Granite-Switch Architecture #25107
CISC
merged
48 commits into
ggml-org:master
from
barvhaim:feature/granite-switch
gabe-l-hart
approved these changes
on 2026-07-06
ngxson
commented
on 2026-07-07
ngxson
approved these changes
on 2026-07-12
granite-switch: add llama.cpp backend (POC, CPU)
d9572250
granite-switch: add Mac (Metal) build + mid-sequence switch demo script
6e734e82
granite-switch mac demo: add -no-cnv so each run is one-shot
af48dfa0
granite-switch: replace global sticky index with in-graph router atte…
16c9ad56
granite-switch: drop scratch tests and mac demo for upstream PR
30ffe08f
granite-switch: trim comments to match native llama.cpp style
0ac7c96b
granite-switch: trim conversion comments to match native style
b17c0bca
granite-switch: drop unused adapter_ranks metadata
7626399e
granite-switch: rename arch to graniteswitch and drop obid alias
e193ea55
granite-switch: fix non-ASCII comments and document router gain assum…
2b179f08
granite-switch: drop section comments from constants.py to match nati…
3041ba8f
granite-switch: add functional tensor block comments matching Granite…
8c353071
granite-switch: clarify n_expert_used comment
36936ada
granite-switch: note n_layer_nextn reuse has no MTP
dba81e38
granite-switch: rename source file and apply review nits
26f2843a
granite-switch: don't force LoRA tensors to F16, follow --outtype ins…
116c6112
granite-switch: drop redundant _permute_qk wrapper, call LlamaModel.p…
6be5d0ca
granite-switch: read router gain from GGUF (control_token_gain) inste…
059401c9
granite-switch: derive n_slots()
8d71a5c4
granite-switch: move llm_graph_input_switch into granite-switch.cpp
afa1b7e4
granite-switch: cut AI-style narration comments
5b2cb93c
granite-switch: collapse multi-line comments
d385fc24
granite-switch: rename control_token_* maps to adapter_token_*
768e6087
granite-switch: cut noise comments
03e46c11
granite-switch: rename embedded LoRA tensors to <base>.lora_a/lora_b
24e06e96
granite-switch: GGML_ASSERT token input to avoid UB on embeddings
613bc199
granite-switch: TODO for raw embedding input support
472d1c9c
granite-switch: collapse LoRA tensor constants to .lora_a/.lora_b suffix
c6014846
granite-switch: drop n_expert_used hack, guard mul_mat_id buft probe
622eb21e
granite-switch: stop forcing dense expert counts, read from config
ad04ce88
granite-switch: renamed control_token_gain metadata key to router_gain
1d15ea35
granite-switch: trim header comments to match native style
e99c0563
granite-switch: collapse LoRA tensors to base name + suffix
46d882ed
granite-switch: inline suffix checks in tensor op resolution
d99296fb
granite-switch: drop switch-lora struct comment
c57b00f9
granite-switch: guard router layer index and inline n_slots
d2ca90b3
granite-switch: group adapter metadata under {arch}.adapters.* namespace
11890d3e
granite-switch: add hparams.has_rope(il) for KV-shift rope skipping
ae2ae69b
granite-switch: skip arch in test-llama-archs (adapter fixture missin…
d8860737
granite-switch: Keys.Adapters namespace + simplify n_slots
2204806a
granite-switch: validate substitute token ids against n_vocab
56698576
granite-switch: bound adapter count and lora rank from GGUF
a7271ce8
granite-switch: reject MTP context type when router_layer is set
64278bb1
granite-switch: throw on bad adapter metadata instead of GGML_ASSERT
34f5607c
granite-switch: use ASCII +/- in router K signal comment
7cc8ba12
granite-switch: document n_layer_nextn repurpose and its leak points
118cf8a8
granite-switch: gate lora_a/lora_b op mapping on router_layer
65cb2018
granite-switch: label all three preview model sizes
15e473d6
CISC
approved these changes
on 2026-08-04
Reviewers
Assignees
No one assigned
Labels
model
testing
conversion
Milestone
No milestone
Login to write a write a comment.
Login via GitHub