#0: fix model and sweeps

tenstorrent · Nov 1, 2024 · 0fa3566 · 0fa3566
1 parent 32c62b9
commit 0fa3566
Show file tree

Hide file tree

Showing 3 changed files with 4 additions and 23 deletions.
diff --git a/...bert_large_performant/unit_tests/test_bert_large_matmuls_and_bmms_with_mixed_precision.py b/...bert_large_performant/unit_tests/test_bert_large_matmuls_and_bmms_with_mixed_precision.py
@@ -101,11 +101,8 @@ def run_bert_large_matmul_test(
     )
 
     if bias_mem_config is not None:
-        bias_t = (
-            ttnn.Tensor(BIAS, bias_dtype)
-            .pad(bias_pad_shape, [0, 0, 0, 0], 0)
-            .to(ttnn.TILE_LAYOUT)
-            .to(device, bias_mem_config)
+        bias_t = ttnn.from_torch(
+            BIAS, dtype=bias_dtype, layout=ttnn.TILE_LAYOUT, memory_config=bias_mem_config, device=device
         )
     else:
         bias_t = None

diff --git a/tests/tt_eager/python_api_testing/sweep_tests/pytests/tt_dnn/test_softmax_sharded.py b/tests/tt_eager/python_api_testing/sweep_tests/pytests/tt_dnn/test_softmax_sharded.py
@@ -130,10 +130,7 @@ def test_softmax(device, in_dtype, causal_mask, grid_size, seq_len, scale_mask):
     else:
         tt_output_sharded = ttnn.softmax_in_place(in1_t_shard, program_config=program_config)
 
-    tt_output = ttnn.sharded_to_interleaved(tt_output_sharded, in0_mem_config)
-    tt_output_tensor = tt_output.cpu().to_torch().float()
-    tt_output_tensor = torch.Tensor(tt_output_tensor).reshape(input_shape)
-    tt_output_tensor = untilize(tt_output_tensor)
+    tt_output_tensor = ttnn.to_torch(tt_output_sharded)
 
     if causal_mask == False:
         attention_mask = attention_mask.reshape(batch, 1, 1, seq_len)

diff --git a/ttnn/cpp/ttnn/tensor/tensor_impl.cpp b/ttnn/cpp/ttnn/tensor/tensor_impl.cpp
@@ -732,20 +732,7 @@ DeviceBuffer to_device_buffer(
             using StorageType = std::decay_t<decltype(storage)>;
             if constexpr (std::is_same_v<StorageType, OwnedStorage> or std::is_same_v<StorageType, BorrowedStorage>) {
                 auto data_to_write = host_buffer::get_as<T>(storage.buffer);
-                auto buffer_size = compute_buffer_size(shape, data_type, tile);
-                TT_ASSERT(
-                    buffer_size == data_to_write.size(),
-                        "Tensor buffer size and number of data elements does not match: {} != {}",
-                        buffer_size,
-                        data_to_write.size());
-                if (layout == Layout::TILE) {
-                    auto tile_shape = tile.value_or(Tile{{constants::TILE_HEIGHT, constants::TILE_WIDTH}}).get_tile_shape();
-                    TT_ASSERT(
-                        (shape[-2] % tile_shape[0] == 0 && shape[-1] % tile_shape[1] == 0),
-                        "Tensor shape incompatible for specified layout");
-                }
-                return initialize_data_on_device<T>(
-                    data_to_write, device, shape, data_type, layout, memory_config, shard_spec, tile);
+                return initialize_data_on_device<T>(data_to_write, device, shape, tensor_layout, queue);
             } else if constexpr (std::is_same_v<StorageType, DeviceStorage>) {
                 TT_THROW("Device storage doesn't support to_device_buffer");
             } else if constexpr (std::is_same_v<StorageType, MultiDeviceStorage>) {