[Bug]: unexpected bits config for gguf #1585

Problem Description

embedding and ffn_up bits are different, I don't understand, especially for ffn_up

python3 -m auto_round Qwen/Qwen3-1.7B --scheme gguf:q2_k_s --layer_config "{embedding:{bits:4},lm_head:{bits:4}}" --iters 1 --nsamples 1
2026-03-20 18:13:04 INFO convert.py L626: token_embd.weight, torch.bfloat16 --> Q2_K, shape = {2048, 151936}
2026-03-20 18:13:04 INFO convert.py L626: blk.0.attn_q.weight, torch.bfloat16 --> Q2_K, shape = {2048, 2048}
2026-03-20 18:13:04 INFO convert.py L626: blk.0.attn_k.weight, torch.bfloat16 --> Q2_K, shape = {2048, 1024}
2026-03-20 18:13:04 INFO convert.py L626: blk.0.attn_v.weight, torch.bfloat16 --> Q2_K, shape = {2048, 1024}
2026-03-20 18:13:05 INFO convert.py L626: blk.0.attn_output.weight, torch.bfloat16 --> Q2_K, shape = {2048, 2048}
2026-03-20 18:13:05 INFO convert.py L626: blk.0.attn_q_norm.weight, torch.bfloat16 --> F32, shape = {128}
2026-03-20 18:13:05 INFO convert.py L626: blk.0.attn_k_norm.weight, torch.bfloat16 --> F32, shape = {128}
2026-03-20 18:13:05 INFO convert.py L626: blk.0.ffn_gate.weight, torch.bfloat16 --> Q2_K, shape = {2048, 6144}
2026-03-20 18:13:05 INFO convert.py L626: blk.0.ffn_up.weight, torch.bfloat16 --> Q2_K, shape = {2048, 6144}
2026-03-20 18:13:06 INFO convert.py L626: blk.0.ffn_down.weight, torch.bfloat16 --> Q2_K, shape = {6144, 2048}

python3 -m auto_round Qwen/Qwen3-1.7B --scheme gguf:q2_k_s --layer_config "{embed:{bits:4},lm_head:{bits:4}}" --iters 1 --nsamples 1

2026-03-20 18:16:13 INFO convert.py L626: token_embd.weight, torch.bfloat16 --> Q4_K, shape = {2048, 151936}
026-03-20 18:16:13 INFO convert.py L626: blk.0.attn_q.weight, torch.bfloat16 --> Q2_K, shape = {2048, 2048}
2026-03-20 18:16:13 INFO convert.py L626: blk.0.attn_k.weight, torch.bfloat16 --> Q2_K, shape = {2048, 1024}
2026-03-20 18:16:14 INFO convert.py L626: blk.0.attn_v.weight, torch.bfloat16 --> Q2_K, shape = {2048, 1024}
2026-03-20 18:16:14 INFO convert.py L626: blk.0.attn_output.weight, torch.bfloat16 --> Q2_K, shape = {2048, 2048}
2026-03-20 18:16:14 INFO convert.py L626: blk.0.attn_q_norm.weight, torch.bfloat16 --> F32, shape = {128}
2026-03-20 18:16:14 INFO convert.py L626: blk.0.attn_k_norm.weight, torch.bfloat16 --> F32, shape = {128}
2026-03-20 18:16:14 INFO convert.py L626: blk.0.ffn_gate.weight, torch.bfloat16 --> Q2_K, shape = {2048, 6144}
2026-03-20 18:16:15 INFO convert.py L626: blk.0.ffn_up.weight, torch.bfloat16 --> Q2_K, shape = {2048, 6144}
2026-03-20 18:16:15 INFO convert.py L626: blk.0.ffn_down.weight, torch.bfloat16 --> Q4_K, shape = {6144, 2048}
2026-03-20 18:16:15 INFO convert.py L626: blk.0.attn_norm.weight, torch.bfloat16 --> F32, shape = {2048}
2026-03-20 18:16:15 INFO convert.py L626: blk.0.ffn_norm.weight, torch.bfloat16 --> F32, shape = {2048}

Reproduction Steps

~

Environment Information

~

Error Logs

Additional Context

No response