https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507 imatrix

#1186
by matatonic - opened

https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507
imatrix quants please, in particular the iQ3_M and iQ3_S versions
thanks!

It's already in the queue, but it can take a while due to the size of the model (and potential issues in llama.cpp). We'll try top keep you posted.

Sorry for not working on this model quicker. We have currently quite a lot of massive models in the backlog using up all my storage. Now that your requested it I will try my best to get it done.

No worries, Thanks for all your hard work and compute!
You may want to keep the fires warm though, it looks like qwen3 coder is about to drop.

I just checked and the source model is completely BF16 so let's use this model as experiment to see what happens when we use BF16 for the source GGUF. The reason I choose this to test this is because it is large enough to require manual handling but with 470 GB it is hopefully small enough to fit on StormPeak for imatrix computation and doesn’t require RPC:

venv/bin/python convert_hf_to_gguf.py --outtype auto --outfile /mradermacher/tmp/quant/Qwen3-235B-A22B-Instruct-2507.gguf /bpool/Qwen3-235B-A22B-Instruct-2507

token_embd.weight,           BF16 --> BF16 {4096, 151936}
blk.0.attn_k_norm.weight,    BF16 --> F32 {128}
blk.0.attn_k.weight,         BF16 --> BF16 {4096, 512}
blk.0.attn_output.weight,    BF16 --> BF16 {8192, 4096}
blk.0.attn_q_norm.weight,    BF16 --> F32 {128}
blk.0.attn_q.weight,         BF16 --> BF16 {4096, 8192}
blk.0.attn_v.weight,         BF16 --> BF16 {4096, 512}
blk.0.attn_norm.weight,      BF16 --> F32 {4096}
blk.0.ffn_down_exps.weight,  BF16 --> BF16 {1536, 4096, 128}
blk.0.ffn_gate_exps.weight,  BF16 --> BF16 {4096, 1536, 128}
blk.0.ffn_up_exps.weight,    BF16 --> BF16 {4096, 1536, 128}
blk.0.ffn_gate_inp.weight,   BF16 --> F32 {4096, 128}
blk.0.ffn_norm.weight,       BF16 --> F32 {4096}
blk.1.attn_k_norm.weight,    BF16 --> F32 {128}
blk.1.attn_k.weight,         BF16 --> BF16 {4096, 512}
blk.1.attn_output.weight,    BF16 --> BF16 {8192, 4096}
blk.1.attn_q_norm.weight,    BF16 --> F32 {128}
blk.1.attn_q.weight,         BF16 --> BF16 {4096, 8192}
blk.1.attn_v.weight,         BF16 --> BF16 {4096, 512}
blk.1.attn_norm.weight,      BF16 --> F32 {4096}
blk.1.ffn_down_exps.weight,  BF16 --> BF16 {1536, 4096, 128}
blk.1.ffn_gate_exps.weight,  BF16 --> BF16 {4096, 1536, 128}
blk.1.ffn_up_exps.weight,    BF16 --> BF16 {4096, 1536, 128}
blk.1.ffn_gate_inp.weight,   BF16 --> F32 {4096, 128}
blk.1.ffn_norm.weight,       BF16 --> F32 {4096}
blk.2.attn_k_norm.weight,    BF16 --> F32 {128}
blk.2.attn_k.weight,         BF16 --> BF16 {4096, 512}
blk.2.attn_output.weight,    BF16 --> BF16 {8192, 4096}
blk.2.attn_q_norm.weight,    BF16 --> F32 {128}
blk.2.attn_q.weight,         BF16 --> BF16 {4096, 8192}
blk.2.attn_v.weight,         BF16 --> BF16 {4096, 512}
blk.2.attn_norm.weight,      BF16 --> F32 {4096}
blk.2.ffn_down_exps.weight,  BF16 --> BF16 {1536, 4096, 128}
blk.2.ffn_gate_exps.weight,  BF16 --> BF16 {4096, 1536, 128}
blk.2.ffn_up_exps.weight,    BF16 --> BF16 {4096, 1536, 128}
blk.2.ffn_gate_inp.weight,   BF16 --> F32 {4096, 128}
blk.2.ffn_norm.weight,       BF16 --> F32 {4096}
blk.3.attn_k_norm.weight,    BF16 --> F32 {128}
blk.3.attn_k.weight,         BF16 --> BF16 {4096, 512}
blk.3.attn_output.weight,    BF16 --> BF16 {8192, 4096}
blk.3.attn_q_norm.weight,    BF16 --> F32 {128}
blk.3.attn_q.weight,         BF16 --> BF16 {4096, 8192}
blk.3.attn_v.weight,         BF16 --> BF16 {4096, 512}
blk.3.attn_norm.weight,      BF16 --> F32 {4096}
blk.3.ffn_down_exps.weight,  BF16 --> BF16 {1536, 4096, 128}
blk.3.ffn_gate_exps.weight,  BF16 --> BF16 {4096, 1536, 128}
blk.3.ffn_up_exps.weight,    BF16 --> BF16 {4096, 1536, 128}
blk.3.ffn_gate_inp.weight,   BF16 --> F32 {4096, 128}
blk.3.ffn_norm.weight,       BF16 --> F32 {4096}
blk.4.attn_k_norm.weight,    BF16 --> F32 {128}
blk.4.attn_k.weight,         BF16 --> BF16 {4096, 512}
blk.4.attn_output.weight,    BF16 --> BF16 {8192, 4096}
blk.4.attn_q_norm.weight,    BF16 --> F32 {128}
blk.4.attn_q.weight,         BF16 --> BF16 {4096, 8192}
blk.4.attn_v.weight,         BF16 --> BF16 {4096, 512}
blk.4.attn_norm.weight,      BF16 --> F32 {4096}
blk.4.ffn_down_exps.weight,  BF16 --> BF16 {1536, 4096, 128}
blk.4.ffn_gate_exps.weight,  BF16 --> BF16 {4096, 1536, 128}
blk.4.ffn_up_exps.weight,    BF16 --> BF16 {4096, 1536, 128}
blk.4.ffn_gate_inp.weight,   BF16 --> F32 {4096, 128}
blk.4.ffn_norm.weight,       BF16 --> F32 {4096}
blk.5.attn_k_norm.weight,    BF16 --> F32 {128}
blk.5.attn_k.weight,         BF16 --> BF16 {4096, 512}
blk.5.attn_output.weight,    BF16 --> BF16 {8192, 4096}
blk.5.attn_q_norm.weight,    BF16 --> F32 {128}
blk.5.attn_q.weight,         BF16 --> BF16 {4096, 8192}
blk.5.attn_v.weight,         BF16 --> BF16 {4096, 512}
blk.5.attn_norm.weight,      BF16 --> F32 {4096}
blk.5.ffn_down_exps.weight,  BF16 --> BF16 {1536, 4096, 128}
blk.5.ffn_gate_exps.weight,  BF16 --> BF16 {4096, 1536, 128}
blk.5.ffn_up_exps.weight,    BF16 --> BF16 {4096, 1536, 128}
blk.5.ffn_gate_inp.weight,   BF16 --> F32 {4096, 128}
blk.5.ffn_norm.weight,       BF16 --> F32 {4096}
blk.6.attn_k_norm.weight,    BF16 --> F32 {128}
blk.6.attn_k.weight,         BF16 --> BF16 {4096, 512}
blk.6.attn_output.weight,    BF16 --> BF16 {8192, 4096}
blk.6.attn_q_norm.weight,    BF16 --> F32 {128}
blk.6.attn_q.weight,         BF16 --> BF16 {4096, 8192}
blk.6.attn_v.weight,         BF16 --> BF16 {4096, 512}
blk.6.attn_norm.weight,      BF16 --> F32 {4096}
blk.6.ffn_down_exps.weight,  BF16 --> BF16 {1536, 4096, 128}
blk.6.ffn_gate_exps.weight,  BF16 --> BF16 {4096, 1536, 128}
blk.6.ffn_up_exps.weight,    BF16 --> BF16 {4096, 1536, 128}
blk.6.ffn_gate_inp.weight,   BF16 --> F32 {4096, 128}
blk.6.ffn_norm.weight,       BF16 --> F32 {4096}
blk.7.attn_k_norm.weight,    BF16 --> F32 {128}
blk.7.attn_k.weight,         BF16 --> BF16 {4096, 512}
blk.7.attn_output.weight,    BF16 --> BF16 {8192, 4096}
blk.7.attn_q_norm.weight,    BF16 --> F32 {128}
blk.7.attn_q.weight,         BF16 --> BF16 {4096, 8192}
blk.7.attn_v.weight,         BF16 --> BF16 {4096, 512}
blk.7.attn_norm.weight,      BF16 --> F32 {4096}
blk.7.ffn_down_exps.weight,  BF16 --> BF16 {1536, 4096, 128}
blk.7.ffn_gate_exps.weight,  BF16 --> BF16 {4096, 1536, 128}
blk.7.ffn_up_exps.weight,    BF16 --> BF16 {4096, 1536, 128}
blk.7.ffn_gate_inp.weight,   BF16 --> F32 {4096, 128}
blk.7.ffn_norm.weight,       BF16 --> F32 {4096}
blk.8.attn_k_norm.weight,    BF16 --> F32 {128}
blk.8.attn_k.weight,         BF16 --> BF16 {4096, 512}
blk.8.attn_output.weight,    BF16 --> BF16 {8192, 4096}
blk.8.attn_q_norm.weight,    BF16 --> F32 {128}
blk.8.attn_q.weight,         BF16 --> BF16 {4096, 8192}
blk.8.attn_v.weight,         BF16 --> BF16 {4096, 512}
blk.8.attn_norm.weight,      BF16 --> F32 {4096}
blk.8.ffn_down_exps.weight,  BF16 --> BF16 {1536, 4096, 128}
blk.8.ffn_gate_exps.weight,  BF16 --> BF16 {4096, 1536, 128}
blk.8.ffn_up_exps.weight,    BF16 --> BF16 {4096, 1536, 128}
blk.8.ffn_gate_inp.weight,   BF16 --> F32 {4096, 128}
blk.8.ffn_norm.weight,       BF16 --> F32 {4096}
blk.9.attn_k_norm.weight,    BF16 --> F32 {128}
blk.9.attn_k.weight,         BF16 --> BF16 {4096, 512}
blk.9.attn_output.weight,    BF16 --> BF16 {8192, 4096}
blk.9.attn_q_norm.weight,    BF16 --> F32 {128}
blk.9.attn_q.weight,         BF16 --> BF16 {4096, 8192}
blk.9.attn_v.weight,         BF16 --> BF16 {4096, 512}
blk.10.attn_k_norm.weight,   BF16 --> F32 {128}
blk.10.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.10.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.10.attn_q_norm.weight,   BF16 --> F32 {128}
blk.10.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.10.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.9.attn_norm.weight,      BF16 --> F32 {4096}
blk.9.ffn_down_exps.weight,  BF16 --> BF16 {1536, 4096, 128}
blk.9.ffn_gate_exps.weight,  BF16 --> BF16 {4096, 1536, 128}
blk.9.ffn_up_exps.weight,    BF16 --> BF16 {4096, 1536, 128}
blk.9.ffn_gate_inp.weight,   BF16 --> F32 {4096, 128}
blk.9.ffn_norm.weight,       BF16 --> F32 {4096}
blk.10.attn_norm.weight,     BF16 --> F32 {4096}
blk.10.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.10.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.10.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.10.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.10.ffn_norm.weight,      BF16 --> F32 {4096}
blk.11.attn_k_norm.weight,   BF16 --> F32 {128}
blk.11.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.11.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.11.attn_q_norm.weight,   BF16 --> F32 {128}
blk.11.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.11.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.11.attn_norm.weight,     BF16 --> F32 {4096}
