v0.13.1 regression: Intel AutoRound (auto-round/INCConfig) checkpoints rejected by Qwen4Exp PLE embedding quant whitelist

[xiongyang]

Environment

• Image: lazymio/vllm-backport:v0.13.1-sm80 (A100 80G x2, TP2 + EP)
• Model: Intel/Qwen3.8-Flash-Next-W4A16-AutoRound (config quant_method: auto-round, no ignore entries for PLE)
• Worked on v0.12.0-sm80 with the same launch flags; fails on v0.13.1-sm80.

Error

(Worker_TP0_EP0) NotImplementedError: Qwen4Exp PLE embedding does not support quantization config INCConfig

Root cause (code diff between v0.12.0 and v0.13.1)

• v0.12.0 vllm/models/qwen4_exp/nvidia/ple_layer.py::_get_ple_embedding_quant_method:

  if not isinstance(quant_config, Fp8Config):
      return None   # -> falls back to unquantized PLE loading

  Unknown quant config types (INCConfig for auto-round checkpoints) are treated as unquantized — and this works, because AutoRound keeps the PLE embedding weights themselves in BF16 (only quant_method is declared globally with no per-layer ignore list).

• v0.13.1 vllm/models/qwen4_exp/nvidia/ngram_embedding.py::Qwen4ExpPLEEmbeddingMethod.from_quant_config:
  rewritten as an explicit whitelist (None / ModelOpt / CompressedTensors / Fp8). Anything else now hits:

  if not isinstance(quant_config, Fp8Config):
      raise NotImplementedError(...)

Suggested fix

Add a permissive branch for INCConfig (or any unhandled config whose PLE weights are BF16 in the checkpoint), e.g. treat INCConfig like the unquantized path — matching v0.12.0 behavior:

if isinstance(quant_config, INCConfig):
    return Qwen4ExpPLEUnquantizedEmbeddingMethod()

Alternatively, a friendlier error message could suggest adding PLE layers to ignore as a workaround (Intel ships the checkpoint, so users cannot edit it easily).

Happy to test a patched image if needed. Thanks for the great fork!