Initial release: LLaVA-OneVision2-8B-Instruct
Browse files- .gitattributes +1 -0
- README.md +101 -3
- added_tokens.json +24 -0
- chat_template.jinja +7 -0
- config.json +115 -0
- configuration_llava_onevision2.py +111 -0
- demo_inference.py +269 -0
- generation_config.json +6 -0
- merges.txt +0 -0
- model-00001-of-00004.safetensors +3 -0
- model-00002-of-00004.safetensors +3 -0
- model-00003-of-00004.safetensors +3 -0
- model-00004-of-00004.safetensors +3 -0
- model.safetensors.index.json +703 -0
- modeling_llava_onevision2.py +1607 -0
- preprocessor_config.json +35 -0
- processing_llava_onevision2.py +406 -0
- special_tokens_map.json +31 -0
- tokenizer.json +3 -0
- tokenizer_config.json +209 -0
- video_preprocessor_config.json +17 -0
- video_processing_llava_onevision2.py +694 -0
- vocab.json +0 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
README.md
CHANGED
|
@@ -1,3 +1,101 @@
|
|
| 1 |
-
---
|
| 2 |
-
|
| 3 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: transformers
|
| 3 |
+
pipeline_tag: image-text-to-text
|
| 4 |
+
license: apache-2.0
|
| 5 |
+
tags:
|
| 6 |
+
- multimodal
|
| 7 |
+
- vision-language
|
| 8 |
+
- image-text-to-text
|
| 9 |
+
- video-text-to-text
|
| 10 |
+
- llava
|
| 11 |
+
- llava-onevision
|
| 12 |
+
- qwen3
|
| 13 |
+
language:
|
| 14 |
+
- en
|
| 15 |
+
- zh
|
| 16 |
+
---
|
| 17 |
+
|
| 18 |
+
# LLaVA-OneVision2-8B-Instruct
|
| 19 |
+
|
| 20 |
+
A multimodal vision-language model that handles **single images, multi-image, and video** inputs, built on a Qwen3-8B language backbone with a OneVision-style vision encoder.
|
| 21 |
+
|
| 22 |
+
The model is distributed as a HuggingFace `transformers` checkpoint with custom code (`trust_remote_code=True`).
|
| 23 |
+
|
| 24 |
+
## Requirements
|
| 25 |
+
|
| 26 |
+
```bash
|
| 27 |
+
pip install "transformers>=5.7.0" "torch>=2.4" pillow requests decord
|
| 28 |
+
```
|
| 29 |
+
|
| 30 |
+
## Quick start
|
| 31 |
+
|
| 32 |
+
The repository ships a ready-to-run `demo_inference.py` that covers both image and video paths.
|
| 33 |
+
|
| 34 |
+
```bash
|
| 35 |
+
# Image (default sample image; no auth required)
|
| 36 |
+
python demo_inference.py
|
| 37 |
+
|
| 38 |
+
# Image, custom file + prompt
|
| 39 |
+
python demo_inference.py --mode image --media /path/to/cat.jpg \
|
| 40 |
+
--prompt "What is the cat doing?"
|
| 41 |
+
|
| 42 |
+
# Video (16 uniformly-sampled frames; max-pixels caps per-frame resolution for memory)
|
| 43 |
+
python demo_inference.py --mode video --media /path/to/clip.mp4 \
|
| 44 |
+
--num-frames 16 --max-pixels 200704 \
|
| 45 |
+
--prompt "Describe what happens in this video."
|
| 46 |
+
```
|
| 47 |
+
|
| 48 |
+
## Programmatic use
|
| 49 |
+
|
| 50 |
+
```python
|
| 51 |
+
import torch
|
| 52 |
+
from transformers import AutoProcessor, AutoModelForImageTextToText
|
| 53 |
+
from PIL import Image
|
| 54 |
+
|
| 55 |
+
MODEL_ID = "lmms-lab-encoder/LLaVA-OneVision2-8B-Instruct"
|
| 56 |
+
|
| 57 |
+
processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
|
| 58 |
+
model = AutoModelForImageTextToText.from_pretrained(
|
| 59 |
+
MODEL_ID, trust_remote_code=True, dtype=torch.bfloat16, device_map="cuda",
|
| 60 |
+
).eval()
|
| 61 |
+
|
| 62 |
+
# ----- Image -----
|
| 63 |
+
image = Image.open("cat.jpg").convert("RGB")
|
| 64 |
+
messages = [{"role": "user", "content": [
|
| 65 |
+
{"type": "image"},
|
| 66 |
+
{"type": "text", "text": "Describe this image in detail."},
|
| 67 |
+
]}]
|
| 68 |
+
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 69 |
+
inputs = processor(text=[text], images=[image], return_tensors="pt", padding=True)
|
| 70 |
+
inputs = {k: v.to("cuda") if hasattr(v, "to") else v for k, v in inputs.items()}
|
| 71 |
+
|
| 72 |
+
out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
|
| 73 |
+
print(processor.tokenizer.decode(out[0, inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
|
| 74 |
+
|
| 75 |
+
# ----- Video -----
|
| 76 |
+
# Lower max_pixels if you hit OOM on long videos.
|
| 77 |
+
processor.video_processor.max_pixels = 200704
|
| 78 |
+
|
| 79 |
+
messages = [{"role": "user", "content": [
|
| 80 |
+
{"type": "video"},
|
| 81 |
+
{"type": "text", "text": "Describe what happens in this video."},
|
| 82 |
+
]}]
|
| 83 |
+
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 84 |
+
inputs = processor(
|
| 85 |
+
text=[text], videos=["clip.mp4"], return_tensors="pt", padding=True,
|
| 86 |
+
num_frames=16, # exact frame count; or use target_fps / max_frames
|
| 87 |
+
)
|
| 88 |
+
inputs = {k: v.to("cuda") if hasattr(v, "to") else v for k, v in inputs.items()}
|
| 89 |
+
out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
|
| 90 |
+
print(processor.tokenizer.decode(out[0, inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
|
| 91 |
+
```
|
| 92 |
+
|
| 93 |
+
## Notes
|
| 94 |
+
|
| 95 |
+
- The vision tower is a OneVision-style encoder; the language backbone is **Qwen3-8B**.
|
| 96 |
+
- `chat_template.jinja` follows the Qwen3 chat format and emits `<|vision_start|>...<|vision_end|>` placeholders; the processor expands them per-frame for video.
|
| 97 |
+
- Inference was validated to be bit-exact at the pixel level and prefix-identical at the token level against the original reference implementation.
|
| 98 |
+
|
| 99 |
+
## License
|
| 100 |
+
|
| 101 |
+
Apache-2.0 (model weights and code in this repository). The Qwen3-8B base is subject to its own license — see [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B).
|
added_tokens.json
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"</tool_call>": 151658,
|
| 3 |
+
"<tool_call>": 151657,
|
| 4 |
+
"<|box_end|>": 151649,
|
| 5 |
+
"<|box_start|>": 151648,
|
| 6 |
+
"<|endoftext|>": 151643,
|
| 7 |
+
"<|file_sep|>": 151664,
|
| 8 |
+
"<|fim_middle|>": 151660,
|
| 9 |
+
"<|fim_pad|>": 151662,
|
| 10 |
+
"<|fim_prefix|>": 151659,
|
| 11 |
+
"<|fim_suffix|>": 151661,
|
| 12 |
+
"<|im_end|>": 151645,
|
| 13 |
+
"<|im_start|>": 151644,
|
| 14 |
+
"<|image_pad|>": 151655,
|
| 15 |
+
"<|object_ref_end|>": 151647,
|
| 16 |
+
"<|object_ref_start|>": 151646,
|
| 17 |
+
"<|quad_end|>": 151651,
|
| 18 |
+
"<|quad_start|>": 151650,
|
| 19 |
+
"<|repo_name|>": 151663,
|
| 20 |
+
"<|video_pad|>": 151656,
|
| 21 |
+
"<|vision_end|>": 151653,
|
| 22 |
+
"<|vision_pad|>": 151654,
|
| 23 |
+
"<|vision_start|>": 151652
|
| 24 |
+
}
|
chat_template.jinja
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{% set image_count = namespace(value=0) %}{% set video_count = namespace(value=0) %}{% for message in messages %}{% if loop.first and message['role'] != 'system' %}<|im_start|>system
|
| 2 |
+
You are a helpful assistant.<|im_end|>
|
| 3 |
+
{% endif %}<|im_start|>{{ message['role'] }}
|
| 4 |
+
{% if message['content'] is string %}{{ message['content'] }}<|im_end|>
|
| 5 |
+
{% else %}{% for content in message['content'] %}{% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}{% set image_count.value = image_count.value + 1 %}{% if add_vision_id %}Picture {{ image_count.value }}: {% endif %}<|vision_start|><|image_pad|><|vision_end|>{% elif content['type'] == 'video' or 'video' in content %}{% set video_count.value = video_count.value + 1 %}{% if add_vision_id %}Video {{ video_count.value }}: {% endif %}<|vision_start|><|video_pad|><|vision_end|>{% elif 'text' in content %}{{ content['text'] }}{% endif %}{% endfor %}<|im_end|>
|
| 6 |
+
{% endif %}{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant
|
| 7 |
+
{% endif %}
|
config.json
ADDED
|
@@ -0,0 +1,115 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"LlavaOnevision2ForConditionalGeneration"
|
| 4 |
+
],
|
| 5 |
+
"auto_map": {
|
| 6 |
+
"AutoConfig": "configuration_llava_onevision2.LlavaOnevision2Config",
|
| 7 |
+
"AutoModel": "modeling_llava_onevision2.LlavaOnevision2Model",
|
| 8 |
+
"AutoModelForImageTextToText": "modeling_llava_onevision2.LlavaOnevision2ForConditionalGeneration",
|
| 9 |
+
"AutoProcessor": "processing_llava_onevision2.LlavaOnevision2Processor",
|
| 10 |
+
"AutoVideoProcessor": "video_processing_llava_onevision2.LlavaOnevision2VideoProcessor"
|
| 11 |
+
},
|
| 12 |
+
"bos_token_id": 151643,
|
| 13 |
+
"dtype": "bfloat16",
|
| 14 |
+
"eos_token_id": 151645,
|
| 15 |
+
"image_token_id": 151655,
|
| 16 |
+
"model_type": "llava_onevision2",
|
| 17 |
+
"text_config": {
|
| 18 |
+
"_name_or_path": "Qwen/Qwen3-8B",
|
| 19 |
+
"attention_bias": false,
|
| 20 |
+
"attention_dropout": 0.0,
|
| 21 |
+
"bos_token_id": 151643,
|
| 22 |
+
"dtype": "bfloat16",
|
| 23 |
+
"eos_token_id": 151645,
|
| 24 |
+
"head_dim": 128,
|
| 25 |
+
"hidden_act": "silu",
|
| 26 |
+
"hidden_size": 4096,
|
| 27 |
+
"initializer_range": 0.02,
|
| 28 |
+
"intermediate_size": 12288,
|
| 29 |
+
"layer_types": [
|
| 30 |
+
"full_attention",
|
| 31 |
+
"full_attention",
|
| 32 |
+
"full_attention",
|
| 33 |
+
"full_attention",
|
| 34 |
+
"full_attention",
|
| 35 |
+
"full_attention",
|
| 36 |
+
"full_attention",
|
| 37 |
+
"full_attention",
|
| 38 |
+
"full_attention",
|
| 39 |
+
"full_attention",
|
| 40 |
+
"full_attention",
|
| 41 |
+
"full_attention",
|
| 42 |
+
"full_attention",
|
| 43 |
+
"full_attention",
|
| 44 |
+
"full_attention",
|
| 45 |
+
"full_attention",
|
| 46 |
+
"full_attention",
|
| 47 |
+
"full_attention",
|
| 48 |
+
"full_attention",
|
| 49 |
+
"full_attention",
|
| 50 |
+
"full_attention",
|
| 51 |
+
"full_attention",
|
| 52 |
+
"full_attention",
|
| 53 |
+
"full_attention",
|
| 54 |
+
"full_attention",
|
| 55 |
+
"full_attention",
|
| 56 |
+
"full_attention",
|
| 57 |
+
"full_attention",
|
| 58 |
+
"full_attention",
|
| 59 |
+
"full_attention",
|
| 60 |
+
"full_attention",
|
| 61 |
+
"full_attention",
|
| 62 |
+
"full_attention",
|
| 63 |
+
"full_attention",
|
| 64 |
+
"full_attention",
|
| 65 |
+
"full_attention"
|
| 66 |
+
],
|
| 67 |
+
"max_position_embeddings": 262144,
|
| 68 |
+
"max_window_layers": 36,
|
| 69 |
+
"model_type": "qwen3",
|
| 70 |
+
"num_attention_heads": 32,
|
| 71 |
+
"num_hidden_layers": 36,
|
| 72 |
+
"num_key_value_heads": 8,
|
| 73 |
+
"pad_token_id": null,
|
| 74 |
+
"rms_norm_eps": 1e-06,
|
| 75 |
+
"rope_parameters": {
|
| 76 |
+
"rope_theta": 8000000,
|
| 77 |
+
"rope_type": "default"
|
| 78 |
+
},
|
| 79 |
+
"sliding_window": null,
|
| 80 |
+
"tie_word_embeddings": false,
|
| 81 |
+
"use_cache": true,
|
| 82 |
+
"use_sliding_window": false,
|
| 83 |
+
"vocab_size": 151936
|
| 84 |
+
},
|
| 85 |
+
"tie_word_embeddings": false,
|
| 86 |
+
"transformers_version": "5.7.0",
|
| 87 |
+
"video_token_id": 151656,
|
| 88 |
+
"vision_config": {
|
| 89 |
+
"attention_dropout": 0.0,
|
| 90 |
+
"frame_windows_size": 4,
|
| 91 |
+
"hidden_act": "gelu",
|
| 92 |
+
"hidden_size": 1024,
|
| 93 |
+
"image_size": 448,
|
| 94 |
+
"initializer_range": 0.02,
|
| 95 |
+
"intermediate_size": 4096,
|
| 96 |
+
"layer_norm_eps": 1e-06,
|
| 97 |
+
"layer_norm_type": "layer_norm",
|
| 98 |
+
"max_position_embeddings": 8192,
|
| 99 |
+
"model_type": "onevision_encoder",
|
| 100 |
+
"num_attention_heads": 16,
|
| 101 |
+
"num_channels": 3,
|
| 102 |
+
"num_hidden_layers": 24,
|
| 103 |
+
"out_hidden_size": 4096,
|
| 104 |
+
"patch_position_encoding_type": "absolute",
|
| 105 |
+
"patch_size": 14,
|
| 106 |
+
"rope_theta": 10000.0,
|
| 107 |
+
"spatial_merge_size": 2,
|
| 108 |
+
"text_hidden_size": 4096,
|
| 109 |
+
"tokens_per_second": 1,
|
| 110 |
+
"use_head": false,
|
| 111 |
+
"use_patch_position_encoding": false
|
| 112 |
+
},
|
| 113 |
+
"vision_end_token_id": 151653,
|
| 114 |
+
"vision_start_token_id": 151652
|
| 115 |
+
}
|
configuration_llava_onevision2.py
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from huggingface_hub.dataclasses import strict
|
| 2 |
+
|
| 3 |
+
from transformers import CONFIG_MAPPING, AutoConfig
|
| 4 |
+
from transformers.configuration_utils import PreTrainedConfig
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
@strict
|
| 8 |
+
class LlavaOnevision2VisionConfig(PreTrainedConfig):
|
| 9 |
+
model_type = "onevision_encoder"
|
| 10 |
+
base_config_key = "vision_config"
|
| 11 |
+
|
| 12 |
+
hidden_size: int = 1024
|
| 13 |
+
intermediate_size: int = 4096
|
| 14 |
+
num_hidden_layers: int = 24
|
| 15 |
+
num_attention_heads: int = 16
|
| 16 |
+
num_channels: int = 3
|
| 17 |
+
image_size: int = 448
|
| 18 |
+
patch_size: int = 14
|
| 19 |
+
hidden_act: str = "gelu"
|
| 20 |
+
layer_norm_eps: float = 1e-6
|
| 21 |
+
layer_norm_type: str = "layer_norm"
|
| 22 |
+
attention_dropout: float = 0.0
|
| 23 |
+
initializer_range: float = 0.02
|
| 24 |
+
rope_theta: float = 10000.0
|
| 25 |
+
use_head: bool = False
|
| 26 |
+
out_hidden_size: int = 1024
|
| 27 |
+
spatial_merge_size: int = 2
|
| 28 |
+
tokens_per_second: int = 1
|
| 29 |
+
frame_windows_size: int = 4
|
| 30 |
+
use_patch_position_encoding: bool = False
|
| 31 |
+
patch_position_encoding_type: str = "absolute"
|
| 32 |
+
max_position_embeddings: int = 8192
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
@strict
|
| 36 |
+
class LlavaOnevision2Config(PreTrainedConfig):
|
| 37 |
+
r"""
|
| 38 |
+
This is the configuration class to store the configuration of a [`LlavaOnevision2Model`]. It is used to instantiate a
|
| 39 |
+
LlavaOnevision2Model model according to the specified arguments, defining the model architecture. Instantiating a configuration
|
| 40 |
+
with the defaults will yield a similar configuration to that of
|
| 41 |
+
Llava-Onevision 1.5 [lmms-lab/LLaVA-OneVision-1.5-8B-Instruct](https://huggingface.co/lmms-lab/LLaVA-OneVision-1.5-8B-Instruct).
|
| 42 |
+
|
| 43 |
+
Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
|
| 44 |
+
documentation from [`PreTrainedConfig`] for more information.
|
| 45 |
+
|
| 46 |
+
Args:
|
| 47 |
+
text_config (`Union[PreTrainedConfig, dict]`, *optional*, defaults to `Qwen3Config`):
|
| 48 |
+
The config object or dictionary of the text backbone.
|
| 49 |
+
vision_config (`Union[PreTrainedConfig, dict]`, *optional*, defaults to `LlavaOnevision2VisionConfig`):
|
| 50 |
+
The config object or dictionary of the vision backbone.
|
| 51 |
+
image_token_id (`int`, *optional*, defaults to 151655):
|
| 52 |
+
The image token index to encode the image prompt.
|
| 53 |
+
video_token_id (`int`, *optional*, defaults to 151656):
|
| 54 |
+
The video token index to encode the image prompt.
|
| 55 |
+
vision_start_token_id (`int`, *optional*, defaults to 151652):
|
| 56 |
+
The token index to denote start of vision input.
|
| 57 |
+
vision_end_token_id (`int`, *optional*, defaults to 151653):
|
| 58 |
+
The token index to denote end of vision input.
|
| 59 |
+
"""
|
| 60 |
+
|
| 61 |
+
model_type = "llava_onevision2"
|
| 62 |
+
# `text_config` is resolved dynamically based on its `model_type` (defaults to `qwen3`),
|
| 63 |
+
# so we use `AutoConfig` here as a placeholder; `__post_init__` swaps it for the
|
| 64 |
+
# concrete config class via `CONFIG_MAPPING`.
|
| 65 |
+
sub_configs = {"vision_config": LlavaOnevision2VisionConfig, "text_config": AutoConfig}
|
| 66 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
| 67 |
+
|
| 68 |
+
text_config: dict | PreTrainedConfig | None = None
|
| 69 |
+
vision_config: dict | PreTrainedConfig | None = None
|
| 70 |
+
image_token_id: int = 151655
|
| 71 |
+
video_token_id: int = 151656
|
| 72 |
+
vision_start_token_id: int = 151652
|
| 73 |
+
vision_end_token_id: int = 151653
|
| 74 |
+
tie_word_embeddings: bool = False
|
| 75 |
+
# Generation-related token ids are mirrored from `text_config` in `__post_init__`
|
| 76 |
+
# so downstream tools (e.g. `generate`, vLLM) that read them at the top level keep working.
|
| 77 |
+
bos_token_id: int | None = None
|
| 78 |
+
eos_token_id: int | list[int] | None = None
|
| 79 |
+
pad_token_id: int | None = None
|
| 80 |
+
|
| 81 |
+
def __post_init__(self, **kwargs):
|
| 82 |
+
# Resolve vision_config
|
| 83 |
+
if isinstance(self.vision_config, dict):
|
| 84 |
+
self.vision_config = self.sub_configs["vision_config"](**self.vision_config)
|
| 85 |
+
elif self.vision_config is None:
|
| 86 |
+
self.vision_config = self.sub_configs["vision_config"]()
|
| 87 |
+
|
| 88 |
+
# Resolve text_config dynamically via CONFIG_MAPPING (defaults to qwen3)
|
| 89 |
+
if isinstance(self.text_config, dict):
|
| 90 |
+
text_model_type = self.text_config.get("model_type", "qwen3")
|
| 91 |
+
self.text_config["model_type"] = text_model_type
|
| 92 |
+
text_config_cls = CONFIG_MAPPING[text_model_type]
|
| 93 |
+
self.sub_configs["text_config"] = text_config_cls
|
| 94 |
+
self.text_config = text_config_cls(**self.text_config)
|
| 95 |
+
elif self.text_config is None:
|
| 96 |
+
text_config_cls = CONFIG_MAPPING["qwen3"]
|
| 97 |
+
self.sub_configs["text_config"] = text_config_cls
|
| 98 |
+
self.text_config = text_config_cls()
|
| 99 |
+
|
| 100 |
+
# Mirror generation-related token ids from text_config to the top level so
|
| 101 |
+
# downstream tools (e.g. `generate`, chat templates, vLLM) that read them
|
| 102 |
+
# from the top-level config keep working.
|
| 103 |
+
for tok_key in ("bos_token_id", "eos_token_id", "pad_token_id"):
|
| 104 |
+
text_val = getattr(self.text_config, tok_key, None)
|
| 105 |
+
if text_val is not None and getattr(self, tok_key, None) is None:
|
| 106 |
+
setattr(self, tok_key, text_val)
|
| 107 |
+
|
| 108 |
+
super().__post_init__(**kwargs)
|
| 109 |
+
|
| 110 |
+
|
| 111 |
+
__all__ = ["LlavaOnevision2Config", "LlavaOnevision2VisionConfig"]
|
demo_inference.py
ADDED
|
@@ -0,0 +1,269 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""End-to-end inference demo for LlavaOnevision2 (image + video).
|
| 2 |
+
|
| 3 |
+
This script shows the two canonical inference paths supported by the model:
|
| 4 |
+
|
| 5 |
+
* Image captioning (``--mode image``, default)
|
| 6 |
+
* Video captioning (``--mode video``)
|
| 7 |
+
|
| 8 |
+
Both modes share the same loading pattern:
|
| 9 |
+
|
| 10 |
+
from transformers import AutoProcessor, AutoModelForImageTextToText
|
| 11 |
+
processor = AutoProcessor.from_pretrained(model_dir, trust_remote_code=True)
|
| 12 |
+
model = AutoModelForImageTextToText.from_pretrained(
|
| 13 |
+
model_dir, trust_remote_code=True, dtype=torch.bfloat16, device_map="cuda",
|
| 14 |
+
)
|
| 15 |
+
|
| 16 |
+
Examples
|
| 17 |
+
--------
|
| 18 |
+
# Image (default sample image from the web)
|
| 19 |
+
python demo_inference.py
|
| 20 |
+
|
| 21 |
+
# Image with a local file and a custom prompt
|
| 22 |
+
python demo_inference.py --mode image --media /path/to/cat.jpg \
|
| 23 |
+
--prompt "What is the cat doing?"
|
| 24 |
+
|
| 25 |
+
# Video
|
| 26 |
+
# - ``--num-frames`` selects exactly N frames (uniform sampling).
|
| 27 |
+
# - ``--max-pixels`` caps each frame's pixel budget. Lower it to fit smaller
|
| 28 |
+
# GPUs; 200704 (=448*448) is a safe default for a single ~80GB card.
|
| 29 |
+
python demo_inference.py --mode video --media /path/to/clip.mp4 \
|
| 30 |
+
--num-frames 16 --max-pixels 200704 \
|
| 31 |
+
--prompt "Describe what happens in this video."
|
| 32 |
+
|
| 33 |
+
Tested with:
|
| 34 |
+
transformers == 5.7.0
|
| 35 |
+
torch >= 2.4
|
| 36 |
+
decord, Pillow, requests
|
| 37 |
+
"""
|
| 38 |
+
|
| 39 |
+
from __future__ import annotations
|
| 40 |
+
|
| 41 |
+
import argparse
|
| 42 |
+
import io
|
| 43 |
+
import os
|
| 44 |
+
import sys
|
| 45 |
+
|
| 46 |
+
import torch
|
| 47 |
+
|
| 48 |
+
# Placeholder constants so the user can swap their own media in easily.
|
| 49 |
+
# (Public sample image from the transformers project; no auth required.)
|
| 50 |
+
DEFAULT_IMAGE_URL = "https://www.ilankelman.org/stopsigns/australia.jpg"
|
| 51 |
+
DEFAULT_VIDEO_PATH = "/path/to/your/video.mp4" # <-- replace me
|
| 52 |
+
|
| 53 |
+
DEFAULT_IMAGE_PROMPT = "Describe this image in detail."
|
| 54 |
+
DEFAULT_VIDEO_PROMPT = "Describe what happens in this video in detail."
|
| 55 |
+
|
| 56 |
+
# Default model. Override with ``--model /local/path`` to use a local checkpoint.
|
| 57 |
+
DEFAULT_MODEL = "lmms-lab-encoder/LLaVA-OneVision2-8B-Instruct"
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
def load_image(source: str):
|
| 61 |
+
"""Load a PIL image from a local path or an http(s) URL."""
|
| 62 |
+
from PIL import Image
|
| 63 |
+
|
| 64 |
+
if source.startswith(("http://", "https://")):
|
| 65 |
+
import requests
|
| 66 |
+
|
| 67 |
+
resp = requests.get(source, stream=True, timeout=30)
|
| 68 |
+
resp.raise_for_status()
|
| 69 |
+
img = Image.open(io.BytesIO(resp.content))
|
| 70 |
+
else:
|
| 71 |
+
img = Image.open(source)
|
| 72 |
+
return img.convert("RGB")
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
def run_image(model, processor, media: str, prompt: str, max_new_tokens: int, device: str) -> str:
|
| 76 |
+
"""Caption a single image."""
|
| 77 |
+
image = load_image(media)
|
| 78 |
+
|
| 79 |
+
messages = [
|
| 80 |
+
{
|
| 81 |
+
"role": "user",
|
| 82 |
+
"content": [
|
| 83 |
+
{"type": "image"},
|
| 84 |
+
{"type": "text", "text": prompt},
|
| 85 |
+
],
|
| 86 |
+
}
|
| 87 |
+
]
|
| 88 |
+
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 89 |
+
|
| 90 |
+
inputs = processor(
|
| 91 |
+
text=[text],
|
| 92 |
+
images=[image],
|
| 93 |
+
return_tensors="pt",
|
| 94 |
+
padding=True,
|
| 95 |
+
)
|
| 96 |
+
inputs = {k: (v.to(device) if isinstance(v, torch.Tensor) else v) for k, v in inputs.items()}
|
| 97 |
+
|
| 98 |
+
tok = processor.tokenizer
|
| 99 |
+
pad_id = tok.pad_token_id or tok.eos_token_id
|
| 100 |
+
with torch.inference_mode():
|
| 101 |
+
out_ids = model.generate(
|
| 102 |
+
**inputs,
|
| 103 |
+
max_new_tokens=max_new_tokens,
|
| 104 |
+
do_sample=False,
|
| 105 |
+
num_beams=1,
|
| 106 |
+
use_cache=True,
|
| 107 |
+
eos_token_id=tok.eos_token_id,
|
| 108 |
+
pad_token_id=pad_id,
|
| 109 |
+
)
|
| 110 |
+
prompt_len = inputs["input_ids"].shape[-1]
|
| 111 |
+
new_ids = out_ids[:, prompt_len:]
|
| 112 |
+
return tok.batch_decode(new_ids, skip_special_tokens=True)[0].strip()
|
| 113 |
+
|
| 114 |
+
|
| 115 |
+
def run_video(
|
| 116 |
+
model,
|
| 117 |
+
processor,
|
| 118 |
+
media: str,
|
| 119 |
+
prompt: str,
|
| 120 |
+
max_new_tokens: int,
|
| 121 |
+
device: str,
|
| 122 |
+
num_frames: int,
|
| 123 |
+
max_pixels: int,
|
| 124 |
+
) -> str:
|
| 125 |
+
"""Caption an mp4/avi/... video file.
|
| 126 |
+
|
| 127 |
+
Key processor knobs (all passed through ``__call__``):
|
| 128 |
+
* ``num_frames`` : force exactly N uniformly-sampled frames.
|
| 129 |
+
* ``max_frames`` : cap on auto-selected frame count (used when num_frames is None).
|
| 130 |
+
* ``target_fps`` : sample at this FPS, capped by ``max_frames``.
|
| 131 |
+
|
| 132 |
+
For memory control, lower the per-frame resolution by overriding
|
| 133 |
+
``processor.video_processor.max_pixels`` before calling the processor.
|
| 134 |
+
"""
|
| 135 |
+
if not os.path.exists(media):
|
| 136 |
+
raise FileNotFoundError(
|
| 137 |
+
f"Video file not found: {media!r}. Pass --media <path/to/video.mp4>."
|
| 138 |
+
)
|
| 139 |
+
|
| 140 |
+
# Constrain per-frame pixel budget (memory-friendly default for a single ~80GB GPU).
|
| 141 |
+
processor.video_processor.max_pixels = max_pixels
|
| 142 |
+
|
| 143 |
+
messages = [
|
| 144 |
+
{
|
| 145 |
+
"role": "user",
|
| 146 |
+
"content": [
|
| 147 |
+
{"type": "video"},
|
| 148 |
+
{"type": "text", "text": prompt},
|
| 149 |
+
],
|
| 150 |
+
}
|
| 151 |
+
]
|
| 152 |
+
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 153 |
+
|
| 154 |
+
inputs = processor(
|
| 155 |
+
text=[text],
|
| 156 |
+
videos=[media],
|
| 157 |
+
return_tensors="pt",
|
| 158 |
+
padding=True,
|
| 159 |
+
num_frames=num_frames, # force exactly N frames
|
| 160 |
+
)
|
| 161 |
+
inputs = {k: (v.to(device) if isinstance(v, torch.Tensor) else v) for k, v in inputs.items()}
|
| 162 |
+
|
| 163 |
+
tok = processor.tokenizer
|
| 164 |
+
pad_id = tok.pad_token_id or tok.eos_token_id
|
| 165 |
+
with torch.inference_mode():
|
| 166 |
+
out_ids = model.generate(
|
| 167 |
+
**inputs,
|
| 168 |
+
max_new_tokens=max_new_tokens,
|
| 169 |
+
do_sample=False,
|
| 170 |
+
num_beams=1,
|
| 171 |
+
use_cache=True,
|
| 172 |
+
eos_token_id=tok.eos_token_id,
|
| 173 |
+
pad_token_id=pad_id,
|
| 174 |
+
)
|
| 175 |
+
prompt_len = inputs["input_ids"].shape[-1]
|
| 176 |
+
new_ids = out_ids[:, prompt_len:]
|
| 177 |
+
return tok.batch_decode(new_ids, skip_special_tokens=True)[0].strip()
|
| 178 |
+
|
| 179 |
+
|
| 180 |
+
def main():
|
| 181 |
+
parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
|
| 182 |
+
parser.add_argument(
|
| 183 |
+
"--model",
|
| 184 |
+
default=DEFAULT_MODEL,
|
| 185 |
+
help=f"HF repo id or local path to the model checkpoint (default: {DEFAULT_MODEL}).",
|
| 186 |
+
)
|
| 187 |
+
parser.add_argument(
|
| 188 |
+
"--mode",
|
| 189 |
+
choices=["image", "video"],
|
| 190 |
+
default="image",
|
| 191 |
+
help="Inference mode (default: image).",
|
| 192 |
+
)
|
| 193 |
+
parser.add_argument(
|
| 194 |
+
"--media",
|
| 195 |
+
default=None,
|
| 196 |
+
help=(
|
| 197 |
+
"Image path/URL (image mode) or video path (video mode). "
|
| 198 |
+
f"Defaults: image={DEFAULT_IMAGE_URL!r}, video={DEFAULT_VIDEO_PATH!r}."
