[Bug]: The Transformers implementation of My Model is not compatible with vLLM.

[SnowCharmQ]

Your current environment

The output of `python collect_env.py`

PyTorch version: 2.6.0+cu124
Is debug build: False
CUDA used to build PyTorch: 12.4
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.5 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 3.22.1
Libc version: glibc-2.35

Python version: 3.11.11 (main, Dec 11 2024, 16:28:39) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.15.0-136-generic-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: 
GPU 0: NVIDIA H100 NVL
  MIG 3g.47gb     Device  0:

Nvidia driver version: 555.42.02
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.1.1
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.1.1
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.1.1
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.1.1
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.1.1
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.1.1
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.1.1
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.1.1
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                         x86_64
CPU op-mode(s):                       32-bit, 64-bit
Address sizes:                        52 bits physical, 57 bits virtual
Byte Order:                           Little Endian
CPU(s):                               128
On-line CPU(s) list:                  0-127
Vendor ID:                            AuthenticAMD
Model name:                           AMD EPYC 9334 32-Core Processor
CPU family:                           25
Model:                                17
Thread(s) per core:                   2
Core(s) per socket:                   32
Socket(s):                            2
Stepping:                             1
Frequency boost:                      enabled
CPU max MHz:                          2700.0000
CPU min MHz:                          1500.0000
BogoMIPS:                             5391.40
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp ibrs_enhanced vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local avx512_bf16 clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid overflow_recov succor smca fsrm flush_l1d
Virtualization:                       AMD-V
L1d cache:                            2 MiB (64 instances)
L1i cache:                            2 MiB (64 instances)
L2 cache:                             64 MiB (64 instances)
L3 cache:                             256 MiB (8 instances)
NUMA node(s):                         2
NUMA node0 CPU(s):                    0-31,64-95
NUMA node1 CPU(s):                    32-63,96-127
Vulnerability Gather data sampling:   Not affected
Vulnerability Itlb multihit:          Not affected
Vulnerability L1tf:                   Not affected
Vulnerability Mds:                    Not affected
Vulnerability Meltdown:               Not affected
Vulnerability Mmio stale data:        Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed:               Not affected
Vulnerability Spec rstack overflow:   Mitigation; safe RET
Vulnerability Spec store bypass:      Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1:             Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:             Mitigation; Enhanced / Automatic IBRS; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

Versions of relevant libraries:
[pip3] numpy==2.2.4
[pip3] nvidia-cublas-cu12==12.4.5.8
[pip3] nvidia-cuda-cupti-cu12==12.4.127
[pip3] nvidia-cuda-nvrtc-cu12==12.4.127
[pip3] nvidia-cuda-runtime-cu12==12.4.127
[pip3] nvidia-cudnn-cu12==9.1.0.70
[pip3] nvidia-cufft-cu12==11.2.1.3
[pip3] nvidia-curand-cu12==10.3.5.147
[pip3] nvidia-cusolver-cu12==11.6.1.9
[pip3] nvidia-cusparse-cu12==12.3.1.170
[pip3] nvidia-cusparselt-cu12==0.6.2
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] pyzmq==26.4.0
[pip3] torch==2.6.0
[pip3] torchaudio==2.6.0
[pip3] torchvision==0.21.0
[pip3] transformers==4.51.3
[pip3] triton==3.2.0
[conda] numpy                     2.2.4                    pypi_0    pypi
[conda] nvidia-cublas-cu12        12.4.5.8                 pypi_0    pypi
[conda] nvidia-cuda-cupti-cu12    12.4.127                 pypi_0    pypi
[conda] nvidia-cuda-nvrtc-cu12    12.4.127                 pypi_0    pypi
[conda] nvidia-cuda-runtime-cu12  12.4.127                 pypi_0    pypi
[conda] nvidia-cudnn-cu12         9.1.0.70                 pypi_0    pypi
[conda] nvidia-cufft-cu12         11.2.1.3                 pypi_0    pypi
[conda] nvidia-curand-cu12        10.3.5.147               pypi_0    pypi
[conda] nvidia-cusolver-cu12      11.6.1.9                 pypi_0    pypi
[conda] nvidia-cusparse-cu12      12.3.1.170               pypi_0    pypi
[conda] nvidia-cusparselt-cu12    0.6.2                    pypi_0    pypi
[conda] nvidia-nccl-cu12          2.21.5                   pypi_0    pypi
[conda] nvidia-nvjitlink-cu12     12.4.127                 pypi_0    pypi
[conda] nvidia-nvtx-cu12          12.4.127                 pypi_0    pypi
[conda] pyzmq                     26.4.0                   pypi_0    pypi
[conda] torch                     2.6.0                    pypi_0    pypi
[conda] torchaudio                2.6.0                    pypi_0    pypi
[conda] torchvision               0.21.0                   pypi_0    pypi
[conda] transformers              4.51.3                   pypi_0    pypi
[conda] triton                    3.2.0                    pypi_0    pypi
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.8.4
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
        GPU0    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      32-39,96-103    1               N/A

