Soroosh-Sa/qwen3-coder-480b-folded-nvfp4-inputscale

Qwen3-Coder-480B-A35B Folded NVFP4 Checkpoint

This checkpoint was created from folded BF16 weights using a streaming tensor-wise NVFP4 converter.

Pipeline:

1. RMSNorm gamma folded into q/k/v and MLP gate/up-related weights.
2. Tensor-wise NVFP4 quantization using the official NVIDIA NVFP4 checkpoint as a layout template.
3. Fused QKV/gate-up weight_scale_2 groups normalized for TensorRT-LLM loading.
4. input_scale adjusted using original RMSNorm gamma statistics.

Validated:

• Safetensors layout validation passed.
• TensorRT-LLM PyTorch backend loads.
• Smoke generation works on 8x GPUs.

Recommended runtime:

• TensorRT-LLM 1.3.0rc16
• transformers 5.5.4
• TP_SIZE=8