|
| 199 |
+
),
|
| 200 |
+
)
|
| 201 |
+
parser.add_argument("--prompt", default=None, help="User prompt sent alongside the media.")
|
| 202 |
+
parser.add_argument("--max-new-tokens", type=int, default=256)
|
| 203 |
+
parser.add_argument(
|
| 204 |
+
"--device",
|
| 205 |
+
default="cuda" if torch.cuda.is_available() else "cpu",
|
| 206 |
+
help="Device to load the model on.",
|
| 207 |
+
)
|
| 208 |
+
parser.add_argument(
|
| 209 |
+
"--dtype",
|
| 210 |
+
default="bfloat16",
|
| 211 |
+
choices=["bfloat16", "float16", "float32"],
|
| 212 |
+
help="Model dtype.",
|
| 213 |
+
)
|
| 214 |
+
# Video-only knobs (ignored in image mode).
|
| 215 |
+
parser.add_argument(
|
| 216 |
+
"--num-frames",
|
| 217 |
+
type=int,
|
| 218 |
+
default=16,
|
| 219 |
+
help="[video] Number of frames to sample (default: 16).",
|
| 220 |
+
)
|
| 221 |
+
parser.add_argument(
|
| 222 |
+
"--max-pixels",
|
| 223 |
+
type=int,
|
| 224 |
+
default=200704,
|
| 225 |
+
help="[video] Per-frame max pixel count (default: 200704 = 448*448).",
|
| 226 |
+
)
|
| 227 |
+
args = parser.parse_args()
|
| 228 |
+
|
| 229 |
+
# Defaults that depend on mode.
|
| 230 |
+
if args.media is None:
|
| 231 |
+
args.media = DEFAULT_IMAGE_URL if args.mode == "image" else DEFAULT_VIDEO_PATH
|
| 232 |
+
if args.prompt is None:
|
| 233 |
+
args.prompt = DEFAULT_IMAGE_PROMPT if args.mode == "image" else DEFAULT_VIDEO_PROMPT
|
| 234 |
+
|
| 235 |
+
dtype = getattr(torch, args.dtype)
|
| 236 |
+
|
| 237 |
+
from transformers import AutoModelForImageTextToText, AutoProcessor
|
| 238 |
+
|
| 239 |
+
print(f"[demo_inference] Loading processor from: {args.model}", flush=True)
|
| 240 |
+
processor = AutoProcessor.from_pretrained(args.model, trust_remote_code=True)
|
| 241 |
+
|
| 242 |
+
print(f"[demo_inference] Loading model on {args.device} ({args.dtype})...", flush=True)
|
| 243 |
+
model = AutoModelForImageTextToText.from_pretrained(
|
| 244 |
+
args.model,
|
| 245 |
+
trust_remote_code=True,
|
| 246 |
+
dtype=dtype,
|
| 247 |
+
device_map=args.device,
|
| 248 |
+
)
|
| 249 |
+
model.eval()
|
| 250 |
+
|
| 251 |
+
print(f"[demo_inference] Mode={args.mode} media={args.media}", flush=True)
|
| 252 |
+
if args.mode == "image":
|
| 253 |
+
caption = run_image(
|
| 254 |
+
model, processor, args.media, args.prompt, args.max_new_tokens, args.device,
|
| 255 |
+
)
|
| 256 |
+
else:
|
| 257 |
+
caption = run_video(
|
| 258 |
+
model, processor, args.media, args.prompt, args.max_new_tokens, args.device,
|
| 259 |
+
num_frames=args.num_frames, max_pixels=args.max_pixels,
|
| 260 |
+
)
|
| 261 |
+
|
| 262 |
+
print("\n========== OUTPUT ==========")
|
| 263 |
+
print(caption)
|
| 264 |
+
print("============================")
|
| 265 |
+
return 0
|
| 266 |
+
|
| 267 |
+
|
| 268 |
+
if __name__ == "__main__":
|
| 269 |
+
sys.exit(main())
|
generation_config.json
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_from_model_config": true,
|
| 3 |
+
"bos_token_id": 151643,
|
| 4 |
+
"eos_token_id": 151645,
|
| 5 |
+
"transformers_version": "5.7.0"
|
| 6 |
+
}
|
merges.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
model-00001-of-00004.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6510d6eb5c999d9f3c869f127016bd8bada90269c1e76d0bea6ab59ffe3ac876
|
| 3 |
+
size 5288213408
|
model-00002-of-00004.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e927a5f586ab4e43b0d13e29c0ae43d9b838850ebce106fc5d290507a8c4b3c3
|
| 3 |
+
size 4546819264
|
model-00003-of-00004.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:608b6d6d544935773de4653cb2c560c12fe2899b1d6bc7602f58ea2dbecb9fdc
|
| 3 |
+
size 5346916160
|
model-00004-of-00004.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:99ef8daea1b127d97b724add790b1dfd28f0ae12a81a22cdfdf2f38c92177538
|
| 3 |
+
size 1872566720
|
model.safetensors.index.json
ADDED
|
@@ -0,0 +1,703 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"metadata": {
|
| 3 |
+
"total_size": 17054427136
|
| 4 |
+
},
|
| 5 |
+
"weight_map": {
|
| 6 |
+
"model.language_model.embed_tokens.weight": "model-00001-of-00004.safetensors",
|
| 7 |
+
"model.language_model.layers.0.input_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 8 |
+
"model.language_model.layers.0.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
|
| 9 |
+
"model.language_model.layers.0.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 10 |
+
"model.language_model.layers.0.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
|
| 11 |
+
"model.language_model.layers.0.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 12 |
+
"model.language_model.layers.0.self_attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 13 |
+
"model.language_model.layers.0.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
|
| 14 |
+
"model.language_model.layers.0.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
|
| 15 |
+
"model.language_model.layers.0.self_attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 16 |
+
"model.language_model.layers.0.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
|
| 17 |
+
"model.language_model.layers.0.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
|
| 18 |
+
"model.language_model.layers.1.input_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 19 |
+
"model.language_model.layers.1.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
|
| 20 |
+
"model.language_model.layers.1.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 21 |
+
"model.language_model.layers.1.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
|
| 22 |
+
"model.language_model.layers.1.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 23 |
+
"model.language_model.layers.1.self_attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 24 |
+
"model.language_model.layers.1.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
|
| 25 |
+
"model.language_model.layers.1.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
|
| 26 |
+
"model.language_model.layers.1.self_attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 27 |
+
"model.language_model.layers.1.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
|
| 28 |
+
"model.language_model.layers.1.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
|
| 29 |
+
"model.language_model.layers.10.input_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 30 |
+
"model.language_model.layers.10.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
|
| 31 |
+
"model.language_model.layers.10.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 32 |
+
"model.language_model.layers.10.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
|
| 33 |
+
"model.language_model.layers.10.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 34 |
+
"model.language_model.layers.10.self_attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 35 |
+
"model.language_model.layers.10.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
|
| 36 |
+
"model.language_model.layers.10.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
|
| 37 |
+
"model.language_model.layers.10.self_attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 38 |
+
"model.language_model.layers.10.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
|
| 39 |
+
"model.language_model.layers.10.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
|
| 40 |
+
"model.language_model.layers.11.input_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 41 |
+
"model.language_model.layers.11.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
|
| 42 |
+
"model.language_model.layers.11.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 43 |
+
"model.language_model.layers.11.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
|
| 44 |
+
"model.language_model.layers.11.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 45 |
+
"model.language_model.layers.11.self_attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 46 |
+
"model.language_model.layers.11.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
|
| 47 |
+
"model.language_model.layers.11.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
|
| 48 |
+
"model.language_model.layers.11.self_attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 49 |
+
"model.language_model.layers.11.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
|
| 50 |
+
"model.language_model.layers.11.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
|
| 51 |
+
"model.language_model.layers.12.input_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 52 |
+
"model.language_model.layers.12.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
|
| 53 |
+
"model.language_model.layers.2.input_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 54 |
+
"model.language_model.layers.2.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
|
| 55 |
+
"model.language_model.layers.2.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 56 |
+
"model.language_model.layers.2.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
|
| 57 |
+
"model.language_model.layers.2.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 58 |
+
"model.language_model.layers.2.self_attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 59 |
+
"model.language_model.layers.2.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
|
| 60 |
+
"model.language_model.layers.2.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
|
| 61 |
+
"model.language_model.layers.2.self_attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 62 |
+
"model.language_model.layers.2.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
|
| 63 |
+
"model.language_model.layers.2.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
|
| 64 |
+
"model.language_model.layers.3.input_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 65 |
+
"model.language_model.layers.3.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
|
| 66 |
+
"model.language_model.layers.3.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 67 |
+
"model.language_model.layers.3.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
|
| 68 |
+
"model.language_model.layers.3.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 69 |
+
"model.language_model.layers.3.self_attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 70 |
+
"model.language_model.layers.3.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
|
| 71 |
+
"model.language_model.layers.3.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
|
| 72 |
+
"model.language_model.layers.3.self_attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 73 |
+
"model.language_model.layers.3.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
|
| 74 |
+
"model.language_model.layers.3.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
|
| 75 |
+
"model.language_model.layers.4.input_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 76 |
+
"model.language_model.layers.4.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
|
| 77 |
+
"model.language_model.layers.4.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 78 |
+
"model.language_model.layers.4.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
|
| 79 |
+
"model.language_model.layers.4.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 80 |
+
"model.language_model.layers.4.self_attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 81 |
+
"model.language_model.layers.4.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
|
| 82 |
+
"model.language_model.layers.4.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
|
| 83 |
+
"model.language_model.layers.4.self_attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 84 |
+
"model.language_model.layers.4.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
|
| 85 |
+
"model.language_model.layers.4.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
|
| 86 |
+
"model.language_model.layers.5.input_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 87 |
+
"model.language_model.layers.5.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
|
| 88 |
+
"model.language_model.layers.5.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 89 |
+
"model.language_model.layers.5.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
|
| 90 |
+
"model.language_model.layers.5.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 91 |
+
"model.language_model.layers.5.self_attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 92 |
+
"model.language_model.layers.5.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
|
| 93 |
+
"model.language_model.layers.5.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
|
| 94 |
+
"model.language_model.layers.5.self_attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 95 |
+
"model.language_model.layers.5.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
|
| 96 |
+
"model.language_model.layers.5.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
|
| 97 |
+
"model.language_model.layers.6.input_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 98 |
+
"model.language_model.layers.6.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
|
| 99 |
+
"model.language_model.layers.6.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 100 |
+
"model.language_model.layers.6.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
|
| 101 |
+
"model.language_model.layers.6.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 102 |
+
"model.language_model.layers.6.self_attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 103 |
+
"model.language_model.layers.6.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
|
| 104 |
+
"model.language_model.layers.6.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
|
| 105 |
+
"model.language_model.layers.6.self_attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 106 |
+
"model.language_model.layers.6.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
|
| 107 |
+
"model.language_model.layers.6.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
|
| 108 |
+
"model.language_model.layers.7.input_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 109 |
+
"model.language_model.layers.7.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 110 |
+
"model.language_model.layers.7.self_attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 111 |
+
"model.language_model.layers.7.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
|
| 112 |
+
"model.language_model.layers.7.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
|
| 113 |
+
"model.language_model.layers.7.self_attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 114 |
+
"model.language_model.layers.7.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
|
| 115 |
+
"model.language_model.layers.7.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
|
| 116 |
+
"model.language_model.layers.8.input_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 117 |
+
"model.language_model.layers.8.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 118 |
+
"model.language_model.layers.8.self_attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 119 |
+
"model.language_model.layers.8.self_attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 120 |
+
"model.language_model.layers.9.input_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 121 |
+
"model.language_model.layers.9.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 122 |
+
"model.language_model.layers.9.self_attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 123 |
+
"model.language_model.layers.9.self_attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 124 |
+
"model.language_model.layers.12.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 125 |
+
"model.language_model.layers.12.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
|
| 126 |
+
"model.language_model.layers.12.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 127 |
+
"model.language_model.layers.12.self_attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 128 |
+
"model.language_model.layers.12.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
|
| 129 |
+
"model.language_model.layers.12.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
|
| 130 |
+
"model.language_model.layers.12.self_attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 131 |
+
"model.language_model.layers.12.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
|
| 132 |
+
"model.language_model.layers.12.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
|
| 133 |
+
"model.language_model.layers.13.input_layernorm.weight": "model-00001-of-00004.safetensors",
|
| 134 |
+
"model.language_model.layers.13.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
|
| 135 |
+
"model.language_model.layers.13.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 136 |
+
"model.language_model.layers.13.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
|
| 137 |
+
"model.language_model.layers.13.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 138 |
+
"model.language_model.layers.13.self_attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 139 |
+
"model.language_model.layers.13.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
|
| 140 |
+
"model.language_model.layers.13.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
|
| 141 |
+
"model.language_model.layers.13.self_attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 142 |
+
"model.language_model.layers.13.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
|
| 143 |
+
"model.language_model.layers.13.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
|
| 144 |
+
"model.language_model.layers.14.input_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 145 |
+
"model.language_model.layers.14.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
|
| 146 |
+
"model.language_model.layers.14.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 147 |
+
"model.language_model.layers.14.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
|
| 148 |
+
"model.language_model.layers.14.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 149 |
+
"model.language_model.layers.14.self_attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 150 |
+
"model.language_model.layers.14.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
|
| 151 |
+
"model.language_model.layers.14.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
|
| 152 |
+
"model.language_model.layers.14.self_attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 153 |
+
"model.language_model.layers.14.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
|
| 154 |
+
"model.language_model.layers.14.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
|
| 155 |
+
"model.language_model.layers.15.input_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 156 |
+
"model.language_model.layers.15.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
|
| 157 |
+
"model.language_model.layers.15.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 158 |
+
"model.language_model.layers.15.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
|
| 159 |
+
"model.language_model.layers.15.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 160 |
+
"model.language_model.layers.15.self_attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 161 |
+
"model.language_model.layers.15.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
|
| 162 |
+
"model.language_model.layers.15.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
|
| 163 |
+
"model.language_model.layers.15.self_attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 164 |
+
"model.language_model.layers.15.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
|
| 165 |
+
"model.language_model.layers.15.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
|
| 166 |
+
"model.language_model.layers.16.input_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 167 |
+
"model.language_model.layers.16.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
|
| 168 |
+
"model.language_model.layers.16.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 169 |
+
"model.language_model.layers.16.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
|
| 170 |
+
"model.language_model.layers.16.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 171 |
+
"model.language_model.layers.16.self_attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 172 |
+
"model.language_model.layers.16.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
|
| 173 |
+
"model.language_model.layers.16.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
|
| 174 |
+
"model.language_model.layers.16.self_attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 175 |
+
"model.language_model.layers.16.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
|
| 176 |
+
"model.language_model.layers.16.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
|
| 177 |
+
"model.language_model.layers.17.input_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 178 |
+
"model.language_model.layers.17.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
|
| 179 |
+
"model.language_model.layers.17.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 180 |
+
"model.language_model.layers.17.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
|
| 181 |
+
"model.language_model.layers.17.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 182 |
+
"model.language_model.layers.17.self_attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 183 |
+
"model.language_model.layers.17.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
|
| 184 |
+
"model.language_model.layers.17.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
|
| 185 |
+
"model.language_model.layers.17.self_attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 186 |
+
"model.language_model.layers.17.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
|
| 187 |
+
"model.language_model.layers.17.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
|
| 188 |
+
"model.language_model.layers.18.input_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 189 |
+
"model.language_model.layers.18.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
|
| 190 |
+
"model.language_model.layers.18.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 191 |
+
"model.language_model.layers.18.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
|
| 192 |
+
"model.language_model.layers.18.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 193 |
+
"model.language_model.layers.18.self_attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 194 |
+
"model.language_model.layers.18.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
|
| 195 |
+
"model.language_model.layers.18.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
|
| 196 |
+
"model.language_model.layers.18.self_attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 197 |
+
"model.language_model.layers.18.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
|
| 198 |
+
"model.language_model.layers.18.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
|
| 199 |
+
"model.language_model.layers.19.input_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 200 |
+
"model.language_model.layers.19.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
|
| 201 |
+
"model.language_model.layers.19.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 202 |
+
"model.language_model.layers.19.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
|
| 203 |
+
"model.language_model.layers.19.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 204 |
+
"model.language_model.layers.19.self_attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 205 |
+
"model.language_model.layers.19.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
|
| 206 |
+
"model.language_model.layers.19.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
|
| 207 |
+
"model.language_model.layers.19.self_attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 208 |
+
"model.language_model.layers.19.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
|
| 209 |
+
"model.language_model.layers.19.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
|
| 210 |
+
"model.language_model.layers.20.input_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 211 |
+
"model.language_model.layers.20.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
|
| 212 |
+
"model.language_model.layers.20.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 213 |
+
"model.language_model.layers.20.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
|
| 214 |
+
"model.language_model.layers.20.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 215 |
+
"model.language_model.layers.20.self_attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 216 |
+
"model.language_model.layers.20.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
|
| 217 |
+
"model.language_model.layers.20.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
|
| 218 |
+
"model.language_model.layers.20.self_attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 219 |
+
"model.language_model.layers.20.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
|
| 220 |
+
"model.language_model.layers.20.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
|
| 221 |
+
"model.language_model.layers.21.input_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 222 |
+
"model.language_model.layers.21.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
|
| 223 |
+
"model.language_model.layers.21.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 224 |
+
"model.language_model.layers.21.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
|
| 225 |
+
"model.language_model.layers.21.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 226 |
+
"model.language_model.layers.21.self_attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 227 |
+
"model.language_model.layers.21.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
|
| 228 |
+
"model.language_model.layers.21.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
|
| 229 |
+
"model.language_model.layers.21.self_attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 230 |
+
"model.language_model.layers.21.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
|
| 231 |
+
"model.language_model.layers.21.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
|
| 232 |
+
"model.language_model.layers.22.input_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 233 |
+
"model.language_model.layers.22.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
|
| 234 |
+
"model.language_model.layers.22.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 235 |
+
"model.language_model.layers.22.self_attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 236 |
+
"model.language_model.layers.22.self_attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 237 |
+
"model.language_model.layers.23.input_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 238 |
+
"model.language_model.layers.23.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
|
| 239 |
+
"model.language_model.layers.23.self_attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 240 |
+
"model.language_model.layers.23.self_attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 241 |
+
"model.language_model.layers.7.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
|
| 242 |
+
"model.language_model.layers.7.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 243 |
+
"model.language_model.layers.7.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
|
| 244 |
+
"model.language_model.layers.8.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
|
| 245 |
+
"model.language_model.layers.8.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 246 |
+
"model.language_model.layers.8.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
|
| 247 |
+
"model.language_model.layers.8.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
|
| 248 |
+
"model.language_model.layers.8.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
|
| 249 |
+
"model.language_model.layers.8.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
|
| 250 |
+
"model.language_model.layers.8.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
|
| 251 |
+
"model.language_model.layers.9.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
|
| 252 |
+
"model.language_model.layers.9.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 253 |
+
"model.language_model.layers.9.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
|
| 254 |
+
"model.language_model.layers.9.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
|
| 255 |
+
"model.language_model.layers.9.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
|
| 256 |
+
"model.language_model.layers.9.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
|
| 257 |
+
"model.language_model.layers.9.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
|
| 258 |
+
"lm_head.weight": "model-00003-of-00004.safetensors",
|
| 259 |
+
"model.language_model.layers.22.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 260 |
+
"model.language_model.layers.22.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
|
| 261 |
+
"model.language_model.layers.22.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
|
| 262 |
+
"model.language_model.layers.22.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
|
| 263 |
+
"model.language_model.layers.22.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
|
| 264 |
+
"model.language_model.layers.22.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
|
| 265 |
+
"model.language_model.layers.23.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
|
| 266 |
+
"model.language_model.layers.23.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 267 |
+
"model.language_model.layers.23.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
|
| 268 |
+
"model.language_model.layers.23.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
|
| 269 |
+
"model.language_model.layers.23.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
|
| 270 |
+
"model.language_model.layers.23.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
|
| 271 |
+
"model.language_model.layers.23.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
|
| 272 |
+
"model.language_model.layers.24.input_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 273 |
+
"model.language_model.layers.24.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
|
| 274 |
+
"model.language_model.layers.24.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 275 |
+
"model.language_model.layers.24.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
|
| 276 |
+
"model.language_model.layers.24.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 277 |
+
"model.language_model.layers.24.self_attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 278 |
+
"model.language_model.layers.24.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
|
| 279 |
+
"model.language_model.layers.24.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
|
| 280 |
+
"model.language_model.layers.24.self_attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 281 |
+
"model.language_model.layers.24.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
|
| 282 |
+
"model.language_model.layers.24.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
|
| 283 |
+
"model.language_model.layers.25.input_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 284 |
+
"model.language_model.layers.25.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
|
| 285 |
+
"model.language_model.layers.25.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 286 |
+
"model.language_model.layers.25.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
|
| 287 |
+
"model.language_model.layers.25.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 288 |
+
"model.language_model.layers.25.self_attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 289 |
+
"model.language_model.layers.25.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
|
| 290 |
+
"model.language_model.layers.25.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
|
| 291 |
+
"model.language_model.layers.25.self_attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 292 |
+
"model.language_model.layers.25.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
|
| 293 |
+
"model.language_model.layers.25.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
|
| 294 |
+
"model.language_model.layers.26.input_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 295 |
+
"model.language_model.layers.26.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
|
| 296 |
+
"model.language_model.layers.26.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 297 |
+
"model.language_model.layers.26.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
|
| 298 |
+
"model.language_model.layers.26.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 299 |
+
"model.language_model.layers.26.self_attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 300 |
+
"model.language_model.layers.26.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
|
| 301 |
+
"model.language_model.layers.26.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
|
| 302 |
+
"model.language_model.layers.26.self_attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 303 |
+
"model.language_model.layers.26.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
|
| 304 |
+
"model.language_model.layers.26.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
|
| 305 |
+
"model.language_model.layers.27.input_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 306 |
+
"model.language_model.layers.27.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
|
| 307 |
+
"model.language_model.layers.27.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 308 |
+
"model.language_model.layers.27.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
|
| 309 |
+
"model.language_model.layers.27.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 310 |
+
"model.language_model.layers.27.self_attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 311 |
+
"model.language_model.layers.27.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
|
| 312 |
+
"model.language_model.layers.27.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
|
| 313 |
+
"model.language_model.layers.27.self_attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 314 |
+
"model.language_model.layers.27.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
|
| 315 |
+
"model.language_model.layers.27.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
|
| 316 |
+
"model.language_model.layers.28.input_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 317 |
+
"model.language_model.layers.28.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
|
| 318 |
+
"model.language_model.layers.28.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 319 |
+
"model.language_model.layers.28.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
|
| 320 |
+
"model.language_model.layers.28.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 321 |
+
"model.language_model.layers.28.self_attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 322 |
+
"model.language_model.layers.28.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
|
| 323 |
+
"model.language_model.layers.28.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
|
| 324 |
+
"model.language_model.layers.28.self_attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 325 |
+
"model.language_model.layers.28.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
|
| 326 |
+
"model.language_model.layers.28.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
|
| 327 |
+
"model.language_model.layers.29.input_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 328 |
+
"model.language_model.layers.29.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
|
| 329 |
+
"model.language_model.layers.29.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 330 |
+
"model.language_model.layers.29.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
|
| 331 |
+
"model.language_model.layers.29.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 332 |
+
"model.language_model.layers.29.self_attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 333 |
+
"model.language_model.layers.29.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
|
| 334 |
+
"model.language_model.layers.29.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
|
| 335 |
+
"model.language_model.layers.29.self_attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 336 |
+
"model.language_model.layers.29.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
|
| 337 |
+
"model.language_model.layers.29.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
|
| 338 |
+
"model.language_model.layers.30.input_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 339 |
+
"model.language_model.layers.30.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
|
| 340 |
+
"model.language_model.layers.30.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 341 |
+
"model.language_model.layers.30.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
|
| 342 |
+
"model.language_model.layers.30.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 343 |
+
"model.language_model.layers.30.self_attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 344 |
+
"model.language_model.layers.30.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
|
| 345 |
+
"model.language_model.layers.30.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
|
| 346 |
+
"model.language_model.layers.30.self_attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 347 |
+
"model.language_model.layers.30.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
|
| 348 |
+
"model.language_model.layers.30.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
|
| 349 |
+
"model.language_model.layers.31.input_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 350 |
+
"model.language_model.layers.31.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
|
| 351 |
+
"model.language_model.layers.31.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 352 |
+
"model.language_model.layers.31.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
|
| 353 |
+
"model.language_model.layers.31.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 354 |
+
"model.language_model.layers.31.self_attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 355 |
+
"model.language_model.layers.31.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
|
| 356 |
+
"model.language_model.layers.31.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
|
| 357 |
+
"model.language_model.layers.31.self_attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 358 |
+
"model.language_model.layers.31.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
|
| 359 |
+
"model.language_model.layers.32.input_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 360 |
+
"model.language_model.layers.32.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 361 |
+
"model.language_model.layers.32.self_attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 362 |
+
"model.language_model.layers.32.self_attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 363 |
+
"model.language_model.layers.33.input_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 364 |
+
"model.language_model.layers.33.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 365 |
+
"model.language_model.layers.33.self_attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 366 |
+
"model.language_model.layers.33.self_attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 367 |
+
"model.language_model.layers.34.input_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 368 |
+
"model.language_model.layers.34.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 369 |
+
"model.language_model.layers.34.self_attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 370 |
+
"model.language_model.layers.34.self_attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 371 |
+
"model.language_model.layers.35.input_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 372 |
+
"model.language_model.layers.35.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
|
| 373 |
+
"model.language_model.layers.35.self_attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 374 |
+
"model.language_model.layers.35.self_attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 375 |
+
"model.language_model.norm.weight": "model-00003-of-00004.safetensors",
|
| 376 |
+
"model.language_model.layers.31.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
|
| 377 |
+
"model.language_model.layers.32.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
|
| 378 |
+
"model.language_model.layers.32.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 379 |
+
"model.language_model.layers.32.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
|
| 380 |
+
"model.language_model.layers.32.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
|
| 381 |
+
"model.language_model.layers.32.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
|
| 382 |
+
"model.language_model.layers.32.self_attn.q_proj.weight": "model-00004-of-00004.safetensors",
|
| 383 |
+
"model.language_model.layers.32.self_attn.v_proj.weight": "model-00004-of-00004.safetensors",
|
| 384 |
+
"model.language_model.layers.33.mlp.down_proj.weight": "model-00004-of-00004.safetensors",
|
| 385 |
+
"model.language_model.layers.33.mlp.gate_proj.weight": "model-00004-of-00004.safetensors",
|
| 386 |
+
"model.language_model.layers.33.mlp.up_proj.weight": "model-00004-of-00004.safetensors",
|
| 387 |
+
