[Bug] Qwen3-235B-A22B (BF16 & FP8) inference fails — streaming loader cannot handle fused MoE expert tensors (Qwen3MoeExperts / FP8Experts)

[itechnixindia]

airllm_github_issue.md

AirLLM v3.0.1 fails at inference time with all Qwen3-235B-A22B variants (BF16, FP8, Instruct-2507), despite the README listing Qwen3-235B as supported ("Run Qwen3-235B on ~3GB"). The root cause is an architectural mismatch between the per-expert tensor layout in the checkpoint files and the fused 3D expert tensor layout used by recent transformers MoE module classes (Qwen3MoeExperts, FP8Experts). AirLLM's streaming loader (set_module_tensor_to_device path) has no conversion logic for this, while transformers.from_pretrained() performs this fusion internally.