refactor(runtime)!: adopt the modern Infini stack - #506

Draft

voltjia wants to merge 54 commits into

mainfrom

refactor/adopt-modern-infini-stack

Commits on Aug 25, 2026

refactor(runtime)!: adopt modern Infini stack
voltjia
committed
04fcda0
docs: define Infini stack repository boundaries
voltjia
committed
e9496ca
refactor: own Infini stack integration build
voltjia
committed
05c92ba
docs: document the InfiniLM stack build
voltjia
committed
bca98df
ci: validate the modern NVIDIA stack
voltjia
committed
f46c074
fix: keep static graph cache metadata dynamic
voltjia
committed
dab847b
refactor(ops): use canonical InfiniOps APIs
voltjia
committed
d78dacb
fix(build): select required linked InfiniOps providers
voltjia
committed
d49427f
fix(build): limit compiled models to Qwen3
voltjia
committed
f128747
feat(distributed): add eager point-to-point wrappers
voltjia
committed
0cbc125
feat(ops): restore prepacked linear execution
voltjia
committed
9f40398
fix(ops): match canonical Gemm call schema
voltjia
committed
28cc18b
fix(sampling): support scalar sample outputs
voltjia
committed
7dd8ac6
fix(nn): pass RoPE output handles by value
voltjia
committed
80af271
fix(attention): include variable-length MHA declaration
voltjia
committed
3af4355
fix(engine): use local static graph assertions
voltjia
committed
0cb41af
fix(config): reject unsupported Qwen3 linear bias
voltjia
committed
6fcd5b7
test: cover modern runtime contracts
voltjia
committed
4e36168
docs: update modern Qwen3 support boundary
voltjia
committed
354372a
refactor(build): select InfiniOps implementations from JSON
voltjia
committed
1a89775
style: format InfiniOps attention adapters
voltjia
committed
2c2d94b
fix(graph): update paged replay metadata through tensors
voltjia
committed
baf3315
fix(graph): fall back to eager paged decode under tensor parallelism
voltjia
committed
1551767
fix: restore modern InfiniCore API compatibility
voltjia
committed
c76b105
fix(runtime): sync post-migration InfiniCore updates
voltjia
committed
9d64e14
style: format migrated sources
voltjia
committed
5964127
fix(runtime): keep receive buffers mutable
voltjia
committed
e4a549b
fix(build): exclude unsupported MXFP4 path
voltjia
committed
1ed454f
fix(packaging): preload Torch shared libraries
voltjia
committed
33f0409

Commits on Aug 26, 2026

Commits on Aug 27, 2026

Commits on Aug 31, 2026

fix(graph): enable safe paged decode capture under TP
voltjia
committed
dcfbebc

Commits on Sep 2, 2026

Commits on Sep 3, 2026

Merge pull request #556 from InfiniTensor/feat/cambricon-tp-flash-attention
voltjia
authored
88e4c6e

Commits on Sep 8, 2026

feat(iluvatar): enable modern Infini stack inference (#557)
gongchensu
authored
7c39376

Commits on Sep 9, 2026

Commits on Sep 10, 2026

fix(ascend): use CommInitAll for single-node TP (#566)
baominghelly
authored
b6fbf94

Commits on Sep 11, 2026

Commits on Sep 21, 2026

fix(runtime): address inference migration review feedback
voltjia
committed
ebf53a6

Commits on Sep 22, 2026

Merge pull request #592 from InfiniTensor/fix/inference-review-feedback
pengcheng888
authored
1d06427