"model.language_model.layers.33.self_attn.k_proj.weight": "model-00004-of-00004.safetensors",
|
| 388 |
+
"model.language_model.layers.33.self_attn.o_proj.weight": "model-00004-of-00004.safetensors",
|
| 389 |
+
"model.language_model.layers.33.self_attn.q_proj.weight": "model-00004-of-00004.safetensors",
|
| 390 |
+
"model.language_model.layers.33.self_attn.v_proj.weight": "model-00004-of-00004.safetensors",
|
| 391 |
+
"model.language_model.layers.34.mlp.down_proj.weight": "model-00004-of-00004.safetensors",
|
| 392 |
+
"model.language_model.layers.34.mlp.gate_proj.weight": "model-00004-of-00004.safetensors",
|
| 393 |
+
"model.language_model.layers.34.mlp.up_proj.weight": "model-00004-of-00004.safetensors",
|
| 394 |
+
"model.language_model.layers.34.self_attn.k_proj.weight": "model-00004-of-00004.safetensors",
|
| 395 |
+
"model.language_model.layers.34.self_attn.o_proj.weight": "model-00004-of-00004.safetensors",
|
| 396 |
+
"model.language_model.layers.34.self_attn.q_proj.weight": "model-00004-of-00004.safetensors",
|
| 397 |
+
"model.language_model.layers.34.self_attn.v_proj.weight": "model-00004-of-00004.safetensors",
|
| 398 |
+
"model.language_model.layers.35.mlp.down_proj.weight": "model-00004-of-00004.safetensors",
|
| 399 |
+
"model.language_model.layers.35.mlp.gate_proj.weight": "model-00004-of-00004.safetensors",
|
| 400 |
+
"model.language_model.layers.35.mlp.up_proj.weight": "model-00004-of-00004.safetensors",
|
| 401 |
+
"model.language_model.layers.35.self_attn.k_proj.weight": "model-00004-of-00004.safetensors",
|
| 402 |
+
"model.language_model.layers.35.self_attn.o_proj.weight": "model-00004-of-00004.safetensors",
|
| 403 |
+
"model.language_model.layers.35.self_attn.q_proj.weight": "model-00004-of-00004.safetensors",
|
| 404 |
+
"model.language_model.layers.35.self_attn.v_proj.weight": "model-00004-of-00004.safetensors",
|
| 405 |
+
"model.visual.embeddings.patch_embedding.weight": "model-00004-of-00004.safetensors",
|
| 406 |
+
"model.visual.encoder.layers.0.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 407 |
+
"model.visual.encoder.layers.0.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 408 |
+
"model.visual.encoder.layers.0.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 409 |
+
"model.visual.encoder.layers.0.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 410 |
+
"model.visual.encoder.layers.0.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 411 |
+
"model.visual.encoder.layers.0.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 412 |
+
"model.visual.encoder.layers.0.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 413 |
+
"model.visual.encoder.layers.0.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 414 |
+
"model.visual.encoder.layers.0.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 415 |
+
"model.visual.encoder.layers.0.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 416 |
+
"model.visual.encoder.layers.0.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 417 |
+
"model.visual.encoder.layers.0.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 418 |
+
"model.visual.encoder.layers.1.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 419 |
+
"model.visual.encoder.layers.1.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 420 |
+
"model.visual.encoder.layers.1.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 421 |
+
"model.visual.encoder.layers.1.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 422 |
+
"model.visual.encoder.layers.1.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 423 |
+
"model.visual.encoder.layers.1.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 424 |
+
"model.visual.encoder.layers.1.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 425 |
+
"model.visual.encoder.layers.1.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 426 |
+
"model.visual.encoder.layers.1.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 427 |
+
"model.visual.encoder.layers.1.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 428 |
+
"model.visual.encoder.layers.1.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 429 |
+
"model.visual.encoder.layers.1.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 430 |
+
"model.visual.encoder.layers.10.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 431 |
+
"model.visual.encoder.layers.10.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 432 |
+
"model.visual.encoder.layers.10.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 433 |
+
"model.visual.encoder.layers.10.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 434 |
+
"model.visual.encoder.layers.10.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 435 |
+
"model.visual.encoder.layers.10.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 436 |
+
"model.visual.encoder.layers.10.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 437 |
+
"model.visual.encoder.layers.10.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 438 |
+
"model.visual.encoder.layers.10.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 439 |
+
"model.visual.encoder.layers.10.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 440 |
+
"model.visual.encoder.layers.10.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 441 |
+
"model.visual.encoder.layers.10.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 442 |
+
"model.visual.encoder.layers.11.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 443 |
+
"model.visual.encoder.layers.11.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 444 |
+
"model.visual.encoder.layers.11.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 445 |
+
"model.visual.encoder.layers.11.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 446 |
+
"model.visual.encoder.layers.11.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 447 |
+
"model.visual.encoder.layers.11.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 448 |
+
"model.visual.encoder.layers.11.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 449 |
+
"model.visual.encoder.layers.11.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 450 |
+
"model.visual.encoder.layers.11.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 451 |
+
"model.visual.encoder.layers.11.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 452 |
+
"model.visual.encoder.layers.11.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 453 |
+
"model.visual.encoder.layers.11.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 454 |
+
"model.visual.encoder.layers.12.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 455 |
+
"model.visual.encoder.layers.12.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 456 |
+
"model.visual.encoder.layers.12.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 457 |
+
"model.visual.encoder.layers.12.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 458 |
+
"model.visual.encoder.layers.12.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 459 |
+
"model.visual.encoder.layers.12.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 460 |
+
"model.visual.encoder.layers.12.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 461 |
+
"model.visual.encoder.layers.12.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 462 |
+
"model.visual.encoder.layers.12.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 463 |
+
"model.visual.encoder.layers.12.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 464 |
+
"model.visual.encoder.layers.12.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 465 |
+
"model.visual.encoder.layers.12.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 466 |
+
"model.visual.encoder.layers.13.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 467 |
+
"model.visual.encoder.layers.13.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 468 |
+
"model.visual.encoder.layers.13.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 469 |
+
"model.visual.encoder.layers.13.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 470 |
+
"model.visual.encoder.layers.13.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 471 |
+
"model.visual.encoder.layers.13.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 472 |
+
"model.visual.encoder.layers.13.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 473 |
+
"model.visual.encoder.layers.13.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 474 |
+
"model.visual.encoder.layers.13.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 475 |
+
"model.visual.encoder.layers.13.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 476 |
+
"model.visual.encoder.layers.13.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 477 |
+
"model.visual.encoder.layers.13.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 478 |
+
"model.visual.encoder.layers.14.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 479 |
+
"model.visual.encoder.layers.14.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 480 |
+
"model.visual.encoder.layers.14.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 481 |
+
"model.visual.encoder.layers.14.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 482 |
+
"model.visual.encoder.layers.14.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 483 |
+
"model.visual.encoder.layers.14.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 484 |
+
"model.visual.encoder.layers.14.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 485 |
+
"model.visual.encoder.layers.14.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 486 |
+
"model.visual.encoder.layers.14.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 487 |
+
"model.visual.encoder.layers.14.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 488 |
+
"model.visual.encoder.layers.14.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 489 |
+
"model.visual.encoder.layers.14.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 490 |
+
"model.visual.encoder.layers.15.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 491 |
+
"model.visual.encoder.layers.15.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 492 |
+
"model.visual.encoder.layers.15.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 493 |
+
"model.visual.encoder.layers.15.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 494 |
+
"model.visual.encoder.layers.15.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 495 |
+
"model.visual.encoder.layers.15.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 496 |
+
"model.visual.encoder.layers.15.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 497 |
+
"model.visual.encoder.layers.15.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 498 |
+
"model.visual.encoder.layers.15.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 499 |
+
"model.visual.encoder.layers.15.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 500 |
+
"model.visual.encoder.layers.15.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 501 |
+
"model.visual.encoder.layers.15.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 502 |
+
"model.visual.encoder.layers.16.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 503 |
+
"model.visual.encoder.layers.16.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 504 |
+
"model.visual.encoder.layers.16.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 505 |
+
"model.visual.encoder.layers.16.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 506 |
+
"model.visual.encoder.layers.16.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 507 |
+
"model.visual.encoder.layers.16.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 508 |
+
"model.visual.encoder.layers.16.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 509 |
+
"model.visual.encoder.layers.16.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 510 |
+
"model.visual.encoder.layers.16.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 511 |
+
"model.visual.encoder.layers.16.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 512 |
+
"model.visual.encoder.layers.16.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 513 |
+
"model.visual.encoder.layers.16.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 514 |
+
"model.visual.encoder.layers.17.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 515 |
+
"model.visual.encoder.layers.17.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 516 |
+
"model.visual.encoder.layers.17.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 517 |
+
"model.visual.encoder.layers.17.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 518 |
+
"model.visual.encoder.layers.17.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 519 |
+
"model.visual.encoder.layers.17.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 520 |
+
"model.visual.encoder.layers.17.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 521 |
+
"model.visual.encoder.layers.17.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 522 |
+
"model.visual.encoder.layers.17.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 523 |
+
"model.visual.encoder.layers.17.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 524 |
+
"model.visual.encoder.layers.17.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 525 |
+
"model.visual.encoder.layers.17.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 526 |
+
"model.visual.encoder.layers.18.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 527 |
+
"model.visual.encoder.layers.18.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 528 |
+
"model.visual.encoder.layers.18.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 529 |
+
"model.visual.encoder.layers.18.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 530 |
+
"model.visual.encoder.layers.18.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 531 |
+
"model.visual.encoder.layers.18.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 532 |
+
"model.visual.encoder.layers.18.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 533 |
+
"model.visual.encoder.layers.18.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 534 |
+
"model.visual.encoder.layers.18.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 535 |
+
"model.visual.encoder.layers.18.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 536 |
+
"model.visual.encoder.layers.18.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 537 |
+
"model.visual.encoder.layers.18.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 538 |
+
"model.visual.encoder.layers.19.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 539 |
+
"model.visual.encoder.layers.19.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 540 |
+
"model.visual.encoder.layers.19.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 541 |
+
"model.visual.encoder.layers.19.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 542 |
+
"model.visual.encoder.layers.19.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 543 |
+
"model.visual.encoder.layers.19.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 544 |
+
"model.visual.encoder.layers.19.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 545 |
+
"model.visual.encoder.layers.19.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 546 |
+
"model.visual.encoder.layers.19.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 547 |
+
"model.visual.encoder.layers.19.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 548 |
+
"model.visual.encoder.layers.19.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 549 |
+
"model.visual.encoder.layers.19.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 550 |
+
"model.visual.encoder.layers.2.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 551 |
+
"model.visual.encoder.layers.2.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 552 |
+
"model.visual.encoder.layers.2.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 553 |
+
"model.visual.encoder.layers.2.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 554 |
+
"model.visual.encoder.layers.2.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 555 |
+
"model.visual.encoder.layers.2.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 556 |
+
"model.visual.encoder.layers.2.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 557 |
+
"model.visual.encoder.layers.2.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 558 |
+
"model.visual.encoder.layers.2.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 559 |
+
"model.visual.encoder.layers.2.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 560 |
+
"model.visual.encoder.layers.2.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 561 |
+
"model.visual.encoder.layers.2.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 562 |
+
"model.visual.encoder.layers.20.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 563 |
+
"model.visual.encoder.layers.20.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 564 |
+
"model.visual.encoder.layers.20.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 565 |
+
"model.visual.encoder.layers.20.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 566 |
+
"model.visual.encoder.layers.20.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 567 |
+
"model.visual.encoder.layers.20.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 568 |
+
"model.visual.encoder.layers.20.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 569 |
+
"model.visual.encoder.layers.20.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 570 |
+
"model.visual.encoder.layers.20.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 571 |
+
"model.visual.encoder.layers.20.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 572 |
+
"model.visual.encoder.layers.20.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 573 |
+
"model.visual.encoder.layers.20.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 574 |
+
"model.visual.encoder.layers.21.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 575 |
+
"model.visual.encoder.layers.21.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 576 |
+
"model.visual.encoder.layers.21.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 577 |
+
"model.visual.encoder.layers.21.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 578 |
+
"model.visual.encoder.layers.21.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 579 |
+
"model.visual.encoder.layers.21.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 580 |
+
"model.visual.encoder.layers.21.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 581 |
+
"model.visual.encoder.layers.21.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 582 |
+
"model.visual.encoder.layers.21.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 583 |
+
"model.visual.encoder.layers.21.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 584 |
+
"model.visual.encoder.layers.21.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 585 |
+
"model.visual.encoder.layers.21.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 586 |
+
"model.visual.encoder.layers.22.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 587 |
+
"model.visual.encoder.layers.22.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 588 |
+
"model.visual.encoder.layers.22.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 589 |
+
"model.visual.encoder.layers.22.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 590 |
+
"model.visual.encoder.layers.22.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 591 |
+
"model.visual.encoder.layers.22.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 592 |
+
"model.visual.encoder.layers.22.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 593 |
+
"model.visual.encoder.layers.22.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 594 |
+
"model.visual.encoder.layers.22.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 595 |
+
"model.visual.encoder.layers.22.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 596 |
+
"model.visual.encoder.layers.22.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 597 |
+
"model.visual.encoder.layers.22.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 598 |
+
"model.visual.encoder.layers.23.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 599 |
+
"model.visual.encoder.layers.23.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 600 |
+
"model.visual.encoder.layers.23.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 601 |
+
"model.visual.encoder.layers.23.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 602 |
+
"model.visual.encoder.layers.23.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 603 |
+
"model.visual.encoder.layers.23.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 604 |
+
"model.visual.encoder.layers.23.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 605 |
+
"model.visual.encoder.layers.23.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 606 |
+
"model.visual.encoder.layers.23.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 607 |
+
"model.visual.encoder.layers.23.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 608 |
+
"model.visual.encoder.layers.23.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 609 |
+
"model.visual.encoder.layers.23.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 610 |
+
"model.visual.encoder.layers.3.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 611 |
+
"model.visual.encoder.layers.3.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 612 |
+
"model.visual.encoder.layers.3.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 613 |
+
"model.visual.encoder.layers.3.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 614 |
+
"model.visual.encoder.layers.3.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 615 |
+
"model.visual.encoder.layers.3.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 616 |
+
"model.visual.encoder.layers.3.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 617 |
+
"model.visual.encoder.layers.3.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 618 |
+
"model.visual.encoder.layers.3.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 619 |
+
"model.visual.encoder.layers.3.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 620 |
+
"model.visual.encoder.layers.3.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 621 |
+
"model.visual.encoder.layers.3.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 622 |
+
"model.visual.encoder.layers.4.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 623 |
+
"model.visual.encoder.layers.4.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 624 |
+
"model.visual.encoder.layers.4.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 625 |
+
"model.visual.encoder.layers.4.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 626 |
+
"model.visual.encoder.layers.4.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 627 |
+
"model.visual.encoder.layers.4.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 628 |
+
"model.visual.encoder.layers.4.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 629 |
+
"model.visual.encoder.layers.4.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 630 |
+
"model.visual.encoder.layers.4.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 631 |
+
"model.visual.encoder.layers.4.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 632 |
+
"model.visual.encoder.layers.4.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 633 |
+
"model.visual.encoder.layers.4.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 634 |
+
"model.visual.encoder.layers.5.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 635 |
+
"model.visual.encoder.layers.5.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 636 |
+
"model.visual.encoder.layers.5.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 637 |
+
"model.visual.encoder.layers.5.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 638 |
+
"model.visual.encoder.layers.5.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 639 |
+
"model.visual.encoder.layers.5.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 640 |
+
"model.visual.encoder.layers.5.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 641 |
+
"model.visual.encoder.layers.5.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 642 |
+
"model.visual.encoder.layers.5.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 643 |
+
"model.visual.encoder.layers.5.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 644 |
+
"model.visual.encoder.layers.5.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 645 |
+
"model.visual.encoder.layers.5.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 646 |
+
"model.visual.encoder.layers.6.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 647 |
+
"model.visual.encoder.layers.6.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 648 |
+
"model.visual.encoder.layers.6.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 649 |
+
"model.visual.encoder.layers.6.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 650 |
+
"model.visual.encoder.layers.6.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 651 |
+
"model.visual.encoder.layers.6.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 652 |
+
"model.visual.encoder.layers.6.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 653 |
+
"model.visual.encoder.layers.6.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 654 |
+
"model.visual.encoder.layers.6.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 655 |
+
"model.visual.encoder.layers.6.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 656 |
+
"model.visual.encoder.layers.6.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 657 |
+
"model.visual.encoder.layers.6.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 658 |
+
"model.visual.encoder.layers.7.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 659 |
+
"model.visual.encoder.layers.7.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 660 |
+
"model.visual.encoder.layers.7.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 661 |
+
"model.visual.encoder.layers.7.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 662 |
+
"model.visual.encoder.layers.7.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 663 |
+
"model.visual.encoder.layers.7.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 664 |
+
"model.visual.encoder.layers.7.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 665 |
+
"model.visual.encoder.layers.7.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 666 |
+
"model.visual.encoder.layers.7.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 667 |
+
"model.visual.encoder.layers.7.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 668 |
+
"model.visual.encoder.layers.7.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 669 |
+
"model.visual.encoder.layers.7.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 670 |
+
"model.visual.encoder.layers.8.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 671 |
+
"model.visual.encoder.layers.8.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 672 |
+
"model.visual.encoder.layers.8.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 673 |
+
"model.visual.encoder.layers.8.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 674 |
+
"model.visual.encoder.layers.8.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 675 |
+
"model.visual.encoder.layers.8.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 676 |
+
"model.visual.encoder.layers.8.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 677 |
+
"model.visual.encoder.layers.8.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 678 |
+
"model.visual.encoder.layers.8.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 679 |
+
"model.visual.encoder.layers.8.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 680 |
+
"model.visual.encoder.layers.8.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 681 |
+
"model.visual.encoder.layers.8.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 682 |
+
"model.visual.encoder.layers.9.layer_norm1.bias": "model-00004-of-00004.safetensors",
|
| 683 |
+
"model.visual.encoder.layers.9.layer_norm1.weight": "model-00004-of-00004.safetensors",
|
| 684 |
+
"model.visual.encoder.layers.9.layer_norm2.bias": "model-00004-of-00004.safetensors",
|
| 685 |
+
"model.visual.encoder.layers.9.layer_norm2.weight": "model-00004-of-00004.safetensors",
|
| 686 |
+
"model.visual.encoder.layers.9.mlp.fc1.bias": "model-00004-of-00004.safetensors",
|
| 687 |
+
"model.visual.encoder.layers.9.mlp.fc1.weight": "model-00004-of-00004.safetensors",
|
| 688 |
+
"model.visual.encoder.layers.9.mlp.fc2.bias": "model-00004-of-00004.safetensors",
|
| 689 |
+
"model.visual.encoder.layers.9.mlp.fc2.weight": "model-00004-of-00004.safetensors",
|
| 690 |
+
"model.visual.encoder.layers.9.self_attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 691 |
+
"model.visual.encoder.layers.9.self_attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 692 |
+
"model.visual.encoder.layers.9.self_attn.qkv.bias": "model-00004-of-00004.safetensors",
|
| 693 |
+
"model.visual.encoder.layers.9.self_attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 694 |
+
"model.visual.layernorm_pre.bias": "model-00004-of-00004.safetensors",
|
| 695 |
+
"model.visual.layernorm_pre.weight": "model-00004-of-00004.safetensors",
|
| 696 |
+
"model.visual.merger.ln_q.bias": "model-00004-of-00004.safetensors",
|
| 697 |
+
"model.visual.merger.ln_q.weight": "model-00004-of-00004.safetensors",
|
| 698 |
+
"model.visual.merger.mlp.0.bias": "model-00004-of-00004.safetensors",
|
| 699 |
+
"model.visual.merger.mlp.0.weight": "model-00004-of-00004.safetensors",
|
| 700 |
+
"model.visual.merger.mlp.2.bias": "model-00004-of-00004.safetensors",
|
| 701 |
+
"model.visual.merger.mlp.2.weight": "model-00004-of-00004.safetensors"
|
| 702 |
+
}
|
| 703 |
+
}
|
modeling_llava_onevision2.py
ADDED
|
@@ -0,0 +1,1607 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from collections.abc import Callable
|
| 2 |
+
from dataclasses import dataclass
|
| 3 |
+
from typing import Any, Optional, Union
|
| 4 |
+
|
| 5 |
+
import torch
|
| 6 |
+
import torch.nn as nn
|
| 7 |
+
from torch.nn import LayerNorm
|
| 8 |
+
|
| 9 |
+
from transformers import AutoModel
|
| 10 |
+
from transformers.cache_utils import Cache
|
| 11 |
+
from transformers.generation import GenerationMixin
|
| 12 |
+
from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling, ModelOutput
|
| 13 |
+
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
|
| 14 |
+
from transformers.models.siglip.modeling_siglip import SiglipMLP
|
| 15 |
+
from transformers.processing_utils import Unpack
|
| 16 |
+
from transformers.utils import (
|
| 17 |
+
TransformersKwargs,
|
| 18 |
+
auto_docstring,
|
| 19 |
+
can_return_tuple,
|
| 20 |
+
replace_return_docstrings,
|
| 21 |
+
)
|
| 22 |
+
from transformers.utils.generic import is_flash_attention_requested
|
| 23 |
+
|
| 24 |
+
from .configuration_llava_onevision2 import LlavaOnevision2Config, LlavaOnevision2VisionConfig
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
@dataclass
|
| 28 |
+
@auto_docstring(
|
| 29 |
+
custom_intro="""
|
| 30 |
+
Base class for Llava-Onevision-1.5 outputs, with hidden states and attentions.
|
| 31 |
+
"""
|
| 32 |
+
)
|
| 33 |
+
class LlavaOnevision2ModelOutputWithPast(ModelOutput):
|
| 34 |
+
r"""
|
| 35 |
+
past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
|
| 36 |
+
It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).
|
| 37 |
+
|
| 38 |
+
Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
|
| 39 |
+
`past_key_values` input) to speed up sequential decoding.
|
| 40 |
+
"""
|
| 41 |
+
|
| 42 |
+
last_hidden_state: Optional[torch.FloatTensor] = None
|
| 43 |
+
past_key_values: Optional[Cache] = None
|
| 44 |
+
hidden_states: Optional[tuple[torch.FloatTensor]] = None
|
| 45 |
+
attentions: Optional[tuple[torch.FloatTensor]] = None
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
@dataclass
|
| 49 |
+
@auto_docstring(
|
| 50 |
+
custom_intro="""
|
| 51 |
+
Base class for Llava-Onevision-1.5 causal language model (or autoregressive) outputs.
|
| 52 |
+
"""
|
| 53 |
+
)
|
| 54 |
+
class LlavaOnevision2CausalLMOutputWithPast(ModelOutput):
|
| 55 |
+
r"""
|
| 56 |
+
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
|
| 57 |
+
Language modeling loss (for next-token prediction).
|
| 58 |
+
logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
|
| 59 |
+
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
|
| 60 |
+
past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
|
| 61 |
+
It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).
|
| 62 |
+
|
| 63 |
+
Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
|
| 64 |
+
`past_key_values` input) to speed up sequential decoding.
|
| 65 |
+
"""
|
| 66 |
+
|
| 67 |
+
loss: Optional[torch.FloatTensor] = None
|
| 68 |
+
logits: Optional[torch.FloatTensor] = None
|
| 69 |
+
past_key_values: Optional[Cache] = None
|
| 70 |
+
hidden_states: Optional[tuple[torch.FloatTensor]] = None
|
| 71 |
+
attentions: Optional[tuple[torch.FloatTensor]] = None
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
# ---------------------------------------------------------------------------
|
| 75 |
+
# Vision Rotary Embedding
|
| 76 |
+
# ---------------------------------------------------------------------------
|
| 77 |
+
|
| 78 |
+
|
| 79 |
+
class VisionRotaryEmbedding(nn.Module):
|
| 80 |
+
"""
|
| 81 |
+
3D (T,H,W) Rotary frequency constructor with 4:6:6 split.
|
| 82 |
+
Supports both grid_thw-based and explicit position-based RoPE computation.
|
| 83 |
+
"""
|
| 84 |
+
|
| 85 |
+
def __init__(self, config: LlavaOnevision2VisionConfig):
|
| 86 |
+
super().__init__()
|
| 87 |
+
head_dim = config.hidden_size // config.num_attention_heads
|
| 88 |
+
base = config.rope_theta
|
| 89 |
+
|
| 90 |
+
assert head_dim % 2 == 0, "head_dim must be even for rotary."
|
| 91 |
+
assert head_dim % 16 == 0, "head_dim must be divisible by 16."
|
| 92 |
+
half = head_dim // 2
|
| 93 |
+
assert half % 16 == 0, "head_dim//2 must also be divisible by 16 to split into 4:6:6."
|
| 94 |
+
|
| 95 |
+
self.head_dim = head_dim
|
| 96 |
+
self.half = half
|
| 97 |
+
self.base = base
|
| 98 |
+
|
| 99 |
+
# 4:6:6 split for T:H:W
|
| 100 |
+
unit = half // 16
|
| 101 |
+
self.t_size = 4 * unit
|
| 102 |
+
self.h_size = 6 * unit
|
| 103 |
+
self.w_size = 6 * unit
|
| 104 |
+
|
| 105 |
+
self.register_buffer(
|
| 106 |
+
"inv_freq_t",
|
| 107 |
+
1.0 / (base ** (torch.arange(self.t_size, dtype=torch.float32) / self.t_size)),
|
| 108 |
+
persistent=False,
|
| 109 |
+
)
|
| 110 |
+
self.register_buffer(
|
| 111 |
+
"inv_freq_h",
|
| 112 |
+
1.0 / (base ** (torch.arange(self.h_size, dtype=torch.float32) / self.h_size)),
|
| 113 |
+
persistent=False,
|
| 114 |
+
)
|
| 115 |
+
self.register_buffer(
|
| 116 |
+
"inv_freq_w",
|
| 117 |
+
1.0 / (base ** (torch.arange(self.w_size, dtype=torch.float32) / self.w_size)),
|
| 118 |
+
persistent=False,
|
| 119 |
+
)
|
| 120 |
+
|
| 121 |
+
def forward(self, grid_thw: torch.Tensor) -> torch.Tensor:
|
| 122 |
+
"""
|
| 123 |
+
Compute rotary position embeddings from grid_thw (Qwen2VL style).
|
| 124 |
+
|
| 125 |
+
Args:
|
| 126 |
+
grid_thw: [num_samples, 3] tensor with [t, h, w] for each sample
|
| 127 |
+
|
| 128 |
+
Returns:
|
| 129 |
+
freqs: [total_seq_len, half] tensor of position frequencies
|
| 130 |
+
"""
|
| 131 |
+
device = grid_thw.device
|
| 132 |
+
inv_t = self.inv_freq_t.to(device=device)
|
| 133 |
+
inv_h = self.inv_freq_h.to(device=device)
|
| 134 |
+
inv_w = self.inv_freq_w.to(device=device)
|
| 135 |
+
|
| 136 |
+
all_freqs = []
|
| 137 |
+
for sample_thw in grid_thw:
|
| 138 |
+
t, h, w = sample_thw[0].item(), sample_thw[1].item(), sample_thw[2].item()
|
| 139 |
+
|
| 140 |
+
# Compute frequency tables
|
| 141 |
+
ft = torch.outer(torch.arange(t, device=device, dtype=torch.float32), inv_t)
|
| 142 |
+
fh = torch.outer(torch.arange(h, device=device, dtype=torch.float32), inv_h)
|
| 143 |
+
fw = torch.outer(torch.arange(w, device=device, dtype=torch.float32), inv_w)
|
| 144 |
+
|
| 145 |
+
# Build position indices for this sample
|
| 146 |
+
t_ids = torch.arange(t, device=device).repeat_interleave(h * w)
|
| 147 |
+
h_ids = torch.arange(h, device=device).repeat_interleave(w).repeat(t)
|
| 148 |
+
w_ids = torch.arange(w, device=device).repeat(h).repeat(t)
|
| 149 |
+
|
| 150 |
+
# Concatenate frequencies: [seq_len, half]
|
| 151 |
+
sample_freqs = torch.cat([ft[t_ids], fh[h_ids], fw[w_ids]], dim=-1)
|
| 152 |
+
all_freqs.append(sample_freqs)
|
| 153 |
+
|
| 154 |
+
return torch.cat(all_freqs, dim=0)
|
| 155 |
+
|
| 156 |
+
def forward_from_positions(self, patch_positions: torch.Tensor) -> torch.Tensor:
|
| 157 |
+
"""
|
| 158 |
+
Compute rotary position embeddings from explicit patch positions.
|
| 159 |
+
|
| 160 |
+
Args:
|
| 161 |
+
patch_positions: [seq_len, 3] tensor with [t, h, w] positions for each patch
|
| 162 |
+
|
| 163 |
+
Returns:
|
| 164 |
+
freqs: [seq_len, half] tensor of position frequencies
|
| 165 |
+
"""
|
| 166 |
+
device = patch_positions.device
|
| 167 |
+
inv_t = self.inv_freq_t.to(device=device)
|
| 168 |
+
inv_h = self.inv_freq_h.to(device=device)
|
| 169 |
+
inv_w = self.inv_freq_w.to(device=device)
|
| 170 |
+
|
| 171 |
+
t_pos = patch_positions[:, 0].float()
|
| 172 |
+
h_pos = patch_positions[:, 1].float()
|
| 173 |
+
w_pos = patch_positions[:, 2].float()
|
| 174 |
+
|
| 175 |
+
ft = torch.outer(t_pos, inv_t)
|
| 176 |
+
fh = torch.outer(h_pos, inv_h)
|
| 177 |
+
fw = torch.outer(w_pos, inv_w)
|
| 178 |
+
|
| 179 |
+
return torch.cat([ft, fh, fw], dim=-1)
|
| 180 |
+
|
| 181 |
+
def forward_with_thw(self, t: int, h: int, w: int, device=None) -> torch.Tensor:
|
| 182 |
+
"""
|
| 183 |
+
Compute rotary position embeddings from explicit t, h, w dimensions.
|
| 184 |
+
|
| 185 |
+
Args:
|
| 186 |
+
t: Number of temporal frames
|
| 187 |
+
h: Number of height patches
|
| 188 |
+
w: Number of width patches
|
| 189 |
+
device: Target device
|
| 190 |
+
|
| 191 |
+
Returns:
|
| 192 |
+
freqs: [t*h*w, half] tensor of position frequencies
|
| 193 |
+
"""
|
| 194 |
+
if device is None:
|
| 195 |
+
device = self.inv_freq_t.device
|
| 196 |
+
|
| 197 |
+
inv_t = self.inv_freq_t.to(device=device)
|
| 198 |
+
inv_h = self.inv_freq_h.to(device=device)
|
| 199 |
+
inv_w = self.inv_freq_w.to(device=device)
|
| 200 |
+
|
| 201 |
+
ft = torch.outer(torch.arange(t, device=device, dtype=torch.float32), inv_t)
|
| 202 |
+
fh = torch.outer(torch.arange(h, device=device, dtype=torch.float32), inv_h)
|
| 203 |
+
fw = torch.outer(torch.arange(w, device=device, dtype=torch.float32), inv_w)
|
| 204 |
+
|
| 205 |
+
t_ids = torch.arange(t, device=device).repeat_interleave(h * w)
|
| 206 |
+
h_ids = torch.arange(h, device=device).repeat_interleave(w).repeat(t)
|
| 207 |
+
w_ids = torch.arange(w, device=device).repeat(h).repeat(t)
|
| 208 |
+
|
| 209 |
+
freqs = torch.cat([ft[t_ids], fh[h_ids], fw[w_ids]], dim=-1)
|
| 210 |
+
return freqs
|
| 211 |
+
|
| 212 |
+
|
| 213 |
+
# ---------------------------------------------------------------------------
|
| 214 |
+
# Patch Embedding
|
| 215 |
+
# ---------------------------------------------------------------------------
|
| 216 |
+
|
| 217 |
+
|
| 218 |
+
class OneVisionEncoderEmbeddings(nn.Module):
|
| 219 |
+
"""
|
| 220 |
+
Patch embedding layer that converts pre-processed patches to embeddings.
|
| 221 |
+
|
| 222 |
+
This module is designed to receive patches that have already been extracted
|
| 223 |
+
and arranged by the Qwen2VL image processor in 2x2 block spatial order.
|
| 224 |
+
|
| 225 |
+
Input format: [total_patches, num_channels, patch_size, patch_size]
|
| 226 |
+
Output format: [total_patches, embed_dim]
|
| 227 |
+
"""
|
| 228 |
+
|
| 229 |
+
def __init__(self, config: LlavaOnevision2VisionConfig):
|
| 230 |
+
super().__init__()
|
| 231 |
+
self.config = config
|
| 232 |
+
self.embed_dim = config.hidden_size
|
| 233 |
+
self.image_size = config.image_size
|
| 234 |
+
self.patch_size = config.patch_size
|
| 235 |
+
self.in_channels = config.num_channels
|
| 236 |
+
|
| 237 |
+
self.patch_embedding = nn.Conv2d(
|
| 238 |
+
in_channels=config.num_channels,
|
| 239 |
+
out_channels=self.embed_dim,
|
| 240 |
+
kernel_size=self.patch_size,
|
| 241 |
+
stride=self.patch_size,
|
| 242 |
+
bias=False,
|
| 243 |
+
)
|
| 244 |
+
|
| 245 |
+
def forward(self, hidden_states: torch.FloatTensor) -> torch.Tensor:
|
| 246 |
+
target_dtype = self.patch_embedding.weight.dtype
|
| 247 |
+
hidden_states = hidden_states.view(-1, self.in_channels, self.patch_size, self.patch_size)
|
| 248 |
+
hidden_states = self.patch_embedding(hidden_states.to(dtype=target_dtype)).view(-1, self.embed_dim)
|
| 249 |
+
|
| 250 |
+
return hidden_states
|
| 251 |
+
|
| 252 |
+
|
| 253 |
+
# ---------------------------------------------------------------------------
|
| 254 |
+
# Patch Merger
|
| 255 |
+
# ---------------------------------------------------------------------------
|
| 256 |
+
|
| 257 |
+
|
| 258 |
+
class LlavaOnevision2VisionPatchMerger(nn.Module):
|
| 259 |
+
"""
|
| 260 |
+
Patch merger that merges spatial_merge_size x spatial_merge_size patches into one.
|
| 261 |
+
|
| 262 |
+
This module is designed to work with Qwen2VL-style patch processing where patches
|
| 263 |
+
are already arranged in 2x2 block order by the image processor.
|
| 264 |
+
"""
|
| 265 |
+
|
| 266 |
+
def __init__(
|
| 267 |
+
self,
|
| 268 |
+
dim: int,
|
| 269 |
+
context_dim: int,
|
| 270 |
+
spatial_merge_size: int = 2,
|
| 271 |
+
layer_norm_eps: float = 1e-05,
|
| 272 |
+
use_patch_position_encoding: bool = False,
|
| 273 |
+
patch_position_encoding_type: str = "absolute",
|
| 274 |
+
max_position_embeddings: int = 8192,
|
| 275 |
+
) -> None:
|
| 276 |
+
super().__init__()
|
| 277 |
+
self.hidden_size = context_dim * (spatial_merge_size**2)
|
| 278 |
+
self.ln_q = LayerNorm(context_dim, eps=layer_norm_eps)
|
| 279 |
+
self.mlp = nn.Sequential(
|
| 280 |
+
nn.Linear(self.hidden_size, self.hidden_size),
|
| 281 |
+
nn.GELU(),
|
| 282 |
+
nn.Linear(self.hidden_size, dim),
|
| 283 |
+
)
|
| 284 |
+
self.spatial_merge_size = spatial_merge_size
|
| 285 |
+
self.use_patch_position_encoding = use_patch_position_encoding
|
| 286 |
+
self.patch_position_encoding_type = patch_position_encoding_type
|
| 287 |
+
|
| 288 |
+
if self.use_patch_position_encoding:
|
| 289 |
+
if self.patch_position_encoding_type != "absolute":
|
| 290 |
+
raise ValueError(
|
| 291 |
+
f"Unknown patch_position_encoding_type: {self.patch_position_encoding_type}. "
|
| 292 |
+
"Only 'absolute' is supported."
|
| 293 |
+
)
|
| 294 |
+
self.pos_emb_h = nn.Embedding(max_position_embeddings, dim)
|
| 295 |
+
self.pos_emb_w = nn.Embedding(max_position_embeddings, dim)
|
| 296 |
+
|
| 297 |
+
def forward(self, x: torch.Tensor, patch_positions: Optional[torch.Tensor] = None) -> torch.Tensor:
|
| 298 |
+
"""
|
| 299 |
+
Merge patches from Qwen2VL-style input.
|
| 300 |
+
|
| 301 |
+
The input patches are already arranged in 2x2 block order by the image processor,
|
| 302 |
+
so we simply need to apply LayerNorm, reshape, and project through MLP.
|
| 303 |
+
|
| 304 |
+
Args:
|
| 305 |
+
x: Input tensor of shape [batch_size, seq_len, hidden_size] or [seq_len, hidden_size]
|
| 306 |
+
where seq_len = t * h * w (patches in 2x2 block order)
|
| 307 |
+
|
| 308 |
+
Returns:
|
| 309 |
+
Merged tensor of shape [batch_size, seq_len // spatial_merge_size^2, dim]
|
| 310 |
+
or [seq_len // spatial_merge_size^2, dim]
|
| 311 |
+
"""
|
| 312 |
+
if patch_positions is not None and patch_positions.dim() == 3:
|
| 313 |
+
patch_positions = patch_positions.squeeze(0)
|
| 314 |
+
|
| 315 |
+
x = self.ln_q(x).view(-1, self.hidden_size)
|
| 316 |
+
x = self.mlp(x)
|
| 317 |
+
|
| 318 |
+
if self.use_patch_position_encoding and patch_positions is not None:
|
| 319 |
+
pp = patch_positions.view(-1, self.spatial_merge_size**2, 3)
|
| 320 |
+
pp = pp[:, 0, :]
|
| 321 |
+
pp = (pp // self.spatial_merge_size).long()
|
| 322 |
+
|
| 323 |
+
x = x + self.pos_emb_h(pp[:, 1]) + self.pos_emb_w(pp[:, 2])
|
| 324 |
+
|
| 325 |
+
return x
|
| 326 |
+
|
| 327 |
+
|
| 328 |
+
def rotate_half(x):
|
| 329 |
+
"""
|
| 330 |
+
Interleaved rotation to match Source model's implementation.
|
| 331 |
+
(x1, x2, x3, x4) -> (-x2, x1, -x4, x3)
|
| 332 |
+
"""
|
| 333 |
+
x_even = x[..., ::2]
|
| 334 |
+
x_odd = x[..., 1::2]
|
| 335 |
+
return torch.stack((-x_odd, x_even), dim=-1).flatten(-2)
|
| 336 |
+
|
| 337 |
+
|
| 338 |
+
def get_norm_layer(config):
|
| 339 |
+
if config.layer_norm_type == "rms_norm":
|
| 340 |
+
return nn.RMSNorm(config.hidden_size, eps=config.layer_norm_eps)
|
| 341 |
+
else:
|
| 342 |
+
return nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
| 343 |
+
|
| 344 |
+
|
| 345 |
+
def apply_rotary_pos_emb(q, k, freqs):
|
| 346 |
+
# q, k: (B, H, L, D)
|
| 347 |
+
# freqs: (B, L, D)
|
| 348 |
+
orig_q_dtype = q.dtype
|
| 349 |
+
orig_k_dtype = k.dtype
|
| 350 |
+
q, k = q.float(), k.float()
|
| 351 |
+
# We need to broadcast freqs to match heads
|
| 352 |
+
# (B, L, D) -> (B, 1, L, D)
|
| 353 |
+
# Keep the same dtype as q, k to avoid memory doubling from float32 promotion
|
| 354 |
+
cos = freqs.cos().unsqueeze(1).float()
|
| 355 |
+
sin = freqs.sin().unsqueeze(1).float()
|
| 356 |
+
|
| 357 |
+
q_embed = (q * cos) + (rotate_half(q) * sin)
|
| 358 |
+
k_embed = (k * cos) + (rotate_half(k) * sin)
|
| 359 |
+
q_embed = q_embed.to(orig_q_dtype)
|
| 360 |
+
k_embed = k_embed.to(orig_k_dtype)
|
| 361 |
+
return q_embed, k_embed
|
| 362 |
+
|
| 363 |
+
|
| 364 |
+
def eager_attention_forward(
|
| 365 |
+
module: nn.Module,
|
| 366 |
+
query: torch.Tensor,
|
| 367 |
+
key: torch.Tensor,
|
| 368 |
+
value: torch.Tensor,
|
| 369 |
+
attention_mask: Optional[torch.Tensor],
|
| 370 |
+
scaling: float,
|
| 371 |
+
dropout: float = 0.0,
|
| 372 |
+
**kwargs,
|
| 373 |
+
):
|
| 374 |
+
"""Eager attention; query/key/value are expected as ``(B, H, L, D)``."""
|
| 375 |
+
attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling
|
| 376 |
+
if attention_mask is not None:
|
| 377 |
+
attn_weights = attn_weights + attention_mask
|
| 378 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
|
| 379 |
+
attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
|
| 380 |
+
attn_output = torch.matmul(attn_weights, value)
|
| 381 |
+
attn_output = attn_output.transpose(1, 2).contiguous() # (B, L, H, D)
|
| 382 |
+
return attn_output, attn_weights
|
| 383 |
+
|
| 384 |
+
|
| 385 |
+
class OneVisionEncoderAttention(nn.Module):
|
| 386 |
+
"""
|
| 387 |
+
Multi-headed attention with RoPE support, dispatched through
|
| 388 |
+
:data:`ALL_ATTENTION_FUNCTIONS` (``eager`` / ``sdpa`` / ``flash_attention_2``)
|
| 389 |
+
based on ``config._attn_implementation``.
|
| 390 |
+
"""
|
| 391 |
+
|
| 392 |
+
def __init__(self, config: LlavaOnevision2VisionConfig):
|
| 393 |
+
super().__init__()
|
| 394 |
+
self.config = config
|
| 395 |
+
self.embed_dim = config.hidden_size
|
| 396 |
+
self.num_heads = config.num_attention_heads
|
| 397 |
+
self.head_dim = self.embed_dim // self.num_heads
|
| 398 |
+
if self.head_dim * self.num_heads != self.embed_dim:
|
| 399 |
+
raise ValueError(
|
| 400 |
+
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {self.num_heads})."
|
| 401 |
+
)
|
| 402 |
+
|
| 403 |
+
self.num_key_value_groups = 1 # required by repeat_kv-aware eager paths
|
| 404 |
+
self.scale = self.head_dim**-0.5
|
| 405 |
+
self.scaling = self.scale # alias expected by some attention interfaces
|
| 406 |
+
self.attention_dropout = config.attention_dropout
|
| 407 |
+
self.is_causal = False
|
| 408 |
+
self.qkv = nn.Linear(self.embed_dim, self.embed_dim * 3)
|
| 409 |
+
self.proj = nn.Linear(self.embed_dim, self.embed_dim)
|
| 410 |
+
|
| 411 |
+
def forward(
|
| 412 |
+
self,
|
| 413 |
+
hidden_states: torch.Tensor,
|
| 414 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 415 |
+
rotary_pos_emb: Optional[torch.Tensor] = None,
|
| 416 |
+
output_attentions: bool = False,
|
| 417 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
| 418 |
+
max_seqlen: Optional[int] = None,
|
| 419 |
+
**kwargs,
|
| 420 |
+
) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
|
| 421 |
+
batch_size, q_len, _ = hidden_states.size()
|
| 422 |
+
# (B, L, 3*H*D) -> (B, L, 3, H, D) -> 3 x (B, L, H, D) -> 3 x (B, H, L, D)
|
| 423 |
+
q, k, v = (
|
| 424 |
+
self.qkv(hidden_states)
|
| 425 |
+
.reshape(batch_size, q_len, 3, self.num_heads, self.head_dim)
|
| 426 |
+
.permute(2, 0, 1, 3, 4)
|
| 427 |
+
.unbind(0)
|
| 428 |
+
)
|
| 429 |
+
query_states = q.transpose(1, 2)
|
| 430 |
+
key_states = k.transpose(1, 2)
|
| 431 |
+
value_states = v.transpose(1, 2)
|
| 432 |
+
|
| 433 |
+
if rotary_pos_emb is not None:
|
| 434 |
+
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, rotary_pos_emb)
|
| 435 |
+
|
| 436 |
+
attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
|
| 437 |
+
self.config._attn_implementation, eager_attention_forward
|
| 438 |
+
)
|
| 439 |
+
dropout = 0.0 if not self.training else self.attention_dropout
|
| 440 |
+
|
| 441 |
+
if cu_seqlens is not None and is_flash_attention_requested(self.config):
|
| 442 |
+
# Flash Attention varlen path: pass cu_seq_lens / max_length kwargs.
|
| 443 |
+
if max_seqlen is None:
|
| 444 |
+
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
|
| 445 |
+
attn_output, _ = attention_interface(
|
| 446 |
+
self,
|
| 447 |
+
query_states,
|
| 448 |
+
key_states,
|
| 449 |
+
value_states,
|
| 450 |
+
attention_mask=None,
|
| 451 |
+
scaling=self.scale,
|
| 452 |
+
dropout=dropout,
|
| 453 |
+
cu_seq_lens_q=cu_seqlens,
|
| 454 |
+
cu_seq_lens_k=cu_seqlens,
|
| 455 |
+
max_length_q=max_seqlen,
|
| 456 |
+
max_length_k=max_seqlen,
|
| 457 |
+
is_causal=False,
|
| 458 |
+
**kwargs,
|
| 459 |
+
)
|
| 460 |
+
elif cu_seqlens is not None:
|
| 461 |
+
# Non-FA implementations do not understand cu_seqlens directly; mirror
|
| 462 |
+
# Qwen3-VL by splitting the packed sequence into per-sample chunks
|
| 463 |
+
# along the L dim of (B, H, L, D) and running attention per chunk.
|
| 464 |
+
lengths = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
|
| 465 |
+
splits = [torch.split(t, lengths, dim=2) for t in (query_states, key_states, value_states)]
|
| 466 |
+
attn_outputs = [
|
| 467 |
+
attention_interface(
|
| 468 |
+
self,
|
| 469 |
+
q_chunk,
|
| 470 |
+
k_chunk,
|
| 471 |
+
v_chunk,
|
| 472 |
+
attention_mask=None,
|
| 473 |
+
scaling=self.scale,
|
| 474 |
+
dropout=dropout,
|
| 475 |
+
is_causal=False,
|
| 476 |
+
**kwargs,
|
| 477 |
+
)[0]
|
| 478 |
+
for q_chunk, k_chunk, v_chunk in zip(*splits)
|
| 479 |
+
]
|
| 480 |
+
# interface output is (B, l_i, H, D); concat along the L axis
|
| 481 |
+
attn_output = torch.cat(attn_outputs, dim=1)
|
| 482 |
+
else:
|
| 483 |
+
attn_mask = None
|
| 484 |
+
if attention_mask is not None:
|
| 485 |
+
attn_mask = attention_mask
|
| 486 |
+
if attn_mask.dim() == 2:
|
| 487 |
+
attn_mask = attn_mask.unsqueeze(0)
|
| 488 |
+
if attn_mask.shape[0] == 1 and batch_size > 1:
|
| 489 |
+
attn_mask = attn_mask.expand(batch_size, -1, -1)
|
| 490 |
+
attn_mask = attn_mask.unsqueeze(1) # (B, 1, L, L)
|
| 491 |
+
attn_output, _ = attention_interface(
|
| 492 |
+
self,
|
| 493 |
+
query_states,
|
| 494 |
+
key_states,
|
| 495 |
+
value_states,
|
| 496 |
+
attention_mask=attn_mask,
|
| 497 |
+
scaling=self.scale,
|
| 498 |
+
dropout=dropout,
|
| 499 |
+
is_causal=False,
|
| 500 |
+
**kwargs,
|
| 501 |
+
)
|
| 502 |
+
|
| 503 |
+
attn_output = attn_output.reshape(batch_size, q_len, self.embed_dim)
|
| 504 |
+
attn_output = self.proj(attn_output)
|
| 505 |
+
|
| 506 |
+
return attn_output, None
|
| 507 |
+
|
| 508 |
+
|
| 509 |
+
class OneVisionEncoderEncoderLayer(nn.Module):
|
| 510 |
+
"""Vision encoder layer with pre-norm and Flash Attention 2."""
|
| 511 |
+
|
| 512 |
+
def __init__(self, config: LlavaOnevision2VisionConfig):
|
| 513 |
+
super().__init__()
|
| 514 |
+
self.embed_dim = config.hidden_size
|
| 515 |
+
self.self_attn = OneVisionEncoderAttention(config)
|
| 516 |
+
self.layer_norm1 = get_norm_layer(config)
|
| 517 |
+
self.mlp = SiglipMLP(config)
|
| 518 |
+
self.layer_norm2 = get_norm_layer(config)
|
| 519 |
+
|
| 520 |
+
def forward(
|
| 521 |
+
self,
|
| 522 |
+
hidden_states: torch.Tensor,
|
| 523 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 524 |
+
rotary_pos_emb: Optional[torch.Tensor] = None,
|
| 525 |
+
output_attentions: bool = False,
|
| 526 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
| 527 |
+
max_seqlen: Optional[int] = None,
|
| 528 |
+
) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
|
| 529 |
+
residual = hidden_states
|
| 530 |
+
hidden_states = self.layer_norm1(hidden_states)
|
| 531 |
+
|
| 532 |
+
hidden_states, attn_weights = self.self_attn(
|
| 533 |
+
hidden_states=hidden_states,
|
| 534 |
+
attention_mask=attention_mask,
|
| 535 |
+
rotary_pos_emb=rotary_pos_emb,
|
| 536 |
+
output_attentions=output_attentions,
|
| 537 |
+
cu_seqlens=cu_seqlens,
|
| 538 |
+
max_seqlen=max_seqlen,
|
| 539 |
+
)
|
| 540 |
+
hidden_states = residual + hidden_states
|
| 541 |
+
|
| 542 |
+
residual = hidden_states
|
| 543 |
+
hidden_states = self.layer_norm2(hidden_states)
|
| 544 |
+
hidden_states = self.mlp(hidden_states)
|
| 545 |
+
hidden_states = residual + hidden_states
|
| 546 |
+
|
| 547 |
+
outputs = (hidden_states, attn_weights) if output_attentions else (hidden_states,)
|
| 548 |
+
return outputs
|
| 549 |
+
|
| 550 |
+
|
| 551 |
+
class OneVisionEncoderEncoder(nn.Module):
|
| 552 |
+
def __init__(self, config: LlavaOnevision2VisionConfig):
|
| 553 |
+
super().__init__()
|
| 554 |
+
self.config = config
|
| 555 |
+
self.layers = nn.ModuleList([OneVisionEncoderEncoderLayer(config) for _ in range(config.num_hidden_layers)])
|
| 556 |
+
# Gradient checkpointing support
|
| 557 |
+
self.gradient_checkpointing = False
|
| 558 |
+
|
| 559 |
+
def forward(
|
| 560 |
+
self,
|
| 561 |
+
hidden_states: torch.Tensor,
|
| 562 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 563 |
+
rotary_pos_emb: Optional[torch.Tensor] = None,
|
| 564 |
+
output_attentions: bool = False,
|
| 565 |
+
output_hidden_states: bool = False,
|
| 566 |
+
return_dict: bool = True,
|
| 567 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
| 568 |
+
max_seqlen: Optional[int] = None,
|
| 569 |
+
) -> Union[tuple, BaseModelOutput]:
|
| 570 |
+
all_hidden_states = () if output_hidden_states else None
|
| 571 |
+
all_self_attentions = () if output_attentions else None
|
| 572 |
+
|
| 573 |
+
for layer in self.layers:
|
| 574 |
+
if output_hidden_states:
|
| 575 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
| 576 |
+
|
| 577 |
+
if self.gradient_checkpointing and self.training:
|
| 578 |
+
layer_outputs = self._gradient_checkpointing_func(
|
| 579 |
+
layer.__call__,
|
| 580 |
+
hidden_states,
|
| 581 |
+
attention_mask,
|
| 582 |
+
rotary_pos_emb,
|
| 583 |
+
output_attentions,
|
| 584 |
+
cu_seqlens,
|
| 585 |
+
max_seqlen,
|
| 586 |
+
)
|
| 587 |
+
else:
|
| 588 |
+
layer_outputs = layer(
|
| 589 |
+
hidden_states,
|
| 590 |
+
attention_mask=attention_mask,
|
| 591 |
+
rotary_pos_emb=rotary_pos_emb,
|
| 592 |
+
output_attentions=output_attentions,
|
| 593 |
+
cu_seqlens=cu_seqlens,
|
| 594 |
+
max_seqlen=max_seqlen,
|
| 595 |
+
)
|
| 596 |
+
|
| 597 |
+
hidden_states = layer_outputs[0]
|
| 598 |
+
|
| 599 |
+
if output_attentions:
|
| 600 |
+
all_self_attentions = all_self_attentions + (layer_outputs[1],)
|
| 601 |
+
|
| 602 |
+
if output_hidden_states:
|
| 603 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
| 604 |
+
|
| 605 |
+
if not return_dict:
|
| 606 |
+
return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None)
|
| 607 |
+
|
| 608 |
+
return BaseModelOutput(
|
| 609 |
+
last_hidden_state=hidden_states,
|
| 610 |
+
hidden_states=all_hidden_states,
|
| 611 |
+
attentions=all_self_attentions,
|
| 612 |
+
)
|
| 613 |
+
|
| 614 |
+
|
| 615 |
+
class LlavaOnevision2PreTrainedModel(PreTrainedModel):
|
| 616 |
+
config_class = LlavaOnevision2Config
|
| 617 |
+
base_model_prefix = "model"
|
| 618 |
+
input_modalities = ("image", "video", "text")
|
| 619 |
+
supports_gradient_checkpointing = True
|
| 620 |
+
_no_split_modules = ["OneVisionEncoderEncoderLayer", "Qwen3DecoderLayer"]
|
| 621 |
+
_skip_keys_device_placement = "past_key_values"
|
| 622 |
+
_supports_flash_attn = True
|
| 623 |
+
_supports_sdpa = True
|
| 624 |
+
|
| 625 |
+
def _init_weights(self, module):
|
| 626 |
+
super()._init_weights(module)
|
| 627 |
+
# Re-initialize VisionRotaryEmbedding inv_freq buffers.
|
| 628 |
+
# These are registered with persistent=False, so they are not in the checkpoint
|
| 629 |
+
# state_dict. When ``from_pretrained`` materializes the model from meta tensors,
|
| 630 |
+
# the values in these buffers end up uninitialized. Mirror Qwen3-VL by explicitly
|
| 631 |
+
# filling them here so RoPE produces the correct frequencies post-load.
|
| 632 |
+
if isinstance(module, VisionRotaryEmbedding):
|
| 633 |
+
base = module.base
|
| 634 |
+
with torch.no_grad():
|
| 635 |
+
inv_t = 1.0 / (base ** (torch.arange(module.t_size, dtype=torch.float32) / module.t_size))
|
| 636 |
+
inv_h = 1.0 / (base ** (torch.arange(module.h_size, dtype=torch.float32) / module.h_size))
|
| 637 |
+
inv_w = 1.0 / (base ** (torch.arange(module.w_size, dtype=torch.float32) / module.w_size))
|
| 638 |
+
module.inv_freq_t.copy_(inv_t.to(module.inv_freq_t.device))
|
| 639 |
+
module.inv_freq_h.copy_(inv_h.to(module.inv_freq_h.device))
|
| 640 |
+
module.inv_freq_w.copy_(inv_w.to(module.inv_freq_w.device))
|
| 641 |
+
|
| 642 |
+
|
| 643 |
+
class Siglip2MultiheadAttentionPoolingHead(nn.Module):
|
| 644 |
+
"""
|
| 645 |
+
Multi-Head Attention Pooling with a learned probe (PMA-style).
|
| 646 |
+
"""
|
| 647 |
+
|
| 648 |
+
def __init__(self, config: LlavaOnevision2VisionConfig):
|
| 649 |
+
super().__init__()
|
| 650 |
+
self.embed_dim = config.hidden_size
|
| 651 |
+
self.probe = nn.Parameter(torch.randn(1, 1, config.hidden_size))
|
| 652 |
+
self.attention = nn.MultiheadAttention(config.hidden_size, config.num_attention_heads, batch_first=True)
|
| 653 |
+
self.norm = nn.RMSNorm(config.hidden_size, eps=config.layer_norm_eps)
|
| 654 |
+
self.mlp = SiglipMLP(config)
|
| 655 |
+
|
| 656 |
+
def forward(self, hidden_states):
|
| 657 |
+
batch_size = hidden_states.shape[0]
|
| 658 |
+
probe = self.probe.repeat(batch_size, 1, 1)
|
| 659 |
+
|
| 660 |
+
attn_output, _ = self.attention(probe, hidden_states, hidden_states)
|
| 661 |
+
|
| 662 |
+
residual = attn_output
|
| 663 |
+
attn_output = self.norm(attn_output)
|
| 664 |
+
attn_output = residual + self.mlp(attn_output)
|
| 665 |
+
|
| 666 |
+
return attn_output[:, 0]
|
| 667 |
+
|
| 668 |
+
|
| 669 |
+
# ---------------------------------------------------------------------------
|
| 670 |
+
# Vision Model
|
| 671 |
+
# ---------------------------------------------------------------------------
|
| 672 |
+
|
| 673 |
+
|
| 674 |
+
class LlavaOnevision2VisionPretrainedModel(LlavaOnevision2PreTrainedModel):
|
| 675 |
+
"""
|
| 676 |
+
LLaVA-OneVision 2.0 Vision Model.
|
| 677 |
+
|
| 678 |
+
This vision model is designed to work with Qwen2VL-style image processing:
|
| 679 |
+
- Receives pre-processed patches in 2x2 block spatial order
|
| 680 |
+
- Applies RoPE with matching 2x2 block layout conversion
|
| 681 |
+
- Accepts explicit patch_positions for RoPE computation
|
| 682 |
+
|
| 683 |
+
Input format:
|
| 684 |
+
hidden_state: [total_patches, num_channels, patch_size, patch_size]
|
| 685 |
+
grid_thw: [num_samples, 3] with [t, h, w] for each sample
|
| 686 |
+
"""
|
| 687 |
+
|
| 688 |
+
def __init__(self, config: LlavaOnevision2VisionConfig):
|
| 689 |
+
super().__init__(config)
|
| 690 |
+
self.config = config
|
| 691 |
+
self.spatial_merge_size = config.spatial_merge_size
|
| 692 |
+
|
| 693 |
+
# Vision components
|
| 694 |
+
self.embeddings = OneVisionEncoderEmbeddings(config)
|
| 695 |
+
self.layernorm_pre = get_norm_layer(config)
|
| 696 |
+
self.encoder = OneVisionEncoderEncoder(config)
|
| 697 |
+
self.video_rope = VisionRotaryEmbedding(config)
|
| 698 |
+
|
| 699 |
+
if config.use_head:
|
| 700 |
+
self.layernorm_post = get_norm_layer(config)
|
| 701 |
+
self.head = Siglip2MultiheadAttentionPoolingHead(config)
|
| 702 |
+
else:
|
| 703 |
+
self.layernorm_post = None
|
| 704 |
+
self.head = None
|
| 705 |
+
|
| 706 |
+
self.merger = LlavaOnevision2VisionPatchMerger(
|
| 707 |
+
dim=config.out_hidden_size,
|
| 708 |
+
context_dim=config.hidden_size,
|
| 709 |
+
spatial_merge_size=config.spatial_merge_size,
|
| 710 |
+
layer_norm_eps=config.layer_norm_eps,
|
| 711 |
+
use_patch_position_encoding=getattr(config, "use_patch_position_encoding", False),
|
| 712 |
+
patch_position_encoding_type=getattr(config, "patch_position_encoding_type", "absolute"),
|
| 713 |
+
max_position_embeddings=getattr(config, "max_position_embeddings", 8192),
|
| 714 |
+
)
|
| 715 |
+
|
| 716 |
+
self.post_init()
|
| 717 |
+
|
| 718 |
+
def _build_cu_seqlens(
|
| 719 |
+
self,
|
| 720 |
+
grid_thw: torch.Tensor,
|
| 721 |
+
total_patches: int,
|
| 722 |
+
fixed_t: Optional[int] = 4,
|
| 723 |
+
device: Optional[torch.device] = None,
|
| 724 |
+
) -> tuple[torch.Tensor, int]:
|
| 725 |
+
if grid_thw is None or grid_thw.numel() == 0:
|
| 726 |
+
# Fallback for no grid_thw: treat as single sequence
|
| 727 |
+
return torch.tensor([0, total_patches], dtype=torch.int32, device=device), total_patches
|
| 728 |
+
|
| 729 |
+
if device is None:
|
| 730 |
+
device = grid_thw.device
|
| 731 |
+
|
| 732 |
+
cu_seqlens = [0]
|
| 733 |
+
max_seqlen = 0
|
| 734 |
+
total_entries = grid_thw.shape[0]
|
| 735 |
+
current_len = 0
|
| 736 |
+
|
| 737 |
+
# Calculate cumulative lengths: split sequences based on fixed_t if provided
|
| 738 |
+
for idx in range(total_entries):
|
| 739 |
+
t_val = grid_thw[idx, 0].item()
|
| 740 |
+
h_val = grid_thw[idx, 1].item()
|
| 741 |
+
w_val = grid_thw[idx, 2].item()
|
| 742 |
+
|
| 743 |
+
if fixed_t is not None and fixed_t > 0 and t_val > fixed_t:
|
| 744 |
+
# Split large t into chunks of fixed_t
|
| 745 |
+
num_full_windows = t_val // fixed_t
|
| 746 |
+
remainder = t_val % fixed_t
|
| 747 |
+
|
| 748 |
+
# Add full windows
|
| 749 |
+
for _ in range(num_full_windows):
|
| 750 |
+
chunk_patches = fixed_t * int(h_val) * int(w_val)
|
| 751 |
+
current_len += chunk_patches
|
| 752 |
+
max_seqlen = max(max_seqlen, chunk_patches)
|
| 753 |
+
cu_seqlens.append(current_len)
|
| 754 |
+
|
| 755 |
+
# Add remainder if any
|
| 756 |
+
if remainder > 0:
|
| 757 |
+
chunk_patches = remainder * int(h_val) * int(w_val)
|
| 758 |
+
current_len += chunk_patches
|
| 759 |
+
max_seqlen = max(max_seqlen, chunk_patches)
|
| 760 |
+
cu_seqlens.append(current_len)
|
| 761 |
+
else:
|
| 762 |
+
# Standard case: add as one chunk
|
| 763 |
+
chunk_patches = t_val * int(h_val) * int(w_val)
|
| 764 |
+
current_len += chunk_patches
|
| 765 |
+
max_seqlen = max(max_seqlen, chunk_patches)
|
| 766 |
+
cu_seqlens.append(current_len)
|
| 767 |
+
|
| 768 |
+
last_len = cu_seqlens[-1]
|
| 769 |
+
if last_len != total_patches:
|
| 770 |
+
raise ValueError(
|
| 771 |
+
"cu_seqlens calculation mismatch:\n"
|
| 772 |
+
f"- total_patches: {total_patches}\n"
|
| 773 |
+
f"- calculated total: {last_len}\n"
|
| 774 |
+
f"- grid_thw: {grid_thw}"
|
| 775 |
+
)
|
| 776 |
+
|
| 777 |
+
return torch.tensor(cu_seqlens, dtype=torch.int32, device=device), max_seqlen
|
| 778 |
+
|
| 779 |
+
def _build_block_attention_mask(
|
| 780 |
+
self,
|
| 781 |
+
grid_thw: torch.Tensor,
|
| 782 |
+
total_patches: int,
|
| 783 |
+
fixed_t: Optional[int] = 4,
|
| 784 |
+
device: Optional[torch.device] = None,
|
| 785 |
+
) -> Optional[torch.Tensor]:
|
| 786 |
+
if grid_thw is None or grid_thw.numel() == 0:
|
| 787 |
+
return None
|
| 788 |
+
|
| 789 |
+
if device is None:
|
| 790 |
+
device = grid_thw.device
|
| 791 |
+
|
| 792 |
+
lengths = []
|
| 793 |
+
total_entries = grid_thw.shape[0]
|
| 794 |
+
|
| 795 |
+
for idx in range(total_entries):
|
| 796 |
+
t_val = grid_thw[idx, 0].item()
|
| 797 |
+
h_val = grid_thw[idx, 1].item()
|
| 798 |
+
w_val = grid_thw[idx, 2].item()
|
| 799 |
+
|
| 800 |
+
if fixed_t is not None and fixed_t > 0 and t_val > fixed_t:
|
| 801 |
+
# Split large t into chunks of fixed_t
|
| 802 |
+
num_full_windows = t_val // fixed_t
|
| 803 |
+
remainder = t_val % fixed_t
|
| 804 |
+
|
| 805 |
+
# Add full windows
|
| 806 |
+
for _ in range(num_full_windows):
|
| 807 |
+
lengths.append(fixed_t * int(h_val) * int(w_val))
|
| 808 |
+
|
| 809 |
+
# Add remainder if any
|
| 810 |
+
if remainder > 0:
|
| 811 |
+
lengths.append(remainder * int(h_val) * int(w_val))
|
| 812 |
+
else:
|
| 813 |
+
lengths.append(t_val * int(h_val) * int(w_val))
|
| 814 |
+
|
| 815 |
+
total_len = sum(lengths)
|
| 816 |
+
if total_len != total_patches:
|
| 817 |
+
raise ValueError(
|
| 818 |
+
"Block attention mask length mismatch:\n"
|
| 819 |
+
f"- total_patches: {total_patches}\n"
|
| 820 |
+
f"- total_len: {total_len}\n"
|
| 821 |
+
f"- grid_thw: {grid_thw}"
|
| 822 |
+
)
|
| 823 |
+
|
| 824 |
+
attn_mask = torch.ones((total_len, total_len), dtype=torch.bool, device=device)
|
| 825 |
+
start = 0
|
| 826 |
+
for size in lengths:
|
| 827 |
+
end = start + size
|
| 828 |
+
attn_mask[start:end, start:end] = False
|
| 829 |
+
start = end
|
| 830 |
+
|
| 831 |
+
return attn_mask
|
| 832 |
+
|
| 833 |
+
@replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=LlavaOnevision2VisionConfig)
|
| 834 |
+
def forward(
|
| 835 |
+
self,
|
| 836 |
+
hidden_state: torch.Tensor,
|
| 837 |
+
grid_thw: Optional[torch.Tensor] = None,
|
| 838 |
+
patch_positions: Optional[torch.Tensor] = None,
|
| 839 |
+
output_attentions: Optional[bool] = None,
|
| 840 |
+
output_hidden_states: Optional[bool] = None,
|
| 841 |
+
return_dict: Optional[bool] = None,
|
| 842 |
+
skip_merger: Optional[bool] = False,
|
| 843 |
+
) -> Union[tuple, BaseModelOutputWithPooling]:
|
| 844 |
+
r"""
|
| 845 |
+
Forward pass for vision model.
