chore: update doc by replacing use_cuda_graph with cuda_graph_config (#5680)

Signed-off-by: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com>

docs/source/blogs/Best_perf_practice_on_DeepSeek-R1_in_TensorRT-LLM.md

@@ -137,7 +137,7 @@ To do the benchmark, run the following command:
 YOUR_DATA_PATH=<your dataset file following the format>
 
 cat >./extra-llm-api-config.yml<<EOF
-use_cuda_graph: true
+cuda_graph_config: {}
 moe_backend: TRTLLM
 speculative_config:
     decoding_type: MTP
@@ -317,7 +317,7 @@ To do the benchmark, run the following command:
 YOUR_DATA_PATH=<your dataset file following the format>
 
 cat >./extra-llm-api-config.yml<<EOF
-use_cuda_graph: true
+cuda_graph_config: {}
 speculative_config:
     decoding_type: MTP
     num_nextn_predict_layers: 3
@@ -366,9 +366,9 @@ python ${YOUR_WORK_PATH}/benchmarks/cpp/prepare_dataset.py \
 YOUR_DATA_PATH=./dataset.txt
 
 cat >./extra-llm-api-config.yml<<EOF
-use_cuda_graph: true
-cuda_graph_batch_sizes:
-- 128
+cuda_graph_config:
+  batch_sizes:
+  - 128
 enable_attention_dp: true
 EOF
 

docs/source/blogs/tech_blog/blog2_DeepSeek_R1_MTP_Implementation_and_Optimization.md

@@ -123,7 +123,7 @@ To benchmark min-latency performance with MTP, you need to follow [this document
 YOUR_DATA_PATH=<your dataset file following the format>
 
 cat >./extra-llm-api-config.yml<<EOF
-use_cuda_graph: true
+cuda_graph_config: {}
 moe_backend: TRTLLM
 speculative_config:
     decoding_type: MTP
@@ -178,7 +178,7 @@ To benchmark min-latency performance with MTP Relaxed Acceptance, you need to fo
 YOUR_DATA_PATH=<your dataset file following the format>
 
 cat >./extra-llm-api-config.yml<<EOF
-use_cuda_graph: true
+cuda_graph_config: {}
 moe_backend: TRTLLM
 speculative_config:
     decoding_type: MTP

docs/source/blogs/tech_blog/blog4_Scaling_Expert_Parallelism_in_TensorRT-LLM.md

@@ -541,7 +541,7 @@ Run 32-way expert parallelism inference on the prepared dataset. Please refer to
 ```bash
 cat > ./extra_llm_api_options.yaml <<EOF
 enable_attention_dp: true
-use_cuda_graph: true
+cuda_graph_config: {}
 EOF
 
 trtllm-llmapi-launch \
@@ -622,7 +622,7 @@ Run 36-way expert parallelism inference with the EPLB configuration incorporated
 ```bash
 cat > ./extra_llm_api_options_eplb.yaml <<EOF
 enable_attention_dp: true
-use_cuda_graph: true
+cuda_graph_config: {}
 moe_load_balancer: ./moe_load_balancer.yaml
 EOF
 

docs/source/scripts/disaggregated/gen_yaml.py

@@ -189,9 +189,10 @@ def gen_config_file(config_path: str,
             'max_num_tokens': gen_max_num_tokens,
             'max_seq_len': 8576,
             'free_gpu_memory_fraction': gen_gpu_memory_fraction,
-            'use_cuda_graph': True,
-            'cuda_graph_padding_enabled': True,
-            'cuda_graph_batch_sizes': gen_cuda_graph_batch_sizes,
+            'cuda_graph_config': {
+                'padding_enabled': True,
+                'batch_sizes': gen_cuda_graph_batch_sizes,
+            },
             'print_iter_log': True,
             'kv_cache_dtype': 'fp8',
             'moe_backend': 'TRTLLM',

examples/disaggregated/README.md

@@ -58,12 +58,11 @@ since `TP` and `PP` is 1 for the two context and one generation server.
 The `disagg_config.yaml` file must now contain the configuration parameters of the context and generation servers. For example,
 it could look like:
 
-```
+```yaml
 hostname: localhost
 port: 8000
 model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
 backend: "pytorch"
-use_cuda_graph: False
 disable_overlap_scheduler: True
 context_servers:
   num_instances: 2

examples/ep_load_balancer/README.md

@@ -26,7 +26,7 @@ Run 32-way expert parallelism inference on the prepared dataset. Please refer to
 ```bash
 cat > ./extra_llm_api_options.yaml <<EOF
 enable_attention_dp: true
-use_cuda_graph: true
+cuda_graph_config: {}
 moe_backend: WideEP
 moe_max_num_tokens: 8192
 EOF
@@ -116,7 +116,7 @@ Run 36-way expert parallelism inference with the EPLB configuration incorporated
 ```bash
 cat > ./extra_llm_api_options_eplb.yaml <<EOF
 enable_attention_dp: true
-use_cuda_graph: true
+cuda_graph_config: {}
 moe_backend: WideEP
 moe_max_num_tokens: 9216
 moe_load_balancer: ./moe_load_balancer.yaml
@@ -182,7 +182,7 @@ Run 36-way expert parallelism inference with the EPLB configuration incorporated
 ```bash
 cat > ./extra_llm_api_options_eplb.yaml <<EOF
 enable_attention_dp: true
-use_cuda_graph: true
+cuda_graph_config: {}
 moe_backend: WideEP
 moe_max_num_tokens: 9216
 moe_load_balancer: ./moe_load_balancer.yaml

examples/llm-eval/lm-eval-harness/lm_eval_tensorrt_llm.py

@@ -96,7 +96,7 @@ class TRTLLMEvalBase(TemplateLM):
                 tp = torch.cuda.device_count()
 
             pytorch_config_params = {
-                "use_cuda_graph": use_cuda_graph,
+                'cuda_graph_config': {} if use_cuda_graph else None,
                 "print_iter_log": False,
             }
             if hasattr(PyTorchConfig, "moe_backend"):

examples/models/core/deepseek_v3/README.md

@@ -192,7 +192,7 @@ Evaluate the model accuracy using `trtllm-eval`.
 1. (Optional) Prepare an advanced configuration file:
 ```bash
 cat >./extra-llm-api-config.yml <<EOF
-use_cuda_graph: true
+cuda_graph_config: {}
 enable_attention_dp: true
 EOF
 ```

examples/models/core/qwen/README.md

@@ -692,7 +692,7 @@ concurrency=128
 path_data=./aa_prompt_isl_1k_osl_2k_qwen3_10000samples.txt
 
 # Setup the extra configuration for llm-api
-echo -e "disable_overlap_scheduler: false\nuse_cuda_graph: true\nprint_iter_log: true\ncuda_graph_batch_sizes: [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,32,64,128]\nenable_attention_dp: true " > ${path_config}
+echo -e "disable_overlap_scheduler: false\cuda_graph_config: {}\nprint_iter_log: true\ncuda_graph_batch_sizes: [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,32,64,128]\nenable_attention_dp: true " > ${path_config}
 
 # Run trtllm-bench with pytorch backend
 mpirun --allow-run-as-root --oversubscribe -n 1 \