Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Dev/thang/ola #561

Closed
wants to merge 9 commits into from
+34 −29
Closed

Dev/thang/ola #561

Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
f645f4d
[Model] add Ola audio
Devininthelab Feb 24, 2025
cc01202
[Model] add Ola
Devininthelab Feb 24, 2025
d2cabd1
[Model] Ola
Devininthelab Feb 25, 2025
4f47fc6
[Model] add Ola
Devininthelab Feb 26, 2025
184fa2d
[Model] add Ola
Devininthelab Feb 26, 2025
2364199
[Model] add Ola
Devininthelab Feb 26, 2025
eba3bcf
[Model] add Ola
Devininthelab Feb 26, 2025
86120b9
fix Ola Audio path for multi GPUs
Devininthelab Feb 28, 2025
70189ec
Merge branch 'main' into dev/thang/Ola
Devininthelab Feb 28, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
63 changes: 34 additions & 29 deletions lmms_eval/models/ola.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -71,8 +71,10 @@
AutoConfig.register("ola_qwen", OlaConfigQwen)
except:
eval_logger.debug("")
import moviepy.editor as mp
from moviepy.video.io.VideoFileClip import VideoFileClip


if "USE_SPEECH" in os.environ:
USE_SPEECH = os.environ["USE_SPEECH"]
print("USE_SPEECH is set")
Expand Down Expand Up @@ -176,7 +178,7 @@ def __init__(
)

self._config = self._model.config
self.model.to("cuda").eval().bfloat16()
self.model.to(self.device).eval().bfloat16()
self.model.tie_weights()
self.truncation = truncation
self.batch_size_per_gpu = int(batch_size)
Expand Down Expand Up @@ -207,10 +209,9 @@ def __init__(
elif accelerator.num_processes == 1 and device_map == "auto":
eval_logger.info(f"Using {accelerator.num_processes} devices with tensor parallelism")
self._rank = 0
self._word_size = 1
self._world_size = 1
else:
eval_logger.info(f"Using single device: {self._device}")
self.model.to(self._device)
self._rank = 0
self._world_size = 1
self.accelerator = accelerator
Expand Down Expand Up @@ -354,7 +355,7 @@ def flatten(self, input):
return new_list

def extract_audio(self, videos_file_path):
my_clip = VideoFileClip(videos_file_path)
my_clip = mp.VideoFileClip(videos_file_path)
return my_clip.audio

def load_audio(self, audio_file_name):
Expand Down Expand Up @@ -407,6 +408,8 @@ def process_audio(self, audio_array, sampling_rate):
CHUNK_LIM = 480000
import librosa


speech_wav = audio_array
if sampling_rate != target_sr:
speech_wav = librosa.resample(audio_array, orig_sr=sampling_rate, target_sr=target_sr).astype(np.float32)
speechs = []
Expand Down Expand Up @@ -485,13 +488,13 @@ def _collate(x):
eval_logger.info(f"Video {visuals} can not load, check the source")
continue
audio = self.extract_audio(visual)
audio.write_audiofile("./video_audio.wav")
video_audio_path = "./video_audio.wav"
audio.write_audiofile(f"./video_audio_{self.rank}.wav")
video_audio_path = f"./video_audio_{self.rank}.wav"
speech, speech_length, speech_chunk, speech_wav = self.load_audio(video_audio_path)
speechs.append(speech.bfloat16().to("cuda"))
speech_lengths.append(speech_length.to("cuda"))
speech_chunks.append(speech_chunk.to("cuda"))
speech_wavs.append(speech_wav.to("cuda"))
speechs.append(speech.bfloat16().to(self.device))
speech_lengths.append(speech_length.to(self.device))
speech_chunks.append(speech_chunk.to(self.device))
speech_wavs.append(speech_wav.to(self.device))
os.remove(video_audio_path)

