Repository navigation

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

Open

voltjia wants to merge 58 commits into

mainfrom

refactor/adopt-modern-infini-stack

Commits on Sep 27, 2026
- refactor(runtime)!: adopt modern Infini stack
  
  voltjia
  authored and
  wooway777
  committed
  54b797c
- docs: define Infini stack repository boundaries
  
  voltjia
  authored and
  wooway777
  committed
  93a161f
- refactor: own Infini stack integration build
  
  voltjia
  authored and
  wooway777
  committed
  da76f3f
- docs: document the InfiniLM stack build
  
  voltjia
  authored and
  wooway777
  committed
  05fad31
- ci: validate the modern NVIDIA stack
  
  voltjia
  authored and
  wooway777
  committed
  b15909a
- fix: keep static graph cache metadata dynamic
  
  voltjia
  authored and
  wooway777
  committed
  2153f7d
- refactor(ops): use canonical InfiniOps APIs
  
  voltjia
  authored and
  wooway777
  committed
  1a8d3d4
- fix(build): select required linked InfiniOps providers
  
  voltjia
  authored and
  wooway777
  committed
  2d34490
- fix(build): limit compiled models to Qwen3
  
  voltjia
  authored and
  wooway777
  committed
  d9aaaf4
- feat(distributed): add eager point-to-point wrappers
  
  voltjia
  authored and
  wooway777
  committed
  f282577
- feat(ops): restore prepacked linear execution
  
  voltjia
  authored and
  wooway777
  committed
  fea3dad
- fix(ops): match canonical Gemm call schema
  
  voltjia
  authored and
  wooway777
  committed
  69d0ff2
- fix(sampling): support scalar sample outputs
  
  voltjia
  authored and
  wooway777
  committed
  4c541c9
- fix(nn): pass RoPE output handles by value
  
  voltjia
  authored and
  wooway777
  committed
  c84a862
- fix(attention): include variable-length MHA declaration
  
  voltjia
  authored and
  wooway777
  committed
  9e380f7
- fix(engine): use local static graph assertions
  
  voltjia
  authored and
  wooway777
  committed
  30d90e8
- fix(config): reject unsupported Qwen3 linear bias
  
  voltjia
  authored and
  wooway777
  committed
  5662890
- test: cover modern runtime contracts
  
  voltjia
  authored and
  wooway777
  committed
  54d95b3
- docs: update modern Qwen3 support boundary
  
  voltjia
  authored and
  wooway777
  committed
  87a4087
- refactor(build): select InfiniOps implementations from JSON
  
  voltjia
  authored and
  wooway777
  committed
  93ae2c6
- style: format InfiniOps attention adapters
  
  voltjia
  authored and
  wooway777
  committed
  2a1880b
- fix(graph): update paged replay metadata through tensors
  
  voltjia
  authored and
  wooway777
  committed
  2809bec
- fix(graph): fall back to eager paged decode under tensor parallelism
  
  voltjia
  authored and
  wooway777
  committed
  9b00e5f
- fix: restore modern InfiniCore API compatibility
  
  voltjia
  authored and
  wooway777
  committed
  3b4543b
- fix(runtime): sync post-migration InfiniCore updates
  
  voltjia
  authored and
  wooway777
  committed
  a6f8c59
- style: format migrated sources
  
  voltjia
  authored and
  wooway777
  committed
  52c5073
- fix(runtime): keep receive buffers mutable
  
  voltjia
  authored and
  wooway777
  committed
  92d809f
- fix(build): exclude unsupported MXFP4 path
  
  voltjia
  authored and
  wooway777
  committed
  2792bad
- fix(packaging): preload Torch shared libraries
  
  voltjia
  authored and
  wooway777
  committed
  eeb55f0
- refactor(runtime): replace deprecated causal softmax backend
  
  voltjia
  authored and
  wooway777
  committed
  8cc81f3
- style: format causal softmax adapter
  
  voltjia
  authored and
  wooway777
  committed
  49a7f26
- feat(runtime): unlock validated InfiniOps capabilities
  
  voltjia
  authored and
  wooway777
  committed
  55e1622
- fix(runtime): select available InfiniOps implementations
  
  voltjia
  authored and
  wooway777
  committed
  bc31217
- fix(runtime): enable available Moore InfiniOps paths
  
  voltjia
  authored and
  wooway777
  committed
  2b4e9c8
- fix(runtime): accept native MetaX device name
  
  voltjia
  authored and
  wooway777
  committed
  7e6d100
- fix(core): validate Cat inputs against tensor rank
  
  voltjia
  authored and
  wooway777
  committed
  6261df2
- feat(models): enable validated MiniCPM Eagle path
  
  voltjia
  authored and
  wooway777
  committed
  ebce722
- style(runtime): group ModelRunner imports
  
  voltjia
  authored and
  wooway777
  committed
  9445436
- fix(graph): enable safe paged decode capture under TP
  
  voltjia
  authored and
  wooway777
  committed
  b700358
- feat(moore): enable modern flash attention
  
  voltjia
  authored and
  wooway777
  committed
  cf92e39
- style(moore): format attention integration
  
  voltjia
  authored and
  wooway777
  committed
  fbcad61
- build(moore): configure the modern stack
  
  voltjia
  authored and
  wooway777
  committed
  822cd66
- fix(distributed): avoid transient runtime teardown
  
  voltjia
  authored and
  wooway777
  committed
  8a2f6e2
- feat(cambricon): enable tensor-parallel flash attention
  
  baominghelly
  authored and
  wooway777
  committed
  56accff
- feat(iluvatar): enable modern Infini stack inference (#557)
  
  gongchensu
  authored and
  wooway777
  committed
  ddc27f1
- feat(hygon): enable the modern Infini stack (#564)
  
  gongchensu
  authored and
  wooway777
  committed
  a8b4c1e
- feat(ascend): enable InfiniOps flash attention (#558)
  
  baominghelly
  authored and
  wooway777
  committed
  70b5154
- perf(nvidia): reduce refactored inference overhead (#565)
  
  voltjia
  authored and
  wooway777
  committed
  90ddf65
- fix(ascend): use CommInitAll for single-node TP (#566)
  
  baominghelly
  authored and
  wooway777
  committed
  dcbf3b6
- fix(ascend): keep unsafe operators out of device graphs (#568)
  
  baominghelly
  authored and
  wooway777
  committed
  a84c58a
- fix(allocator): reuse pinned free blocks (#569)
  
  voltjia
  authored and
  wooway777
  committed
  1e261e5
- fix(runtime): address inference migration review feedback
  
  voltjia
  authored and
  wooway777
  committed
  9ef8a20
- fix(attention): remove paged attn
  
  pengcheng888
  authored and
  wooway777
  committed
  6661765
- fix(inference): preserve migration interfaces and backend selection
  
  voltjia
  authored and
  wooway777
  committed
  1b7f61f
- fix-inference-keep-thead-as-canonical-ppu-device
  wooway777
  committed
  03dab96
- fix(attention): select active flash attention provider
  wooway777
  committed
  e649abc
- fix(attention): remove legacy flash bridges
  wooway777
  committed
  d2a8504
- style: format attention adapter
  wooway777
  committed
  f94d83d

Commits

Commits on Sep 27, 2026