reduce ROCm gpu memory gpu_utilization

dtrifiro · dtrifiro · commit 736fd4769159 · 2025-05-13T11:27:44.000+02:00
diff --git a/Qwen/Qwen2.5-7B-Instruct/accuracy/server-rocm.yml b/Qwen/Qwen2.5-7B-Instruct/accuracy/server-rocm.yml
@@ -0,0 +1 @@
+gpu_memory_utilization: 0.6
diff --git a/RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8-dynamic/accuracy/server-rocm.yml b/RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8-dynamic/accuracy/server-rocm.yml
@@ -0,0 +1 @@
+gpu_memory_utilization: 0.6
diff --git a/RedHatAI/Mistral-Small-3.1-24B-Instruct-2503-FP8-dynamic/accuracy/server-rocm.yml b/RedHatAI/Mistral-Small-3.1-24B-Instruct-2503-FP8-dynamic/accuracy/server-rocm.yml
@@ -0,0 +1 @@
+gpu_memory_utilization: 0.6
diff --git a/RedHatAI/Mistral-Small-3.1-24B-Instruct-2503-quantized.w8a8/accuracy/server-rocm.yml b/RedHatAI/Mistral-Small-3.1-24B-Instruct-2503-quantized.w8a8/accuracy/server-rocm.yml
@@ -0,0 +1 @@
+gpu_memory_utilization: 0.6
diff --git a/RedHatAI/Qwen2.5-7B-Instruct-FP8-dynamic/accuracy/server-rocm.yml b/RedHatAI/Qwen2.5-7B-Instruct-FP8-dynamic/accuracy/server-rocm.yml
@@ -0,0 +1 @@
+gpu_memory_utilization: 0.6
diff --git a/RedHatAI/phi-4-FP8-dynamic/accuracy/server-rocm.yml b/RedHatAI/phi-4-FP8-dynamic/accuracy/server-rocm.yml
@@ -1 +1 @@
-gpu_memory_utilization: 0.8
+gpu_memory_utilization: 0.6
diff --git a/ibm-granite/granite-3.1-8b-instruct/accuracy/server-rocm.yml b/ibm-granite/granite-3.1-8b-instruct/accuracy/server-rocm.yml
@@ -0,0 +1 @@
+gpu_memory_utilization: 0.6
diff --git a/mistralai/Mistral-Small-24B-Instruct-2501/accuracy/server-rocm.yml b/mistralai/Mistral-Small-24B-Instruct-2501/accuracy/server-rocm.yml
@@ -1 +1 @@
-gpu_memory_utilization: 0.8
+gpu_memory_utilization: 0.6
diff --git a/mistralai/Mistral-Small-3.1-24B-Instruct-2503/accuracy/server-rocm.yml b/mistralai/Mistral-Small-3.1-24B-Instruct-2503/accuracy/server-rocm.yml
@@ -1 +1 @@
-gpu_memory_utilization: 0.8
+gpu_memory_utilization: 0.6

Original file line number	Diff line number	Diff line change
`@@ -1 +1 @@`
`1`		`-gpu_memory_utilization: 0.8`
	`1`	`+gpu_memory_utilization: 0.6`