Skip to content

mixture optimization batches - #217

Merged
isLinXu merged 11 commits into
mainfrom
codex/mixture-optimization-batches
Aug 2, 2026
Merged

mixture optimization batches#217
isLinXu merged 11 commits into
mainfrom
codex/mixture-optimization-batches

Conversation

@isLinXu

@isLinXu isLinXu commented Aug 2, 2026

Copy link
Copy Markdown
Collaborator

No description provided.

isLinXu added 11 commits August 2, 2026 18:49
Replace the RoutedModule protocol with minimal dataclass-based adapters:
RoutingMetrics, normalize_routing_snapshot, routing_metrics, usage_gini.
Export the new helpers from the moe facade and cover them with
test_moe_facade_protocol.py.
Extend _LatentAuxMixin and LatentMixture with additional aux-loss
handling and routing snapshot fields; expand latent mixture tests.
…rnels

Add _can_vectorize_linear_experts/_compute_vectorized_linear_experts to
run same-rank Linear experts in two batched kernels; update sparse
dispatch and routing-aware merge tests.
Introduce PlannerResult dataclass with to_dict/from_dict plus adapters
from legacy decisions and placement plans; export from vpeft facade and
extend placement schema and V-PEFT e2e tests.
…lback reasons

Attach PlannerResult via _attach_planner_result and record serializable
fallback reasons via _record_vpeft_fallback in apply_lora; extend
LoRAConfig fields, default.yaml options, and config drift detector
coverage with matching tests.
MoABlock and C2fMoA referenced sparse_inference/sparse_inference_threshold
without declaring them, and default.yaml lacked the moa_sparse_inference
keys mapped by moe/config.py.
Carry over the sorted-form Gini improvement from main's schedule.py into
the canonical protocol implementation.
@isLinXu
isLinXu merged commit a1cc71e into main Aug 2, 2026
12 of 17 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant