operators/idefics2_op_demo.py

from dora import DoraStatus
import pyarrow as pa
from transformers import AutoProcessor, AutoModelForVision2Seq, AwqConfig
import torch
import gc

CAMERA_WIDTH = 1280
CAMERA_HEIGHT = 720
PROCESSOR = AutoProcessor.from_pretrained("HuggingFaceM4/idefics2-tfrm-compatible")
BAD_WORDS_IDS = PROCESSOR.tokenizer(
    ["<image>", "<fake_token_around_image>"], add_special_tokens=False
).input_ids
EOS_WORDS_IDS = PROCESSOR.tokenizer(
    "<end_of_utterance>", add_special_tokens=False
).input_ids + [PROCESSOR.tokenizer.eos_token_id]
model = AutoModelForVision2Seq.from_pretrained(
    "HuggingFaceM4/idefics2-tfrm-compatible-AWQ",
    quantization_config=AwqConfig(
        bits=4,
        fuse_max_seq_len=4096,
        modules_to_fuse={
            "attention": ["q_proj", "k_proj", "v_proj", "o_proj"],
            "mlp": ["gate_proj", "up_proj", "down_proj"],
            "layernorm": ["input_layernorm", "post_attention_layernorm", "norm"],
            "use_alibi": False,
            "num_attention_heads": 32,
            "num_key_value_heads": 8,
            "hidden_size": 4096,
        },
    ),
    trust_remote_code=True,
).to("cuda")


def reset_awq_cache(model):
    """
    Simple method to reset the AWQ fused modules cache
    """
    from awq.modules.fused.attn import QuantAttentionFused

    for name, module in model.named_modules():
        if isinstance(module, QuantAttentionFused):
            module.start_pos = 0


def ask_vlm(image, instruction):
    global model
    prompts = [
        "User:",
        image,
        f"{instruction}.<end_of_utterance>\n",
        "Assistant:",
    ]
    inputs = {k: torch.tensor(v).to("cuda") for k, v in PROCESSOR(prompts).items()}

    generated_ids = model.generate(
        **inputs,
        bad_words_ids=BAD_WORDS_IDS,
        max_new_tokens=25,
        repetition_penalty=1.2,
    )
    generated_texts = PROCESSOR.batch_decode(generated_ids, skip_special_tokens=True)
    reset_awq_cache(model)

    gc.collect()
    torch.cuda.empty_cache()
    return generated_texts[0].split("\nAssistant: ")[1]


import time


class Operator:
    def __init__(self):
        self.image = None
        self.text = None

    def on_event(
        self,
        dora_event,
        send_output,
    ) -> DoraStatus:
        if dora_event["type"] == "INPUT":
            if dora_event["id"] == "image":
                self.image = (
                    dora_event["value"]
                    .to_numpy()
                    .reshape((CAMERA_HEIGHT, CAMERA_WIDTH, 3))
                )
            elif dora_event["id"] == "text":
                self.text = dora_event["value"][0].as_py()
                output = ask_vlm(self.image, self.text).lower()
                send_output(
                    "speak",
                    pa.array([output]),
                )
                if "yes" in output:
                    send_output(
                        "control",
                        pa.array([0.0, 0.0, 0.0, 0.0, 0.0, 50.0, 0.0]),
                    )
                    time.sleep(2)
                    send_output(
                        "control",
                        pa.array([0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0]),
                    )
                elif "no" in output:
                    send_output(
                        "control",
                        pa.array([0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 50.0]),
                    )
                    time.sleep(2)
                    send_output(
                        "control",
                        pa.array([0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0]),
                    )

        return DoraStatus.CONTINUE