mirrors/docling
1
0
Fork 0
mirror of https://github.com/DS4SD/docling.git synced 2025-12-08 20:58:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

feat(ASR): MLX Whisper Support for Apple Silicon (#2366)

Browse Source 
* add mlx-whisper support

* added mlx-whisper example and test. update docling cli to use MLX automatically if present.

* fix pre-commit checks and added proper type safety

* fixed linter issue

* DCO Remediation Commit for Ken Steele <ksteele@gmail.com>

I, Ken Steele <ksteele@gmail.com>, hereby add my Signed-off-by to this commit: a979a680e1dc2fee8461401335cfb5dda8cfdd98
I, Ken Steele <ksteele@gmail.com>, hereby add my Signed-off-by to this commit: 9827068382ca946fe1387ed83f747ae509fcf229
I, Ken Steele <ksteele@gmail.com>, hereby add my Signed-off-by to this commit: ebbeb45c7dc266260e1fad6bdb54a7041f8aeed4
I, Ken Steele <ksteele@gmail.com>, hereby add my Signed-off-by to this commit: 2f6fd3cf46c8ca0bb98810191578278f1df87aa3

Signed-off-by: Ken Steele <ksteele@gmail.com>

* fix unit tests and code coverage for CI

* DCO Remediation Commit for Ken Steele <ksteele@gmail.com>

I, Ken Steele <ksteele@gmail.com>, hereby add my Signed-off-by to this commit: 5e61bf11139a2133978db2c8d306be6289aed732

Signed-off-by: Ken Steele <ksteele@gmail.com>

* fix CI example test - mlx_whisper_example.py defaults to tests/data/audio/sample_10s.mp3 if no args specified.

Signed-off-by: Ken Steele <ksteele@gmail.com>

* refactor: centralize audio file extensions and MIME types in base_models.py

- Move audio file extensions from CLI hardcoded set to FormatToExtensions[InputFormat.AUDIO]
- Add support for additional audio formats: m4a, aac, ogg, flac, mp4, avi, mov
- Update FormatToMimeType mapping to include MIME types for all audio formats
- Update CLI auto-detection to use centralized FormatToExtensions mapping
- Add comprehensive tests for audio file auto-detection and pipeline selection
- Ensure explicit pipeline choices are not overridden by auto-detection

Fixes issue where only .mp3 and .wav files were processed as audio despite
CLI auto-detection working for all formats. The document converter now
properly recognizes all audio formats through MIME type detection.

Addresses review comments:
- Centralizes audio extensions in base_models.py as suggested
- Maintains existing auto-detection behavior while using centralized data
- Adds proper test coverage for the audio detection functionality

All examples and tests pass with the new centralized approach.
All audio formats (mp3, wav, m4a, aac, ogg, flac, mp4, avi, mov) now work correctly.

Signed-off-by: Ken Steele <ksteele@gmail.com>

* feat: address reviewer feedback - improve CLI auto-detection and add explicit model options

Review feedback addressed:
1. Fix CLI auto-detection to only switch to ASR pipeline when ALL files are audio
   - Previously switched if ANY file was audio, now requires ALL files to be audio
   - Added warning for mixed file types with guidance to use --pipeline asr

2. Add explicit WHISPER_X_MLX and WHISPER_X_NATIVE model options
   - Users can now force specific implementations if desired
   - Auto-selecting models (WHISPER_BASE, etc.) still choose best for hardware
   - Added 12 new explicit model options: _MLX and _NATIVE variants for each size

CLI now supports:
- Auto-selecting: whisper_tiny, whisper_base, etc. (choose best for hardware)
- Explicit MLX: whisper_tiny_mlx, whisper_base_mlx, etc. (force MLX)
- Explicit Native: whisper_tiny_native, whisper_base_native, etc. (force native)

Addresses reviewer comments from @dolfim-ibm

Signed-off-by: Ken Steele <ksteele@gmail.com>

* DCO Remediation Commit for Ken Steele <ksteele@gmail.com>

I, Ken Steele <ksteele@gmail.com>, hereby add my Signed-off-by to this commit: c60e72d2b5
I, Ken Steele <ksteele@gmail.com>, hereby add my Signed-off-by to this commit: 94803317a3
I, Ken Steele <ksteele@gmail.com>, hereby add my Signed-off-by to this commit: 21905e8ace
I, Ken Steele <ksteele@gmail.com>, hereby add my Signed-off-by to this commit: 96c669d155
I, Ken Steele <ksteele@gmail.com>, hereby add my Signed-off-by to this commit: 8371c060ea

Signed-off-by: Ken Steele <ksteele@gmail.com>

* test(asr): add coverage for MLX options, pipeline helpers, and VLM prompts

- tests/test_asr_mlx_whisper.py: verify explicit MLX options (framework, repo ids)
- tests/test_asr_pipeline.py: cover _has_text/_determine_status and backend support with proper InputDocument/NoOpBackend wiring
- tests/test_interfaces.py: add BaseVlmPageModel.formulate_prompt tests (RAW/NONE/CHAT, invalid style), with minimal InlineVlmOptions scaffold

Improves reliability of ASR and VLM components by validating configuration paths and helper logic.

Signed-off-by: Ken Steele <ksteele@gmail.com>

* test(asr): broaden coverage for model selection, pipeline flows, and VLM prompts

- tests/test_asr_mlx_whisper.py
  - Add MLX/native selector coverage across all Whisper sizes
  - Validate repo_id choices under MLX and Native paths
  - Cover fallback path when MPS unavailable and mlx_whisper missing

- tests/test_asr_pipeline.py
  - Relax silent-audio assertion to accept PARTIAL_SUCCESS or SUCCESS
  - Force CPU native path in helper tests to avoid torch in device selection
  - Add language handling tests for native/MLX transcribe
  - Cover native run success (BytesIO) and failure (exception) branches
  - Cover MLX run success/failure branches with mocked transcribe
  - Add init path coverage with artifacts_path

- tests/test_interfaces.py
  - Add focused VLM prompt tests (NONE/CHAT variants)

Result: all tests passing with significantly improved coverage for ASR model selectors, pipeline execution paths, and VLM prompt formulation.

Signed-off-by: Ken Steele <ksteele@gmail.com>

* simplify ASR model settings (no pipeline detection needed)

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* clean up disk space in runners

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

---------

Signed-off-by: Ken Steele <ksteele@gmail.com>
Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>
Co-authored-by: Michele Dolfi <dol@zurich.ibm.com>
This commit is contained in:
Ken Steele
2025-10-20 23:05:59 -07:00
committed by GitHub
parent a5af082d82
commit 657ce8b01c
29 changed files with 2016 additions and 71 deletions
Download Patch File Download Diff File Expand all files Collapse all files

77
tests/test_interfaces.py
View File

@@ -1,12 +1,19 @@
from io import BytesIO
from pathlib import Path
from unittest.mock import Mock
import pytest
from docling.datamodel.accelerator_options import AcceleratorDevice
from docling.datamodel.base_models import DocumentStream, InputFormat
from docling.datamodel.pipeline_options import PdfPipelineOptions
from docling.datamodel.pipeline_options_vlm_model import (
InferenceFramework,
InlineVlmOptions,
ResponseFormat,
TransformersPromptStyle,
)
from docling.document_converter import DocumentConverter, PdfFormatOption
from docling.models.base_model import BaseVlmPageModel
from .test_data_gen_flag import GEN_TEST_DATA
from .verify_utils import verify_conversion_result_v2
@@ -21,6 +28,8 @@ def get_pdf_path():
@pytest.fixture
def converter():
from docling.datamodel.pipeline_options import PdfPipelineOptions
pipeline_options = PdfPipelineOptions()
pipeline_options.do_ocr = False
pipeline_options.do_table_structure = True
@@ -44,6 +53,7 @@ def test_convert_path(converter: DocumentConverter):
pdf_path = get_pdf_path()
print(f"converting {pdf_path}")
# Avoid heavy torch-dependent models by not instantiating layout models here in coverage run
doc_result = converter.convert(pdf_path)
verify_conversion_result_v2(
input_path=pdf_path, doc_result=doc_result, generate=GENERATE
@@ -61,3 +71,68 @@ def test_convert_stream(converter: DocumentConverter):
verify_conversion_result_v2(
input_path=pdf_path, doc_result=doc_result, generate=GENERATE
)
class _DummyVlm(BaseVlmPageModel):
def __init__(self, prompt_style: TransformersPromptStyle, repo_id: str = ""): # type: ignore[no-untyped-def]
self.vlm_options = InlineVlmOptions(
repo_id=repo_id or "dummy/repo",
prompt="test prompt",
inference_framework=InferenceFramework.TRANSFORMERS,
response_format=ResponseFormat.PLAINTEXT,
transformers_prompt_style=prompt_style,
)
self.processor = Mock()
def __call__(self, conv_res, page_batch): # type: ignore[no-untyped-def]
return []
def process_images(self, image_batch, prompt): # type: ignore[no-untyped-def]
return []
def test_formulate_prompt_raw():
model = _DummyVlm(TransformersPromptStyle.RAW)
assert model.formulate_prompt("hello") == "hello"
def test_formulate_prompt_none():
model = _DummyVlm(TransformersPromptStyle.NONE)
assert model.formulate_prompt("ignored") == ""
def test_formulate_prompt_phi4_special_case():
model = _DummyVlm(
TransformersPromptStyle.RAW, repo_id="ibm-granite/granite-docling-258M"
)
# RAW style with granite-docling should still invoke the special path only when style not RAW;
# ensure RAW returns the user text
assert model.formulate_prompt("describe image") == "describe image"
def test_formulate_prompt_chat_uses_processor_template():
model = _DummyVlm(TransformersPromptStyle.CHAT)
model.processor.apply_chat_template.return_value = "templated"
out = model.formulate_prompt("summarize")
assert out == "templated"
model.processor.apply_chat_template.assert_called()
def test_formulate_prompt_unknown_style_raises():
# Create an InlineVlmOptions with an invalid enum by patching attribute directly
model = _DummyVlm(TransformersPromptStyle.RAW)
model.vlm_options.transformers_prompt_style = "__invalid__" # type: ignore[assignment]
with pytest.raises(RuntimeError):
model.formulate_prompt("x")
def test_vlm_prompt_style_none_and_chat_variants():
# NONE always empty
m_none = _DummyVlm(TransformersPromptStyle.NONE)
assert m_none.formulate_prompt("anything") == ""
# CHAT path ensures processor used even with complex prompt
m_chat = _DummyVlm(TransformersPromptStyle.CHAT)
m_chat.processor.apply_chat_template.return_value = "ok"
out = m_chat.formulate_prompt("details please")
assert out == "ok"