OGA export generating an invalid model when a lm_head unquantized model is exported

[uday610]

Describe the bug
OGA export generating an invalid model when a lm_head unquantized model is exported

To Reproduce
Steps to reproduce the behavior:

1. Please access the quantized model from here https://huggingface.co/amd/granite-4.0-1b-awq-grp128-exclude_lmhead_quark
2. Create OGA environment, I have tested from 0.12.1 to latest 0.13.2. I have been using transformers==4.57.6 and torch==2.12.0
3. Export using this command

python -m onnxruntime_genai.models.builder `
    -i "granite_awq_lmhead_exclude" `
    -o "granite_oga_int4_lmhead_exclude" `
    -p int4 `
    -e cpu `
    --extra_options exclude_embeds=true int4_nodes_to_exclude=/lm_head/MatMul

4. See the message of invalid model

2026-05-26 13:21:43,041 onnxruntime.quantization.matmul_nbits_quantizer [INFO] - skip to quantize /model/constant_nodes/FLOAT/0.125 ...
2026-05-26 13:21:43,041 onnxruntime.quantization.matmul_nbits_quantizer [INFO] - skip to quantize /lm_head/Mul ...
2026-05-26 13:21:43,057 onnx_ir.serde [WARNING] - Input 'lm_head.MatMul.weight_Q4G32' of node '/model/embed_tokens/Reshape' (::Reshape:) cannot be found in any scope. The model is invalid but we will still create a new input for the node (current depth: 1)
2026-05-26 13:21:43,057 onnx_ir.serde [WARNING] - Input 'lm_head.MatMul.weight_scale' of node '/model/embed_tokens/GatherBlockQuantized' (com.microsoft::GatherBlockQuantized:) cannot be found in any scope. The model is invalid but we will still create a new input for the node (current depth: 1)

[uday610]

Hi kunal-vaishnavi ,
I still see the issue. I have tried with latest main branch. You see "invalid model" message below. I cloned onnxruntime-genai repo today, build wheel (onnxruntime-genai-directml 0.15.0.dev0) and installed.

2026-06-18 12:46:04,922 onnxruntime.quantization.matmul_nbits_quantizer [INFO] - skip to quantize 
/model/layers.31/mlp/down_proj/MatMulNBits ...
2026-06-18 12:46:04,922 onnxruntime.quantization.matmul_nbits_quantizer [INFO] - skip to quantize 
/model/layers.32/final_norm_layernorm/SkipLayerNorm ...
2026-06-18 12:46:04,922 onnxruntime.quantization.matmul_nbits_quantizer [INFO] - exclude to quantize /lm_head/MatMul 
as specified by nodes_to_exclude...
2026-06-18 12:46:04,969 onnx_ir.serde [WARNING] - Input 'lm_head.MatMul.weight_Q4G32' of node 
'/model/embed_tokens/Reshape' (::Reshape:) cannot be found in any scope. The model is invalid but we will still create 
a new input for the node (current depth: 1)
2026-06-18 12:46:04,969 onnx_ir.serde [WARNING] - Input 'lm_head.MatMul.weight_scale' of node 
'/model/embed_tokens/GatherBlockQuantized' (com.microsoft::GatherBlockQuantized:) cannot be found in any scope. The 
model is invalid but we will still create a new input for the node (current depth: 1)

This time I tried with phi-4-mini-instruct model Quantized by Quark but lm_head is excluded. You can check with this model: https://huggingface.co/amd/phi-4-mini-instruct-gptq-grp128-lmheadex

So my OGA conversion command is

python -m onnxruntime_genai.models.builder `
    -i "granite_awq_lmhead_exclude" `
    -o "granite_oga_int4_lmhead_exclude" `
    -p int4 `
    -e dml `
    --extra_options int4_nodes_to_exclude=/lm_head/MatMul