|
| 846 |
+
|
| 847 |
+
This method accepts pre-processed patches from Qwen2VL image processor and applies
|
| 848 |
+
RoPE (Rotary Position Embedding) in 2x2 block layout to match the spatial arrangement
|
| 849 |
+
of patches.
|
| 850 |
+
|
| 851 |
+
Args:
|
| 852 |
+
hidden_state: Pre-processed patches from Qwen2VL processor.
|
| 853 |
+
Shape: [total_patches, num_channels, patch_size, patch_size]
|
| 854 |
+
grid_thw: Grid sizes tensor of shape [num_samples, 3] with [t, h, w] for each sample.
|
| 855 |
+
Required for computing RoPE and handling visible indices.
|
| 856 |
+
patch_positions: Optional explicit patch positions for RoPE computation.
|
| 857 |
+
output_attentions: Whether to return attention weights.
|
| 858 |
+
output_hidden_states: Whether to return all hidden states.
|
| 859 |
+
return_dict: Whether to return a ModelOutput instead of tuple.
|
| 860 |
+
skip_merger: If True, skip patch merger (useful for consistency checking).
|
| 861 |
+
|
| 862 |
+
Returns:
|
| 863 |
+
BaseModelOutputWithPooling with last_hidden_state containing merged features.
|
| 864 |
+
"""
|
| 865 |
+
output_attentions = (
|
| 866 |
+
output_attentions if output_attentions is not None else getattr(self.config, "output_attentions", False)
|
| 867 |
+
)
|
| 868 |
+
output_hidden_states = (
|
| 869 |
+
output_hidden_states
|
| 870 |
+
if output_hidden_states is not None
|
| 871 |
+
else getattr(self.config, "output_hidden_states", False)
|
| 872 |
+
)
|
| 873 |
+
return_dict = True if return_dict is None else return_dict
|
| 874 |
+
|
| 875 |
+
# 1. Embeddings
|
| 876 |
+
# Note: embeddings returns [total_patches, embed_dim], we need to add batch dimension
|
| 877 |
+
hidden_states = self.embeddings(hidden_state)
|
| 878 |
+
if hidden_states.dim() == 2:
|
| 879 |
+
hidden_states = hidden_states.unsqueeze(0) # [1, total_patches, embed_dim]
|
| 880 |
+
batch_size, total_patches, _ = hidden_states.shape
|
| 881 |
+
|
| 882 |
+
# 2. RoPE Construction
|
| 883 |
+
if patch_positions is not None and patch_positions.dim() == 3:
|
| 884 |
+
patch_positions = patch_positions.squeeze(0)
|
| 885 |
+
freqs_visible = self.video_rope.forward_from_positions(patch_positions)
|
| 886 |
+
|
| 887 |
+
# Concatenate D/2 + D/2 -> D for applying rope
|
| 888 |
+
freqs_visible = torch.cat([freqs_visible, freqs_visible], dim=-1)
|
| 889 |
+
if freqs_visible.dim() == 2:
|
| 890 |
+
freqs_visible = freqs_visible.unsqueeze(0)
|
| 891 |
+
|
| 892 |
+
# 3. Pre-Norm & Encoder
|
| 893 |
+
hidden_states = self.layernorm_pre(hidden_states)
|
| 894 |
+
|
| 895 |
+
cu_seqlens, max_seqlen = self._build_cu_seqlens(
|
| 896 |
+
grid_thw=grid_thw,
|
| 897 |
+
total_patches=total_patches,
|
| 898 |
+
fixed_t=getattr(self.config, "frame_windows_size", 4),
|
| 899 |
+
device=hidden_states.device,
|
| 900 |
+
)
|
| 901 |
+
|
| 902 |
+
encoder_outputs = self.encoder(
|
| 903 |
+
hidden_states,
|
| 904 |
+
attention_mask=None,
|
| 905 |
+
rotary_pos_emb=freqs_visible,
|
| 906 |
+
output_attentions=output_attentions,
|
| 907 |
+
output_hidden_states=True, # Always get hidden states to use -2 layer
|
| 908 |
+
return_dict=True,
|
| 909 |
+
cu_seqlens=cu_seqlens,
|
| 910 |
+
max_seqlen=max_seqlen,
|
| 911 |
+
)
|
| 912 |
+
|
| 913 |
+
# Use second-to-last layer output for better feature representation
|
| 914 |
+
if encoder_outputs.hidden_states is not None and len(encoder_outputs.hidden_states) >= 2 and not skip_merger:
|
| 915 |
+
sequence_output = encoder_outputs.hidden_states[-1]
|
| 916 |
+
else:
|
| 917 |
+
sequence_output = encoder_outputs[0]
|
| 918 |
+
|
| 919 |
+
# Post-Norm
|
| 920 |
+
if self.layernorm_post is not None:
|
| 921 |
+
sequence_output = self.layernorm_post(sequence_output)
|
| 922 |
+
|
| 923 |
+
# Skip merger for consistency check with original ViT
|
| 924 |
+
if skip_merger:
|
| 925 |
+
pooled_output = None
|
| 926 |
+
if self.head is not None:
|
| 927 |
+
pooled_output = self.head(sequence_output)
|
| 928 |
+
|
| 929 |
+
if not return_dict:
|
| 930 |
+
return (sequence_output, pooled_output) + (
|
| 931 |
+
encoder_outputs.hidden_states if output_hidden_states else None,
|
| 932 |
+
)
|
| 933 |
+
return BaseModelOutputWithPooling(
|
| 934 |
+
last_hidden_state=sequence_output,
|
| 935 |
+
pooler_output=pooled_output,
|
| 936 |
+
hidden_states=encoder_outputs.hidden_states if output_hidden_states else None,
|
| 937 |
+
attentions=encoder_outputs.attentions if output_attentions else None,
|
| 938 |
+
)
|
| 939 |
+
|
| 940 |
+
# Patch merger: input patches are already in 2x2 block order from Qwen2VL processor
|
| 941 |
+
merged_output = self.merger(sequence_output, patch_positions=patch_positions)
|
| 942 |
+
|
| 943 |
+
if not return_dict:
|
| 944 |
+
return (merged_output,) + (encoder_outputs.hidden_states if output_hidden_states else None,)
|
| 945 |
+
|
| 946 |
+
return BaseModelOutputWithPooling(
|
| 947 |
+
last_hidden_state=merged_output,
|
| 948 |
+
pooler_output=None,
|
| 949 |
+
hidden_states=encoder_outputs.hidden_states if output_hidden_states else None,
|
| 950 |
+
attentions=encoder_outputs.attentions if output_attentions else None,
|
| 951 |
+
)
|
| 952 |
+
|
| 953 |
+
|
| 954 |
+
@auto_docstring
|
| 955 |
+
class LlavaOnevision2Model(LlavaOnevision2PreTrainedModel):
|
| 956 |
+
base_model_prefix = ""
|
| 957 |
+
# Reference: fix gemma3 grad acc #37208
|
| 958 |
+
accepts_loss_kwargs = False
|
| 959 |
+
config: LlavaOnevision2Config
|
| 960 |
+
_no_split_modules = ["OneVisionEncoderEncoderLayer", "Qwen3DecoderLayer"]
|
| 961 |
+
|
| 962 |
+
def __init__(self, config: LlavaOnevision2Config):
|
| 963 |
+
super().__init__(config)
|
| 964 |
+
self.visual = LlavaOnevision2VisionPretrainedModel._from_config(config.vision_config)
|
| 965 |
+
self.language_model = AutoModel.from_config(config.text_config)
|
| 966 |
+
|
| 967 |
+
# Initialize weights and apply final processing
|
| 968 |
+
self.post_init()
|
| 969 |
+
|
| 970 |
+
def get_input_embeddings(self):
|
| 971 |
+
return self.language_model.get_input_embeddings()
|
| 972 |
+
|
| 973 |
+
def set_input_embeddings(self, value):
|
| 974 |
+
self.language_model.set_input_embeddings(value)
|
| 975 |
+
|
| 976 |
+
def set_decoder(self, decoder):
|
| 977 |
+
self.language_model = decoder
|
| 978 |
+
|
| 979 |
+
def get_decoder(self):
|
| 980 |
+
return self.language_model
|
| 981 |
+
|
| 982 |
+
def get_video_features(
|
| 983 |
+
self,
|
| 984 |
+
pixel_values_videos: torch.FloatTensor,
|
| 985 |
+
video_grid_thw: Optional[torch.LongTensor] = None,
|
| 986 |
+
patch_positions=None,
|
| 987 |
+
):
|
| 988 |
+
"""
|
| 989 |
+
Encodes videos into continuous embeddings that can be forwarded to the language model.
|
| 990 |
+
|
| 991 |
+
Args:
|
| 992 |
+
pixel_values_videos: Pre-processed patches from Qwen2VL processor.
|
| 993 |
+
`torch.FloatTensor` of shape `(total_patches, num_channels, patch_size, patch_size)`
|
| 994 |
+
video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
|
| 995 |
+
The temporal, height and width of feature shape of each video in LLM.
|
| 996 |
+
"""
|
| 997 |
+
# Convert to correct dtype
|
| 998 |
+
pixel_values_videos = pixel_values_videos.type(self.visual.embeddings.patch_embedding.weight.dtype)
|
| 999 |
+
|
| 1000 |
+
# Forward through vision model with grid_thw
|
| 1001 |
+
vision_output = self.visual(pixel_values_videos, grid_thw=video_grid_thw, patch_positions=patch_positions)
|
| 1002 |
+
|
| 1003 |
+
# Extract the actual tensor from BaseModelOutputWithPooling
|
| 1004 |
+
if hasattr(vision_output, "last_hidden_state"):
|
| 1005 |
+
video_embeds = vision_output.last_hidden_state
|
| 1006 |
+
else:
|
| 1007 |
+
video_embeds = vision_output[0] # Fallback for tuple output
|
| 1008 |
+
|
| 1009 |
+
# Compute split sizes from video_grid_thw or from input shape
|
| 1010 |
+
if video_grid_thw is not None:
|
| 1011 |
+
split_sizes = (video_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
|
| 1012 |
+
else:
|
| 1013 |
+
# Compute from input shape
|
| 1014 |
+
batch_size = pixel_values_videos.shape[0]
|
| 1015 |
+
split_sizes = [video_embeds.shape[1]] * batch_size
|
| 1016 |
+
|
| 1017 |
+
# Split embeddings per video
|
| 1018 |
+
if len(split_sizes) > 1:
|
| 1019 |
+
video_embeds = torch.split(video_embeds.view(-1, video_embeds.shape[-1]), split_sizes)
|
| 1020 |
+
else:
|
| 1021 |
+
video_embeds = [video_embeds.view(-1, video_embeds.shape[-1])]
|
| 1022 |
+
|
| 1023 |
+
return video_embeds
|
| 1024 |
+
|
| 1025 |
+
def get_image_features(
|
| 1026 |
+
self, pixel_values, image_grid_thw: Optional[torch.LongTensor] = None, patch_positions=None
|
| 1027 |
+
):
|
| 1028 |
+
"""
|
| 1029 |
+
Encodes images into continuous embeddings that can be forwarded to the language model.
|
| 1030 |
+
|
| 1031 |
+
Args:
|
| 1032 |
+
pixel_values: Pre-processed patches from Qwen2VL processor.
|
| 1033 |
+
- `torch.FloatTensor` of shape `(total_patches, num_channels, patch_size, patch_size)`
|
| 1034 |
+
image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
|
| 1035 |
+
The temporal, height and width of feature shape of each image in LLM.
|
| 1036 |
+
"""
|
| 1037 |
+
# Standard format from Qwen2VL processor
|
| 1038 |
+
if pixel_values.dim() == 2:
|
| 1039 |
+
# Convert to correct dtype
|
| 1040 |
+
pixel_values = pixel_values.type(self.visual.embeddings.patch_embedding.weight.dtype)
|
| 1041 |
+
|
| 1042 |
+
# Forward through vision model with grid_thw
|
| 1043 |
+
vision_output = self.visual(pixel_values, grid_thw=image_grid_thw, patch_positions=patch_positions)
|
| 1044 |
+
|
| 1045 |
+
# Extract the actual tensor from BaseModelOutputWithPooling
|
| 1046 |
+
if hasattr(vision_output, "last_hidden_state"):
|
| 1047 |
+
image_embeds = vision_output.last_hidden_state
|
| 1048 |
+
else:
|
| 1049 |
+
image_embeds = vision_output[0]
|
| 1050 |
+
|
| 1051 |
+
# Compute split sizes from grid_thw
|
| 1052 |
+
if image_grid_thw is not None:
|
| 1053 |
+
split_sizes = (image_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
|
| 1054 |
+
else:
|
| 1055 |
+
# Fallback: assume single image
|
| 1056 |
+
split_sizes = [image_embeds.shape[0] if image_embeds.dim() == 2 else image_embeds.shape[1]]
|
| 1057 |
+
|
| 1058 |
+
# Split embeddings per image
|
| 1059 |
+
image_embeds_flat = image_embeds.view(-1, image_embeds.shape[-1])
|
| 1060 |
+
if len(split_sizes) > 1:
|
| 1061 |
+
image_embeds = list(torch.split(image_embeds_flat, split_sizes))
|
| 1062 |
+
else:
|
| 1063 |
+
image_embeds = [image_embeds_flat]
|
| 1064 |
+
|
| 1065 |
+
return image_embeds
|
| 1066 |
+
else:
|
| 1067 |
+
raise ValueError(
|
| 1068 |
+
f"Unsupported pixel_values shape: expected 4D tensor [total_patches, C, H, W], "
|
| 1069 |
+
f"got {pixel_values.shape if hasattr(pixel_values, 'shape') else type(pixel_values)}"
|
| 1070 |
+
)
|
| 1071 |
+
|
| 1072 |
+
def get_placeholder_mask(
|
| 1073 |
+
self,
|
| 1074 |
+
input_ids: torch.LongTensor,
|
| 1075 |
+
inputs_embeds: torch.FloatTensor,
|
| 1076 |
+
image_features: Optional[torch.FloatTensor] = None,
|
| 1077 |
+
video_features: Optional[torch.FloatTensor] = None,
|
| 1078 |
+
):
|
| 1079 |
+
"""
|
| 1080 |
+
Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is
|
| 1081 |
+
equal to the length of multimodal features. If the lengths are different, an error is raised.
|
| 1082 |
+
"""
|
| 1083 |
+
if input_ids is None:
|
| 1084 |
+
special_image_mask = inputs_embeds == self.get_input_embeddings()(
|
| 1085 |
+
torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device)
|
| 1086 |
+
)
|
| 1087 |
+
special_image_mask = special_image_mask.all(-1)
|
| 1088 |
+
special_video_mask = inputs_embeds == self.get_input_embeddings()(
|
| 1089 |
+
torch.tensor(self.config.video_token_id, dtype=torch.long, device=inputs_embeds.device)
|
| 1090 |
+
)
|
| 1091 |
+
special_video_mask = special_video_mask.all(-1)
|
| 1092 |
+
else:
|
| 1093 |
+
special_image_mask = input_ids == self.config.image_token_id
|
| 1094 |
+
special_video_mask = input_ids == self.config.video_token_id
|
| 1095 |
+
|
| 1096 |
+
n_image_tokens = special_image_mask.sum()
|
| 1097 |
+
special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
|
| 1098 |
+
if image_features is not None and inputs_embeds[special_image_mask].numel() != image_features.numel():
|
| 1099 |
+
raise ValueError(
|
| 1100 |
+
f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {image_features.shape[0]}"
|
| 1101 |
+
)
|
| 1102 |
+
|
| 1103 |
+
n_video_tokens = special_video_mask.sum()
|
| 1104 |
+
special_video_mask = special_video_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
|
| 1105 |
+
if video_features is not None and inputs_embeds[special_video_mask].numel() != video_features.numel():
|
| 1106 |
+
raise ValueError(
|
| 1107 |
+
f"Videos features and video tokens do not match: tokens: {n_video_tokens}, features {video_features.shape[0]}"
|
| 1108 |
+
)
|
| 1109 |
+
|
| 1110 |
+
return special_image_mask, special_video_mask
|
| 1111 |
+
|
| 1112 |
+
@auto_docstring
|
| 1113 |
+
def forward(
|
| 1114 |
+
self,
|
| 1115 |
+
input_ids: Optional[torch.LongTensor] = None,
|
| 1116 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1117 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 1118 |
+
past_key_values: Optional[Cache] = None,
|
| 1119 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1120 |
+
use_cache: Optional[bool] = None,
|
| 1121 |
+
output_attentions: Optional[bool] = None,
|
| 1122 |
+
output_hidden_states: Optional[bool] = None,
|
| 1123 |
+
return_dict: Optional[bool] = None,
|
| 1124 |
+
pixel_values: Optional[torch.Tensor] = None,
|
| 1125 |
+
pixel_values_videos: Optional[torch.FloatTensor] = None,
|
| 1126 |
+
image_grid_thw: Optional[torch.LongTensor] = None,
|
| 1127 |
+
patch_positions: Optional[torch.LongTensor] = None,
|
| 1128 |
+
video_grid_thw: Optional[torch.LongTensor] = None,
|
| 1129 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 1130 |
+
second_per_grid_ts: Optional[torch.Tensor] = None,
|
| 1131 |
+
**kwargs: Unpack[TransformersKwargs],
|
| 1132 |
+
) -> Union[tuple, LlavaOnevision2ModelOutputWithPast]:
|
| 1133 |
+
r"""
|
| 1134 |
+
image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
|
| 1135 |
+
The temporal, height and width of feature shape of each image in LLM.
|
| 1136 |
+
video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
|
| 1137 |
+
The temporal, height and width of feature shape of each video in LLM.
|
| 1138 |
+
patch_positions (`torch.LongTensor` of shape `(total_patches, 3)` or `(1, total_patches, 3)`, *optional*):
|
| 1139 |
+
Explicit per-patch `(t, h, w)` position indices used by the vision tower to compute 3D rotary
|
| 1140 |
+
position embeddings (and the optional absolute position embedding inside the patch merger).
|
| 1141 |
+
`total_patches` is the sum of `t * h * w` across all images and videos in the batch, matching
|
| 1142 |
+
the layout produced by the Qwen2VL-style image processor.
|
| 1143 |
+
second_per_grid_ts (`torch.Tensor` of shape `(num_videos)`, *optional*):
|
| 1144 |
+
The time interval (in seconds) for each grid along the temporal dimension in the 3D position IDs.
|
| 1145 |
+
cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
|
| 1146 |
+
Indices depicting the position of the input sequence tokens in the sequence. Contrarily to
|
| 1147 |
+
`position_ids`, this tensor is not affected by padding.
|
| 1148 |
+
|
| 1149 |
+
Note: see the top-level ``LlavaOnevision2ForConditionalGeneration.forward``
|
| 1150 |
+
docstring; currently video flows in via the ``image_grid_thw`` / ``pixel_values``
|
| 1151 |
+
alias, so ``pixel_values_videos`` / ``video_grid_thw`` /
|
| 1152 |
+
``second_per_grid_ts`` are unused at this layer.
|
| 1153 |
+
"""
|
| 1154 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 1155 |
+
output_hidden_states = (
|
| 1156 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 1157 |
+
)
|
| 1158 |
+
return_dict = True if return_dict is None else return_dict
|
| 1159 |
+
|
| 1160 |
+
if inputs_embeds is None:
|
| 1161 |
+
inputs_embeds = self.get_input_embeddings()(input_ids)
|
| 1162 |
+
|
| 1163 |
+
image_embeds = None
|
| 1164 |
+
|
| 1165 |
+
if pixel_values is not None:
|
| 1166 |
+
image_embeds = self.get_image_features(pixel_values, image_grid_thw, patch_positions=patch_positions)
|
| 1167 |
+
|
| 1168 |
+
if image_embeds is not None:
|
| 1169 |
+
image_embeds = torch.cat(image_embeds, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
|
| 1170 |
+
image_mask, _ = self.get_placeholder_mask(
|
| 1171 |
+
input_ids, inputs_embeds=inputs_embeds, image_features=image_embeds
|
| 1172 |
+
)
|
| 1173 |
+
inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
|
| 1174 |
+
|
| 1175 |
+
if pixel_values_videos is not None:
|
| 1176 |
+
video_embeds = self.get_video_features(
|
| 1177 |
+
pixel_values_videos, video_grid_thw, patch_positions=patch_positions
|
| 1178 |
+
)
|
| 1179 |
+
video_embeds = torch.cat(video_embeds, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
|
| 1180 |
+
_, video_mask = self.get_placeholder_mask(
|
| 1181 |
+
input_ids, inputs_embeds=inputs_embeds, video_features=video_embeds
|
| 1182 |
+
)
|
| 1183 |
+
inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)
|
| 1184 |
+
|
| 1185 |
+
# Use simple 1D position_ids
|
| 1186 |
+
if position_ids is None:
|
| 1187 |
+
batch_size, seq_length, _ = inputs_embeds.shape
|
| 1188 |
+
if attention_mask is not None:
|
| 1189 |
+
position_ids = attention_mask.long().cumsum(-1) - 1
|
| 1190 |
+
position_ids.masked_fill_(attention_mask == 0, 1)
|
| 1191 |
+
else:
|
| 1192 |
+
position_ids = (
|
| 1193 |
+
torch.arange(seq_length, device=inputs_embeds.device).unsqueeze(0).expand(batch_size, -1)
|
| 1194 |
+
)
|
| 1195 |
+
|
| 1196 |
+
# Handle cache_position for generation
|
| 1197 |
+
if cache_position is not None and cache_position[0] != 0:
|
| 1198 |
+
position_ids = position_ids + cache_position[0]
|
| 1199 |
+
|
| 1200 |
+
outputs = self.language_model(
|
| 1201 |
+
input_ids=None,
|
| 1202 |
+
position_ids=position_ids,
|
| 1203 |
+
attention_mask=attention_mask,
|
| 1204 |
+
past_key_values=past_key_values,
|
| 1205 |
+
inputs_embeds=inputs_embeds,
|
| 1206 |
+
use_cache=use_cache,
|
| 1207 |
+
output_attentions=output_attentions,
|
| 1208 |
+
output_hidden_states=output_hidden_states,
|
| 1209 |
+
return_dict=True,
|
| 1210 |
+
cache_position=cache_position,
|
| 1211 |
+
**kwargs,
|
| 1212 |
+
)
|
| 1213 |
+
|
| 1214 |
+
output = LlavaOnevision2ModelOutputWithPast(
|
| 1215 |
+
last_hidden_state=outputs.last_hidden_state,
|
| 1216 |
+
past_key_values=outputs.past_key_values,
|
| 1217 |
+
hidden_states=outputs.hidden_states,
|
| 1218 |
+
attentions=outputs.attentions,
|
| 1219 |
+
)
|
| 1220 |
+
return output if return_dict else output.to_tuple()
|
| 1221 |
+
|
| 1222 |
+
|
| 1223 |
+
@auto_docstring
|
| 1224 |
+
class LlavaOnevision2ForConditionalGeneration(LlavaOnevision2PreTrainedModel, GenerationMixin):
|
| 1225 |
+
_tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}
|
| 1226 |
+
# Reference: fix gemma3 grad acc #37208
|
| 1227 |
+
accepts_loss_kwargs = False
|
| 1228 |
+
|
| 1229 |
+
def __init__(self, config):
|
| 1230 |
+
super().__init__(config)
|
| 1231 |
+
self.model = LlavaOnevision2Model(config)
|
| 1232 |
+
self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
|
| 1233 |
+
self.post_init()
|
| 1234 |
+
|
| 1235 |
+
def get_input_embeddings(self):
|
| 1236 |
+
return self.model.get_input_embeddings()
|
| 1237 |
+
|
| 1238 |
+
def set_input_embeddings(self, value):
|
| 1239 |
+
self.model.set_input_embeddings(value)
|
| 1240 |
+
|
| 1241 |
+
def set_decoder(self, decoder):
|
| 1242 |
+
self.model.set_decoder(decoder)
|
| 1243 |
+
|
| 1244 |
+
def get_decoder(self):
|
| 1245 |
+
return self.model.get_decoder()
|
| 1246 |
+
|
| 1247 |
+
def get_video_features(
|
| 1248 |
+
self,
|
| 1249 |
+
pixel_values_videos: torch.FloatTensor,
|
| 1250 |
+
video_grid_thw: Optional[torch.LongTensor] = None,
|
| 1251 |
+
patch_positions=None,
|
| 1252 |
+
):
|
| 1253 |
+
return self.model.get_video_features(pixel_values_videos, video_grid_thw, patch_positions=patch_positions)
|
| 1254 |
+
|
| 1255 |
+
def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: Optional[torch.LongTensor] = None):
|
| 1256 |
+
return self.model.get_image_features(pixel_values, image_grid_thw)
|
| 1257 |
+
|
| 1258 |
+
# Make modules available through conditional class for BC
|
| 1259 |
+
@property
|
| 1260 |
+
def language_model(self):
|
| 1261 |
+
return self.model.language_model
|
| 1262 |
+
|
| 1263 |
+
@property
|
| 1264 |
+
def visual(self):
|
| 1265 |
+
return self.model.visual
|
| 1266 |
+
|
| 1267 |
+
@can_return_tuple
|
| 1268 |
+
@auto_docstring
|
| 1269 |
+
def forward(
|
| 1270 |
+
self,
|
| 1271 |
+
input_ids: Optional[torch.LongTensor] = None,
|
| 1272 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1273 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 1274 |
+
past_key_values: Optional[Cache] = None,
|
| 1275 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1276 |
+
labels: Optional[torch.LongTensor] = None,
|
| 1277 |
+
use_cache: Optional[bool] = None,
|
| 1278 |
+
output_attentions: Optional[bool] = None,
|
| 1279 |
+
output_hidden_states: Optional[bool] = None,
|
| 1280 |
+
pixel_values: Optional[torch.Tensor] = None,
|
| 1281 |
+
pixel_values_videos: Optional[torch.FloatTensor] = None,
|
| 1282 |
+
image_grid_thw: Optional[torch.LongTensor] = None,
|
| 1283 |
+
patch_positions: Optional[torch.LongTensor] = None,
|
| 1284 |
+
video_grid_thw: Optional[torch.LongTensor] = None,
|
| 1285 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 1286 |
+
second_per_grid_ts: Optional[torch.Tensor] = None,
|
| 1287 |
+
logits_to_keep: Union[int, torch.Tensor] = 0,
|
| 1288 |
+
**kwargs: Unpack[TransformersKwargs],
|
| 1289 |
+
) -> Union[tuple, LlavaOnevision2CausalLMOutputWithPast]:
|
| 1290 |
+
r"""
|
| 1291 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 1292 |
+
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
| 1293 |
+
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
| 1294 |
+
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
| 1295 |
+
image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
|
| 1296 |
+
The temporal, height and width of feature shape of each image in LLM.
|
| 1297 |
+
video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
|
| 1298 |
+
The temporal, height and width of feature shape of each video in LLM.
|
| 1299 |
+
patch_positions (`torch.LongTensor` of shape `(total_patches, 3)` or `(1, total_patches, 3)`, *optional*):
|
| 1300 |
+
Explicit per-patch `(t, h, w)` position indices used by the vision tower to compute 3D rotary
|
| 1301 |
+
position embeddings (and the optional absolute position embedding inside the patch merger).
|
| 1302 |
+
`total_patches` is the sum of `t * h * w` across all images and videos in the batch, matching
|
| 1303 |
+
the layout produced by the Qwen2VL-style image processor.
|
| 1304 |
+
second_per_grid_ts (`torch.Tensor` of shape `(num_videos)`, *optional*):
|
| 1305 |
+
The time interval (in seconds) for each grid along the temporal dimension in the 3D position IDs.
|
| 1306 |
+
cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
|
| 1307 |
+
Indices depicting the position of the input sequence tokens in the sequence. Contrarily to
|
| 1308 |
+
`position_ids`, this tensor is not affected by padding.
|
| 1309 |
+
|
| 1310 |
+
Note (native-video alias):
|
| 1311 |
+
The companion ``LlavaOnevision2Processor.__call__(videos=...)`` does NOT
|
| 1312 |
+
pass ``pixel_values_videos`` / ``video_grid_thw`` / ``second_per_grid_ts``
|
| 1313 |
+
to this forward. Instead it aliases the video patch tensor as
|
| 1314 |
+
``pixel_values=`` and ``image_grid_thw=``, so video inputs share the
|
| 1315 |
+
same code path as multi-image inputs (the OneVision encoder is purely
|
| 1316 |
+
spatial; temporal information is carried by per-frame ``<X.X seconds>``
|
| 1317 |
+
text tags emitted by the processor). The ``*_videos`` and
|
| 1318 |
+
``second_per_grid_ts`` kwargs are kept declared here only for API
|
| 1319 |
+
completeness and future use (e.g. 3D mRoPE / ``get_rope_index``); they
|
| 1320 |
+
are NOT consumed by the current OneVision encoder.
|
| 1321 |
+
|
| 1322 |
+
Example:
|
| 1323 |
+
|
| 1324 |
+
```python
|
| 1325 |
+
>>> from PIL import Image
|
| 1326 |
+
>>> import requests
|
| 1327 |
+
>>> from transformers import AutoProcessor, LlavaOnevision2ForConditionalGeneration
|
| 1328 |
+
|
| 1329 |
+
>>> model = LlavaOnevision2ForConditionalGeneration.from_pretrained("lmms-lab-encoder/LLaVA-OneVision2-8B-Instruct", trust_remote_code=True)
|
| 1330 |
+
>>> processor = AutoProcessor.from_pretrained("lmms-lab-encoder/LLaVA-OneVision2-8B-Instruct", trust_remote_code=True)
|
| 1331 |
+
|
| 1332 |
+
>>> messages = [
|
| 1333 |
+
{
|
| 1334 |
+
"role": "user",
|
| 1335 |
+
"content": [
|
| 1336 |
+
{"type": "image"},
|
| 1337 |
+
{"type": "text", "text": "What is shown in this image?"},
|
| 1338 |
+
],
|
| 1339 |
+
},
|
| 1340 |
+
]
|
| 1341 |
+
>>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
|
| 1342 |
+
>>> image = Image.open(requests.get(url, stream=True).raw)
|
| 1343 |
+
|
| 1344 |
+
>>> text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 1345 |
+
>>> inputs = processor(text=[text], images=[image], return_tensors="pt")
|
| 1346 |
+
|
| 1347 |
+
>>> # Generate
|
| 1348 |
+
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
| 1349 |
+
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
| 1350 |
+
"The image shows a street scene with a red stop sign in the foreground. In the background, there is a large red gate with Chinese characters ..."