blk.11.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.11.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.11.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.11.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.11.ffn_norm.weight,      BF16 --> F32 {4096}
blk.12.attn_k_norm.weight,   BF16 --> F32 {128}
blk.12.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.12.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.12.attn_q_norm.weight,   BF16 --> F32 {128}
blk.12.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.12.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.12.attn_norm.weight,     BF16 --> F32 {4096}
blk.12.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.12.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.12.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.12.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.12.ffn_norm.weight,      BF16 --> F32 {4096}
blk.13.attn_k_norm.weight,   BF16 --> F32 {128}
blk.13.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.13.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.13.attn_q_norm.weight,   BF16 --> F32 {128}
blk.13.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.13.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.13.attn_norm.weight,     BF16 --> F32 {4096}
blk.13.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.13.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.13.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.13.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.13.ffn_norm.weight,      BF16 --> F32 {4096}
blk.14.attn_k_norm.weight,   BF16 --> F32 {128}
blk.14.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.14.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.14.attn_q_norm.weight,   BF16 --> F32 {128}
blk.14.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.14.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.14.attn_norm.weight,     BF16 --> F32 {4096}
blk.14.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.14.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.14.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.14.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.14.ffn_norm.weight,      BF16 --> F32 {4096}
blk.15.attn_k_norm.weight,   BF16 --> F32 {128}
blk.15.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.15.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.15.attn_q_norm.weight,   BF16 --> F32 {128}
blk.15.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.15.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.15.attn_norm.weight,     BF16 --> F32 {4096}
blk.15.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.15.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.15.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.15.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.15.ffn_norm.weight,      BF16 --> F32 {4096}
blk.16.attn_k_norm.weight,   BF16 --> F32 {128}
blk.16.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.16.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.16.attn_q_norm.weight,   BF16 --> F32 {128}
blk.16.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.16.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.16.attn_norm.weight,     BF16 --> F32 {4096}
blk.16.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.16.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.16.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.16.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.16.ffn_norm.weight,      BF16 --> F32 {4096}
blk.17.attn_k_norm.weight,   BF16 --> F32 {128}
blk.17.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.17.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.17.attn_q_norm.weight,   BF16 --> F32 {128}
blk.17.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.17.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.17.attn_norm.weight,     BF16 --> F32 {4096}
blk.17.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.17.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.17.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.17.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.17.ffn_norm.weight,      BF16 --> F32 {4096}
blk.18.attn_k_norm.weight,   BF16 --> F32 {128}
blk.18.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.18.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.18.attn_q_norm.weight,   BF16 --> F32 {128}
blk.18.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.18.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.18.attn_norm.weight,     BF16 --> F32 {4096}
blk.18.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.18.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.18.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.18.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.18.ffn_norm.weight,      BF16 --> F32 {4096}
blk.19.attn_k_norm.weight,   BF16 --> F32 {128}
blk.19.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.19.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.19.attn_q_norm.weight,   BF16 --> F32 {128}
blk.19.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.19.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.19.attn_norm.weight,     BF16 --> F32 {4096}
blk.19.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.19.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.19.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.19.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.19.ffn_norm.weight,      BF16 --> F32 {4096}
blk.20.attn_k_norm.weight,   BF16 --> F32 {128}
blk.20.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.20.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.20.attn_q_norm.weight,   BF16 --> F32 {128}
blk.20.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.20.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.20.attn_norm.weight,     BF16 --> F32 {4096}
blk.20.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.20.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.20.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.20.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.20.ffn_norm.weight,      BF16 --> F32 {4096}
blk.21.attn_k_norm.weight,   BF16 --> F32 {128}
blk.21.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.21.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.21.attn_q_norm.weight,   BF16 --> F32 {128}
blk.21.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.21.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.21.attn_norm.weight,     BF16 --> F32 {4096}
blk.21.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.21.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.21.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.21.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.21.ffn_norm.weight,      BF16 --> F32 {4096}
blk.22.attn_k_norm.weight,   BF16 --> F32 {128}
blk.22.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.22.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.22.attn_q_norm.weight,   BF16 --> F32 {128}
blk.22.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.22.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.22.attn_norm.weight,     BF16 --> F32 {4096}
blk.22.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.22.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.22.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.22.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.22.ffn_norm.weight,      BF16 --> F32 {4096}
blk.23.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.23.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.23.attn_k_norm.weight,   BF16 --> F32 {128}
blk.23.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.23.attn_q_norm.weight,   BF16 --> F32 {128}
blk.23.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.23.attn_norm.weight,     BF16 --> F32 {4096}
blk.23.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.23.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.23.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.23.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.23.ffn_norm.weight,      BF16 --> F32 {4096}
blk.24.attn_k_norm.weight,   BF16 --> F32 {128}
blk.24.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.24.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.24.attn_q_norm.weight,   BF16 --> F32 {128}
blk.24.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.24.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.24.attn_norm.weight,     BF16 --> F32 {4096}
blk.24.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.24.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.24.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.24.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.24.ffn_norm.weight,      BF16 --> F32 {4096}
blk.25.attn_k_norm.weight,   BF16 --> F32 {128}
blk.25.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.25.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.25.attn_q_norm.weight,   BF16 --> F32 {128}
