Add CPU test for Whisper(large_v3_turbo) Model (#752)

forge/test/models/pytorch/audio/whisper/test_whisper_3.py

@@ -0,0 +1,61 @@
+# SPDX-FileCopyrightText: © 2024 Tenstorrent AI ULC
+
+# SPDX-License-Identifier: Apache-2.0
+# Whisper Large V3 turbo - Automatic speech recognition and speech translation
+# Model link : https://huggingface.co/openai/whisper-large-v3-turbo
+# By default, Transformers uses the sequential algorithm.
+# To enable the chunked algorithm, pass the chunk_length_s parameter to the pipeline.
+# For large-v3, a chunk length of 30-seconds is optimal. To activate batching over long audio files, pass the argument batch_size
+
+import pytest
+import torch
+from transformers import WhisperConfig, WhisperForConditionalGeneration, WhisperProcessor
+import forge
+
+
+class Wrapper(torch.nn.Module):
+    def __init__(self, model):
+        super().__init__()
+        self.model = model
+        self.decoder_attention_mask = torch.ones((1, 1))
+
+    def forward(self, decoder_input_ids, encoder_hidden_states):
+        dec_out = self.model.model.decoder(
+            decoder_input_ids,
+            self.decoder_attention_mask,
+            encoder_hidden_states,
+        )
+        lin_out = self.model.proj_out(dec_out[0])
+        return lin_out
+
+
+@pytest.mark.nightly
+@pytest.mark.model_analysis
+@pytest.mark.xfail(
+    reason='RuntimeError: TT_ASSERT @ /tt-forge-fe/forge/csrc/passes/commute_utils.cpp:1103: reshape->op_name() == "reshape"'
+)
+@pytest.mark.parametrize("variant", ["openai/whisper-large-v3-turbo"])
+def test_whisper_large_v3_speech_translation(variant):
+    processor = WhisperProcessor.from_pretrained(variant)
+    framework_model = WhisperForConditionalGeneration.from_pretrained(variant)
+    model_config = WhisperConfig.from_pretrained(variant)
+    model = Wrapper(framework_model)
+
+    sample = torch.load("forge/test/models/files/samples/audio/1272-128104-0000.pt")
+    sample_audio = sample["audio"]["array"]
+    inputs = processor(sample_audio, return_tensors="pt", sampling_rate=16000)
+    input_features = inputs.input_features
+
+    # Get decoder inputs
+    decoder_input_ids = torch.tensor([[1, 1]]) * model_config.decoder_start_token_id
+    decoder_input_ids = decoder_input_ids.to(torch.int32)
+    encoder_outputs = model.model.model.encoder(input_features)[0].detach()
+    encoder_outputs = encoder_outputs.to(torch.float32)
+    data_input = [decoder_input_ids, encoder_outputs]
+
+    # Compiler test
+    compiled_model = forge.compile(
+        model, sample_inputs=data_input, module_name="pt_" + str(variant.split("/")[-1].replace("-", "_"))
+    )
+
+    verify(data_input, model, compiled_model, VerifyConfig(verify_allclose=False))