|
| 1351 |
+
```"""
|
| 1352 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 1353 |
+
output_hidden_states = (
|
| 1354 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 1355 |
+
)
|
| 1356 |
+
outputs = self.model(
|
| 1357 |
+
input_ids=input_ids,
|
| 1358 |
+
pixel_values=pixel_values,
|
| 1359 |
+
pixel_values_videos=pixel_values_videos,
|
| 1360 |
+
image_grid_thw=image_grid_thw,
|
| 1361 |
+
patch_positions=patch_positions,
|
| 1362 |
+
video_grid_thw=video_grid_thw,
|
| 1363 |
+
second_per_grid_ts=second_per_grid_ts,
|
| 1364 |
+
position_ids=position_ids,
|
| 1365 |
+
attention_mask=attention_mask,
|
| 1366 |
+
past_key_values=past_key_values,
|
| 1367 |
+
inputs_embeds=inputs_embeds,
|
| 1368 |
+
use_cache=use_cache,
|
| 1369 |
+
output_attentions=output_attentions,
|
| 1370 |
+
output_hidden_states=output_hidden_states,
|
| 1371 |
+
return_dict=True,
|
| 1372 |
+
cache_position=cache_position,
|
| 1373 |
+
**kwargs,
|
| 1374 |
+
)
|
| 1375 |
+
|
| 1376 |
+
hidden_states = outputs[0]
|
| 1377 |
+
|
| 1378 |
+
# Only compute necessary logits, and do not upcast them to float if we are not computing the loss
|
| 1379 |
+
slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
|
| 1380 |
+
logits = self.lm_head(hidden_states[:, slice_indices, :])
|
| 1381 |
+
|
| 1382 |
+
loss = None
|
| 1383 |
+
if labels is not None:
|
| 1384 |
+
loss = self.loss_function(
|
| 1385 |
+
logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **kwargs
|
| 1386 |
+
)
|
| 1387 |
+
|
| 1388 |
+
return LlavaOnevision2CausalLMOutputWithPast(
|
| 1389 |
+
loss=loss,
|
| 1390 |
+
logits=logits,
|
| 1391 |
+
past_key_values=outputs.past_key_values,
|
| 1392 |
+
hidden_states=outputs.hidden_states,
|
| 1393 |
+
attentions=outputs.attentions,
|
| 1394 |
+
)
|
| 1395 |
+
|
| 1396 |
+
def prepare_inputs_for_generation(
|
| 1397 |
+
self,
|
| 1398 |
+
input_ids,
|
| 1399 |
+
past_key_values=None,
|
| 1400 |
+
attention_mask=None,
|
| 1401 |
+
inputs_embeds=None,
|
| 1402 |
+
cache_position=None,
|
| 1403 |
+
position_ids=None,
|
| 1404 |
+
use_cache=True,
|
| 1405 |
+
pixel_values=None,
|
| 1406 |
+
pixel_values_videos=None,
|
| 1407 |
+
image_grid_thw=None,
|
| 1408 |
+
patch_positions=None,
|
| 1409 |
+
video_grid_thw=None,
|
| 1410 |
+
second_per_grid_ts=None,
|
| 1411 |
+
is_first_iteration=False,
|
| 1412 |
+
**kwargs,
|
| 1413 |
+
):
|
| 1414 |
+
# Overwritten -- in specific circumstances we don't want to forward image inputs to the model
|
| 1415 |
+
model_inputs = super().prepare_inputs_for_generation(
|
| 1416 |
+
input_ids,
|
| 1417 |
+
past_key_values=past_key_values,
|
| 1418 |
+
attention_mask=attention_mask,
|
| 1419 |
+
inputs_embeds=inputs_embeds,
|
| 1420 |
+
cache_position=cache_position,
|
| 1421 |
+
position_ids=position_ids,
|
| 1422 |
+
pixel_values=pixel_values,
|
| 1423 |
+
pixel_values_videos=pixel_values_videos,
|
| 1424 |
+
image_grid_thw=image_grid_thw,
|
| 1425 |
+
video_grid_thw=video_grid_thw,
|
| 1426 |
+
second_per_grid_ts=second_per_grid_ts,
|
| 1427 |
+
patch_positions=patch_positions,
|
| 1428 |
+
use_cache=use_cache,
|
| 1429 |
+
is_first_iteration=is_first_iteration,
|
| 1430 |
+
**kwargs,
|
| 1431 |
+
)
|
| 1432 |
+
|
| 1433 |
+
# After the prefill iteration, drop image inputs so the vision tower
|
| 1434 |
+
# isn't re-run on decode steps. Gating on `is_first_iteration` (the
|
| 1435 |
+
# Qwen3-VL convention) is the only reliable signal in transformers
|
| 1436 |
+
# 5.x: `past_key_values` is non-None even on the first call (an empty
|
| 1437 |
+
# DynamicCache is created up-front by `generate`), and `cache_position`
|
| 1438 |
+
# may be `None` for remote-code models.
|
| 1439 |
+
if not is_first_iteration and use_cache:
|
| 1440 |
+
model_inputs["pixel_values"] = None
|
| 1441 |
+
model_inputs["pixel_values_videos"] = None
|
| 1442 |
+
|
| 1443 |
+
return model_inputs
|
| 1444 |
+
|
| 1445 |
+
def _get_image_nums_and_video_nums(
|
| 1446 |
+
self,
|
| 1447 |
+
input_ids: Optional[torch.LongTensor],
|
| 1448 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
| 1449 |
+
) -> tuple[torch.Tensor, torch.Tensor]:
|
| 1450 |
+
"""
|
| 1451 |
+
Get the number of images and videos for each sample to calculate the separation length of the sample tensor.
|
| 1452 |
+
These parameters are not passed through the processor to avoid unpredictable impacts from interface modifications.
|
| 1453 |
+
|
| 1454 |
+
Args:
|
| 1455 |
+
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
| 1456 |
+
Indices of input sequence tokens in the vocabulary.
|
| 1457 |
+
|
| 1458 |
+
Returns:
|
| 1459 |
+
image_nums (`torch.LongTensor` of shape `(batch_size, num_images_sample)`)
|
| 1460 |
+
video_nums (`torch.LongTensor` of shape `(batch_size, num_videos_sample)`)
|
| 1461 |
+
"""
|
| 1462 |
+
image_token_id = self.config.image_token_id
|
| 1463 |
+
video_token_id = self.config.video_token_id
|
| 1464 |
+
vision_start_token_id = self.config.vision_start_token_id
|
| 1465 |
+
|
| 1466 |
+
if inputs_embeds is not None:
|
| 1467 |
+
vision_start_mask = (
|
| 1468 |
+
inputs_embeds
|
| 1469 |
+
== self.get_input_embeddings()(
|
| 1470 |
+
torch.tensor(vision_start_token_id, dtype=torch.long, device=inputs_embeds.device)
|
| 1471 |
+
)
|
| 1472 |
+
)[..., 0]
|
| 1473 |
+
image_mask = (
|
| 1474 |
+
inputs_embeds
|
| 1475 |
+
== self.get_input_embeddings()(
|
| 1476 |
+
torch.tensor(image_token_id, dtype=torch.long, device=inputs_embeds.device)
|
| 1477 |
+
)
|
| 1478 |
+
)[..., 0]
|
| 1479 |
+
video_mask = (
|
| 1480 |
+
inputs_embeds
|
| 1481 |
+
== self.get_input_embeddings()(
|
| 1482 |
+
torch.tensor(video_token_id, dtype=torch.long, device=inputs_embeds.device)
|
| 1483 |
+
)
|
| 1484 |
+
)[..., 0]
|
| 1485 |
+
else:
|
| 1486 |
+
vision_start_mask = input_ids == vision_start_token_id
|
| 1487 |
+
image_mask = input_ids == image_token_id
|
| 1488 |
+
video_mask = input_ids == video_token_id
|
| 1489 |
+
|
| 1490 |
+
vision_first_mask = torch.roll(vision_start_mask, shifts=1, dims=1)
|
| 1491 |
+
image_nums = torch.sum(vision_first_mask & image_mask, dim=1)
|
| 1492 |
+
video_nums = torch.sum(vision_first_mask & video_mask, dim=1)
|
| 1493 |
+
|
| 1494 |
+
return image_nums, video_nums
|
| 1495 |
+
|
| 1496 |
+
def _expand_inputs_for_generation(
|
| 1497 |
+
self,
|
| 1498 |
+
expand_size: int = 1,
|
| 1499 |
+
is_encoder_decoder: bool = False,
|
| 1500 |
+
input_ids: Optional[torch.LongTensor] = None,
|
| 1501 |
+
**model_kwargs,
|
| 1502 |
+
) -> tuple[torch.LongTensor, dict[str, Any]]:
|
| 1503 |
+
# Overwritten -- Support for expanding tensors without a batch size dimension
|
| 1504 |
+
# e.g., pixel_values, image_grid_thw, pixel_values_videos, video_grid_thw, second_per_grid_t
|
| 1505 |
+
# pixel_values.shape[0] is sum(seqlen_images for samples)
|
| 1506 |
+
# image_grid_thw.shape[0] is sum(num_images for samples)
|
| 1507 |
+
|
| 1508 |
+
if expand_size == 1:
|
| 1509 |
+
return input_ids, model_kwargs
|
| 1510 |
+
|
| 1511 |
+
visual_keys = [
|
| 1512 |
+
"pixel_values",
|
| 1513 |
+
"image_grid_thw",
|
| 1514 |
+
"pixel_values_videos",
|
| 1515 |
+
"video_grid_thw",
|
| 1516 |
+
"second_per_grid_ts",
|
| 1517 |
+
"patch_positions",
|
| 1518 |
+
]
|
| 1519 |
+
|
| 1520 |
+
def _expand_dict_for_generation_visual(dict_to_expand):
|
| 1521 |
+
image_grid_thw = model_kwargs.get("image_grid_thw", None)
|
| 1522 |
+
video_grid_thw = model_kwargs.get("video_grid_thw", None)
|
| 1523 |
+
image_nums, video_nums = self._get_image_nums_and_video_nums(
|
| 1524 |
+
input_ids, inputs_embeds=model_kwargs.get("inputs_embeds", None)
|
| 1525 |
+
)
|
| 1526 |
+
|
| 1527 |
+
def _repeat_interleave_samples(x, lengths, repeat_times):
|
| 1528 |
+
samples = torch.split(x, lengths)
|
| 1529 |
+
repeat_args = [repeat_times] + [1] * (x.dim() - 1)
|
| 1530 |
+
result = torch.cat([sample.repeat(*repeat_args) for sample in samples], dim=0)
|
| 1531 |
+
return result
|
| 1532 |
+
|
| 1533 |
+
for key in dict_to_expand:
|
| 1534 |
+
if key == "pixel_values":
|
| 1535 |
+
# split images into samples
|
| 1536 |
+
samples = torch.split(image_grid_thw, list(image_nums))
|
| 1537 |
+
# compute the sequence length of images for each sample
|
| 1538 |
+
lengths = [torch.prod(sample, dim=1).sum() for sample in samples]
|
| 1539 |
+
dict_to_expand[key] = _repeat_interleave_samples(
|
| 1540 |
+
dict_to_expand[key], lengths=lengths, repeat_times=expand_size
|
| 1541 |
+
)
|
| 1542 |
+
elif key == "image_grid_thw":
|
| 1543 |
+
# get the num of images for each sample
|
| 1544 |
+
lengths = list(image_nums)
|
| 1545 |
+
dict_to_expand[key] = _repeat_interleave_samples(
|
| 1546 |
+
dict_to_expand[key], lengths=lengths, repeat_times=expand_size
|
| 1547 |
+
)
|
| 1548 |
+
elif key == "pixel_values_videos":
|
| 1549 |
+
samples = torch.split(video_grid_thw, list(video_nums))
|
| 1550 |
+
lengths = [torch.prod(sample, dim=1).sum() for sample in samples]
|
| 1551 |
+
dict_to_expand[key] = _repeat_interleave_samples(
|
| 1552 |
+
dict_to_expand[key], lengths=lengths, repeat_times=expand_size
|
| 1553 |
+
)
|
| 1554 |
+
elif key == "video_grid_thw":
|
| 1555 |
+
lengths = list(video_nums)
|
| 1556 |
+
dict_to_expand[key] = _repeat_interleave_samples(
|
| 1557 |
+
dict_to_expand[key], lengths=lengths, repeat_times=expand_size
|
| 1558 |
+
)
|
| 1559 |
+
elif key == "second_per_grid_ts":
|
| 1560 |
+
dict_to_expand[key] = _repeat_interleave_samples(
|
| 1561 |
+
dict_to_expand[key], lengths=list(video_nums), repeat_times=expand_size
|
| 1562 |
+
)
|
| 1563 |
+
elif key == "patch_positions":
|
| 1564 |
+
if image_grid_thw is not None and image_grid_thw.numel() > 0 and image_nums.sum() > 0:
|
| 1565 |
+
samples = torch.split(image_grid_thw, list(image_nums))
|
| 1566 |
+
lengths = [torch.prod(sample, dim=1).sum() for sample in samples]
|
| 1567 |
+
elif video_grid_thw is not None and video_grid_thw.numel() > 0 and video_nums.sum() > 0:
|
| 1568 |
+
samples = torch.split(video_grid_thw, list(video_nums))
|
| 1569 |
+
lengths = [torch.prod(sample, dim=1).sum() for sample in samples]
|
| 1570 |
+
else:
|
| 1571 |
+
continue
|
| 1572 |
+
dict_to_expand[key] = _repeat_interleave_samples(
|
| 1573 |
+
dict_to_expand[key], lengths=lengths, repeat_times=expand_size
|
| 1574 |
+
)
|
| 1575 |
+
return dict_to_expand
|
| 1576 |
+
|
| 1577 |
+
def _expand_dict_for_generation(dict_to_expand):
|
| 1578 |
+
for key in dict_to_expand:
|
| 1579 |
+
if (
|
| 1580 |
+
key != "cache_position"
|
| 1581 |
+
and dict_to_expand[key] is not None
|
| 1582 |
+
and isinstance(dict_to_expand[key], torch.Tensor)
|
| 1583 |
+
and key not in visual_keys
|
| 1584 |
+
):
|
| 1585 |
+
dict_to_expand[key] = dict_to_expand[key].repeat_interleave(expand_size, dim=0)
|
| 1586 |
+
return dict_to_expand
|
| 1587 |
+
|
| 1588 |
+
model_kwargs = _expand_dict_for_generation_visual(model_kwargs)
|
| 1589 |
+
|
| 1590 |
+
if input_ids is not None:
|
| 1591 |
+
input_ids = input_ids.repeat_interleave(expand_size, dim=0)
|
| 1592 |
+
|
| 1593 |
+
model_kwargs = _expand_dict_for_generation(model_kwargs)
|
| 1594 |
+
|
| 1595 |
+
if is_encoder_decoder:
|
| 1596 |
+
if model_kwargs.get("encoder_outputs") is None:
|
| 1597 |
+
raise ValueError("If `is_encoder_decoder` is True, make sure that `encoder_outputs` is defined.")
|
| 1598 |
+
model_kwargs["encoder_outputs"] = _expand_dict_for_generation(model_kwargs["encoder_outputs"])
|
| 1599 |
+
|
| 1600 |
+
return input_ids, model_kwargs
|
| 1601 |
+
|
| 1602 |
+
|
| 1603 |
+
__all__ = [
|
| 1604 |
+
"LlavaOnevision2ForConditionalGeneration",
|
| 1605 |
+
"LlavaOnevision2Model",
|
| 1606 |
+
"LlavaOnevision2PreTrainedModel",
|
| 1607 |
+
]
|
preprocessor_config.json
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"data_format": "channels_first",
|
| 3 |
+
"default_to_square": true,
|
| 4 |
+
"do_convert_rgb": true,
|
| 5 |
+
"do_normalize": true,
|
| 6 |
+
"do_rescale": true,
|
| 7 |
+
"do_resize": true,
|
| 8 |
+
"image_mean": [
|
| 9 |
+
0.48145466,
|
| 10 |
+
0.4578275,
|
| 11 |
+
0.40821073
|
| 12 |
+
],
|
| 13 |
+
"image_processor_type": "Qwen2VLImageProcessor",
|
| 14 |
+
"image_std": [
|
| 15 |
+
0.26862954,
|
| 16 |
+
0.26130258,
|
| 17 |
+
0.27577711
|
| 18 |
+
],
|
| 19 |
+
"max_pixels": 4000000,
|
| 20 |
+
"merge_size": 2,
|
| 21 |
+
"min_pixels": 3136,
|
| 22 |
+
"patch_size": 14,
|
| 23 |
+
"processor_class": "LlavaOnevision2Processor",
|
| 24 |
+
"auto_map": {
|
| 25 |
+
"AutoProcessor": "processing_llava_onevision2.LlavaOnevision2Processor",
|
| 26 |
+
"AutoVideoProcessor": "video_processing_llava_onevision2.LlavaOnevision2VideoProcessor"
|
| 27 |
+
},
|
| 28 |
+
"resample": 3,
|
| 29 |
+
"rescale_factor": 0.00392156862745098,
|
| 30 |
+
"size": {
|
| 31 |
+
"longest_edge": 4000000,
|
| 32 |
+
"shortest_edge": 3136
|
| 33 |
+
},
|
| 34 |
+
"temporal_patch_size": 1
|
| 35 |
+
}
|
processing_llava_onevision2.py
ADDED
|
@@ -0,0 +1,406 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""LlavaOnevision2 multi-modal processor.
|
| 2 |
+
|
| 3 |
+
Combines:
|
| 4 |
+
- ``Qwen2VLImageProcessor[Fast]`` (existing in checkpoint preprocessor_config)
|
| 5 |
+
- ``LlavaOnevision2VideoProcessor`` (this checkpoint, video_processing_*)
|
| 6 |
+
- ``AutoTokenizer`` (existing tokenizer.json)
|
| 7 |
+
- ``chat_template.jinja`` (existing, emits <|video_pad|>)
|
| 8 |
+
|
| 9 |
+
Public API:
|
| 10 |
+
proc = LlavaOnevision2Processor(image_processor, tokenizer, video_processor)
|
| 11 |
+
text = proc.apply_chat_template(messages, add_generation_prompt=True)
|
| 12 |
+
inputs = proc(text=[text], videos=[mp4_or_frames], return_tensors="pt")
|
| 13 |
+
out = model.generate(**inputs)
|
| 14 |
+
|
| 15 |
+
Design choices:
|
| 16 |
+
- Video path is "in-processor, transformed to multi-image + per-frame
|
| 17 |
+
timestamps" — model.forward sees the image path only.
|
| 18 |
+
- The chat_template's <|vision_start|><|video_pad|><|vision_end|> placeholder
|
| 19 |
+
is rewritten in __call__ to per-frame blocks:
|
| 20 |
+
<X.X seconds><|vision_start|><|image_pad|>*n<|vision_end|>\n
|
| 21 |
+
- We DO NOT emit `second_per_grid_ts`; see plan §0.5.
|
| 22 |
+
- Backward-compatible: `images=...` / pure-text usage matches the existing
|
| 23 |
+
Qwen2_5_VLProcessor output.
|
| 24 |
+
"""
|
| 25 |
+
|
| 26 |
+
from __future__ import annotations
|
| 27 |
+
|
| 28 |
+
import re
|
| 29 |
+
from typing import List, Optional, Sequence, Union
|
| 30 |
+
|
| 31 |
+
import torch
|
| 32 |
+
|
| 33 |
+
# Special-token strings used by the checkpoint's tokenizer / chat_template.
|
| 34 |
+
VISION_START = "<|vision_start|>"
|
| 35 |
+
VISION_END = "<|vision_end|>"
|
| 36 |
+
IMAGE_PAD = "<|image_pad|>"
|
| 37 |
+
VIDEO_PAD = "<|video_pad|>"
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
def _format_seconds_tag(seconds: float) -> str:
|
| 41 |
+
"""Match training format: ``<X.X seconds>`` (one decimal place)."""
|
| 42 |
+
return f"<{float(seconds):.1f} seconds>"
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
def _expand_video_block_for_frames(
|
| 46 |
+
n_per_frame: int,
|
| 47 |
+
frame_seconds: Sequence[float],
|
| 48 |
+
) -> str:
|
| 49 |
+
"""Build the per-frame expanded text that replaces a single
|
| 50 |
+
``<|vision_start|><|video_pad|><|vision_end|>`` block.
|
| 51 |
+
|
| 52 |
+
Output (one block per frame, newline-separated):
|
| 53 |
+
``<X.X seconds><|vision_start|><|image_pad|>*n_per_frame<|vision_end|>\\n``
|
| 54 |
+
"""
|
| 55 |
+
parts: List[str] = []
|
| 56 |
+
for sec in frame_seconds:
|
| 57 |
+
parts.append(_format_seconds_tag(sec))
|
| 58 |
+
parts.append(VISION_START)
|
| 59 |
+
parts.append(IMAGE_PAD * n_per_frame)
|
| 60 |
+
parts.append(VISION_END)
|
| 61 |
+
return "".join(parts)
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
class LlavaOnevision2Processor:
|
| 65 |
+
"""Native multi-modal processor for LlavaOnevision2.
|
| 66 |
+
|
| 67 |
+
NOTE: We deliberately do NOT inherit ``transformers.ProcessorMixin``.
|
| 68 |
+
This class is registered via ``auto_map`` so
|
| 69 |
+
``AutoProcessor.from_pretrained(..., trust_remote_code=True)`` returns it.
|
| 70 |
+
"""
|
| 71 |
+
|
| 72 |
+
attributes = ["image_processor", "video_processor", "tokenizer"]
|
| 73 |
+
image_processor_class = "AutoImageProcessor"
|
| 74 |
+
tokenizer_class = "AutoTokenizer"
|
| 75 |
+
|
| 76 |
+
def __init__(
|
| 77 |
+
self,
|
| 78 |
+
image_processor=None,
|
| 79 |
+
tokenizer=None,
|
| 80 |
+
video_processor=None,
|
| 81 |
+
chat_template: Optional[str] = None,
|
| 82 |
+
):
|
| 83 |
+
self.image_processor = image_processor
|
| 84 |
+
self.tokenizer = tokenizer
|
| 85 |
+
self.video_processor = video_processor
|
| 86 |
+
|
| 87 |
+
# Inherit chat_template from the tokenizer if not given (matches Qwen2_5_VLProcessor).
|
| 88 |
+
if chat_template is None and tokenizer is not None:
|
| 89 |
+
chat_template = getattr(tokenizer, "chat_template", None)
|
| 90 |
+
self.chat_template = chat_template
|
| 91 |
+
|
| 92 |
+
# Cache the merge size from image_processor for token-count math.
|
| 93 |
+
self.spatial_merge_size = int(
|
| 94 |
+
getattr(image_processor, "merge_size", 2) if image_processor is not None else 2
|
| 95 |
+
)
|
| 96 |
+
|
| 97 |
+
# ------------------------------------------------------------------ utils
|
| 98 |
+
|
| 99 |
+
@classmethod
|
| 100 |
+
def register_for_auto_class(cls, auto_class="AutoProcessor"):
|
| 101 |
+
"""No-op stub so ``AutoProcessor.from_pretrained(..., trust_remote_code=True)``
|
| 102 |
+
can call this on the dynamically-loaded class without erroring.
|
| 103 |
+
Real ``ProcessorMixin`` uses this to remember the auto-class for
|
| 104 |
+
``push_to_hub``; we don't need that for inference-only use."""
|
| 105 |
+
cls._auto_class = auto_class
|
| 106 |
+
|
| 107 |
+
@classmethod
|
| 108 |
+
def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
|
| 109 |
+
"""Convenience builder mirroring HF's ``from_pretrained`` pattern."""
|
| 110 |
+
from transformers import AutoTokenizer, Qwen2VLImageProcessor
|
| 111 |
+
|
| 112 |
+
# Drop kwargs that AutoProcessor injects but downstream constructors
|
| 113 |
+
# don't accept (e.g. _from_auto / trust_remote_code propagation).
|
| 114 |
+
kwargs.pop("_from_auto", None)
|
| 115 |
+
kwargs.pop("trust_remote_code", None)
|
| 116 |
+
kwargs.pop("code_revision", None)
|
| 117 |
+
|
| 118 |
+
# Use the SLOW Qwen2VLImageProcessor: the Fast variant has small
|
| 119 |
+
# normalization rounding differences that change pixel_values bit-for-bit.
|
| 120 |
+
image_processor = Qwen2VLImageProcessor.from_pretrained(
|
| 121 |
+
pretrained_model_name_or_path, **kwargs
|
| 122 |
+
)
|
| 123 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
| 124 |
+
pretrained_model_name_or_path, **kwargs
|
| 125 |
+
)
|
| 126 |
+
|
| 127 |
+
# Use the bundled VideoProcessor. Try a relative import first (when
|
| 128 |
+
# this module is loaded as part of a remote_code package), and fall
|
| 129 |
+
# back to a top-level import (when loaded as a standalone file via
|
| 130 |
+
# ``get_class_from_dynamic_module``, which places sibling files on
|
| 131 |
+
# ``sys.path``).
|
| 132 |
+
try:
|
| 133 |
+
from .video_processing_llava_onevision2 import LlavaOnevision2VideoProcessor
|
| 134 |
+
except ImportError:
|
| 135 |
+
from video_processing_llava_onevision2 import LlavaOnevision2VideoProcessor
|
| 136 |
+
|
| 137 |
+
video_processor = LlavaOnevision2VideoProcessor(
|
| 138 |
+
image_processor=image_processor,
|
| 139 |
+
min_pixels=getattr(image_processor, "min_pixels", 256 * 28 * 28),
|
| 140 |
+
max_pixels=getattr(image_processor, "max_pixels", 1605632),
|
| 141 |
+
patch_size=getattr(image_processor, "patch_size", 14),
|
| 142 |
+
spatial_merge_size=getattr(image_processor, "merge_size", 2),
|
| 143 |
+
)
|
| 144 |
+
return cls(
|
| 145 |
+
image_processor=image_processor,
|
| 146 |
+
tokenizer=tokenizer,
|
| 147 |
+
video_processor=video_processor,
|
| 148 |
+
)
|
| 149 |
+
|
| 150 |
+
# ------------------------------------------------------------- chat helpers
|
| 151 |
+
|
| 152 |
+
def apply_chat_template(self, messages, **kwargs):
|
| 153 |
+
"""Delegate to the tokenizer (which already has ``chat_template``)."""
|
| 154 |
+
if self.chat_template and "chat_template" not in kwargs:
|
| 155 |
+
kwargs["chat_template"] = self.chat_template
|
| 156 |
+
return self.tokenizer.apply_chat_template(messages, **kwargs)
|
| 157 |
+
|
| 158 |
+
# ----------------------------------------------------------- main __call__
|
| 159 |
+
|
| 160 |
+
def __call__(
|
| 161 |
+
self,
|
| 162 |
+
text: Optional[Union[str, List[str]]] = None,
|
| 163 |
+
images=None,
|
| 164 |
+
videos=None,
|
| 165 |
+
return_tensors: Optional[str] = "pt",
|
| 166 |
+
padding: Union[bool, str] = False,
|
| 167 |
+
num_frames: Optional[int] = None,
|
| 168 |
+
max_frames: Optional[int] = None,
|
| 169 |
+
target_fps: Optional[float] = None,
|
| 170 |
+
**kwargs,
|
| 171 |
+
):
|
| 172 |
+
"""Process an aligned (text, images, videos) batch.
|
| 173 |
+
|
| 174 |
+
Behaviour:
|
| 175 |
+
* ``videos is not None``: run the VideoProcessor, rewrite each
|
| 176 |
+
``<|video_pad|>`` block in ``text`` to per-frame ``<X.X seconds>``
|
| 177 |
+
blocks, then alias the video patches as ``pixel_values`` /
|
| 178 |
+
``image_grid_thw`` so the model's image path consumes them.
|
| 179 |
+
* ``images is not None``: passed through to the underlying
|
| 180 |
+
``image_processor``. (May coexist with ``videos``; expansion order
|
| 181 |
+
in the prompt is determined by the chat_template / placeholders.)
|
| 182 |
+
* Pure text: tokenize and return.
|
| 183 |
+
|
| 184 |
+
Per-call frame-sampling overrides (apply only to ``videos`` path; do
|
| 185 |
+
not mutate the underlying VideoProcessor's defaults):
|
| 186 |
+
* ``num_frames`` : force exactly N frames per video
|
| 187 |
+
(alias of ``fixed_num_frames``).
|
| 188 |
+
* ``max_frames`` : cap on auto-selected frame count (long videos).
|
| 189 |
+
* ``target_fps`` : sample at this FPS (capped by ``max_frames``).
|
| 190 |
+
|
| 191 |
+
Returns a ``BatchFeature`` with at minimum ``input_ids`` and
|
| 192 |
+
``attention_mask``; plus ``pixel_values`` / ``image_grid_thw`` /
|
| 193 |
+
``patch_positions`` when visuals are present.
|
| 194 |
+
"""
|
| 195 |
+
if text is None:
|
| 196 |
+
raise ValueError("`text` is required.")
|
| 197 |
+
if isinstance(text, str):
|
| 198 |
+
text = [text]
|
| 199 |
+
text = list(text)
|
| 200 |
+
|
| 201 |
+
out: dict = {}
|
| 202 |
+
|
| 203 |
+
# ---------------- VIDEO PATH ----------------
|
| 204 |
+
# Process videos first so we can rewrite their placeholders into the
|
| 205 |
+
# text before tokenization.
|
| 206 |
+
video_outputs = None
|
| 207 |
+
if videos is not None:
|
| 208 |
+
if self.video_processor is None:
|
| 209 |
+
raise ValueError("videos passed but no video_processor configured.")
|
| 210 |
+
# Normalise to a list of videos.
|
| 211 |
+
if isinstance(videos, (str,)):
|
| 212 |
+
videos_list = [videos]
|
| 213 |
+
elif isinstance(videos, list) and len(videos) > 0 and not isinstance(
|
| 214 |
+
videos[0], (list, str)
|
| 215 |
+
):
|
| 216 |
+
# list[PIL]/[np.ndarray] = single video
|
| 217 |
+
videos_list = [videos]
|
| 218 |
+
else:
|
| 219 |
+
videos_list = list(videos)
|
| 220 |
+
|
| 221 |
+
# Per-call sampling overrides: temporarily swap the
|
| 222 |
+
# VideoProcessor's attributes, then restore. Lets users do
|
| 223 |
+
# processor(videos=[mp4], num_frames=8)
|
| 224 |
+
# without mutating processor.video_processor.
|
| 225 |
+
vp = self.video_processor
|
| 226 |
+
saved = (vp.fixed_num_frames, vp.max_frames, vp.target_fps)
|
| 227 |
+
try:
|
| 228 |
+
if num_frames is not None:
|
| 229 |
+
vp.fixed_num_frames = int(num_frames)
|
| 230 |
+
if max_frames is not None:
|
| 231 |
+
vp.max_frames = int(max_frames)
|
| 232 |
+
if target_fps is not None:
|
| 233 |
+
vp.target_fps = float(target_fps)
|
| 234 |
+
video_outputs = vp(videos=videos_list, return_tensors="pt")
|
| 235 |
+
finally:
|
| 236 |
+
vp.fixed_num_frames, vp.max_frames, vp.target_fps = saved
|
| 237 |
+
|
| 238 |
+
# Rewrite each <|video_pad|> in `text` into per-frame blocks.
|
| 239 |
+
video_grid_thw = video_outputs["video_grid_thw"] # [num_videos, 3]
|
| 240 |
+
frame_timestamps = video_outputs["frame_timestamps"]
|
| 241 |
+
sms = self.spatial_merge_size
|
| 242 |
+
|
| 243 |
+
# We iterate placeholders globally across all texts (matching how
|
| 244 |
+
# Qwen2_5_VLProcessor sources `image_grid_thw` rows).
|
| 245 |
+
video_idx = 0
|
| 246 |
+
|
| 247 |
+
def _rewrite_one_text(s: str) -> str:
|
| 248 |
+
nonlocal video_idx
|
| 249 |
+
pattern = re.compile(
|
| 250 |
+
re.escape(VISION_START) + r"\s*" + re.escape(VIDEO_PAD) + r"\s*" + re.escape(VISION_END)
|
| 251 |
+
)
|
| 252 |
+
|
| 253 |
+
def _sub(_match):
|
| 254 |
+
nonlocal video_idx
|
| 255 |
+
if video_idx >= video_grid_thw.shape[0]:
|
| 256 |
+
raise ValueError(
|
| 257 |
+
"More <|video_pad|> placeholders in text than videos provided."
