Skip to content

Batches: accept the models served through Amazon Bedrock Mantle #148

Description

@JGoutin

/v1/batches refuses any model served through Amazon Bedrock Mantle: _resolve_model
forces the resolution to None for a Mantle model, and the caller gets a 400 saying the
model is not available for batched requests.

That was the safe assumption when the Batch API landed, but it is wrong. AWS's own
supported-models list for batch inference includes DeepSeek, Qwen 3, gpt-oss, Gemma 3,
MiniMax, Moonshot, NVIDIA Nemotron and Z.AI — which are precisely the families this
gateway routes through Mantle. So the largest gap between what the gateway allows in a
batch and what the backend allows is one the gateway imposes on itself.

Establish empirically, per family, whether a CreateModelInvocationJob naming the
Mantle-served model id is accepted and returns usable output, then widen the resolution
to whatever actually works. Where a family is genuinely unbatchable, the backend's own
rejection already produces a clean 400, so the client-side refusal earns nothing.

Note that batch requests are translated to the Converse shape before submission, so part
of this is establishing what shape a Mantle-served batch job expects.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    Status
    Backlog

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions