Use all 8 GPUs for A4 TRTLLM DeepSeek-R1 recipe; Adjust values.yaml

Author: hmhv1222

inference/a4/single-host-serving/tensorrt-llm/values.yaml

@@ -37,7 +37,7 @@ service:
 workload:
   model:
     name:
-  gpus: 4
+  gpus: 8 # Adjustable
   image:
   framework:
   configFile: serving-args.yaml
@@ -55,9 +55,9 @@ workload:
       value: "/usr/local/nvidia/lib64:/usr/local/lib/"
   benchmarks:
     experiments:
-      - isl: 1024 # Input sequence length
-        osl: 4096 # Output sequence length
-        num_requests: 1000
+      - isl: 1024 # Input sequence length, adjustable
+        osl: 4096 # Output sequence length, adjustable
+        num_requests: 1000 # Adjustable
 
 network:
   subnetworks[]:

inference/a4x/single-host-serving/tensorrt-llm/values.yaml

@@ -37,7 +37,7 @@ service:
 workload:
   model:
     name:
-  gpus: 4
+  gpus: 4 # Adjustable
   image:
   framework:
   configFile: serving-args.yaml
@@ -51,9 +51,9 @@ workload:
       value: "/workload/configs/serving-args.yaml"
   benchmarks:
     experiments:
-      - isl: 1024 # Input sequence length
-        osl: 4096 # Output sequence length
-        num_requests: 1000
+      - isl: 1024 # Input sequence length, adjustable
+        osl: 4096 # Output sequence length, adjustable
+        num_requests: 1000 # Adjustable
 
 network:
   subnetworks[]:

src/frameworks/a4/trtllm-configs/deepseek-r1-nvfp4.yaml

@@ -1,5 +1,5 @@
-tp_size: 4
-ep_size: 4
+tp_size: 8
+ep_size: 8
 pp_size: 1
 backend: pytorch
 kv_cache_free_gpu_mem_fraction: 0.85