|
| 258 |
+
)
|
| 259 |
+
T_eff = int(video_grid_thw[video_idx, 0].item())
|
| 260 |
+
H_p = int(video_grid_thw[video_idx, 1].item())
|
| 261 |
+
W_p = int(video_grid_thw[video_idx, 2].item())
|
| 262 |
+
n_per_frame = (H_p * W_p) // (sms * sms)
|
| 263 |
+
frame_seconds = frame_timestamps[video_idx]
|
| 264 |
+
if len(frame_seconds) != T_eff:
|
| 265 |
+
# Defensive: pad/truncate so the count matches the grid.
|
| 266 |
+
if len(frame_seconds) < T_eff:
|
| 267 |
+
frame_seconds = list(frame_seconds) + [
|
| 268 |
+
frame_seconds[-1] if frame_seconds else 0.0
|
| 269 |
+
] * (T_eff - len(frame_seconds))
|
| 270 |
+
else:
|
| 271 |
+
frame_seconds = list(frame_seconds[:T_eff])
|
| 272 |
+
expanded = _expand_video_block_for_frames(
|
| 273 |
+
n_per_frame, frame_seconds
|
| 274 |
+
)
|
| 275 |
+
video_idx += 1
|
| 276 |
+
# Strip trailing newline so we don't double-newline existing prompts.
|
| 277 |
+
return expanded.rstrip("\n")
|
| 278 |
+
|
| 279 |
+
return pattern.sub(_sub, s)
|
| 280 |
+
|
| 281 |
+
text = [_rewrite_one_text(s) for s in text]
|
| 282 |
+
|
| 283 |
+
if video_idx != video_grid_thw.shape[0]:
|
| 284 |
+
raise ValueError(
|
| 285 |
+
f"Provided {video_grid_thw.shape[0]} videos but only "
|
| 286 |
+
f"{video_idx} <|video_pad|> placeholders were found in text."
|
| 287 |
+
)
|
| 288 |
+
|
| 289 |
+
# Alias video tensors into the image path (NEW model only consumes the image path).
|
| 290 |
+
# Option 1 (multi-image semantics, training-aligned): expand each
|
| 291 |
+
# video_grid_thw row [T, H, W] into T rows of [1, H, W]. The
|
| 292 |
+
# pixel_values rows are already laid out frame-by-frame (T*H*W per
|
| 293 |
+
# video, with temporal_patch_size=1), so this row-expansion of
|
| 294 |
+
# image_grid_thw is the only adjustment needed for the model's
|
| 295 |
+
# forward to treat each frame as a separate image (matching the
|
| 296 |
+
# multi-image inference path).
|
| 297 |
+
out["pixel_values"] = video_outputs["pixel_values_videos"]
|
| 298 |
+
vgthw = video_outputs["video_grid_thw"]
|
| 299 |
+
expanded_rows = []
|
| 300 |
+
for row in vgthw:
|
| 301 |
+
T_v, H_v, W_v = int(row[0]), int(row[1]), int(row[2])
|
| 302 |
+
expanded_rows.extend([[1, H_v, W_v]] * T_v)
|
| 303 |
+
out["image_grid_thw"] = torch.tensor(expanded_rows, dtype=vgthw.dtype)
|
| 304 |
+
out["patch_positions"] = video_outputs["patch_positions"]
|
| 305 |
+
|
| 306 |
+
# ---------------- IMAGE PATH ----------------
|
| 307 |
+
if images is not None:
|
| 308 |
+
if self.image_processor is None:
|
| 309 |
+
raise ValueError("images passed but no image_processor configured.")
|
| 310 |
+
image_outputs = self.image_processor(
|
| 311 |
+
images=images, return_tensors="pt"
|
| 312 |
+
)
|
| 313 |
+
image_grid_thw = image_outputs["image_grid_thw"]
|
| 314 |
+
|
| 315 |
+
# Expand each <|image_pad|> placeholder to the number of merged tokens.
|
| 316 |
+
sms = self.spatial_merge_size
|
| 317 |
+
merge_factor = sms * sms
|
| 318 |
+
image_token_counts = (
|
| 319 |
+
(image_grid_thw[:, 0] * image_grid_thw[:, 1] * image_grid_thw[:, 2])
|
| 320 |
+
// merge_factor
|
| 321 |
+
).tolist()
|
| 322 |
+
img_idx = 0
|
| 323 |
+
|
| 324 |
+
def _expand_image_pads(s: str) -> str:
|
| 325 |
+
nonlocal img_idx
|
| 326 |
+
while IMAGE_PAD in s:
|
| 327 |
+
if img_idx >= len(image_token_counts):
|
| 328 |
+
break
|
| 329 |
+
n = int(image_token_counts[img_idx])
|
| 330 |
+
s = s.replace(IMAGE_PAD, "<|placeholder|>" * n, 1)
|
| 331 |
+
img_idx += 1
|
| 332 |
+
return s.replace("<|placeholder|>", IMAGE_PAD)
|
| 333 |
+
|
| 334 |
+
text = [_expand_image_pads(s) for s in text]
|
| 335 |
+
|
| 336 |
+
# If videos and images coexist, prefer concatenation of patch tensors.
|
| 337 |
+
if "pixel_values" in out:
|
| 338 |
+
out["pixel_values"] = torch.cat(
|
| 339 |
+
[out["pixel_values"], image_outputs["pixel_values"]], dim=0
|
| 340 |
+
)
|
| 341 |
+
out["image_grid_thw"] = torch.cat(
|
| 342 |
+
[out["image_grid_thw"], image_outputs["image_grid_thw"]], dim=0
|
| 343 |
+
)
|
| 344 |
+
# Build image patch_positions and concat.
|
| 345 |
+
from .video_processing_llava_onevision2 import build_patch_positions
|
| 346 |
+
image_pp = build_patch_positions(
|
| 347 |
+
image_outputs["image_grid_thw"], spatial_merge_size=sms
|
| 348 |
+
)
|
| 349 |
+
out["patch_positions"] = torch.cat(
|
| 350 |
+
[out["patch_positions"], image_pp], dim=0
|
| 351 |
+
)
|
| 352 |
+
else:
|
| 353 |
+
out["pixel_values"] = image_outputs["pixel_values"]
|
| 354 |
+
out["image_grid_thw"] = image_outputs["image_grid_thw"]
|
| 355 |
+
from .video_processing_llava_onevision2 import build_patch_positions
|
| 356 |
+
out["patch_positions"] = build_patch_positions(
|
| 357 |
+
image_outputs["image_grid_thw"], spatial_merge_size=sms
|
| 358 |
+
)
|
| 359 |
+
|
| 360 |
+
# ---------------- VIDEO PATH FINAL EXPANSION ----------------
|
| 361 |
+
# When `videos` was given (and possibly without `images`), the per-frame
|
| 362 |
+
# rewrite above already produced runs of <|image_pad|> that need to be
|
| 363 |
+
# treated like image placeholders (one per merged token). Because the
|
| 364 |
+
# rewrite directly emits ``IMAGE_PAD * n_per_frame``, the texts are
|
| 365 |
+
# already in their tokenize-ready form for the video portion. So nothing
|
| 366 |
+
# more to do here — fall through to tokenize.
|
| 367 |
+
|
| 368 |
+
# ---------------- TOKENIZE ----------------
|
| 369 |
+
encoding = self.tokenizer(
|
| 370 |
+
text,
|
| 371 |
+
padding=padding,
|
| 372 |
+
return_tensors=return_tensors,
|
| 373 |
+
**{k: v for k, v in kwargs.items() if k in (
|
| 374 |
+
"max_length", "truncation", "add_special_tokens",
|
| 375 |
+
"return_attention_mask", "return_token_type_ids",
|
| 376 |
+
)},
|
| 377 |
+
)
|
| 378 |
+
out["input_ids"] = encoding["input_ids"]
|
| 379 |
+
out["attention_mask"] = encoding.get(
|
| 380 |
+
"attention_mask",
|
| 381 |
+
torch.ones_like(encoding["input_ids"]),
|
| 382 |
+
)
|
| 383 |
+
|
| 384 |
+
try:
|
| 385 |
+
from transformers.feature_extraction_utils import BatchFeature
|
| 386 |
+
|
| 387 |
+
return BatchFeature(data=out)
|
| 388 |
+
except Exception:
|
| 389 |
+
return out
|
| 390 |
+
|
| 391 |
+
# ---------------------------------------------------------------- decoding
|
| 392 |
+
|
| 393 |
+
def batch_decode(self, *args, **kwargs):
|
| 394 |
+
return self.tokenizer.batch_decode(*args, **kwargs)
|
| 395 |
+
|
| 396 |
+
def decode(self, *args, **kwargs):
|
| 397 |
+
return self.tokenizer.decode(*args, **kwargs)
|
| 398 |
+
|
| 399 |
+
|
| 400 |
+
__all__ = [
|
| 401 |
+
"LlavaOnevision2Processor",
|
| 402 |
+
"VISION_START",
|
| 403 |
+
"VISION_END",
|
| 404 |
+
"IMAGE_PAD",
|
| 405 |
+
"VIDEO_PAD",
|
| 406 |
+
]
|
special_tokens_map.json
ADDED
|
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"additional_special_tokens": [
|
| 3 |
+
"<|im_start|>",
|
| 4 |
+
"<|im_end|>",
|
| 5 |
+
"<|object_ref_start|>",
|
| 6 |
+
"<|object_ref_end|>",
|
| 7 |
+
"<|box_start|>",
|
| 8 |
+
"<|box_end|>",
|
| 9 |
+
"<|quad_start|>",
|
| 10 |
+
"<|quad_end|>",
|
| 11 |
+
"<|vision_start|>",
|
| 12 |
+
"<|vision_end|>",
|
| 13 |
+
"<|vision_pad|>",
|
| 14 |
+
"<|image_pad|>",
|
| 15 |
+
"<|video_pad|>"
|
| 16 |
+
],
|
| 17 |
+
"eos_token": {
|
| 18 |
+
"content": "<|im_end|>",
|
| 19 |
+
"lstrip": false,
|
| 20 |
+
"normalized": false,
|
| 21 |
+
"rstrip": false,
|
| 22 |
+
"single_word": false
|
| 23 |
+
},
|
| 24 |
+
"pad_token": {
|
| 25 |
+
"content": "<|endoftext|>",
|
| 26 |
+
"lstrip": false,
|
| 27 |
+
"normalized": false,
|
| 28 |
+
"rstrip": false,
|
| 29 |
+
"single_word": false
|
| 30 |
+
}
|
| 31 |
+
}
|
tokenizer.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ba0c439f7be467bf47d12a7e6f9adc6116201056fc60c67f431c679b7c16afc8
|
| 3 |
+
size 11422064
|
tokenizer_config.json
ADDED
|
@@ -0,0 +1,209 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_bos_token": false,
|
| 3 |
+
"add_prefix_space": false,
|
| 4 |
+
"added_tokens_decoder": {
|
| 5 |
+
"151643": {
|
| 6 |
+
"content": "<|endoftext|>",
|
| 7 |
+
"lstrip": false,
|
| 8 |
+
"normalized": false,
|
| 9 |
+
"rstrip": false,
|
| 10 |
+
"single_word": false,
|
| 11 |
+
"special": true
|
| 12 |
+
},
|
| 13 |
+
"151644": {
|
| 14 |
+
"content": "<|im_start|>",
|
| 15 |
+
"lstrip": false,
|
| 16 |
+
"normalized": false,
|
| 17 |
+
"rstrip": false,
|
| 18 |
+
"single_word": false,
|
| 19 |
+
"special": true
|
| 20 |
+
},
|
| 21 |
+
"151645": {
|
| 22 |
+
"content": "<|im_end|>",
|
| 23 |
+
"lstrip": false,
|
| 24 |
+
"normalized": false,
|
| 25 |
+
"rstrip": false,
|
| 26 |
+
"single_word": false,
|
| 27 |
+
"special": true
|
| 28 |
+
},
|
| 29 |
+
"151646": {
|
| 30 |
+
"content": "<|object_ref_start|>",
|
| 31 |
+
"lstrip": false,
|
| 32 |
+
"normalized": false,
|
| 33 |
+
"rstrip": false,
|
| 34 |
+
"single_word": false,
|
| 35 |
+
"special": true
|
| 36 |
+
},
|
| 37 |
+
"151647": {
|
| 38 |
+
"content": "<|object_ref_end|>",
|
| 39 |
+
"lstrip": false,
|
| 40 |
+
"normalized": false,
|
| 41 |
+
"rstrip": false,
|
| 42 |
+
"single_word": false,
|
| 43 |
+
"special": true
|
| 44 |
+
},
|
| 45 |
+
"151648": {
|
| 46 |
+
"content": "<|box_start|>",
|
| 47 |
+
"lstrip": false,
|
| 48 |
+
"normalized": false,
|
| 49 |
+
"rstrip": false,
|
| 50 |
+
"single_word": false,
|
| 51 |
+
"special": true
|
| 52 |
+
},
|
| 53 |
+
"151649": {
|
| 54 |
+
"content": "<|box_end|>",
|
| 55 |
+
"lstrip": false,
|
| 56 |
+
"normalized": false,
|
| 57 |
+
"rstrip": false,
|
| 58 |
+
"single_word": false,
|
| 59 |
+
"special": true
|
| 60 |
+
},
|
| 61 |
+
"151650": {
|
| 62 |
+
"content": "<|quad_start|>",
|
| 63 |
+
"lstrip": false,
|
| 64 |
+
"normalized": false,
|
| 65 |
+
"rstrip": false,
|
| 66 |
+
"single_word": false,
|
| 67 |
+
"special": true
|
| 68 |
+
},
|
| 69 |
+
"151651": {
|
| 70 |
+
"content": "<|quad_end|>",
|
| 71 |
+
"lstrip": false,
|
| 72 |
+
"normalized": false,
|
| 73 |
+
"rstrip": false,
|
| 74 |
+
"single_word": false,
|
| 75 |
+
"special": true
|
| 76 |
+
},
|
| 77 |
+
"151652": {
|
| 78 |
+
"content": "<|vision_start|>",
|
| 79 |
+
"lstrip": false,
|
| 80 |
+
"normalized": false,
|
| 81 |
+
"rstrip": false,
|
| 82 |
+
"single_word": false,
|
| 83 |
+
"special": true
|
| 84 |
+
},
|
| 85 |
+
"151653": {
|
| 86 |
+
"content": "<|vision_end|>",
|
| 87 |
+
"lstrip": false,
|
| 88 |
+
"normalized": false,
|
| 89 |
+
"rstrip": false,
|
| 90 |
+
"single_word": false,
|
| 91 |
+
"special": true
|
| 92 |
+
},
|
| 93 |
+
"151654": {
|
| 94 |
+
"content": "<|vision_pad|>",
|
| 95 |
+
"lstrip": false,
|
| 96 |
+
"normalized": false,
|
| 97 |
+
"rstrip": false,
|
| 98 |
+
"single_word": false,
|
| 99 |
+
"special": true
|
| 100 |
+
},
|
| 101 |
+
"151655": {
|
| 102 |
+
"content": "<|image_pad|>",
|
| 103 |
+
"lstrip": false,
|
| 104 |
+
"normalized": false,
|
| 105 |
+
"rstrip": false,
|
| 106 |
+
"single_word": false,
|
| 107 |
+
"special": true
|
| 108 |
+
},
|
| 109 |
+
"151656": {
|
| 110 |
+
"content": "<|video_pad|>",
|
| 111 |
+
"lstrip": false,
|
| 112 |
+
"normalized": false,
|
| 113 |
+
"rstrip": false,
|
| 114 |
+
"single_word": false,
|
| 115 |
+
"special": true
|
| 116 |
+
},
|
| 117 |
+
"151657": {
|
| 118 |
+
"content": "<tool_call>",
|
| 119 |
+
"lstrip": false,
|
| 120 |
+
"normalized": false,
|
| 121 |
+
"rstrip": false,
|
| 122 |
+
"single_word": false,
|
| 123 |
+
"special": false
|
| 124 |
+
},
|
| 125 |
+
"151658": {
|
| 126 |
+
"content": "</tool_call>",
|
| 127 |
+
"lstrip": false,
|
| 128 |
+
"normalized": false,
|
| 129 |
+
"rstrip": false,
|
| 130 |
+
"single_word": false,
|
| 131 |
+
"special": false
|
| 132 |
+
},
|
| 133 |
+
"151659": {
|
| 134 |
+
"content": "<|fim_prefix|>",
|
| 135 |
+
"lstrip": false,
|
| 136 |
+
"normalized": false,
|
| 137 |
+
"rstrip": false,
|
| 138 |
+
"single_word": false,
|
| 139 |
+
"special": false
|
| 140 |
+
},
|
| 141 |
+
"151660": {
|
| 142 |
+
"content": "<|fim_middle|>",
|
| 143 |
+
"lstrip": false,
|
| 144 |
+
"normalized": false,
|
| 145 |
+
"rstrip": false,
|
| 146 |
+
"single_word": false,
|
| 147 |
+
"special": false
|
| 148 |
+
},
|
| 149 |
+
"151661": {
|
| 150 |
+
"content": "<|fim_suffix|>",
|
| 151 |
+
"lstrip": false,
|
| 152 |
+
"normalized": false,
|
| 153 |
+
"rstrip": false,
|
| 154 |
+
"single_word": false,
|
| 155 |
+
"special": false
|
| 156 |
+
},
|
| 157 |
+
"151662": {
|
| 158 |
+
"content": "<|fim_pad|>",
|
| 159 |
+
"lstrip": false,
|
| 160 |
+
"normalized": false,
|
| 161 |
+
"rstrip": false,
|
| 162 |
+
"single_word": false,
|
| 163 |
+
"special": false
|
| 164 |
+
},
|
| 165 |
+
"151663": {
|
| 166 |
+
"content": "<|repo_name|>",
|
| 167 |
+
"lstrip": false,
|
| 168 |
+
"normalized": false,
|
| 169 |
+
"rstrip": false,
|
| 170 |
+
"single_word": false,
|
| 171 |
+
"special": false
|
| 172 |
+
},
|
| 173 |
+
"151664": {
|
| 174 |
+
"content": "<|file_sep|>",
|
| 175 |
+
"lstrip": false,
|
| 176 |
+
"normalized": false,
|
| 177 |
+
"rstrip": false,
|
| 178 |
+
"single_word": false,
|
| 179 |
+
"special": false
|
| 180 |
+
}
|
| 181 |
+
},
|
| 182 |
+
"additional_special_tokens": [
|
| 183 |
+
"<|im_start|>",
|
| 184 |
+
"<|im_end|>",
|
| 185 |
+
"<|object_ref_start|>",
|
| 186 |
+
"<|object_ref_end|>",
|
| 187 |
+
"<|box_start|>",
|
| 188 |
+
"<|box_end|>",
|
| 189 |
+
"<|quad_start|>",
|
| 190 |
+
"<|quad_end|>",
|
| 191 |
+
"<|vision_start|>",
|
| 192 |
+
"<|vision_end|>",
|
| 193 |
+
"<|vision_pad|>",
|
| 194 |
+
"<|image_pad|>",
|
| 195 |
+
"<|video_pad|>"
|
| 196 |
+
],
|
| 197 |
+
"bos_token": null,
|
| 198 |
+
"clean_up_tokenization_spaces": false,
|
| 199 |
+
"eos_token": "<|im_end|>",
|
| 200 |
+
"errors": "replace",
|
| 201 |
+
"extra_special_tokens": {},
|
| 202 |
+
"model_max_length": 131072,
|
| 203 |
+
"pad_token": "<|endoftext|>",
|
| 204 |
+
"processor_class": "Qwen2_5_VLProcessor",
|
| 205 |
+
"split_special_tokens": false,
|
| 206 |
+
"tokenizer_class": "Qwen2Tokenizer",
|
| 207 |
+
"unk_token": null,
|
| 208 |
+
"use_fast": true
|
| 209 |
+
}
|
video_preprocessor_config.json
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"video_processor_type": "LlavaOnevision2VideoProcessor",
|
| 3 |
+
"processor_class": "LlavaOnevision2Processor",
|
| 4 |
+
"auto_map": {
|
| 5 |
+
"AutoProcessor": "processing_llava_onevision2.LlavaOnevision2Processor",
|
| 6 |
+
"AutoVideoProcessor": "video_processing_llava_onevision2.LlavaOnevision2VideoProcessor"
|
| 7 |
+
},
|
| 8 |
+
"max_frames": 768,
|
| 9 |
+
"fixed_num_frames": null,
|
| 10 |
+
"target_fps": null,
|
| 11 |
+
"min_pixels": 3136,
|
| 12 |
+
"max_pixels": 12845056,
|
| 13 |
+
"patch_size": 14,
|
| 14 |
+
"spatial_merge_size": 2,
|
| 15 |
+
"temporal_patch_size": 1,
|
| 16 |
+
"resize_frames": true
|
| 17 |
+
}
|
video_processing_llava_onevision2.py
ADDED
|
@@ -0,0 +1,694 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Video frame extraction helpers for LlavaOnevision2 native video input.
|
| 2 |
+
|
| 3 |
+
Helpers (decord-first / opencv-fallback decoding) are used by
|
| 4 |
+
``LlavaOnevision2VideoProcessor`` defined below.
|
| 5 |
+
|
| 6 |
+
The helpers were ported from the training pipeline with minor cleanups:
|
| 7 |
+
- dropped wrapper-only imports
|
| 8 |
+
- consolidated timestamp helpers
|
| 9 |
+
- kept decord-first / opencv-fallback decoding identical
|
| 10 |
+
|
| 11 |
+
Public API:
|
| 12 |
+
- format_timestamp(seconds) -> "MM:SS.xx"
|
| 13 |
+
- choose_target_frames(duration, max_frames, fixed_num_frames=None,
|
| 14 |
+
target_fps=None) -> int
|
| 15 |
+
- select_frame_indices(frame_count, target_count) -> list[int]
|
| 16 |
+
- smart_resize(h, w, patch_size=14, min_pixels=None, max_pixels=None,
|
| 17 |
+
align_patch_size=None) -> (h, w)
|
| 18 |
+
- extract_video_frames(video_path, ...) -> (frames_np, frame_indices,
|
| 19 |
+
timestamps_dict)
|
| 20 |
+
- extract_video_frames_to_pil(video_path, ...) -> (frames_pil, frame_indices,
|
| 21 |
+
timestamps_dict)
|
| 22 |
+
"""
|
| 23 |
+
|
| 24 |
+
from __future__ import annotations
|
| 25 |
+
|
| 26 |
+
import logging
|
| 27 |
+
import math
|
| 28 |
+
from typing import List, Optional, Tuple
|
| 29 |
+
|
| 30 |
+
import numpy as np
|
| 31 |
+
import torch
|
| 32 |
+
|
| 33 |
+
logger = logging.getLogger(__name__)
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
# =============================================================================
|
| 37 |
+
# Timestamp helpers
|
| 38 |
+
# =============================================================================
|
| 39 |
+
|
| 40 |
+
def format_timestamp(seconds: float) -> str:
|
| 41 |
+
minutes = int(seconds // 60)
|
| 42 |
+
sec = seconds - minutes * 60
|
| 43 |
+
return f"{minutes:02d}:{sec:09.6f}"
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
def time_str_to_seconds(t: str) -> float:
|
| 47 |
+
"""Convert ``MM:SS.xx`` back to a float number of seconds.
|
| 48 |
+
|
| 49 |
+
Inverse of :func:`format_timestamp`.
|
| 50 |
+
"""
|
| 51 |
+
minute, sec = t.split(":")
|
| 52 |
+
return int(minute) * 60 + float(sec)
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
# =============================================================================
|
| 56 |
+
# Frame-count / index selection
|
| 57 |
+
# =============================================================================
|
| 58 |
+
|
| 59 |
+
def choose_target_frames(
|
| 60 |
+
duration_seconds: float,
|
| 61 |
+
max_frames: int,
|
| 62 |
+
fixed_num_frames: Optional[int] = None,
|
| 63 |
+
target_fps: Optional[float] = None,
|
| 64 |
+
) -> int:
|
| 65 |
+
"""Choose target frame count based on video duration in seconds.
|
| 66 |
+
|
| 67 |
+
Sampling strategy:
|
| 68 |
+
- if ``target_fps`` is set, sample at that fps (capped by ``max_frames``)
|
| 69 |
+
- elif ``fixed_num_frames`` is set, use that exact count
|
| 70 |
+
- else duration < 10s -> 8 frames
|
| 71 |
+
- duration < 30s -> 16 frames
|
| 72 |
+
- otherwise -> ``max_frames`` (default 32)
|
| 73 |
+
"""
|
| 74 |
+
if target_fps is not None and target_fps > 0:
|
| 75 |
+
return min(max(1, int(duration_seconds * target_fps)), max_frames)
|
| 76 |
+
if fixed_num_frames is not None:
|
| 77 |
+
return fixed_num_frames
|
| 78 |
+
if duration_seconds < 10:
|
| 79 |
+
return 8
|
| 80 |
+
if duration_seconds < 30:
|
| 81 |
+
return 16
|
| 82 |
+
return max_frames
|
| 83 |
+
|
| 84 |
+
|
| 85 |
+
def select_frame_indices(frame_count: int, target_count: int) -> List[int]:
|
| 86 |
+
if frame_count <= target_count:
|
| 87 |
+
return list(range(frame_count))
|
| 88 |
+
return torch.linspace(0, frame_count - 1, target_count).round().long().tolist()
|
| 89 |
+
|
| 90 |
+
|
| 91 |
+
# =============================================================================
|
| 92 |
+
# Spatial resize
|
| 93 |
+
# =============================================================================
|
| 94 |
+
|
| 95 |
+
def smart_resize(height, width, patch_size=14, min_pixels=None, max_pixels=None, align_patch_size=None):
|
| 96 |
+
if height <= 0 or width <= 0:
|
| 97 |
+
raise ValueError(f"Invalid size: height={height}, width={width}")
|
| 98 |
+
factor = align_patch_size or patch_size
|
| 99 |
+
h_bar = max(factor, int(round(height / factor) * factor))
|
| 100 |
+
w_bar = max(factor, int(round(width / factor) * factor))
|
| 101 |
+
if max_pixels and h_bar * w_bar > max_pixels:
|
| 102 |
+
beta = math.sqrt((height * width) / max_pixels)
|
| 103 |
+
h_bar = math.floor(height / beta / factor) * factor
|
| 104 |
+
w_bar = math.floor(width / beta / factor) * factor
|
| 105 |
+
elif min_pixels and h_bar * w_bar < min_pixels:
|
| 106 |
+
beta = math.sqrt(min_pixels / (height * width))
|
| 107 |
+
h_bar = math.ceil(height * beta / factor) * factor
|
| 108 |
+
w_bar = math.ceil(width * beta / factor) * factor
|
| 109 |
+
return int(h_bar), int(w_bar)
|
| 110 |
+
|
| 111 |
+
|
| 112 |
+
# =============================================================================
|
| 113 |
+
# Frame extraction (decord first, opencv fallback)
|
| 114 |
+
# =============================================================================
|
| 115 |
+
|
| 116 |
+
def extract_video_frames(
|
| 117 |
+
video_path: str,
|
| 118 |
+
max_frames: int = 32,
|
| 119 |
+
patch_size: int = 14,
|
| 120 |
+
min_pixels: Optional[int] = None,
|
| 121 |
+
max_pixels: Optional[int] = None,
|
| 122 |
+
resize_frames: bool = True,
|
| 123 |
+
fixed_num_frames: Optional[int] = None,
|
| 124 |
+
target_fps: Optional[float] = None,
|
| 125 |
+
) -> Tuple[List[np.ndarray], torch.Tensor, dict]:
|
| 126 |
+
"""Extract frames from a video.
|
| 127 |
+
|
| 128 |
+
Sampling rule matches :func:`choose_target_frames`. Decoding tries decord
|
| 129 |
+
first (better codec coverage) and falls back to OpenCV.
|
| 130 |
+
|
| 131 |
+
Args:
|
| 132 |
+
video_path: path to the input video file.
|
| 133 |
+
max_frames: cap for long videos.
|
| 134 |
+
patch_size: vision tower patch size for alignment.
|
| 135 |
+
min_pixels: minimum pixel budget for resize.
|
| 136 |
+
max_pixels: maximum pixel budget for resize.
|
| 137 |
+
resize_frames: whether to apply :func:`smart_resize` (with
|
| 138 |
+
``align_patch_size = patch_size * 2``, i.e. 28 for spatial_merge=2).
|
| 139 |
+
fixed_num_frames: see :func:`choose_target_frames`.
|
| 140 |
+
target_fps: see :func:`choose_target_frames`.
|
| 141 |
+
|
| 142 |
+
Returns:
|
| 143 |
+
Tuple of:
|
| 144 |
+
- ``frames`` : list of RGB ``np.ndarray`` (H, W, 3), dtype uint8.
|
| 145 |
+
- ``frame_indices`` : 1D ``torch.Tensor[int64]`` of selected indices.
|
| 146 |
+
- ``timestamps`` : ``dict[str(frame_idx) -> "MM:SS.xx"]``.
|
| 147 |
+
|
| 148 |
+
Notes:
|
| 149 |
+
Lazy imports of ``decord`` and ``cv2`` keep the module importable in
|
| 150 |
+
environments where neither is installed (e.g. unit tests that only
|
| 151 |
+
exercise the helpers above).
