dylanneve1/npuw-mha-test-model

npuw-synth-mha-llama-12L

Synthetic LLM test model — MHA variant (num_kv_heads == num_heads). Companion to npuw-gqa-test-model for isolating GQA-specific code paths on NPU.

Random weights — not for inference quality.

Config

field	value
layers	12
hidden_size	256
num_heads	8
num_kv_heads	8 (MHA)
head_dim	32
intermediate_size	1024
vocab_size	32000
ffn	swiglu
norm	rms
rope	half
weight	fp32
position_ids	2d
kv_cache	yes

Source

Generated with npuw_model_generator_demo from dylanneve1/openvino@gqa-fix:

npuw_model_generator_demo --type llm \
  -t Meta-Llama-3.1-8B-Instruct \
  -o out -n npuw-synth-mha-llama-12L \
  --num-layers 12