Legend:

  X    = Self
  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
  PIX  = Connection traversing at most a single PCIe bridge
  NV#  = Connection traversing a bonded set of # NVLinks

CUDA_VISIBLE_DEVICES=0
CUDA_VISIBLE_DEVICES=0
NCCL_CUMEM_ENABLE=0
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

🐛 Describe the bug

Hi there,

I saw the post from X (https://x.com/vllm_project/status/1912958639633277218) that the vllm library currently supports self-implemented transformers models. This is an amazing progress and I think it is pretty important. However, when I try to use this function with the latest vllm library and refer to the huggingface/transformers#36934 to update my model accordingly, it seems that my model is still not compatible with vllm. My model is provided as follows:

import os
import re
import torch
import torch.nn as nn

from torch.nn import CrossEntropyLoss
from typing import Optional, List, Tuple, Union
from transformers import Qwen2ForCausalLM, PretrainedConfig
from transformers.modeling_outputs import CausalLMOutputWithPast


HIDDEN_SIZE = 1536
MULT_K = 4

class Qwen2ForCausalPersonalLM(Qwen2ForCausalLM):
    _supports_attention_backend = True

    def __init__(self, config, **kwargs):
        super().__init__(config, **kwargs)
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.align_mlp_his = nn.Sequential(nn.Linear(HIDDEN_SIZE, config.hidden_size * MULT_K), 
                                        nn.GELU(), 
                                        nn.Linear(config.hidden_size * MULT_K, config.hidden_size))
        self.align_mlp_his.to_empty(device=device)

    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        num_logits_to_keep: int = 0,
        his_emb: Optional[torch.FloatTensor] = None,
        **kwargs,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        inputs_embs = self.model.get_input_embeddings()(input_ids)

        his_emb = his_emb / (his_emb.norm(dim=-1, keepdim=True) + 1e-6)
        his_emb = self.align_mlp_his(his_emb)

        if his_emb is not None:
            his_emb_len = his_emb.shape[1]
            for bidx in range(inputs_embs.shape[0]):
                new_tokens = [f"[HIS_TOKEN_{i}]" for i in range(his_emb_len)]
                token_ids = self.llm_tokenizer.convert_tokens_to_ids(new_tokens)
                b_his_emb = his_emb[bidx]
                for i in range(his_emb_len):
                    pemb = b_his_emb[i, :]
                    inputs_embs[bidx][input_ids[bidx] == token_ids[i]] = pemb.to(inputs_embs.dtype)

        outputs = self.model(
            inputs_embeds=inputs_embs,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
            **kwargs,
        )

        hidden_states = outputs[0]
        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :]).float()

        loss = None
        if labels is not None:
            logits = logits.float()
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss_fct = CrossEntropyLoss()
            shift_logits = shift_logits.view(-1, self.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            shift_labels = shift_labels.to(shift_logits.device)
            loss = loss_fct(shift_logits, shift_labels)

        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )
    
    @classmethod
    def from_pretrained(
        cls,
        pretrained_model_name_or_path: Optional[Union[str, os.PathLike]],
        *model_args,
        config: Optional[Union[PretrainedConfig, str, os.PathLike]] = None,
        cache_dir: Optional[Union[str, os.PathLike]] = None,
        ignore_mismatched_sizes: bool = False,
        force_download: bool = False,
        local_files_only: bool = False,
        token: Optional[Union[str, bool]] = None,
        revision: str = "main",
        use_safetensors: bool = None,
        val: bool = False,
        **kwargs,
    ):
        model = super().from_pretrained(pretrained_model_name_or_path, 
                                        *model_args, 
                                        config=config, 
                                        cache_dir=cache_dir, 
                                        ignore_mismatched_sizes=ignore_mismatched_sizes, 
                                        force_download=force_download, 
                                        local_files_only=local_files_only, 
                                        token=token, revision=revision, 
                                        use_safetensors=use_safetensors)
        model.llm_tokenizer = kwargs.get("tokenizer", None)
        if not val:
            for layer in model.align_mlp_his:
                if isinstance(layer, nn.Linear):
                    torch.nn.init.xavier_uniform_(layer.weight)
                    if layer.bias is not None:
                        torch.nn.init.zeros_(layer.bias)
        return model

I really need your help, thanks!

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[SnowCharmQ]

Btw, the base model I used for this is Qwen2.5-3B-Instruct.

[hmellor]

Have you uploaded your model as a custom model on the Hugging Face Hub?

The integration uses the json files (like those for Qwen) in a model repo to load your custom model class.

Here is a guide for creating your custom model. Then you can follow vLLM's guide for making it compatible with vLLM.

[SnowCharmQ]

Have you uploaded your model as a custom model on the Hugging Face Hub?

The integration uses the json files (like those for Qwen) in a model repo to load your custom model class.

Here is a guide for creating your custom model. Then you can follow vLLM's guide for making it compatible with vLLM.

So is it necessary to upload the model to the HuggingFace hub? I have already used the save_pretrained method to save the trained model on my local machine with the config.json file as shown below. But I didn't try to upload it.

{
  "architectures": [
    "Qwen2ForCausalPersonalLM"
  ],
  "attention_dropout": 0.0,
  "bos_token_id": 151643,
  "eos_token_id": 151645,
  "hidden_act": "silu",
  "hidden_size": 2048,
  "initializer_range": 0.02,
  "intermediate_size": 11008,
  "max_position_embeddings": 32768,
  "max_window_layers": 70,
  "model_type": "qwen2",
  "num_attention_heads": 16,
  "num_hidden_layers": 36,
  "num_key_value_heads": 2,
  "rms_norm_eps": 1e-06,
  "rope_scaling": null,
  "rope_theta": 1000000.0,
  "sliding_window": 32768,
  "tie_word_embeddings": true,
  "torch_dtype": "bfloat16",
  "transformers_version": "4.51.0",
  "use_cache": true,
  "use_sliding_window": false,
  "vocab_size": 151673
}

[hmellor]

Ok if you have it in HF format locally that should still work.

I notice that you don't have automap in your config.json. This is how Transformers knows which classes to load for config, model and tokenizer. This is mentioned in the custom model guide I linked before.

You can also look at https://huggingface.co/ArthurZ/Ilama-3.2-1B/blob/main/config.json for an example of how this is done with local imports rather than needing a HF repo id.

[SnowCharmQ]

Ok if you have it in HF format locally that should still work.

I notice that you don't have automap in your config.json. This is how Transformers knows which classes to load for config, model and tokenizer. This is mentioned in the custom model guide I linked before.

You can also look at https://huggingface.co/ArthurZ/Ilama-3.2-1B/blob/main/config.json for an example of how this is done with local imports rather than needing a HF repo id.

Hi, I tried your suggestion and also followed the guide you provided, but it still got the same error and failed...
Currently, the dir structure of my custom model is:

.
└── qwen2_personal
    ├── __init__.py
    ├── configuration_qwen2_personal.py
    └── modeling_qwen2_personal.py

The __init__.py file:

from .configuration_qwen2_personal import Qwen2PersonalConfig
from .modeling_qwen2_personal import Qwen2ForCausalPersonalLM

__all__ = ["Qwen2PersonalConfig", "Qwen2ForCausalPersonalLM"]

The configuration_qwen2_personal.py file:

from transformers import Qwen2Config

class Qwen2PersonalConfig(Qwen2Config):
    model_type = "qwen2_personal"

    def __init__(
        self,
        embed_hidden_size=1536,
        mult_k=4,
        **kwargs
    ):
        super().__init__(**kwargs)
        self.embed_hidden_size = embed_hidden_size
        self.mult_k = mult_k

The modeling_qwen2_personal.py file:

import os
import torch
import torch.nn as nn

from torch.nn import CrossEntropyLoss
from typing import Optional, List, Tuple, Union
from transformers import Qwen2ForCausalLM
from transformers.modeling_outputs import CausalLMOutputWithPast

from .configuration_qwen2_personal import Qwen2PersonalConfig

class Qwen2ForCausalPersonalLM(Qwen2ForCausalLM):
    config_class = Qwen2PersonalConfig
    _supports_attention_backend = True
    base_model_prefix = "model"

    def __init__(self, config, **kwargs):
        super().__init__(config, **kwargs)
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.align_mlp_his = nn.Sequential(
            nn.Linear(config.embed_hidden_size, config.hidden_size * config.mult_k), 
            nn.GELU(), 
            nn.Linear(config.hidden_size * config.mult_k, config.hidden_size)
        )
        self.align_mlp_his.to_empty(device=device)
        self.llm_tokenizer = None

