Skip to content

[Model Request]: Enable Qwen4Exp GGUF payload import and execution #699

Description

@justinchuby

Model name

Qwen4Exp / Qwen3.8-Flash-Next

Hugging Face model or repository link

https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF

Evidence revision: d3bc75ee6ccef3efc1e228ec00a6cc2cdb1e2249

Primary task

Text generation

Use case

Complete Mobius support for executing the released Qwen4Exp GGUF checkpoint. Mobius already supports the exact qwen4exp header, configuration, split-shard closure, 1,224-tensor shape/qtype contract, tensor-name mapping, and text graph. Payload conversion currently fails closed because there is no faithful executable import route.

The implementation should add one of these truthful bounded-memory routes:

  • Native quantized execution for the IQ4_NL per-layer token embedding and rank-3 routed expert banks using mixed IQ1_S gate/up and IQ4_NL down tensors; or
  • Streaming/lazy dequantization and serialization that never materializes the approximately 191 GiB dense payload in memory.

Do not reinterpret these tensors as ordinary affine MatMulNBits; that would change their semantics.

Completion requires immutable-artifact weight parity, bounded-memory conversion evidence, ORT execution, cache/decode validation, and multi-token generation parity. Runtime/export metadata must remain fail closed until the required bindings and execution semantics are representable.

Execution providers

  • CPU
  • CUDA
  • QNN
  • OpenVINO
  • MIGraphX
  • MLX (macOS)
  • Other

Target hardware

The conversion path must have bounded host memory and should support machines that cannot hold the approximately 191 GiB dense expansion. CUDA execution is the primary performance target; CPU is required as a correctness reference where practical.

Quantization requirements

Preserve the released GGUF semantics: IQ4_NL per-layer token embedding, rank-3 routed experts with IQ1_S gate/up tensors, and IQ4_NL down tensors. Any dequantized route must stream or serialize incrementally rather than materializing the full dense model.

Contribution

  • I can help test this model.
  • I can help implement this model.
  • I have checked that a duplicate request does not already exist.

Additional context

The current design and fail-closed boundary are documented in docs/design/qwen4-exp-text-core.md. PR #698 only makes preflight/download diagnostics truthful; it does not implement GGUF payload execution.

Known downstream packaging gap: released ONNX Runtime GenAI metadata cannot yet represent ple_input_ids, four-axis position state, and heterogeneous per-layer PLE/QSA cache membership. This must not be reported as runtime-supported until the schema and bindings are faithful.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    model-requestRequests support for a new model

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions