Online video support for VLMs (#10020)

Signed-off-by: DarkLight1337 <[email protected]>
Co-authored-by: litianjian <[email protected]>
Co-authored-by: DarkLight1337 <[email protected]>

docs/source/conf.py

@@ -116,6 +116,7 @@ def setup(app):
     "soundfile",
     "gguf",
     "lark",
+    "decord",
 ]
 
 for mock_target in autodoc_mock_imports:

requirements-test.in

@@ -8,19 +8,21 @@ pytest-shard
 
 # testing utils
 awscli
+decord # required for video tests
 einops # required for MPT, qwen-vl and Mamba
 httpx
 librosa # required for audio tests
 opencv-python # required for video tests
 peft
 requests
 ray[adag]==2.35
-sentence-transformers # required for embedding
-soundfile # required for audio test
+sentence-transformers # required for embedding tests
+soundfile # required for audio tests
 timm # required for internvl test
 torch==2.5.1
 transformers_stream_generator # required for qwen-vl test
 matplotlib # required for qwen-vl test
+mistral_common[opencv] >= 1.4.4 # required for pixtral test
 datamodel_code_generator # required for minicpm3 test
 lm-eval[api]==0.4.4 # required for model evaluation test
 

requirements-test.txt

@@ -1,8 +1,8 @@
 #
-# This file is autogenerated by pip-compile with Python 3.12
+# This file is autogenerated by pip-compile with Python 3.9
 # by the following command:
 #
-#    pip-compile --output-file=requirements-test.txt requirements-test.in
+#    pip-compile requirements-test.in
 #
 absl-py==2.1.0
     # via rouge-score
@@ -28,6 +28,10 @@ anyio==4.6.2.post1
     # via httpx
 argcomplete==3.5.1
     # via datamodel-code-generator
+async-timeout==4.0.3
+    # via
+    #   aiohttp
+    #   redis
 attrs==24.2.0
     # via
     #   aiohttp
@@ -90,6 +94,8 @@ datasets==3.0.2
     #   lm-eval
 decorator==5.1.1
     # via librosa
+decord==0.6.0
+    # via -r requirements-test.in
 dill==0.3.8
     # via
     #   datasets
@@ -106,6 +112,10 @@ email-validator==2.2.0
     # via pydantic
 evaluate==0.4.3
     # via lm-eval
+exceptiongroup==1.2.2
+    # via
+    #   anyio
+    #   pytest
 fastrlock==0.8.2
     # via cupy-cuda12x
 filelock==3.16.1
@@ -156,6 +166,8 @@ idna==3.10
     #   httpx
     #   requests
     #   yarl
+importlib-resources==6.4.5
+    # via matplotlib
 inflect==5.6.2
     # via datamodel-code-generator
 iniconfig==2.0.0
@@ -178,7 +190,9 @@ joblib==1.4.2
 jsonlines==4.0.0
     # via lm-eval
 jsonschema==4.23.0
-    # via ray
+    # via
+    #   mistral-common
+    #   ray
 jsonschema-specifications==2024.10.1
     # via jsonschema
 kiwisolver==1.4.7
@@ -204,6 +218,10 @@ mbstrdecoder==1.1.3
     #   dataproperty
     #   pytablewriter
     #   typepy
+mistral-common[opencv]==1.4.4
+    # via
+    #   -r requirements-test.in
+    #   mistral-common
 more-itertools==10.5.0
     # via lm-eval
 mpmath==1.3.0
@@ -238,12 +256,15 @@ numpy==1.26.4
     #   contourpy
     #   cupy-cuda12x
     #   datasets
+    #   decord
     #   evaluate
     #   librosa
     #   matplotlib
+    #   mistral-common
     #   numba
     #   numexpr
     #   opencv-python
+    #   opencv-python-headless
     #   pandas
     #   peft
     #   rouge-score
@@ -288,6 +309,8 @@ nvidia-nvtx-cu12==12.4.127
     # via torch
 opencv-python==4.10.0.84
     # via -r requirements-test.in
+opencv-python-headless==4.10.0.84
+    # via mistral-common
 packaging==24.1
     # via
     #   accelerate
@@ -317,9 +340,10 @@ peft==0.13.2
     # via
     #   -r requirements-test.in
     #   lm-eval
-pillow==11.0.0
+pillow==10.4.0
     # via
     #   matplotlib
+    #   mistral-common
     #   sentence-transformers
     #   torchvision
 platformdirs==4.3.6
@@ -354,7 +378,9 @@ pybind11==2.13.6
 pycparser==2.22
     # via cffi
 pydantic[email]==2.9.2
-    # via datamodel-code-generator
+    # via
+    #   datamodel-code-generator
+    #   mistral-common
 pydantic-core==2.23.4
     # via pydantic
 pyparsing==3.2.0
@@ -420,6 +446,7 @@ requests==2.32.3
     #   evaluate
     #   huggingface-hub
     #   lm-eval
+    #   mistral-common
     #   pooch
     #   ray
     #   tiktoken
@@ -456,6 +483,8 @@ scipy==1.13.1
     #   sentence-transformers
 sentence-transformers==3.2.1
     # via -r requirements-test.in
+sentencepiece==0.2.0
+    # via mistral-common
 six==1.16.0
     # via
     #   python-dateutil
@@ -486,12 +515,20 @@ tensorizer==2.9.0
     # via -r requirements-test.in
 threadpoolctl==3.5.0
     # via scikit-learn
-tiktoken==0.8.0
-    # via lm-eval
+tiktoken==0.7.0
+    # via
+    #   lm-eval
+    #   mistral-common
 timm==1.0.11
     # via -r requirements-test.in
 tokenizers==0.20.1
     # via transformers
+toml==0.10.2
+    # via datamodel-code-generator
+tomli==2.0.2
+    # via
+    #   black
+    #   pytest
 torch==2.5.1
     # via
     #   -r requirements-test.in
@@ -535,8 +572,12 @@ typepy[datetime]==1.3.2
     #   tabledata
 typing-extensions==4.12.2
     # via
+    #   anyio
+    #   black
     #   huggingface-hub
     #   librosa
+    #   mistral-common
+    #   multidict
     #   pydantic
     #   pydantic-core
     #   torch
@@ -554,6 +595,8 @@ xxhash==3.5.0
     #   evaluate
 yarl==1.17.1
     # via aiohttp
+zipp==3.20.2
+    # via importlib-resources
 zstandard==0.23.0
     # via lm-eval
 

setup.py

@@ -554,7 +554,8 @@ def _read_requirements(filename: str) -> List[str]:
     ext_modules=ext_modules,
     extras_require={
         "tensorizer": ["tensorizer>=2.9.0"],
-        "audio": ["librosa", "soundfile"]  # Required for audio processing
+        "audio": ["librosa", "soundfile"],  # Required for audio processing
+        "video": ["decord"]  # Required for video processing
     },
     cmdclass={"build_ext": cmake_build_ext} if len(ext_modules) > 0 else {},
     package_data=package_data,