llama.cpp
metal : add MoE and SSM_CONV fusion optimizations
#28948
Merged
Go
Login via GitHub
Home
Pricing
FAQ
Install
Login
via GitHub
Overview
Commits
40
Changes
View On
GitHub
metal : add MoE and SSM_CONV fusion optimizations
#28948
ggerganov
merged 40 commits into
master
from
gg/metal-fusion-qwen3.6
github-actions
added
testing
github-actions
added
ggml
github-actions
added
Apple Metal
ggerganov
commented on 2026-09-15
ggerganov
force pushed
from
69a1dfb8
to
996143df
40 days ago
github-actions
added
examples
ggerganov
marked this pull request as ready for review
40 days ago
ggerganov
requested a review
40 days ago
metal : add top-k MoE fusion
078a811c
metal : add MoE weighted reduction fusion
2fecc528
tests : expose MoE weighted reduction in fusion baseline
057f7a60
metal : fuse RMS_NORM + SCALE
627172fe
metal : use function constant for RMS_NORM + SCALE
4788328e
metal : use function constant for top-k MoE with_norm
bef5ed1e
metal : rename moe_weighted_reduction suffix to moe_reduce
9c0768f5
metal : add MUL_MAT + UNARY and MUL_MAT + ADD + UNARY fusion
c318506a
metal : revert MUL_MAT + UNARY and MUL_MAT + ADD + UNARY fusion
1dc1e3db
metal : add SSM_CONV + UNARY (silu) fusion
3e9c3b4b
metal : address fusion review comments
4f387a7d
metal : fix unused parameter in top-k MoE fusion check
990c5d85
metal : guard SSM_CONV fusion lookup behind use_fusion
0264daa3
metal : track all fused outputs in graph reorder
ba10d2f8
metal : keep top-k MoE logits alive until fused output
5120797a
metal : refactor alloc deps to pattern-driven approach
77d81fbc
metal : check fused kernel destination in concurrency tracking
060952ab
meta : forward graph_optimize to underlying backends
ba130719
metal : use vector for fusion table
c5779d6a
meta : keep graph_optimize unimplemented
ec59d32d
parallel : fix non-deterministic prompt selection
33ceb68b
parallel : support dummy models and add global logits run hash
6e06cd46
metal : sync cross-device copies with destination completion event
9e5ce6f2
metal : avoid const_cast in fusion alloc deps
9ee17c2d
metal : skip fusions with aliased sources
da289484
metal : hide fusion pattern definition
b968e349
metal : use vector fusion op sequences
9b0362d0
metal : drop redundant struct keywords
f050011f
metal : add alloc deps comment separator
a4fb973f
metal : generalize fusion output memory ranges
363b4774
metal : rename fusion out_offsets to outs
f4057093
metal : avoid dst vector in memory range check
5c05aa32
metal : optimize fusion matching and multi-output handling
78cf4d12
Revert "parallel : support dummy models and add global logits run hash"
66cafead
fusion : update MTL.csv
551b96c8
metal : unroll constant loops in top-k MoE kernel
22fb9bcf
metal : use function constants for top-k MoE n_expert and top_k
1b13514d
metal : rename fusion kargs to scale and clamp
bde2ee59
metal : use function constants for moe_reduce and ssm_conv
a066528b
fusion : update MTL.csv
9bf10776
ggerganov
force pushed
from
03eddaba
to
9bf10776
40 days ago
ggerganov
merged
5b59b83f
into master
39 days ago
ggerganov
deleted the gg/metal-fusion-qwen3.6 branch
39 days ago
ggerganov
added
highlight
Login to write a write a comment.
Login via GitHub
Reviewers
No reviews
Assignees
No one assigned
Labels
testing
examples
ggml
Apple Metal
highlight
Milestone
No milestone
Login to write a write a comment.
Login via GitHub