Your current environment

• vLLM 0.17.1 (also checked current main @ ba2f0acc, same code)
• transformers 4.57.6
• model: google/gemma-3n-E4B-it
• GPU: NVIDIA A10G (24GB)
• API: /v1/chat/completions with input_audio items

Describe the bug

When two concurrent /v1/chat/completions requests with audio of different durations get batched together, EngineCore crashes:

ValueError: input_features_padded contains inconsistent shapes: torch.Size([496, 128]) (index 0) vs torch.Size([449, 128]) (index 1)

The first dimension (496 vs 449) is the audio sequence length, which naturally varies with duration. Each request has a single input_audio item — these are separate concurrent requests, not one request with multiple audios.

This kills EngineCore and every in-flight request gets a 500.

Root cause

This is the same bug that #31219 reported for Qwen2Audio/AudioFlamingo3/MiniCPM-V, fixed in #31223 — but Gemma3n wasn't included in that fix.

Gemma3nAudioInputs in gemma3n_mm.py declares its shapes without dynamic_dims:

class Gemma3nAudioInputs(TensorSchema):
    input_features_padded: Annotated[torch.Tensor, TensorShape("bn", "s", "f")]
    input_features_mask: Annotated[torch.Tensor, TensorShape("bn", "s")]

The "s" dimension varies with audio length, so TensorSchema.validate() rejects batches where the two items have different sequence lengths.

The padding logic itself works fine (#24052 added it back in Sept 2025). The validator just doesn't know that "s" is allowed to vary across batch items.

Stack trace

File "vllm/utils/tensor_schema.py", in _validate_field
    ValueError: input_features_padded contains inconsistent shapes:
    torch.Size([496, 128]) (index 0) vs torch.Size([449, 128]) (index 1)

Suggested fix

Same pattern as #31223 — add dynamic_dims={"s"}:

input_features_padded: Annotated[torch.Tensor, TensorShape("bn", "s", "f", dynamic_dims={"s"})]
input_features_mask: Annotated[torch.Tensor, TensorShape("bn", "s", dynamic_dims={"s"})]

I can put up a PR for this if it makes sense.