#0: Fix N150 llama3 demo CI tests to properly save perf information t…

…o dashboard
tenstorrent · Nov 19, 2024 · 128daa3 · 128daa3
1 parent 336982c
commit 128daa3
Showing 1 changed file with 21 additions and 15 deletions.
diff --git a/tests/scripts/single_card/run_single_card_demo_tests.sh b/tests/scripts/single_card/run_single_card_demo_tests.sh
@@ -42,21 +42,6 @@ run_common_perf_tests(){
   # Skipping: kills WH cards, check issue #14440
   # WH_ARCH_YAML=wormhole_b0_80_arch_eth_dispatch.yaml pytest -n auto models/demos/wormhole/mistral7b/demo/demo_with_prefill.py --timeout 420; fail+=$?
 
-  # Llama3.1-8B
-  llama8b=/mnt/MLPerf/tt_dnn-models/llama/Meta-Llama-3.1-8B-Instruct/
-  # Llama3.2-1B
-  llama1b=/mnt/MLPerf/tt_dnn-models/llama/Llama3.2-1B-Instruct/
-  # Llama3.2-3B
-  llama3b=/mnt/MLPerf/tt_dnn-models/llama/Llama3.2-3B-Instruct/
-  # Llama3.2-11B
-  llama11b=/mnt/MLPerf/tt_dnn-models/llama/Llama3.2-11B-Vision-Instruct/
-
-  # Run all Llama3 tests for 8B, 1B, and 3B weights
-  for llama_dir in "$llama1b" "$llama3b" "$llama8b" "$llama11b"; do
-    LLAMA_DIR=$llama_dir WH_ARCH_YAML=wormhole_b0_80_arch_eth_dispatch.yaml pytest -n auto models/demos/llama3/demo/demo.py --timeout 600; fail+=$?
-    echo "LOG_METAL: Llama3 tests for $llama_dir completed"
-  done
-
   # Mamba
   WH_ARCH_YAML=wormhole_b0_80_arch_eth_dispatch.yaml pytest -n auto --disable-warnings -q -s --input-method=json --input-path='models/demos/wormhole/mamba/demo/prompts.json' models/demos/wormhole/mamba/demo/demo.py --timeout 420; fail+=$?
 
@@ -91,6 +76,27 @@ run_n300_perf_tests(){
 
   run_common_perf_tests; fail+=$?
 
+  # Llama3.1-8B
+  llama8b=/mnt/MLPerf/tt_dnn-models/llama/Meta-Llama-3.1-8B-Instruct/
+  # Llama3.2-1B
+  llama1b=/mnt/MLPerf/tt_dnn-models/llama/Llama3.2-1B-Instruct/
+  # Llama3.2-3B
+  llama3b=/mnt/MLPerf/tt_dnn-models/llama/Llama3.2-3B-Instruct/
+  # Llama3.2-11B
+  llama11b=/mnt/MLPerf/tt_dnn-models/llama/Llama3.2-11B-Vision-Instruct/
+
+  # Run all Llama3 tests for 1B, 3B, 8B and 11B weights for N150
+  # To ensure a proper perf measurement and dashboard upload of the Llama3 models on a N150, we have to run them on the N300 perf pipeline for now
+  for llama_dir in "$llama1b" "$llama3b" "$llama8b" "$llama11b"; do
+    FAKE_DEVICE=N150 LLAMA_DIR=$llama_dir WH_ARCH_YAML=wormhole_b0_80_arch_eth_dispatch.yaml pytest -n auto models/demos/llama3/demo/demo.py --timeout 600; fail+=$?
+    echo "LOG_METAL: Llama3 tests for $llama_dir completed on N150"
+  done
+  # Run all Llama3 tests for 1B, 3B, 8B and 11B weights
+  for llama_dir in "$llama1b" "$llama3b" "$llama8b" "$llama11b"; do
+    LLAMA_DIR=$llama_dir WH_ARCH_YAML=wormhole_b0_80_arch_eth_dispatch.yaml pytest -n auto models/demos/llama3/demo/demo.py --timeout 600; fail+=$?
+    echo "LOG_METAL: Llama3 tests for $llama_dir completed"
+  done
+
   # Falcon7b (perf verification for 128/1024/2048 seq lens and output token verification)
   WH_ARCH_YAML=wormhole_b0_80_arch_eth_dispatch.yaml pytest -n auto --disable-warnings -q -s --input-method=json --input-path='models/demos/falcon7b_common/demo/input_data.json' models/demos/wormhole/falcon7b/demo_wormhole.py; fail+=$?