Repository navigation

refactor: decouple cache engine with model, remove all hardcode method. - #34

Merged

Hermit-w merged 17 commits into

THU-MIG:mainfrom

Hermit-w:refactor/cache_framework

Jul 14, 2026

Commits on Jul 14, 2026

refactor: migrate Qwen on-GPU feature reuse to the declarative device-backed cache slot
Hermit-w
committed
42801d5
docs: correct CacheStateManager transaction comments to match reality
Hermit-w
committed
82b3d55
refactor: remove unused reserved cache code with no callers
Hermit-w
committed
6df0fd2
refactor: remove dead inject_feature_gpu hook surface and variant_uses_seam
Hermit-w
committed
0780d0a
fix: reject Output-granularity caching under CFG-parallel to avoid silent per-rank skip divergence
Hermit-w
committed
2fe6c17
feat: migrate Qwen MagCache/DiCache to a unified substep loop behind ED_CACHE_SUBSTEP
Hermit-w
committed
773b004
feat: drive Qwen MagCache capture through a tap registry instead of the cache-scope hook
Hermit-w
committed
3b9ded6
feat: drive Qwen DiCache probe metrics through the tap registry
Hermit-w
committed
347777a
feat: migrate Flux MagCache/DiCache to the tap registry
Hermit-w
committed
b0513a8
feat: add Wan cache taps and gate substep on device availability
Hermit-w
committed
2a281f4
feat: complete Wan host-path tap migration for MagCache and DiCache
Hermit-w
committed
d788e27
feat: drive Wan substep reuse inject through the tap registry (dehook)
Hermit-w
committed
d0f8fe0
feat: drive Qwen/Flux/Wan substep reuse inject through the tap registry (dehook)
Hermit-w
committed
b1e8a44
feat: migrate SD3 (mmdit) MagCache/DiCache to the tap registry
Hermit-w
committed
3e167ab
feat: substep-migrate the remaining cache methods (Output + Feature)
Hermit-w
committed
587ab7f
refactor: make the substep loop the sole cache execution path; delete legacy execute
Hermit-w
committed
a2831c6
refactor: remove dead code left after the legacy execute deletion
Hermit-w
committed
b881d61

Navigation Menu

Repository navigation

refactor: decouple cache engine with model, remove all hardcode method. - #34

refactor: decouple cache engine with model, remove all hardcode method.#34
Hermit-w merged 17 commits into
THU-MIG:mainfrom
Hermit-w:refactor/cache_framework

Commits on Jul 14, 2026

refactor: migrate Qwen on-GPU feature reuse to the declarative device-backed cache slot

docs: correct CacheStateManager transaction comments to match reality

refactor: remove unused reserved cache code with no callers

refactor: remove dead inject_feature_gpu hook surface and variant_uses_seam

fix: reject Output-granularity caching under CFG-parallel to avoid silent per-rank skip divergence

feat: migrate Qwen MagCache/DiCache to a unified substep loop behind ED_CACHE_SUBSTEP

feat: drive Qwen MagCache capture through a tap registry instead of the cache-scope hook

feat: drive Qwen DiCache probe metrics through the tap registry

feat: migrate Flux MagCache/DiCache to the tap registry

feat: add Wan cache taps and gate substep on device availability

feat: complete Wan host-path tap migration for MagCache and DiCache

feat: drive Wan substep reuse inject through the tap registry (dehook)

feat: drive Qwen/Flux/Wan substep reuse inject through the tap registry (dehook)

feat: migrate SD3 (mmdit) MagCache/DiCache to the tap registry

feat: substep-migrate the remaining cache methods (Output + Feature)

refactor: make the substep loop the sole cache execution path; delete legacy execute

refactor: remove dead code left after the legacy execute deletion