Granite-Switch Architecture#25107

Open

barvhaim wants to merge 40 commits into

ggml-org:masterfrom

barvhaim:feature/granite-switch

Commits on Jul 12, 2026

granite-switch: add llama.cpp backend (POC, CPU)
barvhaim
committed
granite-switch: add Mac (Metal) build + mid-sequence switch demo script
barvhaim
committed
granite-switch mac demo: add -no-cnv so each run is one-shot
barvhaim
committed
granite-switch: replace global sticky index with in-graph router attention
barvhaim
committed
granite-switch: drop scratch tests and mac demo for upstream PR
barvhaim
committed
granite-switch: trim comments to match native llama.cpp style
barvhaim
committed
granite-switch: trim conversion comments to match native style
barvhaim
committed
granite-switch: drop unused adapter_ranks metadata
barvhaim
committed
granite-switch: rename arch to graniteswitch and drop obid alias
barvhaim
committed
granite-switch: fix non-ASCII comments and document router gain assumption
barvhaim
committed
granite-switch: drop section comments from constants.py to match native style
barvhaim
committed
granite-switch: add functional tensor block comments matching Granite4 Vision style
barvhaim
committed
granite-switch: clarify n_expert_used comment
barvhaim
committed
granite-switch: note n_layer_nextn reuse has no MTP
barvhaim
committed
granite-switch: rename source file and apply review nits
barvhaim
committed
granite-switch: don't force LoRA tensors to F16, follow --outtype instead
barvhaim
committed
granite-switch: drop redundant _permute_qk wrapper, call LlamaModel.permute directly
barvhaim
committed
granite-switch: read router gain from GGUF (control_token_gain) instead of hardcoding 15.0
barvhaim
committed
granite-switch: derive n_slots()
barvhaim
committed
granite-switch: move llm_graph_input_switch into granite-switch.cpp
barvhaim
committed
granite-switch: cut AI-style narration comments
barvhaim
committed
granite-switch: collapse multi-line comments
barvhaim
committed
granite-switch: rename control_token_* maps to adapter_token_*
barvhaim
committed
granite-switch: cut noise comments
barvhaim
committed
granite-switch: rename embedded LoRA tensors to <base>.lora_a/lora_b
barvhaim
committed
granite-switch: GGML_ASSERT token input to avoid UB on embeddings
barvhaim
committed
granite-switch: TODO for raw embedding input support
barvhaim
committed
granite-switch: collapse LoRA tensor constants to .lora_a/.lora_b suffix
barvhaim
committed
granite-switch: drop n_expert_used hack, guard mul_mat_id buft probe
barvhaim
committed
granite-switch: stop forcing dense expert counts, read from config
barvhaim
committed
granite-switch: renamed control_token_gain metadata key to router_gain
barvhaim
committed
granite-switch: trim header comments to match native style
barvhaim
committed
granite-switch: collapse LoRA tensors to base name + suffix
barvhaim
committed
granite-switch: inline suffix checks in tensor op resolution
barvhaim
committed
granite-switch: drop switch-lora struct comment
barvhaim
committed
granite-switch: guard router layer index and inline n_slots
barvhaim
committed
granite-switch: group adapter metadata under {arch}.adapters.* namespace
barvhaim
committed
granite-switch: add hparams.has_rope(il) for KV-shift rope skipping
barvhaim
committed
granite-switch: skip arch in test-llama-archs (adapter fixture missing, TODO)
barvhaim
committed
granite-switch: Keys.Adapters namespace + simplify n_slots
barvhaim
committed