model: GraniteSWAForCausalLM / GraniteMoeSWAForCausalLM #25505
CISC
merged
31 commits into
ggml-org:master
from
gabe-l-hart:GraniteSWAForCausalLM
feat(convert): Add conversion for GraniteSWAForCausalLM
4bbf9aba
feat(llama): Add granite_swa support
c74049a8
feat(conversion): Add conversion infra for rope_pattern array
1f3741a5
fix(conversion): Fix SWA pattern logic and support for non-rope layers
f7d057b7
feat(conversion): Add support for GraniteMoeSWA
44215bb1
feat: Add llama_hparams::has_rope and arch constants
cf315175
feat: Add support for per-layer rope determination
f5d4e9f4
style: Fix failing flake8 for extra newlines
9d75ea8a
test: Write out SLIDING_WINDOW_PATTERN in llama-model-saver
ef14e1c3
fix(convert): Fix missing registration for GraniteMoeSWAForCausalLM
c32fe0bf
fix: Load MoE params as optional
36aa2e02
feat: Handle MoE params in conversion
94d2643c
Merge remote-tracking branch 'origin/master' into GraniteSWAForCausalLM
dc7f155f
style: Remove unnecessary newline
90c3d646
fix: Remove unnecessary tensor additions to GRANITE architecture
0b1df4c5
fix: Correctly handle naming for ffn gate inp
8865f1bd
gabe-l-hart
commented
on 2026-08-12
gabe-l-hart
commented
on 2026-08-12
fix: Always default hparams.rope_pattern to 1s
8e5bcadb
feat: Move to has_rope for all granite model architectures
977437b0
feat: No hacky rope_finetuned fallback in has_rope
e400df6f
gabe-l-hart
commented
on 2026-08-12
fix: Fully remove rope hparam filling in granitemoe
e7695cf3
gabe-l-hart
commented
on 2026-08-12
gabe-l-hart
commented
on 2026-08-12
fix: Save out rope_pattern in model-saver
03067047
fix: Set hparams.rope_finetuned for round trip
93f6cb8c
pwilkin
approved these changes
on 2026-08-12
fix: Code review cleanup
600615dc
refactor: Keep gate/up fused for MoE path
94fc69fd
Merge remote-tracking branch 'origin/master' into GraniteSWAForCausalLM
283c4ec4
fix: Skip GRANITE_SWA in model saver
c94d5ac9
gabe-l-hart
commented
on 2026-08-17
add sliding window pattern for model in test
b5b7d1d6
style: Fix indentation
f7b13112
fix: Fix \r\n
eeb83f3e
feat: Keep shared expert fused
d7a74f1e
CISC
approved these changes
on 2026-08-19
CISC
commented
on 2026-08-19
CISC
approved these changes
on 2026-08-19
style: More indentation fixes
de5f6fd5
CISC
approved these changes
on 2026-08-19
Login to write a write a comment.
Login via GitHub