blk.25.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.25.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.25.attn_norm.weight,     BF16 --> F32 {4096}
blk.25.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.25.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.25.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.25.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.25.ffn_norm.weight,      BF16 --> F32 {4096}
blk.26.attn_k_norm.weight,   BF16 --> F32 {128}
blk.26.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.26.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.26.attn_q_norm.weight,   BF16 --> F32 {128}
blk.26.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.26.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.26.attn_norm.weight,     BF16 --> F32 {4096}
blk.26.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.26.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.26.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.26.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.26.ffn_norm.weight,      BF16 --> F32 {4096}
blk.27.attn_k_norm.weight,   BF16 --> F32 {128}
blk.27.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.27.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.27.attn_q_norm.weight,   BF16 --> F32 {128}
blk.27.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.27.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.27.attn_norm.weight,     BF16 --> F32 {4096}
blk.27.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.27.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.27.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.27.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.27.ffn_norm.weight,      BF16 --> F32 {4096}
blk.28.attn_k_norm.weight,   BF16 --> F32 {128}
blk.28.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.28.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.28.attn_q_norm.weight,   BF16 --> F32 {128}
blk.28.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.28.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.28.attn_norm.weight,     BF16 --> F32 {4096}
blk.28.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.28.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.28.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.28.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.28.ffn_norm.weight,      BF16 --> F32 {4096}
blk.29.attn_k_norm.weight,   BF16 --> F32 {128}
blk.29.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.29.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.29.attn_q_norm.weight,   BF16 --> F32 {128}
blk.29.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.29.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.29.attn_norm.weight,     BF16 --> F32 {4096}
blk.29.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.29.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.29.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.29.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.29.ffn_norm.weight,      BF16 --> F32 {4096}
blk.30.attn_k_norm.weight,   BF16 --> F32 {128}
blk.30.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.30.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.30.attn_q_norm.weight,   BF16 --> F32 {128}
blk.30.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.30.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.30.attn_norm.weight,     BF16 --> F32 {4096}
blk.30.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.30.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.30.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.30.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.30.ffn_norm.weight,      BF16 --> F32 {4096}
blk.31.attn_k_norm.weight,   BF16 --> F32 {128}
blk.31.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.31.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.31.attn_q_norm.weight,   BF16 --> F32 {128}
blk.31.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.31.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.31.attn_norm.weight,     BF16 --> F32 {4096}
blk.31.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.31.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.31.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.31.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.31.ffn_norm.weight,      BF16 --> F32 {4096}
blk.32.attn_k_norm.weight,   BF16 --> F32 {128}
blk.32.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.32.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.32.attn_q_norm.weight,   BF16 --> F32 {128}
blk.32.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.32.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.32.attn_norm.weight,     BF16 --> F32 {4096}
blk.32.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.32.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.32.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.32.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.32.ffn_norm.weight,      BF16 --> F32 {4096}
blk.33.attn_k_norm.weight,   BF16 --> F32 {128}
blk.33.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.33.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.33.attn_q_norm.weight,   BF16 --> F32 {128}
blk.33.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.33.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.33.attn_norm.weight,     BF16 --> F32 {4096}
blk.33.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.33.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.33.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.33.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.33.ffn_norm.weight,      BF16 --> F32 {4096}
blk.34.attn_k_norm.weight,   BF16 --> F32 {128}
blk.34.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.34.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.34.attn_q_norm.weight,   BF16 --> F32 {128}
blk.34.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.34.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.34.attn_norm.weight,     BF16 --> F32 {4096}
blk.34.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.34.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.34.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.34.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.34.ffn_norm.weight,      BF16 --> F32 {4096}
blk.35.attn_k_norm.weight,   BF16 --> F32 {128}
blk.35.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.35.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.35.attn_q_norm.weight,   BF16 --> F32 {128}
blk.35.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.35.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.35.attn_norm.weight,     BF16 --> F32 {4096}
blk.35.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.35.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.35.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.35.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.35.ffn_norm.weight,      BF16 --> F32 {4096}
blk.36.attn_k_norm.weight,   BF16 --> F32 {128}
blk.36.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.36.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.36.attn_q_norm.weight,   BF16 --> F32 {128}
blk.36.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.36.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.36.attn_norm.weight,     BF16 --> F32 {4096}
blk.36.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.36.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.36.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.36.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.36.ffn_norm.weight,      BF16 --> F32 {4096}
blk.37.attn_k_norm.weight,   BF16 --> F32 {128}
blk.37.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.37.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.37.attn_q_norm.weight,   BF16 --> F32 {128}
blk.37.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.37.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.37.attn_norm.weight,     BF16 --> F32 {4096}
blk.37.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.37.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.37.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.37.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.37.ffn_norm.weight,      BF16 --> F32 {4096}
blk.38.attn_k_norm.weight,   BF16 --> F32 {128}
blk.38.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.38.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.38.attn_q_norm.weight,   BF16 --> F32 {128}
blk.38.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.38.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.38.attn_norm.weight,     BF16 --> F32 {4096}
blk.38.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.38.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.38.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.38.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.38.ffn_norm.weight,      BF16 --> F32 {4096}
blk.39.attn_k_norm.weight,   BF16 --> F32 {128}