    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        num_logits_to_keep: int = 0,
        his_emb: Optional[torch.FloatTensor] = None,
        **kwargs,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        inputs_embs = self.model.get_input_embeddings()(input_ids)

        if his_emb is not None:
            his_emb = his_emb / (his_emb.norm(dim=-1, keepdim=True) + 1e-6)
            his_emb = self.align_mlp_his(his_emb)
            
            his_emb_len = his_emb.shape[1]
            for bidx in range(inputs_embs.shape[0]):
                new_tokens = [f"[HIS_TOKEN_{i}]" for i in range(his_emb_len)]
                token_ids = self.llm_tokenizer.convert_tokens_to_ids(new_tokens)
                b_his_emb = his_emb[bidx]
                for i in range(his_emb_len):
                    pemb = b_his_emb[i, :]
                    inputs_embs[bidx][input_ids[bidx] == token_ids[i]] = pemb.to(inputs_embs.dtype)

        outputs = self.model(
            inputs_embeds=inputs_embs,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
            **kwargs,
        )

        hidden_states = outputs[0]
        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :]).float()

        loss = None
        if labels is not None:
            logits = logits.float()
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss_fct = CrossEntropyLoss()
            shift_logits = shift_logits.view(-1, self.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            shift_labels = shift_labels.to(shift_logits.device)
            loss = loss_fct(shift_logits, shift_labels)

        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )
    
    @classmethod
    def from_pretrained(
        cls,
        pretrained_model_name_or_path: Optional[Union[str, os.PathLike]],
        *model_args,
        config=None,
        cache_dir=None,
        ignore_mismatched_sizes=False,
        force_download=False,
        local_files_only=False,
        token=None,
        revision="main",
        use_safetensors=None,
        tokenizer=None,
        **kwargs,
    ):
        model = super().from_pretrained(
            pretrained_model_name_or_path, 
            *model_args, 
            config=config, 
            cache_dir=cache_dir, 
            ignore_mismatched_sizes=ignore_mismatched_sizes, 
            force_download=force_download, 
            local_files_only=local_files_only, 
            token=token, 
            revision=revision, 
            use_safetensors=use_safetensors,
            **kwargs
        )
        
        model.llm_tokenizer = tokenizer
        return model

And in the root path, I implement a file with the following content:

from qwen2_personal import Qwen2ForCausalPersonalLM
from transformers import AutoTokenizer
import torch

from transformers import AutoConfig, AutoModelForCausalLM
from qwen2_personal.configuration_qwen2_personal import Qwen2PersonalConfig
from qwen2_personal.modeling_qwen2_personal import Qwen2ForCausalPersonalLM

AutoConfig.register("qwen2_personal", Qwen2PersonalConfig)
AutoModelForCausalLM.register(Qwen2PersonalConfig, Qwen2ForCausalPersonalLM)

Qwen2PersonalConfig.register_for_auto_class()
Qwen2ForCausalPersonalLM.register_for_auto_class("AutoModelForCausalLM")

llm_model_name = "Qwen/Qwen2.5-0.5B-Instruct"
llm_tokenizer = AutoTokenizer.from_pretrained(llm_model_name)
llm_tokenizer.padding_side = "left"
personal_model = Qwen2ForCausalPersonalLM.from_pretrained(llm_model_name,
                                                          device_map="cuda",
                                                          torch_dtype=torch.bfloat16,
                                                          tokenizer=llm_tokenizer)

config = personal_model.config
personal_model.save_pretrained("test_model/")
llm_tokenizer.save_pretrained("test_model/")
config.save_pretrained("test_model/")

if __name__ == "__main__":
    from vllm import LLM, SamplingParams

    sampling_params = SamplingParams(
        temperature=0.7,
        top_p=0.9,
        max_tokens=100,
        top_k=50,
    )

    llm = LLM(model="test_model", model_impl="transformers")

    prompts = [
        "Hello, how are you?",
        "Tell me a joke.",
    ]

    llm.generate(prompts, sampling_params)

I got the same error, and I tried to examine the config.json stored in the test_model dir, which looks good as follows:

