}")
 
+        sampled_requests = []
         for item in self.data:
             if len(prompts) >= num_requests:
                 break