Merge pull request #37 from arjunsuresh/mlperf-inference

Fixes for mlperf inference intel bert
GATEOverflow · May 30, 2024 · 14f84d3 · 14f84d3
2 parents cc66235 + c616358
commit 14f84d3
Show file tree

Hide file tree

Showing 3 changed files with 15 additions and 3 deletions.
diff --git a/script/app-mlperf-inference-intel/_cm.yaml b/script/app-mlperf-inference-intel/_cm.yaml
@@ -173,6 +173,7 @@ variations:
         inference-results
       version: v4.0
   v3.1:
+    group: version
     env:
       CM_MLPERF_INFERENCE_CODE_VERSION: "v3.1"
     adr:
@@ -184,6 +185,7 @@ variations:
         inference-results
       version: v3.1
 
+
   # Target devices
   cpu:
     group: device
@@ -685,9 +687,16 @@ variations:
     default_env:
       CM_MLPERF_LOADGEN_BATCH_SIZE: 1
 
-  sapphire-rapids.24c,bert-99:
+  sapphire-rapids.24c,bert_:
     env:
       WORKERS_PER_PROC: 1
+  sapphire-rapids.112c,bert_,offline:
+    env:
+      WORKERS_PER_PROC: 4
+  sapphire-rapids.112c,bert_,server:
+    env:
+      WORKERS_PER_PROC: 8
+
 
 docker:
-  docker_real_run: False
+  real_run: False
diff --git a/script/app-mlperf-inference-intel/run_bert_harness.sh b/script/app-mlperf-inference-intel/run_bert_harness.sh
@@ -1,6 +1,7 @@
 #!/bin/bash
 
-THREADS_PER_INSTANCE=$(((4 * ${CM_HOST_CPU_THREADS_PER_CORE}) / ${CM_HOST_CPU_SOCKETS}))
+WORKERS_PER_PROC=${WORKERS_PER_PROC:-4}
+THREADS_PER_INSTANCE=$((( ${WORKERS_PER_PROC} * ${CM_HOST_CPU_THREADS_PER_CORE}) / ${CM_HOST_CPU_SOCKETS}))
 
 export LD_PRELOAD=${CONDA_PREFIX}/lib/libjemalloc.so
 export MALLOC_CONF="oversize_threshold:1,background_thread:true,percpu_arena:percpu,metadata_thp:always,dirty_decay_ms:9000000000,muzzy_decay_ms:9000000000";

diff --git a/script/app-mlperf-inference/_cm.yaml b/script/app-mlperf-inference/_cm.yaml
@@ -1152,6 +1152,8 @@ variations:
       nvidia-inference-server:
         version: r3.1
         tags: _ctuning
+      intel-harness:
+        tags: _v3.1
     default_env:
       CM_SKIP_SYS_UTILS: 'yes'
       CM_REGENERATE_MEASURE_FILES: 'yes'