NeoAnd/Dolphin
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

update doc

Browse Source
This commit is contained in:
yingdong.han 2025-06-30 19:47:10 +08:00
parent f4a6c495a6
commit 0705bc12ce
2 changed files with 6 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
deployment/tensorrt_llm/ReadMe.md
View File

@ -5,7 +5,9 @@
## ✅ Introduction ## ✅ Introduction
The Dolphin model employs a **Swin Encoder + MBart Decoder** architecture. In the HuggingFace Transformers [Config](https://huggingface.co/ByteDance/Dolphin/blob/main/config.json), The Dolphin model employs a **Swin Encoder + MBart Decoder** architecture. In the HuggingFace Transformers [Config](https://huggingface.co/ByteDance/Dolphin/blob/main/config.json),
its architectures field is specified as "VisionEncoderDecoderModel". Dolphin, Nougat, and Donut share the same model architecture. TensorRT-LLM has already supported the Nougat model. its architectures field is specified as "VisionEncoderDecoderModel". Dolphin, Nougat, and Donut share the same model architecture. TensorRT-LLM has already supported the Nougat model.
Following Nougat's conversion script, we have successfully implemented Dolphin on TensorRT-LLM. Note: input_ids MUST be of int32 type, otherwise TensorRT-LLM will produce incorrect results. Following Nougat's conversion script, we have successfully implemented Dolphin on TensorRT-LLM.
**Note:** prompt_ids MUST be of **int32** type, otherwise TensorRT-LLM will produce incorrect results.
## 🛠️ Installation ## 🛠️ Installation
> We only test TensorRT-LLM 0.18.1 on Linux. > We only test TensorRT-LLM 0.18.1 on Linux.

5
deployment/tensorrt_llm/dolphin_runner.py
View File

@ -115,8 +115,9 @@ class DolphinRunner(MultimodalModelRunner):
prompts = [f"<s>{text.strip()} <Answer/>" for text in input_texts] prompts = [f"<s>{text.strip()} <Answer/>" for text in input_texts]
images = self.processor(input_images, return_tensors="pt")['pixel_values'].to("cuda") images = self.processor(input_images, return_tensors="pt")['pixel_values'].to("cuda")
prompt_ids = self.tokenizer(prompts, add_special_tokens=False, return_tensors="pt").input_ids.to("cuda") prompt_ids = self.tokenizer(prompts, add_special_tokens=False, return_tensors="pt").input_ids.to("cuda")
prompt_ids = prompt_ids.to(
torch.int32) # Important! If the type of prompt_ids is not int32, the output will be wrong. # 🚨🚨🚨 Important! If the type of prompt_ids is not int32, the output will be wrong. 🚨🚨🚨
prompt_ids = prompt_ids.to(torch.int32)
logger.info("---------------------------------------------------------") logger.info("---------------------------------------------------------")
logger.info(f"images size: {images.size()}") logger.info(f"images size: {images.size()}")