Collecting environment information...
PyTorch version              : 2.10.0+cu128
Is debug build               : False
CUDA used to build PyTorch   : 12.8
ROCM used to build PyTorch   : N/A
OS                           : Ubuntu 24.04.1 LTS (x86_64)
GCC version                  : (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Python version               : 3.12.3 (64-bit runtime)
Is CUDA available            : True
CUDA runtime version         : 12.8.93
GPU models and configuration :
GPU 0-7: NVIDIA H100 80GB HBM3
Nvidia driver version        : 570.148.08
cuDNN version                : 9.20.0
vLLM Version                 : 0.19.0
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled

[pip3] transformers==5.5.0
[pip3] torch==2.10.0+cu128
[pip3] flash-attn==2.8.3+cu128torch2.10
[pip3] transformer-engine-torch==2.12.0

---

AssertionError: Failed to apply prompt replacement for mm_items['vision_chunk'][0]

---

media_placeholder_token_id: int = 163605

---

def _get_prompt_updates(self, ...):
    hf_config = self.info.get_hf_config()
    media_token_id = hf_config.media_placeholder_token_id  # 163605 = [UNK]
    ...
    return [
        PromptReplacement(
            modality="vision_chunk",
            target=[media_token_id],  # looking for 163605, never found
            replacement=get_replacement,
        ),
    ]

---

from transformers import AutoTokenizer, AutoConfig

tok = AutoTokenizer.from_pretrained("moonshotai/Kimi-K2.5", trust_remote_code=True)
cfg = AutoConfig.from_pretrained("moonshotai/Kimi-K2.5", trust_remote_code=True)

# config.json says 163605, but the tokenizer says 163602
print(f"config.media_placeholder_token_id = {cfg.media_placeholder_token_id}")  # 163605
print(f"tokenizer('<|media_pad|>') = {tok.convert_tokens_to_ids('<|media_pad|>')}")  # 163602
print(f"token at 163605 = {tok.convert_ids_to_tokens(163605)}")  # [UNK]

---

config.media_placeholder_token_id = 163605
tokenizer('<|media_pad|>') = 163602
token at 163605 = [UNK]

---

from transformers import AutoTokenizer, AutoConfig

tok = AutoTokenizer.from_pretrained("moonshotai/Kimi-K2.5", trust_remote_code=True)
cfg = AutoConfig.from_pretrained("moonshotai/Kimi-K2.5", trust_remote_code=True)

# Tokenize a VL prompt using the chat template
messages = [{"role": "user", "content": [
    {"type": "image_url", "image_url": {"url": "placeholder"}},
    {"type": "text", "text": "Describe this image"},
]}]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
input_ids = tok.encode(text, add_special_tokens=False)

config_target = cfg.media_placeholder_token_id  # 163605
actual_media_id = tok.convert_tokens_to_ids("<|media_pad|>")  # 163602

print(f"Config target {config_target} ([UNK]) in input_ids: {config_target in input_ids}")  # False
print(f"Actual <|media_pad|> {actual_media_id} in input_ids: {actual_media_id in input_ids}")  # True

---

Config target 163605 ([UNK]) in input_ids: False
Actual <|media_pad|> 163602 in input_ids: True

---

from vllm import LLM, SamplingParams

llm = LLM(
    model="moonshotai/Kimi-K2.5",
    trust_remote_code=True,
    tensor_parallel_size=8,
)

sampling_params = SamplingParams(temperature=0.7, max_tokens=256)

outputs = llm.generate(
    {
        "prompt": "<|im_user|>user<|im_middle|><|media_pad|>Describe this image<|im_end|><|im_assistant|>assistant<|im_middle|>",
        "multi_modal_data": {
            "image": "https://upload.wikimedia.org/wikipedia/commons/thumb/4/47/PNG_transparency_demonstration_1.png/300px-PNG_transparency_demonstration_1.png",
        },
    },
    sampling_params,
)

---

AssertionError: Failed to apply prompt replacement for mm_items['vision_chunk'][0]

---

def _get_prompt_updates(self, ...):
    media_token_id = self.info.media_token_id  # resolved from tokenizer
    ...

---

def __init__(self, ctx):
    ...
    tokenizer = self.get_tokenizer()
    correct_id = tokenizer.convert_tokens_to_ids("<|media_pad|>")
    if isinstance(correct_id, int) and correct_id != self.media_token_id:
        self.media_token_id = correct_id
        self.media_token = tokenizer.decode(correct_id)
        self.hf_processor.media_token_id = correct_id
        self.hf_config.media_placeholder_token_id = correct_id