llama.cpp
metal : single-source fusion table + fusion debug rework
#28164
Merged

metal : single-source fusion table + fusion debug rework #28164

ggerganov merged 44 commits into master from gg/metal-fuse-rework
ggerganov
github-actions github-actions added testing
github-actions github-actions added ggml
github-actions github-actions added Apple Metal
github-actions github-actions added devops
ggerganov ggerganov force pushed from dea34cf8 to 4cd2466d 13 days ago
ggerganov
ggerganov commented on 2026-09-01
ggerganov
ggerganov commented on 2026-09-01
ggerganov
ggerganov commented on 2026-09-01
ggerganov
ggerganov commented on 2026-09-01
github-actions github-actions added model
ggerganov ggerganov marked this pull request as ready for review 12 days ago
ggerganov ggerganov requested a review from JohannesGaessler JohannesGaessler 12 days ago
ggerganov ggerganov requested a review from CISC CISC 12 days ago
ggerganov ggerganov requested a review 12 days ago
ggerganov ggerganov force pushed from 7d59361e to 4e290cd2 11 days ago
ggerganov ggerganov force pushed from 4e290cd2 to a050bce3 9 days ago
ggerganov ggerganov force pushed from a050bce3 to 0d774833 6 days ago
ggerganov metal : rework fusion patterns into a single table
18292605
ggerganov metal : fix absolute output indices in fusion patterns
0ef37acc
ggerganov metal : fuse gated_delta_net with cache cpy
667e9844
ggerganov metal : drop is_view_consumer mem-range skip
14b322bb
ggerganov metal : rename gated_delta_net fused state output args
cf6e83ff
ggerganov metal : rename raw fusion flag to unsafe
029aa425
ggerganov metal : tidy fusion pattern checks and table
1f936075
ggerganov metal : add generic fusion stats via ad-hoc proc-address API
945c4175
ggerganov tests : add fusion count regression test with per-backend baseline
f75b65ee
ggerganov tests : rename fusion api helpers to match stats_init signature
732f4e2b
ggerganov tests : rename backend to device in fusion test CLI
868f0f8d
ggerganov tests : add --model and --help to fusion test
6caa5ebf
ggerganov tests : use backend base name for fusion baseline output
634ea902
ggerganov tests : run fusion test from ci instead of ctest
eb0a57bc
ggerganov tests : align fusion baseline TSV columns
688e456e
ggerganov tests : widen label column and align fusion TSV header
41520d2e
ggerganov tests : switch fusion baseline from TSV to CSV
01b4bb71
ggerganov cont : rebase + update MTL stats
00cfe64e
ggerganov tests : avoid graph reallocations for some archs
f3282661
ggerganov metal : tidy fusion debugging context and op init
8930c55a
ggerganov tests : dedup fusion baseline into any mode
9043c2f2
ggerganov ci : move fusion model generation to a separate step
d0b75aad
ggerganov tests : bump nmse thold
d875074c
ggerganov models : fix plamo2 graph
8230a310
ggerganov tests : remove "skip" logic from test-fusion
60f2f382
ggerganov tests : set qwen3tts dummy vocab to codec head size
2911254f
ggerganov tests : regen fusion baseline
4569f517
ggerganov ci : skip dummy model generation on OpenVINO
a41d19ff
ggerganov cont : minor
382b4152
ggerganov tests : enable test-llama-archs on windows
619dc677
ggerganov cont : disable on windows + workaround
b34648bf
ggerganov metal : naming nits
74adf022
ggerganov ggerganov force pushed from 0d774833 to 74adf022 5 days ago
ggerganov test-fusion : add instructions to update baseline
5c106d5d
ggerganov context : fix Kimi-K3 graph reserve
9c2f9780
ggerganov fusion : update MTL
67904157
ggerganov cont : fix naming
dda1743f
ggerganov metal : rework fusion info storage
45169520
ggerganov metal : align fusion info API
8d511031
ggerganov metal : use opaque fusion handle in ad-hoc API
f11b7fb4
ggerganov ggerganov changed the title metal : single-source fusion table + gated_delta_net cache fusion metal : single-source fusion table + fusion debug rework 4 days ago
ggerganov ci : move fusion test to dedicated workflow
9be32954
ggerganov ggerganov requested a review 4 days ago
ggerganov cont : run only on ggml changes
8ae00631
ggerganov cont : simplify
f1f5b542
ggerganov fusion : remove multi-output stuff for now
a9e14017
ggerganov ci : fix typo
a200ae82
ggerganov ggerganov merged a2878d30 into master 4 days ago
ggerganov ggerganov deleted the gg/metal-fuse-rework branch 4 days ago

Login to write a write a comment.

Login via GitHub

Assignees
No one assigned
Labels
Milestone