# Process images of video
Expand All @@ -508,9 +511,8 @@ def _collate(x):
if frame_idx is None:
frame_idx = np.arange(0, len(video_processed), dtype=int).tolist()

video_processed = torch.cat(video_processed, dim=0).bfloat16().to("cuda")
video_processed = torch.cat(video_processed, dim=0).bfloat16().to(self.device)
video_processed = (video_processed, video_processed)

video_data = (video_processed, (384, 384), "video")

elif isinstance(visual, PIL.Image.Image): # For Image
Expand All @@ -527,32 +529,34 @@ def _collate(x):
if all(x.shape == image_highres_tensor[0].shape for x in image_highres_tensor):
image_highres_tensor = torch.stack(image_highres_tensor, dim=0)
if type(image_tensor) is list:
image_tensor = [_image.bfloat16().to("cuda") for _image in image_tensor]
image_tensor = [_image.bfloat16().to(self.device) for _image in image_tensor]
else:
image_tensor = image_tensor.bfloat16().to("cuda")
image_tensor = image_tensor.bfloat16().to(self.device)
if type(image_highres_tensor) is list:
image_highres_tensor = [_image.bfloat16().to("cuda") for _image in image_highres_tensor]
image_highres_tensor = [_image.bfloat16().to(self.device) for _image in image_highres_tensor]
else:
image_highres_tensor = image_highres_tensor.bfloat16().to("cuda")
image_highres_tensor = image_highres_tensor.bfloat16().to(self.device)

# Processing dummy audio, as required by model
speechs.append(torch.zeros(1, 3000, 128).bfloat16().to("cuda"))
speech_lengths.append(torch.LongTensor([3000]).to("cuda"))
speech_wavs.append(torch.zeros([1, 480000]).to("cuda"))
speech_chunks.append(torch.LongTensor([1]).to("cuda"))
speechs.append(torch.zeros(1, 3000, 128).bfloat16().to(self.device))
speech_lengths.append(torch.LongTensor([3000]).to(self.device))
speech_wavs.append(torch.zeros([1, 480000]).to(self.device))
speech_chunks.append(torch.LongTensor([1]).to(self.device))


elif isinstance(visual, dict) and "array" in visual: # For Audio
if MODALITY is None:
MODALITY = "AUDIO"
mels, speech_length, speech_chunk, speech_wav = self.process_audio(visual["array"], visual["sampling_rate"])
speechs.append(mels.bfloat16().to("cuda"))
speech_lengths.append(speech_length.to("cuda"))
speech_chunks.append(speech_chunk.to("cuda"))
speech_wavs.append(speech_wav.to("cuda"))

speechs.append(mels.bfloat16().to(self.device))
speech_lengths.append(speech_length.to(self.device))
speech_chunks.append(speech_chunk.to(self.device))
speech_wavs.append(speech_wav.to(self.device))

# Processing dummy images, as required by model
images.append(torch.zeros(1, 3, 224, 224).to(dtype=torch.bfloat16, device="cuda", non_blocking=True))
images_highres.append(torch.zeros(1, 3, 224, 224).to(dtype=torch.bfloat16, device="cuda", non_blocking=True))
images.append(torch.zeros(1, 3, 224, 224).to(dtype=torch.bfloat16, device=self.device, non_blocking=True))
images_highres.append(torch.zeros(1, 3, 224, 224).to(dtype=torch.bfloat16, device=self.device, non_blocking=True))
image_sizes.append((224, 224))

if not video_processed and MODALITY == "VIDEO":
Expand Down Expand Up @@ -601,11 +605,12 @@ def _collate(x):
eval_logger.debug(f"Prompt for doc ID {doc_id[0]}:\n\n{prompt}\n")

if MODALITY == "AUDIO":
input_ids = tokenizer_speech_token(prompt, self.tokenizer, SPEECH_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(self._device)
input_ids = tokenizer_speech_token(prompt, self.tokenizer, SPEECH_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(self.device)
elif MODALITY == "IMAGE":
input_ids = tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(self._device)
input_ids = tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(self.device)
elif MODALITY == "VIDEO":
input_ids = tokenizer_speech_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to("cuda")
input_ids = tokenizer_speech_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(self.device)

pad_token_ids = 151643
attention_masks = input_ids.ne(pad_token_ids).long().to(self.device)

Expand Down
Loading