vllm-project/vllm · error · Error

unsupported multimodal content: {0}

Error message

unsupported multimodal content: {0}

What it means

The nixl EPLB backend registers GPU memory with a NIXL agent for one-sided RDMA-style transfers, which requires a CUDA-like platform and expert weights resident on the device. Either the platform is not CUDA-like or the expert weight tensors are on CPU.

Source

Thrown at rust/src/chat/src/error.rs:22

use thiserror::Error;
use thiserror_ext::{AsReport as _, Macro};

type BoxedError = Box<dyn std::error::Error + Send + Sync>;

#[derive(Debug, Error, Macro)]
#[thiserror_ext(macro(path = "crate::error"))]
pub enum Error {
    #[error("chat request must contain at least one message")]
    EmptyMessages,
    #[error("cannot continue the final message when the last message is not from the assistant")]
    ContinueFinalAssistantWithoutFinalAssistant,
    #[error("chat template is required but none was configured")]
    MissingChatTemplate,
    #[error("chat template error: {0}")]
    ChatTemplate(String),
    #[error("multimodal input is not supported by this chat renderer")]
    UnsupportedMultimodalRenderer,
    #[error("unsupported multimodal content: {0}")]
    UnsupportedMultimodalContent(&'static str),
    #[error("`{modality}` input is not supported by this model")]
    UnsupportedModality { modality: String },
    #[error("At most {limit} {modality}(s) may be provided in one prompt.")]
    MmLimitExceeded { modality: String, limit: usize },
    #[error("multimodal preprocessing error: {0}")]
    Multimodal(#[message] String),
    #[error("{kind} parsing is not available for model `{model_id}`")]
    ParserUnavailableForModel {
        kind: &'static str,
        model_id: String,
    },
    #[error("{kind} parsing is disabled by frontend configuration")]
    ParserDisabled { kind: &'static str },
    #[error(
        "{kind} parser `{name}` is not registered{}",
        available_parser_hint(.available_names)
    )]

View on GitHub (pinned to c794754062)

Solutions

1. Ensure expert weights are on the GPU before the EPLB communicator is created
2. On CPU or unsupported platforms, use 'torch_gloo' (CPU) or 'torch_nccl' instead
3. On ROCM/other accelerators, confirm the platform reports is_cuda_alike() or wait for platform support

Example fix

# before
backend = "nixl"  # expert weights on CPU

# after
weights = [w.cuda() for w in weights]
backend = "nixl"

from vllm.platforms import current_platform
device_ok = current_platform.is_cuda_alike() and expert_weights[0][0].device.type != 'cpu'
backend = 'nixl' if device_ok else 'torch_gloo'

When it happens

Trigger: Requesting backend='nixl' on non-CUDA platforms (e.g. ROCM without cuda-alike platform detection, CPU inference) or when expert weights are still on CPU (offloading or pre-H2D initialization).

Common situations: CPU-offloaded MoE deployments wanting NIXL; running on accelerator platforms not covered by current_platform.is_cuda_alike(); creating the communicator before weights are moved to GPU.

Related errors

• Configuration error: {0}
• Endpoint not ready after {0}s: {1}
• Backend error: {0}
• IO error: {0}
• Image generation should not fail

AI-assisted analysis of vllm-project/vllm@c794754062 (2026-08-14). Data as JSON: /api/errors/abd8998d51e3e887. Report an issue: GitHub.