{
  "architectures": [
    "Qwen2ForCausalPersonalLM"
  ],
  "attention_dropout": 0.0,
  "auto_map": {
    "AutoConfig": "configuration_qwen2_personal.Qwen2PersonalConfig",
    "AutoModelForCausalLM": "modeling_qwen2_personal.Qwen2ForCausalPersonalLM"
  },
  "bos_token_id": 151643,
  "embed_hidden_size": 1536,
  "eos_token_id": 151645,
  "hidden_act": "silu",
  "hidden_size": 2048,
  "initializer_range": 0.02,
  "intermediate_size": 11008,
  "max_position_embeddings": 32768,
  "max_window_layers": 70,
  "model_type": "qwen2_personal",
  "mult_k": 4,
  "num_attention_heads": 16,
  "num_hidden_layers": 36,
  "num_key_value_heads": 2,
  "rms_norm_eps": 1e-06,
  "rope_scaling": null,
  "rope_theta": 1000000.0,
  "sliding_window": 32768,
  "tie_word_embeddings": true,
  "torch_dtype": "bfloat16",
  "transformers_version": "4.51.3",
  "use_cache": true,
  "use_sliding_window": false,
  "vocab_size": 151936
}

I even tried to upload the model to the Huggingface hub, but it still failed. I have no idea how to handle it now...

[SnowCharmQ]

ERROR 04-18 22:14:42 [core.py:387]   File "/home/e/e1351134/miniconda3/envs/v0/lib/python3.11/site-packages/vllm/v1/engine/core.py", line 378, in run_engine_core
ERROR 04-18 22:14:42 [core.py:387]     engine_core = EngineCoreProc(*args, **kwargs)
ERROR 04-18 22:14:42 [core.py:387]                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 04-18 22:14:42 [core.py:387]   File "/home/e/e1351134/miniconda3/envs/v0/lib/python3.11/site-packages/vllm/v1/engine/core.py", line 320, in __init__
ERROR 04-18 22:14:42 [core.py:387]     super().__init__(vllm_config, executor_class, log_stats)ERROR 04-18 22:14:42 [core.py:387]   File "/home/e/e1351134/miniconda3/envs/v0/lib/python3.11/site-packages/vllm/v1/engine/core.py", line 67, in __init__
ERROR 04-18 22:14:42 [core.py:387]     self.model_executor = executor_class(vllm_config)
ERROR 04-18 22:14:42 [core.py:387]                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 04-18 22:14:42 [core.py:387]   File "/home/e/e1351134/miniconda3/envs/v0/lib/python3.11/site-packages/vllm/executor/executor_base.py", line 52, in __init__
ERROR 04-18 22:14:42 [core.py:387]     self._init_executor()
ERROR 04-18 22:14:42 [core.py:387]   File "/home/e/e1351134/miniconda3/envs/v0/lib/python3.11/site-packages/vllm/executor/uniproc_executor.py", line 47, in _init_executor
ERROR 04-18 22:14:42 [core.py:387]     self.collective_rpc("load_model")
ERROR 04-18 22:14:42 [core.py:387]   File "/home/e/e1351134/miniconda3/envs/v0/lib/python3.11/site-packages/vllm/executor/uniproc_executor.py", line 56, in collective_rpc
ERROR 04-18 22:14:42 [core.py:387]     answer = run_method(self.driver_worker, method, args, kwargs)
ERROR 04-18 22:14:42 [core.py:387]              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 04-18 22:14:42 [core.py:387]   File "/home/e/e1351134/miniconda3/envs/v0/lib/python3.11/site-packages/vllm/utils.py", line 2378, in run_method
ERROR 04-18 22:14:42 [core.py:387]     return func(*args, **kwargs)
ERROR 04-18 22:14:42 [core.py:387]            ^^^^^^^^^^^^^^^^^^^^^
ERROR 04-18 22:14:42 [core.py:387]   File "/home/e/e1351134/miniconda3/envs/v0/lib/python3.11/site-packages/vllm/v1/worker/gpu_worker.py", line 136, in load_model
ERROR 04-18 22:14:42 [core.py:387]     self.model_runner.load_model()
ERROR 04-18 22:14:42 [core.py:387]   File "/home/e/e1351134/miniconda3/envs/v0/lib/python3.11/site-packages/vllm/v1/worker/gpu_model_runner.py", line 1279, in load_model
ERROR 04-18 22:14:42 [core.py:387]     self.model = get_model(vllm_config=self.vllm_config)
ERROR 04-18 22:14:42 [core.py:387]                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 04-18 22:14:42 [core.py:387]   File "/home/e/e1351134/miniconda3/envs/v0/lib/python3.11/site-packages/vllm/model_executor/model_loader/__init__.py", line 14, in get_model
ERROR 04-18 22:14:42 [core.py:387]     return loader.load_model(vllm_config=vllm_config)
ERROR 04-18 22:14:42 [core.py:387]            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 04-18 22:14:42 [core.py:387]   File "/home/e/e1351134/miniconda3/envs/v0/lib/python3.11/site-packages/vllm/model_executor/model_loader/loader.py", line 452, in load_model
ERROR 04-18 22:14:42 [core.py:387]     model = _initialize_model(vllm_config=vllm_config)
ERROR 04-18 22:14:42 [core.py:387]             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 04-18 22:14:42 [core.py:387]   File "/home/e/e1351134/miniconda3/envs/v0/lib/python3.11/site-packages/vllm/model_executor/model_loader/loader.py", line 123, in _initialize_model
ERROR 04-18 22:14:42 [core.py:387]     model_class, _ = get_model_architecture(model_config)
ERROR 04-18 22:14:42 [core.py:387]                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 04-18 22:14:42 [core.py:387]   File "/home/e/e1351134/miniconda3/envs/v0/lib/python3.11/site-packages/vllm/model_executor/model_loader/utils.py", line 104, in get_model_architecture
ERROR 04-18 22:14:42 [core.py:387]     architectures = resolve_transformers_arch(model_config, architectures)
ERROR 04-18 22:14:42 [core.py:387]                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 04-18 22:14:42 [core.py:387]   File "/home/e/e1351134/miniconda3/envs/v0/lib/python3.11/site-packages/vllm/model_executor/model_loader/utils.py", line 66, in resolve_transformers_arch
ERROR 04-18 22:14:42 [core.py:387]     raise ValueError(
ERROR 04-18 22:14:42 [core.py:387] ValueError: The Transformers implementation of Qwen2ForCausalPersonalLM is not compatible with vLLM.
ERROR 04-18 22:14:42 [core.py:387] 
CRITICAL 04-18 22:14:42 [core_client.py:359] Got fatal signal from worker processes, shutting down. See stack trace above for root cause issue.
[1]    468135 killed     python test.py

