app.py

import gradio as gr
import os, gc
from datetime import datetime
from transformers import CLIPImageProcessor
from huggingface_hub import hf_hub_download
from typing import List, Dict
from dataclasses import dataclass
DEFAULT_IMAGE_TOKEN = "<image>"


ctx_limit = 3500
num_image_embeddings = 4096
title = "rwkv1b5-vitl336p14-577token_mix665k_rwkv"
vision_tower_name = 'openai/clip-vit-large-patch14-336'

os.environ["RWKV_JIT_ON"] = '1'
os.environ["RWKV_CUDA_ON"] = '0' # if '1' then use CUDA kernel for seq mode (much faster)

from rwkv.model import RWKV
model_path = hf_hub_download(repo_id="howard-hou/visualrwkv-5", filename=f"{title}.pth")
model = RWKV(model=model_path, strategy='cpu fp32')
from rwkv.utils import PIPELINE, PIPELINE_ARGS
pipeline = PIPELINE(model, "rwkv_vocab_v20230424")

##########################################################################
from modeling import VisualEncoder, EmbeddingMixer, VisualEncoderConfig
emb_mixer = EmbeddingMixer(model.w["emb.weight"], 
 num_image_embeddings=num_image_embeddings)
config = VisualEncoderConfig(n_embd=model.args.n_embd, 
 vision_tower_name=vision_tower_name, 
 grid_size=-1)
visual_encoder = VisualEncoder(config)
image_processor = CLIPImageProcessor.from_pretrained(vision_tower_name)
##########################################################################
def generate_prompt(instruction):
 instruction = instruction.strip().replace('\r\n','\n').replace('\n\n','\n')
 input = input.strip().replace('\r\n','\n').replace('\n\n','\n')
 return f"\n{instruction}\n\nAssistant:"

def generate(
 ctx,
 image_ids,
 token_count=200,
 temperature=1.0,
 top_p=0.7,
 presencePenalty = 0.1,
 countPenalty = 0.1,
):
 args = PIPELINE_ARGS(temperature = max(0.2, float(temperature)), top_p = float(top_p),
 alpha_frequency = countPenalty,
 alpha_presence = presencePenalty,
 token_ban = [], # ban the generation of some tokens
 token_stop = [0]) # stop generation whenever you see any token here
 ctx = ctx.strip()
 all_tokens = []
 out_last = 0
 out_str = ''
 occurrence = {}
 state = None
 for i in range(int(token_count)):
 if i == 0:
 input_ids = (image_ids + pipeline.encode(ctx))[-ctx_limit:]
 else:
 input_ids = [token]
 out, state = model.forward(input_ids, state)
 for n in occurrence:
 out[n] -= (args.alpha_presence + occurrence[n] * args.alpha_frequency)

 token = pipeline.sample_logits(out, temperature=args.temperature, top_p=args.top_p)
 if token in args.token_stop:
 break
 all_tokens += [token]
 for xxx in occurrence:
 occurrence[xxx] *= 0.996 
 if token not in occurrence:
 occurrence[token] = 1
 else:
 occurrence[token] += 1
 
 tmp = pipeline.decode(all_tokens[out_last:])
 if '\ufffd' not in tmp:
 out_str += tmp
 yield out_str.strip()
 out_last = i + 1

 del out
 del state
 gc.collect()
 yield out_str.strip()


##########################################################################
cur_dir = os.path.dirname(os.path.abspath(__file__))
examples = [
 [
 f"{cur_dir}/examples_extreme_ironing.jpg",
 "What is unusual about this image?",
 ],
 [
 f"{cur_dir}/examples_waterview.jpg",
 "What are the things I should be cautious about when I visit here?",
 ]
]
def test(image, question):
 image = image_processor(images=image.convert('RGB'), return_tensors='pt')['pixel_values']
 image_features = visual_encoder.encode_images(image.unsqueeze(0))
 emb_mixer.set_image_embeddings(image_features)
 model.w["emb.weight"] = emb_mixer.get_input_embeddings()
 image_ids = [i for i in range(emb_mixer.image_start_index, emb_mixer.image_start_index + len(image_features))]
 input_text = generate_prompt(question)
 for output in generate(input_text, image_ids):
 yield output

demo = gr.ChatInterface(fn=test, 
 inputs=[gr.Image(type='pil'), "text"], 
 outputs="text", 
 examples=examples, 
 title=title, 
 description="VisualRWKV-v5.0")

demo = demo.queue(concurrency_count=1, max_size=10)
demo.launch(share=False, enable_queue=True)