|
| 152 |
+
"""
|
| 153 |
+
frames: List[np.ndarray] = []
|
| 154 |
+
timestamps: dict = {}
|
| 155 |
+
frame_indices: List[int] = []
|
| 156 |
+
|
| 157 |
+
# Prefer decord because of broader codec support.
|
| 158 |
+
try:
|
| 159 |
+
import decord # type: ignore
|
| 160 |
+
|
| 161 |
+
vr = decord.VideoReader(video_path)
|
| 162 |
+
frame_count = len(vr)
|
| 163 |
+
fps = vr.get_avg_fps()
|
| 164 |
+
if not fps or fps <= 0:
|
| 165 |
+
fps = 30.0
|
| 166 |
+
|
| 167 |
+
duration = frame_count / fps
|
| 168 |
+
target_count = choose_target_frames(
|
| 169 |
+
duration, max_frames, fixed_num_frames, target_fps
|
| 170 |
+
)
|
| 171 |
+
selected_indices = select_frame_indices(frame_count, target_count)
|
| 172 |
+
|
| 173 |
+
# One-shot batch decode + torchvision BICUBIC+antialias resize.
|
| 174 |
+
# Mirrors qwen_vl_utils.fetch_video, replacing per-frame cv2 INTER_AREA/LINEAR.
|
| 175 |
+
arr = vr.get_batch(selected_indices).asnumpy() # [N,H,W,3] uint8 RGB
|
| 176 |
+
H, W = arr.shape[1], arr.shape[2]
|
| 177 |
+
if resize_frames and (min_pixels or max_pixels):
|
| 178 |
+
resized_h, resized_w = smart_resize(
|
| 179 |
+
H, W, patch_size,
|
| 180 |
+
min_pixels=min_pixels,
|
| 181 |
+
max_pixels=max_pixels,
|
| 182 |
+
align_patch_size=patch_size * 2,
|
| 183 |
+
)
|
| 184 |
+
if (resized_h, resized_w) != (H, W):
|
| 185 |
+
from torchvision import transforms as _T
|
| 186 |
+
from torchvision.transforms import InterpolationMode as _IM
|
| 187 |
+
video_t = torch.from_numpy(arr).permute(0, 3, 1, 2).contiguous()
|
| 188 |
+
video_t = _T.functional.resize(
|
| 189 |
+
video_t,
|
| 190 |
+
[resized_h, resized_w],
|
| 191 |
+
interpolation=_IM.BICUBIC,
|
| 192 |
+
antialias=True,
|
| 193 |
+
)
|
| 194 |
+
arr = video_t.permute(0, 2, 3, 1).contiguous().numpy()
|
| 195 |
+
|
| 196 |
+
frames = list(arr)
|
| 197 |
+
frame_indices = list(selected_indices)
|
| 198 |
+
for frame_idx in selected_indices:
|
| 199 |
+
timestamps[str(int(frame_idx))] = format_timestamp(int(frame_idx) / fps)
|
| 200 |
+
|
| 201 |
+
return frames, torch.tensor(frame_indices, dtype=torch.int64), timestamps
|
| 202 |
+
except Exception as e:
|
| 203 |
+
logger.warning(
|
| 204 |
+
f"decord failed to open {video_path}: {e}; falling back to OpenCV"
|
| 205 |
+
)
|
| 206 |
+
|
| 207 |
+
# OpenCV fallback.
|
| 208 |
+
import cv2 # type: ignore
|
| 209 |
+
|
| 210 |
+
cap = cv2.VideoCapture(video_path)
|
| 211 |
+
if not cap.isOpened():
|
| 212 |
+
logger.warning(f"OpenCV also failed to open video, skipped: {video_path}")
|
| 213 |
+
return frames, torch.tensor(frame_indices, dtype=torch.int64), timestamps
|
| 214 |
+
|
| 215 |
+
fps = cap.get(cv2.CAP_PROP_FPS)
|
| 216 |
+
if not fps or fps <= 0:
|
| 217 |
+
fps = 30.0
|
| 218 |
+
|
| 219 |
+
frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT) or 0)
|
| 220 |
+
|
| 221 |
+
if frame_count > 0:
|
| 222 |
+
duration = frame_count / fps
|
| 223 |
+
target_count = choose_target_frames(
|
| 224 |
+
duration, max_frames, fixed_num_frames, target_fps
|
| 225 |
+
)
|
| 226 |
+
selected_indices = select_frame_indices(frame_count, target_count)
|
| 227 |
+
|
| 228 |
+
for frame_idx in selected_indices:
|
| 229 |
+
cap.set(cv2.CAP_PROP_POS_FRAMES, frame_idx)
|
| 230 |
+
ret, frame = cap.read()
|
| 231 |
+
if not ret:
|
| 232 |
+
continue
|
| 233 |
+
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
| 234 |
+
|
| 235 |
+
if resize_frames and (min_pixels or max_pixels):
|
| 236 |
+
resized_h, resized_w = smart_resize(
|
| 237 |
+
frame.shape[0],
|
| 238 |
+
frame.shape[1],
|
| 239 |
+
patch_size,
|
| 240 |
+
min_pixels,
|
| 241 |
+
max_pixels,
|
| 242 |
+
align_patch_size=patch_size * 2,
|
| 243 |
+
)
|
| 244 |
+
if (resized_h, resized_w) != (frame.shape[0], frame.shape[1]):
|
| 245 |
+
interp = (
|
| 246 |
+
cv2.INTER_AREA
|
| 247 |
+
if resized_h < frame.shape[0] or resized_w < frame.shape[1]
|
| 248 |
+
else cv2.INTER_LINEAR
|
| 249 |
+
)
|
| 250 |
+
frame = cv2.resize(frame, (resized_w, resized_h), interpolation=interp)
|
| 251 |
+
|
| 252 |
+
frames.append(frame)
|
| 253 |
+
timestamps[str(frame_idx)] = format_timestamp(frame_idx / fps)
|
| 254 |
+
frame_indices.append(frame_idx)
|
| 255 |
+
else:
|
| 256 |
+
# Unknown frame count: read sequentially then sample.
|
| 257 |
+
frame_idx = 0
|
| 258 |
+
temp_frames: List[Tuple[int, np.ndarray]] = []
|
| 259 |
+
while True:
|
| 260 |
+
ret, frame = cap.read()
|
| 261 |
+
if not ret:
|
| 262 |
+
break
|
| 263 |
+
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
| 264 |
+
temp_frames.append((frame_idx, frame))
|
| 265 |
+
frame_idx += 1
|
| 266 |
+
|
| 267 |
+
if temp_frames:
|
| 268 |
+
duration = len(temp_frames) / fps
|
| 269 |
+
target_count = choose_target_frames(
|
| 270 |
+
duration, max_frames, fixed_num_frames, target_fps
|
| 271 |
+
)
|
| 272 |
+
selected_indices = select_frame_indices(len(temp_frames), target_count)
|
| 273 |
+
|
| 274 |
+
for idx in selected_indices:
|
| 275 |
+
frame_idx, frame = temp_frames[idx]
|
| 276 |
+
if resize_frames and (min_pixels or max_pixels):
|
| 277 |
+
resized_h, resized_w = smart_resize(
|
| 278 |
+
frame.shape[0],
|
| 279 |
+
frame.shape[1],
|
| 280 |
+
patch_size,
|
| 281 |
+
min_pixels,
|
| 282 |
+
max_pixels,
|
| 283 |
+
align_patch_size=patch_size * 2,
|
| 284 |
+
)
|
| 285 |
+
if (resized_h, resized_w) != (frame.shape[0], frame.shape[1]):
|
| 286 |
+
interp = (
|
| 287 |
+
cv2.INTER_AREA
|
| 288 |
+
if resized_h < frame.shape[0] or resized_w < frame.shape[1]
|
| 289 |
+
else cv2.INTER_LINEAR
|
| 290 |
+
)
|
| 291 |
+
frame = cv2.resize(frame, (resized_w, resized_h), interpolation=interp)
|
| 292 |
+
|
| 293 |
+
frames.append(frame)
|
| 294 |
+
timestamps[str(frame_idx)] = format_timestamp(frame_idx / fps)
|
| 295 |
+
frame_indices.append(frame_idx)
|
| 296 |
+
|
| 297 |
+
cap.release()
|
| 298 |
+
return frames, torch.tensor(frame_indices, dtype=torch.int64), timestamps
|
| 299 |
+
|
| 300 |
+
|
| 301 |
+
def extract_video_frames_to_pil(
|
| 302 |
+
video_path: str,
|
| 303 |
+
max_frames: int = 32,
|
| 304 |
+
patch_size: int = 14,
|
| 305 |
+
min_pixels: Optional[int] = None,
|
| 306 |
+
max_pixels: Optional[int] = None,
|
| 307 |
+
resize_frames: bool = True,
|
| 308 |
+
fixed_num_frames: Optional[int] = None,
|
| 309 |
+
target_fps: Optional[float] = None,
|
| 310 |
+
):
|
| 311 |
+
"""Same as :func:`extract_video_frames` but returns a list of PIL Images."""
|
| 312 |
+
from PIL import Image # local import: PIL is mandatory for the processor
|
| 313 |
+
|
| 314 |
+
frames_np, frame_indices, timestamps = extract_video_frames(
|
| 315 |
+
video_path=video_path,
|
| 316 |
+
max_frames=max_frames,
|
| 317 |
+
patch_size=patch_size,
|
| 318 |
+
min_pixels=min_pixels,
|
| 319 |
+
max_pixels=max_pixels,
|
| 320 |
+
resize_frames=resize_frames,
|
| 321 |
+
fixed_num_frames=fixed_num_frames,
|
| 322 |
+
target_fps=target_fps,
|
| 323 |
+
)
|
| 324 |
+
frames_pil = [Image.fromarray(frame) for frame in frames_np]
|
| 325 |
+
return frames_pil, frame_indices, timestamps
|
| 326 |
+
|
| 327 |
+
|
| 328 |
+
# =============================================================================
|
| 329 |
+
# patch_positions construction (row-major + 2x2 block-layout reorder)
|
| 330 |
+
# =============================================================================
|
| 331 |
+
# Block-layout reorder mirroring the training pipeline, kept here so the
|
| 332 |
+
# VideoProcessor is self-contained.
|
| 333 |
+
|
| 334 |
+
def _convert_positions_to_block_layout(
|
| 335 |
+
positions: torch.Tensor,
|
| 336 |
+
t: int,
|
| 337 |
+
h: int,
|
| 338 |
+
w: int,
|
| 339 |
+
spatial_merge_size: int = 2,
|
| 340 |
+
) -> torch.Tensor:
|
| 341 |
+
"""Reorder ``[t*h*w, 3]`` row-major positions to 2x2 block layout."""
|
| 342 |
+
sms = spatial_merge_size
|
| 343 |
+
if sms == 1:
|
| 344 |
+
return positions
|
| 345 |
+
device = positions.device
|
| 346 |
+
total = t * h * w
|
| 347 |
+
indices = torch.arange(total, device=device).view(t, h, w)
|
| 348 |
+
h_m, w_m = h // sms, w // sms
|
| 349 |
+
indices = (
|
| 350 |
+
indices.view(t, h_m, sms, w_m, sms)
|
| 351 |
+
.permute(0, 1, 3, 2, 4)
|
| 352 |
+
.contiguous()
|
| 353 |
+
.view(total)
|
| 354 |
+
)
|
| 355 |
+
return positions[indices]
|
| 356 |
+
|
| 357 |
+
|
| 358 |
+
def build_patch_positions(
|
| 359 |
+
grid_thw: torch.Tensor,
|
| 360 |
+
spatial_merge_size: int = 2,
|
| 361 |
+
frame_indices: Optional[List[Optional[torch.Tensor]]] = None,
|
| 362 |
+
) -> torch.Tensor:
|
| 363 |
+
"""Build block-layout ``[t,h,w]`` patch positions for one or many videos/images.
|
| 364 |
+
|
| 365 |
+
Args:
|
| 366 |
+
grid_thw: ``[num_samples, 3]`` LongTensor (T, H_p, W_p) per sample.
|
| 367 |
+
spatial_merge_size: vision tower spatial-merge size (default 2).
|
| 368 |
+
frame_indices: optional list (one entry per row of ``grid_thw``) of
|
| 369 |
+
real frame indices to use as the t-coordinate. Each entry should
|
| 370 |
+
be a 1-D LongTensor of length ``T`` for that sample. When provided
|
| 371 |
+
this matches the training pipeline,
|
| 372 |
+
where ``t`` is the original frame number in the source video so
|
| 373 |
+
the vision tower's 3-D RoPE encodes the actual temporal position
|
| 374 |
+
rather than a 0..T-1 dense index. Pass ``None`` for an entry to
|
| 375 |
+
fall back to dense ``arange(T)`` for that sample.
|
| 376 |
+
|
| 377 |
+
Returns:
|
| 378 |
+
``[sum(T*H_p*W_p), 3]`` Int64Tensor in block layout, ready to feed
|
| 379 |
+
``forward(... patch_positions=...)``.
|
| 380 |
+
"""
|
| 381 |
+
out = []
|
| 382 |
+
for sample_idx, row in enumerate(grid_thw):
|
| 383 |
+
t_v, h_v, w_v = int(row[0]), int(row[1]), int(row[2])
|
| 384 |
+
h_coords = torch.arange(h_v, dtype=torch.int64).repeat_interleave(w_v).repeat(t_v)
|
| 385 |
+
w_coords = torch.arange(w_v, dtype=torch.int64).repeat(h_v).repeat(t_v)
|
| 386 |
+
# t-coords: prefer real frame_indices (training convention) when given.
|
| 387 |
+
sample_frame_idx = None
|
| 388 |
+
if frame_indices is not None and sample_idx < len(frame_indices):
|
| 389 |
+
sample_frame_idx = frame_indices[sample_idx]
|
| 390 |
+
if sample_frame_idx is not None:
|
| 391 |
+
fi = torch.as_tensor(sample_frame_idx, dtype=torch.int64)
|
| 392 |
+
if fi.numel() != t_v:
|
| 393 |
+
raise ValueError(
|
| 394 |
+
f"frame_indices[{sample_idx}] has length {fi.numel()} but "
|
| 395 |
+
f"grid_thw[{sample_idx}, 0] = {t_v}"
|
| 396 |
+
)
|
| 397 |
+
t_coords = fi.repeat_interleave(h_v * w_v)
|
| 398 |
+
else:
|
| 399 |
+
# Each frame's t coordinate runs 0..t_v-1 (each value repeated h_v*w_v).
|
| 400 |
+
t_coords = torch.arange(t_v, dtype=torch.int64).repeat_interleave(h_v * w_v)
|
| 401 |
+
pp = torch.stack([t_coords, h_coords, w_coords], dim=1)
|
| 402 |
+
pp = _convert_positions_to_block_layout(pp, t_v, h_v, w_v, spatial_merge_size)
|
| 403 |
+
out.append(pp)
|
| 404 |
+
return torch.cat(out, dim=0)
|
| 405 |
+
|
| 406 |
+
|
| 407 |
+
# =============================================================================
|
| 408 |
+
# LlavaOnevision2VideoProcessor
|
| 409 |
+
# =============================================================================
|
| 410 |
+
# A thin processor that wraps `Qwen2VLImageProcessor` to convert raw video
|
| 411 |
+
# files (or pre-decoded frame lists) into the tensor bundle needed by the
|
| 412 |
+
# LlavaOnevision2 model.
|
| 413 |
+
#
|
| 414 |
+
# Output (BatchFeature):
|
| 415 |
+
# - pixel_values_videos : [sum(T*H_p*W_p), C, P, P] patch tensor
|
| 416 |
+
# - video_grid_thw : [num_videos, 3] (T_eff, H_p, W_p)
|
| 417 |
+
# - patch_positions : [sum(T*H_p*W_p), 3] block layout
|
| 418 |
+
# - frame_timestamps : list[list[float]] per-video per-frame seconds
|
| 419 |
+
#
|
| 420 |
+
# Aligned with the modeling code, we deliberately
|
| 421 |
+
# DO NOT emit `second_per_grid_ts`.
|
| 422 |
+
|
| 423 |
+
class LlavaOnevision2VideoProcessor:
|
| 424 |
+
"""Decode + sample + patch-ify videos for LlavaOnevision2.
|
| 425 |
+
|
| 426 |
+
Designed to be standalone (does not inherit ``transformers.ProcessorMixin``)
|
| 427 |
+
so it can be unit-tested without the full Processor stack.
|
| 428 |
+
"""
|
| 429 |
+
|
| 430 |
+
# Canonical defaults.
|
| 431 |
+
DEFAULT_MAX_FRAMES = 384
|
| 432 |
+
DEFAULT_PATCH_SIZE = 14
|
| 433 |
+
DEFAULT_SPATIAL_MERGE_SIZE = 2
|
| 434 |
+
DEFAULT_TEMPORAL_PATCH_SIZE = 1 # this checkpoint ships tps=1
|
| 435 |
+
DEFAULT_MIN_PIXELS = 256 * 28 * 28
|
| 436 |
+
DEFAULT_MAX_PIXELS = 1605632
|
| 437 |
+
|
| 438 |
+
def __init__(
|
| 439 |
+
self,
|
| 440 |
+
image_processor=None,
|
| 441 |
+
max_frames: int = DEFAULT_MAX_FRAMES,
|
| 442 |
+
fixed_num_frames: Optional[int] = None,
|
| 443 |
+
target_fps: Optional[float] = None,
|
| 444 |
+
patch_size: int = DEFAULT_PATCH_SIZE,
|
| 445 |
+
spatial_merge_size: int = DEFAULT_SPATIAL_MERGE_SIZE,
|
| 446 |
+
temporal_patch_size: int = DEFAULT_TEMPORAL_PATCH_SIZE,
|
| 447 |
+
min_pixels: int = DEFAULT_MIN_PIXELS,
|
| 448 |
+
max_pixels: int = DEFAULT_MAX_PIXELS,
|
| 449 |
+
resize_frames: bool = True,
|
| 450 |
+
):
|
| 451 |
+
"""
|
| 452 |
+
Args:
|
| 453 |
+
image_processor: a `Qwen2VLImageProcessor` instance. If ``None`` an
|
| 454 |
+
instance is built from the other kwargs at first call.
|
| 455 |
+
max_frames / fixed_num_frames / target_fps: see
|
| 456 |
+
:func:`choose_target_frames`.
|
| 457 |
+
patch_size: vision tower patch size (default 14).
|
| 458 |
+
spatial_merge_size: vision tower spatial merge factor (default 2).
|
| 459 |
+
temporal_patch_size: temporal-patch grouping; this checkpoint
|
| 460 |
+
ships ``temporal_patch_size=1`` so each pv row is one single
|
| 461 |
+
patch (3*14*14=588) and ``Σ t·h·w == total_patches``
|
| 462 |
+
naturally. Override only if loading a non-default processor.
|
| 463 |
+
min_pixels / max_pixels: smart_resize budget.
|
| 464 |
+
resize_frames: whether to resize frames before patching.
|
| 465 |
+
"""
|
| 466 |
+
self._image_processor = image_processor
|
| 467 |
+
self.max_frames = max_frames
|
| 468 |
+
self.fixed_num_frames = fixed_num_frames
|
| 469 |
+
self.target_fps = target_fps
|
| 470 |
+
self.patch_size = patch_size
|
| 471 |
+
self.spatial_merge_size = spatial_merge_size
|
| 472 |
+
self.temporal_patch_size = temporal_patch_size
|
| 473 |
+
self.min_pixels = min_pixels
|
| 474 |
+
self.max_pixels = max_pixels
|
| 475 |
+
self.resize_frames = resize_frames
|
| 476 |
+
|
| 477 |
+
# ------------------------------------------------------------------ utils
|
| 478 |
+
|
| 479 |
+
@property
|
| 480 |
+
def image_processor(self):
|
| 481 |
+
"""Lazy-build the underlying `Qwen2VLImageProcessor`."""
|
| 482 |
+
if self._image_processor is None:
|
| 483 |
+
from transformers import Qwen2VLImageProcessor
|
| 484 |
+
|
| 485 |
+
self._image_processor = Qwen2VLImageProcessor(
|
| 486 |
+
min_pixels=self.min_pixels,
|
| 487 |
+
max_pixels=self.max_pixels,
|
| 488 |
+
patch_size=self.patch_size,
|
| 489 |
+
merge_size=self.spatial_merge_size,
|
| 490 |
+
temporal_patch_size=self.temporal_patch_size,
|
| 491 |
+
)
|
| 492 |
+
return self._image_processor
|
| 493 |
+
|
| 494 |
+
@staticmethod
|
| 495 |
+
def _coerce_video_input(video):
|
| 496 |
+
"""Normalise a single video input to ``(frames_pil, timestamps_seconds)``.
|
| 497 |
+
|
| 498 |
+
Accepts:
|
| 499 |
+
- ``str`` path to a video file,
|
| 500 |
+
- ``list[PIL.Image]`` (already decoded; timestamps default to None),
|
| 501 |
+
- ``list[np.ndarray]`` (RGB uint8; converted to PIL).
|
| 502 |
+
"""
|
| 503 |
+
from PIL import Image
|
| 504 |
+
|
| 505 |
+
if isinstance(video, str):
|
| 506 |
+
return None # signal: use video path through extract_video_frames_to_pil
|
| 507 |
+
if isinstance(video, list) and len(video) > 0:
|
| 508 |
+
first = video[0]
|
| 509 |
+
if isinstance(first, Image.Image):
|
| 510 |
+
return list(video), None
|
| 511 |
+
if isinstance(first, np.ndarray):
|
| 512 |
+
return [Image.fromarray(f) for f in video], None
|
| 513 |
+
raise TypeError(
|
| 514 |
+
f"Unsupported video input type: {type(video).__name__}. "
|
| 515 |
+
"Expected file path, list[PIL.Image], or list[np.ndarray]."
|
| 516 |
+
)
|
| 517 |
+
|
| 518 |
+
# ---------------------------------------------------------------- __call__
|
| 519 |
+
|
| 520 |
+
def __call__(
|
| 521 |
+
self,
|
| 522 |
+
videos,
|
| 523 |
+
return_tensors: Optional[str] = "pt",
|
| 524 |
+
**kwargs,
|
| 525 |
+
):
|
| 526 |
+
"""Process one or several videos.
|
| 527 |
+
|
| 528 |
+
Args:
|
| 529 |
+
videos: a single video or a list of videos. Each video may be a
|
| 530 |
+
path, a list of PIL frames, or a list of np.ndarray RGB frames.
|
| 531 |
+
return_tensors: only ``"pt"`` is supported (mirrors the underlying
|
| 532 |
+
image processor).
|
| 533 |
+
**kwargs: ignored / reserved for transformers ProcessorMixin
|
| 534 |
+
compatibility (e.g. ``do_rescale``).
|
| 535 |
+
|
| 536 |
+
Returns:
|
| 537 |
+
A dict-like object with keys:
|
| 538 |
+
- ``pixel_values_videos`` : Tensor ``[N_total_patches, C, P, P]``
|
| 539 |
+
- ``video_grid_thw`` : Tensor ``[num_videos, 3]`` (T, H_p, W_p)
|
| 540 |
+
- ``patch_positions`` : Tensor ``[N_total_patches, 3]`` block layout
|
| 541 |
+
- ``frame_timestamps`` : ``list[list[float]]`` per video
|
| 542 |
+
"""
|
| 543 |
+
if return_tensors not in (None, "pt"):
|
| 544 |
+
raise ValueError(
|
| 545 |
+
f"return_tensors={return_tensors!r} not supported; only 'pt' is."
|
| 546 |
+
)
|
| 547 |
+
|
| 548 |
+
# Normalise to a list of videos.
|
| 549 |
+
if not isinstance(videos, (list, tuple)) or (
|
| 550 |
+
len(videos) > 0
|
| 551 |
+
and (isinstance(videos[0], str) is False)
|
| 552 |
+
and not isinstance(videos[0], list)
|
| 553 |
+
):
|
| 554 |
+
# Heuristic: a single video as `list[PIL.Image]` should not be
|
| 555 |
+
# treated as a batch of single-frame videos. We detect that case
|
| 556 |
+
# by checking the inner element type.
|
| 557 |
+
from PIL import Image
|
| 558 |
+
|
| 559 |
+
if isinstance(videos, list) and len(videos) > 0 and isinstance(
|
| 560 |
+
videos[0], (Image.Image, np.ndarray)
|
| 561 |
+
):
|
| 562 |
+
videos = [videos]
|
| 563 |
+
elif isinstance(videos, str):
|
| 564 |
+
videos = [videos]
|
| 565 |
+
if not isinstance(videos, (list, tuple)):
|
| 566 |
+
videos = [videos]
|
| 567 |
+
|
| 568 |
+
per_video_pixel_values = []
|
| 569 |
+
per_video_grid_thw = []
|
| 570 |
+
per_video_patch_positions = []
|
| 571 |
+
frame_timestamps_all: List[List[float]] = []
|
| 572 |
+
|
| 573 |
+
for video in videos:
|
| 574 |
+
# 1) Decode + sample
|
| 575 |
+
if isinstance(video, str):
|
| 576 |
+
frames_pil, frame_indices, timestamps = extract_video_frames_to_pil(
|
| 577 |
+
video_path=video,
|
| 578 |
+
max_frames=self.max_frames,
|
| 579 |
+
patch_size=self.patch_size,
|
| 580 |
+
min_pixels=self.min_pixels,
|
| 581 |
+
max_pixels=self.max_pixels,
|
| 582 |
+
resize_frames=self.resize_frames,
|
| 583 |
+
fixed_num_frames=self.fixed_num_frames,
|
| 584 |
+
target_fps=self.target_fps,
|
| 585 |
+
)
|
| 586 |
+
# Reconstruct fps from any two timestamps, fall back to 30.
|
| 587 |
+
seconds_seq: List[float] = []
|
| 588 |
+
if len(frames_pil) > 0:
|
| 589 |
+
fi_list = frame_indices.tolist()
|
| 590 |
+
for fi in fi_list:
|
| 591 |
+
ts = timestamps.get(str(int(fi)))
|
| 592 |
+
if ts is None:
|
| 593 |
+
seconds_seq.append(0.0)
|
| 594 |
+
else:
|
| 595 |
+
seconds_seq.append(time_str_to_seconds(ts))
|
| 596 |
+
# Real frame indices in the source video (training convention
|
| 597 |
+
# for the t-axis of patch_positions).
|
| 598 |
+
frame_indices_t = frame_indices.to(torch.int64)
|
| 599 |
+
else:
|
| 600 |
+
pre_decoded = self._coerce_video_input(video)
|
| 601 |
+
frames_pil, _ = pre_decoded
|
| 602 |
+
seconds_seq = [float(i) for i in range(len(frames_pil))]
|
| 603 |
+
# Without the original video we have no real indices; fall back
|
| 604 |
+
# to dense ``arange(T)``.
|
| 605 |
+
frame_indices_t = torch.arange(len(frames_pil), dtype=torch.int64)
|
| 606 |
+
|
| 607 |
+
if len(frames_pil) == 0:
|
| 608 |
+
raise ValueError(f"No frames decoded from video: {video!r}")
|
| 609 |
+
|
| 610 |
+
# 2) Patch-ify via Qwen2VLImageProcessor.
|
| 611 |
+
# Video frames go
|
| 612 |
+
# through the *image* path, one frame == one image. The
|
| 613 |
+
# resulting `image_grid_thw` has shape ``[N, 3]`` with each row
|
| 614 |
+
# ``[1, H_p, W_p]``. We then merge into a single video grid
|
| 615 |
+
# ``[1, T=N, H_p, W_p]`` (smart_resize guarantees same H/W).
|
| 616 |
+
#
|
| 617 |
+
# Important: this checkpoint ships an image processor with
|
| 618 |
+
# ``temporal_patch_size=1``, so each pv row encodes ONE single
|
| 619 |
+
# patch (3*14*14 = 588). The OneVision encoder's embedding
|
| 620 |
+
# layer reshapes pv via ``view(-1, 3, 14, 14)`` and produces
|
| 621 |
+
# exactly ``pv.shape[0]`` patches, so the cu_seqlens check
|
| 622 |
+
# ``Σ t·h·w == total_patches`` is satisfied with the natural
|
| 623 |
+
# per-frame grid below. The lazy-built fallback in
|
| 624 |
+
# ``image_processor`` honors ``temporal_patch_size=1`` to keep
|
| 625 |
+
# standalone tests aligned with the checkpoint convention.
|
| 626 |
+
ip = self.image_processor
|
| 627 |
+
data = ip(images=frames_pil, return_tensors="pt")
|
| 628 |
+
pixel_values = data["pixel_values"]
|
| 629 |
+
image_grid_thw = data["image_grid_thw"] # [N, 3]
|
| 630 |
+
|
| 631 |
+
if not torch.all(image_grid_thw[:, 1] == image_grid_thw[0, 1]) or not torch.all(
|
| 632 |
+
image_grid_thw[:, 2] == image_grid_thw[0, 2]
|
| 633 |
+
):
|
| 634 |
+
raise RuntimeError(
|
| 635 |
+
"Frames yielded inconsistent (H_p, W_p); smart_resize should "
|
| 636 |
+
f"prevent this. Got grid_thw={image_grid_thw.tolist()}"
|
| 637 |
+
)
|
| 638 |
+
|
| 639 |
+
T_eff = int(image_grid_thw[:, 0].sum().item()) # sum of per-frame t (each is 1)
|
| 640 |
+
H_p = int(image_grid_thw[0, 1].item())
|
| 641 |
+
W_p = int(image_grid_thw[0, 2].item())
|
| 642 |
+
video_grid_thw = torch.tensor(
|
| 643 |
+
[[T_eff, H_p, W_p]], dtype=image_grid_thw.dtype
|
| 644 |
+
)
|
| 645 |
+
pixel_values_videos = pixel_values # already [T_eff*H_p*W_p, C, P, P]
|
| 646 |
+
|
| 647 |
+
# 3) patch_positions in block layout (over the merged video grid).
|
| 648 |
+
# Use REAL frame_indices for the t-axis (training convention).
|
| 649 |
+
patch_positions = build_patch_positions(
|
| 650 |
+
video_grid_thw,
|
| 651 |
+
spatial_merge_size=self.spatial_merge_size,
|
| 652 |
+
frame_indices=[frame_indices_t],
|
| 653 |
+
)
|
| 654 |
+
|
| 655 |
+
per_video_pixel_values.append(pixel_values_videos)
|
| 656 |
+
per_video_grid_thw.append(video_grid_thw)
|
| 657 |
+
per_video_patch_positions.append(patch_positions)
|
| 658 |
+
frame_timestamps_all.append(seconds_seq)
|
| 659 |
+
|
| 660 |
+
out_pixel_values = torch.cat(per_video_pixel_values, dim=0)
|
| 661 |
+
out_grid_thw = torch.cat(per_video_grid_thw, dim=0)
|
| 662 |
+
out_patch_positions = torch.cat(per_video_patch_positions, dim=0)
|
| 663 |
+
|
| 664 |
+
try:
|
| 665 |
+
from transformers.feature_extraction_utils import BatchFeature
|
| 666 |
+
|
| 667 |
+
return BatchFeature(
|
| 668 |
+
data={
|
| 669 |
+
"pixel_values_videos": out_pixel_values,
|
| 670 |
+
"video_grid_thw": out_grid_thw,
|
| 671 |
+
"patch_positions": out_patch_positions,
|
| 672 |
+
"frame_timestamps": frame_timestamps_all,
|
| 673 |
+
}
|
| 674 |
+
)
|
| 675 |
+
except Exception:
|
| 676 |
+
return {
|
| 677 |
+
"pixel_values_videos": out_pixel_values,
|
| 678 |
+
"video_grid_thw": out_grid_thw,
|
| 679 |
+
"patch_positions": out_patch_positions,
|
| 680 |
+
"frame_timestamps": frame_timestamps_all,
|
| 681 |
+
}
|
| 682 |
+
|
| 683 |
+
|
| 684 |
+
__all__ = [
|
| 685 |
+
"format_timestamp",
|
| 686 |
+
"time_str_to_seconds",
|
| 687 |
+
"choose_target_frames",
|
| 688 |
+
"select_frame_indices",
|
| 689 |
+
"smart_resize",
|
| 690 |
+
"extract_video_frames",
|
| 691 |
+
"extract_video_frames_to_pil",
|
| 692 |
+
"build_patch_positions",
|
| 693 |
+
"LlavaOnevision2VideoProcessor",
|
| 694 |
+
]
|
vocab.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|