The error information is here for your reference.

[hmellor]

Thank you for all the extra information!

I won't be able to dig into this until next week but I'm sure we'll be able to figure out what's going on.

Then we'll be able to fix the issue in vLLM (if that's the problem) or document the step that we're stuck at now (if that's the problem) so that future experimenters such as yourself have a better experience next time!

[SnowCharmQ]

Thank you for all the extra information!

I won't be able to dig into this until next week but I'm sure we'll be able to figure out what's going on.

Then we'll be able to fix the issue in vLLM (if that's the problem) or document the step that we're stuck at now (if that's the problem) so that future experimenters such as yourself have a better experience next time!

Thank you for your kind help! I'll also continue to keep track of this issue. If there's anything you need help with or any additional information you'd like me to provide, please feel free to reach out!

[ashutos-cf]

I am also getting the same error when using the vLLM (0.8.4) for inference using custom whisper-model after converting into HF format.

[hmellor]

@SnowCharmQ I have found the reason for the failing compatibility check. It's because your custom model has no custom base class (Qwen2PersonalModel).

Generally, ...ForCausalLM classes are meant to be a composition of a base class (...Model) with an LM head after it. Therefore, any customisation of a model should occur inside the base class. I've made #16988 to make this clearer in the error message and the docs.

From your modelling code it looks like the custom layer could be moved into the Qwen2PersonalModel class so I'd recommend that you try that.

However, I do see a couple of issues with your custom model that my suggestion will not solve:

• You are placing the tokenizer inside the model by adding a custom tokenizer kwarg to from_pretrained:
  • the Transformers backend does not use from_pretrained it uses from_config
  • the Transformers backend cannot pass arbitrary kwargs in model construction (how could it know what they are?)
• You are passing a custom kwarg his_emb in forward:
  • the Transformers backend cannot pass arbitrary kwargs in the forward pass (how could it know what they are?)

---

@ashutos-cf as you haven't shared your modelling code I can't be sure if your error is caused by the same issue. Could you try my suggestion above and, if it doesn't resolve your issue, could you open a new issue with a reproducer?

[eghouti]

I am facing the same error, and I did not understand what we concretly need to do with this base class @hmellor. Thank you in advance for your hep

[hmellor]

https://docs.vllm.ai/en/latest/models/supported_models.html#custom-models (and the section immediately after it called Writing custom models) details exactly what needs to be done to the base class to make it compatible!

Do let me know if anything can be improved about the documentation or the integration!