blk.39.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.39.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.39.attn_q_norm.weight,   BF16 --> F32 {128}
blk.39.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.39.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.39.attn_norm.weight,     BF16 --> F32 {4096}
blk.39.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.39.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.39.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.39.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.39.ffn_norm.weight,      BF16 --> F32 {4096}
blk.40.attn_k_norm.weight,   BF16 --> F32 {128}
blk.40.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.40.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.40.attn_q_norm.weight,   BF16 --> F32 {128}
blk.40.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.40.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.40.attn_norm.weight,     BF16 --> F32 {4096}
blk.40.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.40.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.40.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.40.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.40.ffn_norm.weight,      BF16 --> F32 {4096}
blk.41.attn_k_norm.weight,   BF16 --> F32 {128}
blk.41.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.41.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.41.attn_q_norm.weight,   BF16 --> F32 {128}
blk.41.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.41.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.41.attn_norm.weight,     BF16 --> F32 {4096}
blk.41.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.41.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.41.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.41.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.41.ffn_norm.weight,      BF16 --> F32 {4096}
blk.42.attn_k_norm.weight,   BF16 --> F32 {128}
blk.42.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.42.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.42.attn_q_norm.weight,   BF16 --> F32 {128}
blk.42.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.42.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.42.attn_norm.weight,     BF16 --> F32 {4096}
blk.42.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.42.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.42.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.42.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.42.ffn_norm.weight,      BF16 --> F32 {4096}
blk.43.attn_k_norm.weight,   BF16 --> F32 {128}
blk.43.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.43.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.43.attn_q_norm.weight,   BF16 --> F32 {128}
blk.43.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.43.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.43.attn_norm.weight,     BF16 --> F32 {4096}
blk.43.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.43.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.43.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.43.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.43.ffn_norm.weight,      BF16 --> F32 {4096}
blk.44.attn_k_norm.weight,   BF16 --> F32 {128}
blk.44.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.44.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.44.attn_q_norm.weight,   BF16 --> F32 {128}
blk.44.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.44.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.44.attn_norm.weight,     BF16 --> F32 {4096}
blk.44.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.44.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.44.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.44.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.44.ffn_norm.weight,      BF16 --> F32 {4096}
blk.45.attn_k_norm.weight,   BF16 --> F32 {128}
blk.45.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.45.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.45.attn_q_norm.weight,   BF16 --> F32 {128}
blk.45.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.45.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.45.attn_norm.weight,     BF16 --> F32 {4096}
blk.45.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.45.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.45.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.45.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.45.ffn_norm.weight,      BF16 --> F32 {4096}
blk.46.attn_k_norm.weight,   BF16 --> F32 {128}
blk.46.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.46.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.46.attn_q_norm.weight,   BF16 --> F32 {128}
blk.46.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.46.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.46.attn_norm.weight,     BF16 --> F32 {4096}
blk.46.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.46.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.46.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.46.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.46.ffn_norm.weight,      BF16 --> F32 {4096}
blk.47.attn_k_norm.weight,   BF16 --> F32 {128}
blk.47.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.47.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.47.attn_q_norm.weight,   BF16 --> F32 {128}
blk.47.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.47.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.47.attn_norm.weight,     BF16 --> F32 {4096}
blk.47.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.47.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.47.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.47.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.47.ffn_norm.weight,      BF16 --> F32 {4096}
blk.48.attn_k_norm.weight,   BF16 --> F32 {128}
blk.48.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.48.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.48.attn_q_norm.weight,   BF16 --> F32 {128}
blk.48.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.48.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.48.attn_norm.weight,     BF16 --> F32 {4096}
blk.48.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.48.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.48.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.48.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.48.ffn_norm.weight,      BF16 --> F32 {4096}
blk.49.attn_k_norm.weight,   BF16 --> F32 {128}
blk.49.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.49.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.49.attn_q_norm.weight,   BF16 --> F32 {128}
blk.49.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.49.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.49.attn_norm.weight,     BF16 --> F32 {4096}
blk.49.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.49.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.49.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.49.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.49.ffn_norm.weight,      BF16 --> F32 {4096}
blk.50.attn_k_norm.weight,   BF16 --> F32 {128}
blk.50.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.50.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.50.attn_q_norm.weight,   BF16 --> F32 {128}
blk.50.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.50.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.50.attn_norm.weight,     BF16 --> F32 {4096}
blk.50.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.50.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.50.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.50.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.50.ffn_norm.weight,      BF16 --> F32 {4096}
blk.51.attn_k_norm.weight,   BF16 --> F32 {128}
blk.51.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.51.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.51.attn_q_norm.weight,   BF16 --> F32 {128}
blk.51.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.51.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.51.attn_norm.weight,     BF16 --> F32 {4096}
blk.51.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.51.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.51.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.51.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.51.ffn_norm.weight,      BF16 --> F32 {4096}
blk.52.attn_k_norm.weight,   BF16 --> F32 {128}
blk.52.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.52.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.52.attn_q_norm.weight,   BF16 --> F32 {128}
blk.52.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.52.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.52.attn_norm.weight,     BF16 --> F32 {4096}
blk.52.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.52.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.52.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.52.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.52.ffn_norm.weight,      BF16 --> F32 {4096}
