#0: increase perf bounds

tenstorrent · Mar 1, 2025 · 56d12b0 · 56d12b0
1 parent 4c8f233
commit 56d12b0
Show file tree

Hide file tree

Showing 8 changed files with 13 additions and 13 deletions.
diff --git a/models/demos/bert_tiny/tests/test_performance.py b/models/demos/bert_tiny/tests/test_performance.py
@@ -119,7 +119,7 @@ def test_perf_device_bare_metal(batch_size, expected_perf):
     margin = 0.03
 
     if is_wormhole_b0():
-        expected_perf = 4439.2
+        expected_perf = 4475.2
     else:
         expected_perf = 3460.0
 

diff --git a/models/demos/distilbert/tests/test_perf_distilbert.py b/models/demos/distilbert/tests/test_perf_distilbert.py
@@ -154,7 +154,7 @@ def test_distilbert_perf_device(batch_size, test, reset_seeds):
     if is_grayskull():
         expected_perf = 57.3
     elif is_wormhole_b0():
-        expected_perf = 95.5
+        expected_perf = 97.0
 
     command = f"pytest tests/ttnn/integration_tests/distilbert/test_ttnn_distilbert.py::test_distilbert_for_question_answering[sequence_size=768-batch_size=8-model_name=distilbert-base-uncased-distilled-squad]"
     cols = ["DEVICE FW", "DEVICE KERNEL", "DEVICE BRISC KERNEL"]

diff --git a/models/demos/falcon7b_common/tests/test_falcon_device_perf.py b/models/demos/falcon7b_common/tests/test_falcon_device_perf.py
@@ -90,12 +90,12 @@ def test_device_perf_wh_bare_metal(
 @pytest.mark.parametrize(
     "llm_mode, batch, seq_len, kv_cache_len, model_config_str, samples",
     (
-        ("prefill", 1, 128, 0, "BFLOAT16-DRAM", 2060),
-        ("prefill", 1, 1024, 0, "BFLOAT16-DRAM", 2890),
-        ("prefill", 1, 2048, 0, "BFLOAT16-DRAM", 2680),
-        ("decode", 32, 1, 128, "BFLOAT16-L1_SHARDED", 625),
-        ("decode", 32, 1, 1024, "BFLOAT16-L1_SHARDED", 568),
-        ("decode", 32, 1, 2047, "BFLOAT16-L1_SHARDED", 550),
+        ("prefill", 1, 128, 0, "BFLOAT16-DRAM", 2068),
+        ("prefill", 1, 1024, 0, "BFLOAT16-DRAM", 2895),
+        ("prefill", 1, 2048, 0, "BFLOAT16-DRAM", 2684),
+        ("decode", 32, 1, 128, "BFLOAT16-L1_SHARDED", 629),
+        ("decode", 32, 1, 1024, "BFLOAT16-L1_SHARDED", 572),
+        ("decode", 32, 1, 2047, "BFLOAT16-L1_SHARDED", 553),
     ),
 )
 @skip_for_grayskull()

diff --git a/models/demos/roberta/tests/test_perf_device_roberta.py b/models/demos/roberta/tests/test_perf_device_roberta.py
@@ -18,7 +18,7 @@ def test_perf_device_bare_metal(batch_size, test):
     subdir = "ttnn_roberta"
     num_iterations = 1
     margin = 0.03
-    expected_perf = 154.94 if is_grayskull() else 153.70
+    expected_perf = 154.94 if is_grayskull() else 153.9
 
     command = f"pytest tests/ttnn/integration_tests/roberta/test_ttnn_optimized_roberta.py::test_roberta_for_question_answering[{test}]"
     cols = ["DEVICE FW", "DEVICE KERNEL", "DEVICE BRISC KERNEL"]

diff --git a/models/demos/squeezebert/tests/test_perf_device_squeezebert.py b/models/demos/squeezebert/tests/test_perf_device_squeezebert.py
@@ -18,7 +18,7 @@ def test_perf_device_bare_metal(batch_size, test):
     subdir = "ttnn_squeezebert"
     num_iterations = 1
     margin = 0.03
-    expected_perf = 102.7 if is_grayskull() else 298.7
+    expected_perf = 102.7 if is_grayskull() else 299.7
 
     command = f"pytest tests/ttnn/integration_tests/squeezebert/test_ttnn_squeezebert.py::test_squeezebert_for_question_answering"
     cols = ["DEVICE FW", "DEVICE KERNEL", "DEVICE BRISC KERNEL"]

diff --git a/models/demos/wormhole/resnet50/tests/test_perf_device_resnet50.py b/models/demos/wormhole/resnet50/tests/test_perf_device_resnet50.py
@@ -12,7 +12,7 @@
 @pytest.mark.parametrize(
     "batch_size, test, expected_perf",
     [
-        [16, "16-act_dtype0-weight_dtype0-math_fidelity0-device_params0", 5255],
+        [16, "16-act_dtype0-weight_dtype0-math_fidelity0-device_params0", 5311.0],
     ],
 )
 def test_perf_device(batch_size, test, expected_perf):

diff --git a/models/demos/wormhole/stable_diffusion/tests/test_perf.py b/models/demos/wormhole/stable_diffusion/tests/test_perf.py
@@ -207,7 +207,7 @@ def test_stable_diffusion_perf(device, batch_size, num_inference_steps, expected
 @pytest.mark.models_device_performance_bare_metal
 @pytest.mark.parametrize(
     "expected_kernel_samples_per_second",
-    ((9.5),),
+    ((9.8),),
 )
 def test_stable_diffusion_device_perf(expected_kernel_samples_per_second):
     subdir = "ttnn_stable_diffusion"

diff --git a/models/experimental/functional_unet/tests/test_unet_perf.py b/models/experimental/functional_unet/tests/test_unet_perf.py
@@ -92,7 +92,7 @@ def test_unet_trace_perf(
 @pytest.mark.parametrize(
     "batch, groups, iterations, expected_compile_time, expected_throughput, use_async_mode",
     (
-        (1, 2, 128, 25.0, 1500.0, True),
+        (1, 2, 128, 25.0, 1450.0, True),
         (1, 2, 128, 25.0, 1660.0, False),
     ),
 )