whisper tiny pass

Author: rafaelpierrehf

.gitignore

@@ -3,10 +3,12 @@
 #  please consider a global .gitignore https://help.github.com/articles/ignoring-files
 .gitignore
 .egg-info
+.ruff_cache
 .vagrant*
 .hcl
 .terraform.lock.hcl
 .terraform
+pip-unpack-*
 __pycache__
 bin
 docker/docker

src/huggingface_inference_toolkit/utils.py

@@ -140,6 +140,7 @@ def _load_repository_from_hf(
 
     if framework is None:
         framework = _get_framework()
+    logging.info(f"Framework: {framework}")
 
     if isinstance(target_dir, str):
         target_dir = Path(target_dir)
@@ -149,22 +150,24 @@ def _load_repository_from_hf(
         target_dir.mkdir(parents=True)
 
     # check if safetensors weights are available
-    if framework == "pytorch":
-        files = HfApi().model_info(repository_id).siblings
-        if any(f.rfilename.endswith("safetensors") for f in files):
-            framework = "safetensors"
+    #if framework == "pytorch":
+        #files = HfApi().model_info(repository_id).siblings
+        #if any(f.rfilename.endswith("safetensors") for f in files):
+            #framework = "safetensors"
 
     # create regex to only include the framework specific weights
     ignore_regex = create_artifact_filter(framework)
+    logging.info(f"ignore_regex: {ignore_regex}")
+    logging.info(f"Framework after filtering: {framework}")
     logger.info(f"Ignore regex pattern for files, which are not downloaded: { ', '.join(ignore_regex) }")
 
     # Download the repository to the workdir and filter out non-framework specific weights
     snapshot_download(
-        repository_id,
-        revision=revision,
-        local_dir=str(target_dir),
-        local_dir_use_symlinks=False,
-        ignore_patterns=ignore_regex,
+        repo_id = repository_id,
+        revision = revision,
+        local_dir = str(target_dir),
+        local_dir_use_symlinks = False,
+        ignore_patterns = ignore_regex,
     )
 
     return target_dir
@@ -223,7 +226,12 @@ def get_device():
         return -1
 
 
-def get_pipeline(task: str, model_dir: Path, **kwargs) -> Pipeline:
+def get_pipeline(
+    task: str,
+    model_dir: Path,
+    framework = "pytorch",
+    **kwargs,
+) -> Pipeline:
     """
     create pipeline class for a specific task based on local saved model
     """
@@ -244,6 +252,12 @@ def get_pipeline(task: str, model_dir: Path, **kwargs) -> Pipeline:
         "zero-shot-image-classification",
     }:
         kwargs["feature_extractor"] = model_dir
+        hf_pipeline = pipeline(
+            task=task,
+            model=model_dir,
+            device=device,
+            **kwargs
+        )
     elif task in {"image-to-text"}:
         pass
     else:
@@ -265,12 +279,20 @@ def get_pipeline(task: str, model_dir: Path, **kwargs) -> Pipeline:
         logging.info(f"Model: {model_dir}")
         logging.info(f"Device: {device}")
         logging.info(f"Args: {kwargs}")
-        hf_pipeline = pipeline(task=task, model=model_dir, device=device, **kwargs)
+        hf_pipeline = pipeline(
+            task=task,
+            model=model_dir,
+            device=device,
+            **kwargs
+        )
 
     # wrapp specific pipeline to support better ux
     if task == "conversational":
         hf_pipeline = wrap_conversation_pipeline(hf_pipeline)
-    elif task == "automatic-speech-recognition" and isinstance(hf_pipeline.model, WhisperForConditionalGeneration):
+    elif task == "automatic-speech-recognition" and isinstance(
+        hf_pipeline.model,
+        WhisperForConditionalGeneration
+    ):
         # set chunk length to 30s for whisper to enable long audio files
         hf_pipeline._preprocess_params["chunk_length_s"] = 30
         hf_pipeline._preprocess_params["ignore_warning"] = True

tests/unit/test_utils.py

@@ -17,6 +17,8 @@
     wrap_conversation_pipeline,
 )
 
+import logging
+
 MODEL = "lysandre/tiny-bert-random"
 TASK = "text-classification"
 TASK_MODEL = "sshleifer/tiny-dbmdz-bert-large-cased-finetuned-conll03-english"
@@ -112,17 +114,39 @@ def test_get_framework_tensorflow():
 def test_get_pipeline():
     with tempfile.TemporaryDirectory() as tmpdirname:
         storage_dir = _load_repository_from_hf(MODEL, tmpdirname, framework="pytorch")
-        pipe = get_pipeline(TASK, storage_dir.as_posix())
+        pipe = get_pipeline(
+            task = TASK,
+            model_dir = storage_dir.as_posix(),
+            framework = "pytorch"
+        )
         res = pipe("Life is good, Life is bad")
         assert "score" in res[0]
 
 
 @require_torch
 def test_whisper_long_audio():
     with tempfile.TemporaryDirectory() as tmpdirname:
-        storage_dir = _load_repository_from_hf("openai/whisper-tiny", tmpdirname, framework="pytorch")
-        pipe = get_pipeline("automatic-speech-recognition", storage_dir.as_posix())
-        res = pipe(os.path.join(os.getcwd(), "tests/resources/audio", "long_sample.mp3"))
+        storage_dir = _load_repository_from_hf(
+            repository_id = "openai/whisper-tiny",
+            target_dir = tmpdirname,
+            framework = "pytorch",
+            revision = "be0ba7c2f24f0127b27863a23a08002af4c2c279"
+        )
+        logging.info(f"Temp dir: {tmpdirname}")
+        logging.info(f"POSIX Path: {storage_dir.as_posix()}")
+        logging.info(f"Contents: {os.listdir(tmpdirname)}")
+        pipe = get_pipeline(
+            task = "automatic-speech-recognition",
+            model_dir = storage_dir.as_posix(),
+            framework = "safetensors"
+        )
+        res = pipe(
+            os.path.join(
+                os.getcwd(),
+                "tests/resources/audio",
+                "long_sample.mp3"
+            )
+        )
 
         assert len(res["text"]) > 700
 
@@ -149,7 +173,7 @@ def test_wrap_conversation_pipeline():
 @require_torch
 def test_wrapped_pipeline():
     with tempfile.TemporaryDirectory() as tmpdirname:
-        storage_dir = _load_repository_from_hf("microsoft/DialoGPT-small", tmpdirname, framework="pytorch")
+        storage_dir = _load_repository_from_hf("hf-internal-testing/tiny-random-blenderbot", tmpdirname, framework="pytorch")
         conv_pipe = get_pipeline("conversational", storage_dir.as_posix())
         data = {
             "past_user_inputs": ["Which movie is the best ?"],

tox.ini

@@ -20,14 +20,19 @@ commands = ruff src --fix
 
 [testenv:unit-torch]
 install_command = 
-    pip install -e .
-    pip install -e ".[test,dev,torch,st]"
-allowlist_externals = pytest
+    pip install -e ".[test,torch,st]"
+allowlist_externals = 
+    pytest
 commands =
     pytest -s -v \
     {tty:--color=yes} \
-    tests/unit/ {posargs} \
-    --log-cli-level=ERROR \
+    tests/unit/test_const.py \
+    tests/unit/test_handler.py \
+    tests/unit/test_sentence_transformers.py \
+    tests/unit/test_serializer.py \ 
+    tests/unit/test_utils.py \
+    {posargs} \
+    --log-cli-level=DEBUG \
     --log-format='%(asctime)s %(levelname)s %(module)s:%(lineno)d %(message)s'
 
 [testenv:unit-torch-slow]