blk.53.attn_k_norm.weight,   BF16 --> F32 {128}
blk.53.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.53.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.53.attn_q_norm.weight,   BF16 --> F32 {128}
blk.53.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.53.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.53.attn_norm.weight,     BF16 --> F32 {4096}
blk.53.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.53.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.53.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.53.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.53.ffn_norm.weight,      BF16 --> F32 {4096}
blk.54.attn_k_norm.weight,   BF16 --> F32 {128}
blk.54.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.54.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.54.attn_q_norm.weight,   BF16 --> F32 {128}
blk.54.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.54.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.54.attn_norm.weight,     BF16 --> F32 {4096}
blk.54.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.54.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.54.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.54.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.54.ffn_norm.weight,      BF16 --> F32 {4096}
blk.55.attn_k_norm.weight,   BF16 --> F32 {128}
blk.55.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.55.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.55.attn_q_norm.weight,   BF16 --> F32 {128}
blk.55.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.55.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.55.attn_norm.weight,     BF16 --> F32 {4096}
blk.55.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.55.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.55.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.55.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.55.ffn_norm.weight,      BF16 --> F32 {4096}
blk.56.attn_k_norm.weight,   BF16 --> F32 {128}
blk.56.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.56.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.56.attn_q_norm.weight,   BF16 --> F32 {128}
blk.56.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.56.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.56.attn_norm.weight,     BF16 --> F32 {4096}
blk.56.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.56.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.56.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.56.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.56.ffn_norm.weight,      BF16 --> F32 {4096}
blk.57.attn_k_norm.weight,   BF16 --> F32 {128}
blk.57.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.57.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.57.attn_q_norm.weight,   BF16 --> F32 {128}
blk.57.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.57.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.57.attn_norm.weight,     BF16 --> F32 {4096}
blk.57.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.57.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.57.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.57.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.57.ffn_norm.weight,      BF16 --> F32 {4096}
blk.58.attn_k_norm.weight,   BF16 --> F32 {128}
blk.58.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.58.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.58.attn_q_norm.weight,   BF16 --> F32 {128}
blk.58.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.58.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.58.attn_norm.weight,     BF16 --> F32 {4096}
blk.58.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.58.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.58.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.58.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.58.ffn_norm.weight,      BF16 --> F32 {4096}
blk.59.attn_k_norm.weight,   BF16 --> F32 {128}
blk.59.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.59.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.59.attn_q_norm.weight,   BF16 --> F32 {128}
blk.59.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.59.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.59.attn_norm.weight,     BF16 --> F32 {4096}
blk.59.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.59.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.59.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.59.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.59.ffn_norm.weight,      BF16 --> F32 {4096}
blk.60.attn_k_norm.weight,   BF16 --> F32 {128}
blk.60.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.60.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.60.attn_q_norm.weight,   BF16 --> F32 {128}
blk.60.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.60.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.60.attn_norm.weight,     BF16 --> F32 {4096}
blk.60.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.60.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.60.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.60.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.60.ffn_norm.weight,      BF16 --> F32 {4096}
blk.61.attn_k_norm.weight,   BF16 --> F32 {128}
blk.61.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.61.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.61.attn_q_norm.weight,   BF16 --> F32 {128}
blk.61.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.61.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.61.attn_norm.weight,     BF16 --> F32 {4096}
blk.61.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.61.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.61.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.61.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.61.ffn_norm.weight,      BF16 --> F32 {4096}
blk.62.attn_k_norm.weight,   BF16 --> F32 {128}
blk.62.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.62.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.62.attn_q_norm.weight,   BF16 --> F32 {128}
blk.62.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.62.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.62.attn_norm.weight,     BF16 --> F32 {4096}
blk.62.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.62.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.62.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.62.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.62.ffn_norm.weight,      BF16 --> F32 {4096}
blk.63.attn_k_norm.weight,   BF16 --> F32 {128}
blk.63.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.63.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.63.attn_q_norm.weight,   BF16 --> F32 {128}
blk.63.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.63.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.63.attn_norm.weight,     BF16 --> F32 {4096}
blk.63.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.63.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.63.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.63.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.63.ffn_norm.weight,      BF16 --> F32 {4096}
blk.64.attn_k_norm.weight,   BF16 --> F32 {128}
blk.64.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.64.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.64.attn_q_norm.weight,   BF16 --> F32 {128}
blk.64.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.64.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.64.attn_norm.weight,     BF16 --> F32 {4096}
blk.64.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.64.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.64.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.64.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.64.ffn_norm.weight,      BF16 --> F32 {4096}
blk.65.attn_k_norm.weight,   BF16 --> F32 {128}
blk.65.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.65.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.65.attn_q_norm.weight,   BF16 --> F32 {128}
blk.65.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.65.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.65.attn_norm.weight,     BF16 --> F32 {4096}
blk.65.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.65.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.65.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.65.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.65.ffn_norm.weight,      BF16 --> F32 {4096}
blk.66.attn_k_norm.weight,   BF16 --> F32 {128}
blk.66.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.66.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.66.attn_q_norm.weight,   BF16 --> F32 {128}
blk.66.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.66.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.66.attn_norm.weight,     BF16 --> F32 {4096}
