Repository navigation

Improve model caching and experimental distributed weight access - #1

Merged

whjthu merged 29 commits into

masterfrom

improve-model-cache

Oct 27, 2026

Merged

Improve model caching and experimental distributed weight access#1
whjthu merged 29 commits into
masterfrom
improve-model-cache

Commits on Oct 26, 2026
- Protect sealed weight snapshots and daemon runtime ownership
  whjthu
  committed
  ea57e3c
- Bound IPC requests and cancel failed distributed reads
  whjthu
  committed
  63e7ea4
- Publish verified disk cache generations with lease-aware collection
  whjthu
  committed
  97335f6
- Enforce container memory limits and shared resident budgets
  whjthu
  committed
  8cca2e6
- Add leased commands and cache lifecycle diagnostics
  whjthu
  committed
  e2c2065
- Document cache operations and add Linux regression CI
  whjthu
  committed
  7454e9c
- Add explicit tensor sharding with atomic streaming export
  whjthu
  committed
  48615de
- Share sealed CPU tensors with live topology slicing
  whjthu
  committed
  43d7a75
- Validate distributed TP and PP execution after coordinated preload
  whjthu
  committed
  70c0979
- Manage GPU expert residency with CPU fallback and safe eviction
  whjthu
  committed
  895c5ef
- Document validated history and development priorities
  whjthu
  committed
  33b176e
- Close cached descriptors across interrupted and partial loads
  whjthu
  committed
  8dc5f34
- Normalize runtime paths and reject invalid cache locations
  whjthu
  committed
  2b6522d
- Publish runtime sockets safely and wait for cache shutdown
  whjthu
  committed
  5acb8b7
- Share strict safetensors validation across planning commands
  whjthu
  committed
  8fe11ec
- Separate CLI arguments and derive completion from the parser
  whjthu
  committed
  da2efa2
- Clarify cache usage and archive historical development records
  whjthu
  committed
  f6ba2e5
- Keep rank view metadata out of source file inventories
  whjthu
  committed
  92afa75
- Recheck disk cache usage under the collection lease
  whjthu
  committed
  6693f9b
- Forward run cancellation and support explicit daemon sockets
  whjthu
  committed
  9b0de02
- Discount completed resident files from pending headroom reservations
  whjthu
  committed
  d515001
- Use one tensor layout contract for planning and runtime views
  whjthu
  committed
  5360338
- Keep the latest expert completion event per consumer stream
  whjthu
  committed
  3bd87d5
- Document lease transfer limits and remaining preparation costs
  whjthu
  committed
  4412029
- Preserve inherited ignored signals in foreground consumers
  whjthu
  committed
  31fde8b
- Retain recovery state until runtime cleanup completes
  whjthu
  committed
  df177cf
- Start daemon readiness waits after local cache preparation
  whjthu
  committed
  ca908bb
Commits on Oct 27, 2026
- Release private socket ownership after successful publication
  whjthu
  committed
  146c256
- Isolate default signal tests from inherited dispositions
  whjthu
  committed
  8aabad2

Commits

Commits on Oct 26, 2026

Commits on Oct 27, 2026