#7379: Added proper testing method for program caching

tenstorrent · May 16, 2024 · 17ae8f4 · 17ae8f4 · github-actions · Dec 2, 2024
1 parent 43930e6
commit 17ae8f4
Showing 1 changed file with 31 additions and 4 deletions.
diff --git a/tests/tt_eager/python_api_testing/unit_testing/misc/test_nlp_kv_cache_load_slice.py b/tests/tt_eager/python_api_testing/unit_testing/misc/test_nlp_kv_cache_load_slice.py
@@ -34,15 +34,14 @@ def unpadding_test(
         .to(ttl.tensor.Layout.TILE)
         .to(device)
     )
-    # breakpoint()
     test_tensor_tt = ttl.tensor.nlp_kv_cache_load_slice(test_tensor, seq_len_start, seq_len_end)
 
     test_tensor_pt = test_tensor_tt.cpu().to(ttl.tensor.Layout.ROW_MAJOR).to_torch()
 
     # Pytorch reference
     test_tensor_ref = inp[:, :, seq_len_start:seq_len_end]
 
-    return test_tensor_pt, test_tensor_ref, test_tensor_tt.memory_config()
+    return test_tensor_pt, test_tensor_ref, test_tensor_tt.memory_config(), device.num_program_cache_entries()
 
 
 @pytest.mark.parametrize(
@@ -85,14 +84,15 @@ def test_run_unpadding_test(
         pytest.skip("Skipping test on Grayskull")
 
     for i in range(3):
-        a_pt, a_ref, memory_config = unpadding_test(
+        a_pt, a_ref, memory_config, num_cache_entries = unpadding_test(
             kv_cache_shape,
             seq_len_start,
             seq_len_end,
             device,
             dtype,
         )
         assert a_pt.shape == a_ref.shape
+        assert num_cache_entries == 1
         if dtype == ttl.tensor.DataType.BFLOAT8_B:
             # inevitable precision loss for bfloat8_b
             eq, pcc = comp_pcc(a_pt, a_ref, 0.999)
@@ -107,19 +107,33 @@ def test_run_unpadding_test(
         assert memory_config.shard_spec.shape[0] == a_ref.shape[-2]
         assert memory_config.shard_spec.shape[1] == a_ref.shape[-1]
 
+        # shift input/output tensor by creating very small tensor between loop
+        inp = torch.rand(1, 1, 32, 32)
+        test_tensor = (
+            ttl.tensor.Tensor(
+                inp.reshape(-1).tolist(),
+                inp.shape,
+                dtype,
+                ttl.tensor.Layout.ROW_MAJOR,
+            )
+            .to(ttl.tensor.Layout.TILE)
+            .to(device)
+        )
+
     # hardcoded test 2 to check program caching
     kv_cache_shape = (2, 2, 128, 32)
     seq_len_start = 0
     seq_len_end = 64
     for i in range(2):
-        a_pt, a_ref, memory_config = unpadding_test(
+        a_pt, a_ref, memory_config, num_cache_entries = unpadding_test(
             kv_cache_shape,
             seq_len_start,
             seq_len_end,
             device,
             dtype,
         )
         assert a_pt.shape == a_ref.shape
+        assert num_cache_entries == 2
         if dtype == ttl.tensor.DataType.BFLOAT8_B:
             # inevitable precision loss for bfloat8_b
             eq, pcc = comp_pcc(a_pt, a_ref, 0.999)
@@ -133,3 +147,16 @@ def test_run_unpadding_test(
         assert memory_config.buffer_type == ttl.tensor.BufferType.L1
         assert memory_config.shard_spec.shape[0] == a_ref.shape[-2]
         assert memory_config.shard_spec.shape[1] == a_ref.shape[-1]
+
+        # shift input/output tensor by creating very small tensor between loop
+        inp = torch.rand(1, 1, 32, 32)
+        test_tensor = (
+            ttl.tensor.Tensor(
+                inp.reshape(-1).tolist(),
+                inp.shape,
+                dtype,
+                ttl.tensor.Layout.ROW_MAJOR,
+            )
+            .to(ttl.tensor.Layout.TILE)
+            .to(device)
+        )