blk.66.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.66.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.66.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.66.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.66.ffn_norm.weight,      BF16 --> F32 {4096}
blk.67.attn_k_norm.weight,   BF16 --> F32 {128}
blk.67.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.67.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.67.attn_q_norm.weight,   BF16 --> F32 {128}
blk.67.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.67.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.67.attn_norm.weight,     BF16 --> F32 {4096}
blk.67.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.67.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.67.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.67.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.67.ffn_norm.weight,      BF16 --> F32 {4096}
blk.68.attn_k_norm.weight,   BF16 --> F32 {128}
blk.68.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.68.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.68.attn_q_norm.weight,   BF16 --> F32 {128}
blk.68.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.68.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.68.attn_norm.weight,     BF16 --> F32 {4096}
blk.68.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.68.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.68.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.68.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.68.ffn_norm.weight,      BF16 --> F32 {4096}
blk.69.attn_k_norm.weight,   BF16 --> F32 {128}
blk.69.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.69.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.69.attn_q_norm.weight,   BF16 --> F32 {128}
blk.69.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.69.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.69.attn_norm.weight,     BF16 --> F32 {4096}
blk.69.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.69.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.69.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.69.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.69.ffn_norm.weight,      BF16 --> F32 {4096}
blk.70.attn_k_norm.weight,   BF16 --> F32 {128}
blk.70.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.70.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.70.attn_q_norm.weight,   BF16 --> F32 {128}
blk.70.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.70.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.70.attn_norm.weight,     BF16 --> F32 {4096}
blk.70.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.70.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.70.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.70.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.70.ffn_norm.weight,      BF16 --> F32 {4096}
blk.71.attn_k_norm.weight,   BF16 --> F32 {128}
blk.71.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.71.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.71.attn_q_norm.weight,   BF16 --> F32 {128}
blk.71.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.71.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.71.attn_norm.weight,     BF16 --> F32 {4096}
blk.71.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.71.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.71.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.71.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.71.ffn_norm.weight,      BF16 --> F32 {4096}
blk.72.attn_k_norm.weight,   BF16 --> F32 {128}
blk.72.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.72.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.72.attn_q_norm.weight,   BF16 --> F32 {128}
blk.72.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.72.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.72.attn_norm.weight,     BF16 --> F32 {4096}
blk.72.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.72.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.72.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.72.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.72.ffn_norm.weight,      BF16 --> F32 {4096}
blk.73.attn_k_norm.weight,   BF16 --> F32 {128}
blk.73.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.73.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.73.attn_q_norm.weight,   BF16 --> F32 {128}
blk.73.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.73.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.73.attn_norm.weight,     BF16 --> F32 {4096}
blk.73.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.73.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.73.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.73.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.73.ffn_norm.weight,      BF16 --> F32 {4096}
blk.74.attn_k_norm.weight,   BF16 --> F32 {128}
blk.74.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.74.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.74.attn_q_norm.weight,   BF16 --> F32 {128}
blk.74.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.74.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.74.attn_norm.weight,     BF16 --> F32 {4096}
blk.74.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.74.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.74.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.74.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.74.ffn_norm.weight,      BF16 --> F32 {4096}
blk.75.attn_k_norm.weight,   BF16 --> F32 {128}
blk.75.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.75.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.75.attn_q_norm.weight,   BF16 --> F32 {128}
blk.75.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.75.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.75.attn_norm.weight,     BF16 --> F32 {4096}
blk.75.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.75.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.75.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.75.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.75.ffn_norm.weight,      BF16 --> F32 {4096}
blk.76.attn_k_norm.weight,   BF16 --> F32 {128}
blk.76.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.76.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.76.attn_q_norm.weight,   BF16 --> F32 {128}
blk.76.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.76.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.76.attn_norm.weight,     BF16 --> F32 {4096}
blk.76.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.76.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.76.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.76.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.76.ffn_norm.weight,      BF16 --> F32 {4096}
blk.77.attn_k_norm.weight,   BF16 --> F32 {128}
blk.77.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.77.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.77.attn_q_norm.weight,   BF16 --> F32 {128}
blk.77.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.77.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.77.attn_norm.weight,     BF16 --> F32 {4096}
blk.77.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.77.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.77.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.77.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.77.ffn_norm.weight,      BF16 --> F32 {4096}
blk.78.attn_k_norm.weight,   BF16 --> F32 {128}
blk.78.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.78.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.78.attn_q_norm.weight,   BF16 --> F32 {128}
blk.78.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.78.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.78.attn_norm.weight,     BF16 --> F32 {4096}
blk.78.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.78.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.78.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.78.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.78.ffn_norm.weight,      BF16 --> F32 {4096}
blk.79.attn_k_norm.weight,   BF16 --> F32 {128}
blk.79.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.79.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.79.attn_q_norm.weight,   BF16 --> F32 {128}
blk.79.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.79.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.79.attn_norm.weight,     BF16 --> F32 {4096}
blk.79.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.79.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.79.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.79.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.79.ffn_norm.weight,      BF16 --> F32 {4096}
blk.80.attn_k_norm.weight,   BF16 --> F32 {128}
blk.80.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.80.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.80.attn_q_norm.weight,   BF16 --> F32 {128}
blk.80.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.80.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.80.attn_norm.weight,     BF16 --> F32 {4096}
blk.80.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.80.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.80.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.80.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.80.ffn_norm.weight,      BF16 --> F32 {4096}
blk.81.attn_k_norm.weight,   BF16 --> F32 {128}
blk.81.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.81.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.81.attn_q_norm.weight,   BF16 --> F32 {128}
blk.81.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.81.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.81.attn_norm.weight,     BF16 --> F32 {4096}
blk.81.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.81.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.81.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.81.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.81.ffn_norm.weight,      BF16 --> F32 {4096}
blk.82.attn_k_norm.weight,   BF16 --> F32 {128}
blk.82.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.82.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.82.attn_q_norm.weight,   BF16 --> F32 {128}
blk.82.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.82.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.82.attn_norm.weight,     BF16 --> F32 {4096}
blk.82.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.82.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.82.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.82.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.82.ffn_norm.weight,      BF16 --> F32 {4096}
blk.83.attn_k_norm.weight,   BF16 --> F32 {128}
blk.83.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.83.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.83.attn_q_norm.weight,   BF16 --> F32 {128}
blk.83.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.83.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.83.attn_norm.weight,     BF16 --> F32 {4096}
blk.83.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.83.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.83.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.83.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.83.ffn_norm.weight,      BF16 --> F32 {4096}
blk.84.attn_k_norm.weight,   BF16 --> F32 {128}
blk.84.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.84.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.84.attn_q_norm.weight,   BF16 --> F32 {128}
blk.84.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.84.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.84.attn_norm.weight,     BF16 --> F32 {4096}
blk.84.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.84.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.84.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.84.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.84.ffn_norm.weight,      BF16 --> F32 {4096}
blk.85.attn_k_norm.weight,   BF16 --> F32 {128}
blk.85.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.85.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.85.attn_q_norm.weight,   BF16 --> F32 {128}
blk.85.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.85.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.85.attn_norm.weight,     BF16 --> F32 {4096}
blk.85.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.85.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.85.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.85.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.85.ffn_norm.weight,      BF16 --> F32 {4096}
blk.86.attn_k_norm.weight,   BF16 --> F32 {128}
blk.86.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.86.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.86.attn_q_norm.weight,   BF16 --> F32 {128}
blk.86.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.86.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.86.attn_norm.weight,     BF16 --> F32 {4096}
blk.86.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.86.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.86.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.86.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.86.ffn_norm.weight,      BF16 --> F32 {4096}
blk.87.attn_k_norm.weight,   BF16 --> F32 {128}
blk.87.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.87.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.87.attn_q_norm.weight,   BF16 --> F32 {128}
blk.87.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.87.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.87.attn_norm.weight,     BF16 --> F32 {4096}
blk.87.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.87.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.87.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.87.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.87.ffn_norm.weight,      BF16 --> F32 {4096}
blk.88.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.88.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.88.attn_k_norm.weight,   BF16 --> F32 {128}
blk.88.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.88.attn_q_norm.weight,   BF16 --> F32 {128}
blk.88.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.88.attn_norm.weight,     BF16 --> F32 {4096}
blk.88.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.88.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.88.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.88.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.88.ffn_norm.weight,      BF16 --> F32 {4096}
blk.89.attn_k_norm.weight,   BF16 --> F32 {128}
blk.89.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.89.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.89.attn_q_norm.weight,   BF16 --> F32 {128}
blk.89.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.89.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.89.attn_norm.weight,     BF16 --> F32 {4096}
blk.89.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.89.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.89.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.89.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.89.ffn_norm.weight,      BF16 --> F32 {4096}
blk.90.attn_k_norm.weight,   BF16 --> F32 {128}
blk.90.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.90.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.90.attn_q_norm.weight,   BF16 --> F32 {128}
blk.90.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.90.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.90.attn_norm.weight,     BF16 --> F32 {4096}
blk.90.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.90.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.90.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.90.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.90.ffn_norm.weight,      BF16 --> F32 {4096}
blk.91.attn_k_norm.weight,   BF16 --> F32 {128}
blk.91.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.91.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.91.attn_q_norm.weight,   BF16 --> F32 {128}
blk.91.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.91.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.91.attn_norm.weight,     BF16 --> F32 {4096}
blk.91.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.91.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.91.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.91.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.91.ffn_norm.weight,      BF16 --> F32 {4096}
blk.92.attn_k_norm.weight,   BF16 --> F32 {128}
blk.92.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.92.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.92.attn_q_norm.weight,   BF16 --> F32 {128}
blk.92.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.92.attn_v.weight,        BF16 --> BF16 {4096, 512}
blk.92.attn_norm.weight,     BF16 --> F32 {4096}
blk.92.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.92.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.92.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.92.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.92.ffn_norm.weight,      BF16 --> F32 {4096}
blk.93.attn_k_norm.weight,   BF16 --> F32 {128}
blk.93.attn_k.weight,        BF16 --> BF16 {4096, 512}
blk.93.attn_output.weight,   BF16 --> BF16 {8192, 4096}
blk.93.attn_q_norm.weight,   BF16 --> F32 {128}
blk.93.attn_q.weight,        BF16 --> BF16 {4096, 8192}
blk.93.attn_v.weight,        BF16 --> BF16 {4096, 512}
output.weight,               BF16 --> BF16 {4096, 151936}
blk.93.attn_norm.weight,     BF16 --> F32 {4096}
blk.93.ffn_down_exps.weight, BF16 --> BF16 {1536, 4096, 128}
blk.93.ffn_gate_exps.weight, BF16 --> BF16 {4096, 1536, 128}
blk.93.ffn_up_exps.weight,   BF16 --> BF16 {4096, 1536, 128}
blk.93.ffn_gate_inp.weight,  BF16 --> F32 {4096, 128}
blk.93.ffn_norm.weight,      BF16 --> F32 {4096}
output_norm.weight,          BF16 --> F32 {4096}

