Add direct HuggingFace safetensors loader for Gemma 4 (E2B/E4B)#919

Open

ssfdre38 wants to merge 8 commits into

google:mainfrom

ssfdre38:safetensors-gemma4-loader

Commits on May 21, 2026

Add Gemma 4 E2B and E4B model configs

ssfdre38
and
Copilot
committed
gemma4: fix KV cache indexing for dual-attention (variable qkv_dim per layer)

ssfdre38
and
Copilot
committed
build: fix sentencepiece vendored header missing cstdint include

ssfdre38
and
Copilot
committed
build: remove stray .o object files
ssfdre38
committed
gemma4: add standalone config validation test

ssfdre38
and
Copilot
committed
Add direct HuggingFace safetensors loader for Gemma 4 E2B/E4B

ssfdre38
and
Copilot
committed
run.cc: wire --safetensors/--model_spec flags into Run()

ssfdre38
and
Copilot
committed
load_safetensors: fix tensor names for Gemma 4 multimodal HF checkpoints

ssfdre38
and
Copilot
committed