Model can't inference for Llama3.2-1B when use -d fp16 to convert pte

[WeiMa01]

when we runing Llama3.2-1B fp16.pte which use -d fp16 to convert Llama3.2-1B w/BF16 to pte, It meet an issue:
error log:
I 00:00:00.013206 executorch:main.cpp:69] Resetting threadpool with num threads = 6
I 00:00:00.027952 executorch:runner.cpp:67] Creating LLaMa runner: model_path=llama3_2_fp16_org.pte, tokenizer_path=../tokenizer.model
E 00:00:00.728030 executorch:XNNCompiler.cpp:635] Failed to create multiply node 266 with code: xnn_status_invalid_parameter
E 00:00:00.728090 executorch:XNNPACKBackend.cpp:106] XNNCompiler::compileModel failed: 0x1
E 00:00:00.728099 executorch:method.cpp:110] Init failed for backend XnnpackBackend: 0x1
E 00:00:00.771031 executorch:XNNCompiler.cpp:635] Failed to create multiply node 266 with code: xnn_status_invalid_parameter
E 00:00:00.771096 executorch:XNNPACKBackend.cpp:106] XNNCompiler::compileModel failed: 0x1
E 00:00:00.771104 executorch:method.cpp:110] Init failed for backend XnnpackBackend: 0x1

convert command:
python -m examples.models.llama.export_llama --model "llama3_2" --checkpoint "/model_convert/Llama-3.2-1B/original/consolidated_00.pth" --params "/Llama-3.2-1B/original/params.json" --use_sdpa_with_kv_cache -X --xnnpack-extended-ops --output_name "llama3_2_fp16_direct_convert_runtime.pte" -kv -d fp16 --max_seq_length 256

cc @digantdesai @mcr229 @cbilgin

[JacobSzwejbka]

cc @mcr229

[mcr229]

@WeiMa01 I think there is an issue in the llama dtype conversion (source type bf16 --> target type fp16), in which one of the edges isn't converted, causing a multiply node with one bf16 and one fp16 input, which is likely what is being complained about here.

On the XNNPACK side we are trying to add better error checking here: #9023

But i think there is something with the model transformation code that isn't consistently changing all the dtypes to fp16. I remember @kimishpatel and @jackzhxng talking about this?

[jackzhxng]

Oh nope, the entire model should be getting converted to fp16 here, especially since there is no quantization