It even expands the 1dim-tensors to f32 from bf16. There is a lot to be investigated and potentially experimented here.

btw., to make it a more useful experiment, we should see if problems actually go away for an existing model that failed (both for nans in the source, to see if they are inexplicably introduced during conversions, and during imatrix gen, to see if they get introduced dynamically during inference).

blk.77.ffn_gate_inp.weight, BF16 --> F32 {4096, 128}

not just one-dim ones. that is completely nonsensical, is it not?

Now we know that while BF16 fixes some quantisation issues it does not fix all NaN errors:

[ 309/1131]            blk.25.attn_q_norm.weight - [  128,     1,     1,     1], type =    f32, size =    0.000 MB
[ 310/1131]                 blk.25.attn_v.weight - [ 4096,   512,     1,     1], type =   bf16, converting to q4_K .. size =     4.00 MiB ->     1.12 MiB
[ 311/1131]          blk.25.ffn_down_exps.weight - [ 1536,  4096,   128,     1], type =   bf16, converting to q2_K .. ggml_validate_row_data: found nan value at block 29
ggml_validate_row_data: found nan value at block 59
ggml_validate_row_data: found nan value at block 65
ggml_validate_row_data: found nan value at block 47
ggml_validate_row_data: found nan value at block 23
ggml_validate_row_data: found nan value at block 53
llama_model_quantize: failed to quantize: quantized data validation failed
main: failed to quantize model from './Qwen3-235B-A22B-Instruct-2507.gguf'
job finished, status 47
job-done<0 Qwen3-235B-A22B-Instruct-2507 imatrix 47>

@mradermacher Can you please skip Q2_K_S and let Qwen3-235B-A22B-Instruct-2507 continue with the next one?

done - as a sidenote, i've added the abiltiy to more easily skip individual quants, but I haven't exposed it via audit yet.

Sign up or log in to comment