chore: nightly sync main into dev (22_06_2026)

[svcnvidia-nemo-ci]

Summary

Automated nightly sync of main → dev for 22_06_2026.

• Merged 42 commits from origin/main into origin/dev (merge commit, origin/dev + origin/main parents preserved).
• Python code surface: Python lines: +6198 / -395 across 41 files (excludes uv.lock, golden-value JSON, etc.).
• This run supersedes the earlier same-day PR chore: nightly sync main into dev (22_06_2026) #5429, which was closed because main advanced by 3 commits after it merged. Those 3 newest commits are included here:
  • fc4597c0c Disag MR1: Add inference shard specs and pg-collection building (Disag MR1: Add inference shard specs and pg-collection building #5186)
  • ae2efd53a Disag MR2: Refit into multiple destination pools and tied-embedding + UVM fixes (Disag MR2: Refit into multiple destination pools and tied-embedding + UVM fixes #5187)
  • 57c484e02 Fix Mamba prefix match for chunked prefill (Fix Mamba prefix match for chunked prefill #4758)

Merge strategy

Started from origin/dev, ran git merge origin/main --no-edit, and resolved 16 conflicts surgically (no global -X theirs). The dev branch has a substantially diverged inference stack; conflicts were resolved to preserve dev features. Post-review exceptions are documented below: theoretical_memory_usage.py intentionally favors main's LatentMoE memory-accounting fix from #5145, and #5493 restores main's fused group MLP offload support from #5082.

Dependency triple & governance — kept dev's versions verbatim

• pyproject.toml, uv.lock, docker/Dockerfile.ci.dev — identical to dev (dev's TE revision, emerging_optimizers v0.3.0, nvidia-resiliency-ext, PyPI mamba-ssm). No git-source reconciliation needed.
• .github/CODEOWNERS — identical to dev (never modified).

Files taken from main (override list)

• megatron/training/training.py — main's version (RL-inference import retargeted to megatron.core.inference.shards).
• initialize.py, datasets/data_samplers.py, optimizer/layer_wise_optimizer.py — resolved to dev (already equal to dev; consistent with dev-based arguments.py).

Conflicts resolved toward dev (diverged inference / MLA / offload stacks)

multi_latent_attention.py, absorbed_mla.py, fine_grained_activation_offload.py, rope_utils.py, gpt_model.py, moe/experts.py, moe/router.py, transformer_config.py, checkpointing.py, pretrain_gpt.py, pretrain_hybrid.py, inference config.py / text_generation_controller.py / mamba_slot_allocator.py, and matching unit tests. arguments.py resolved to dev + the new --inference-shards arg.

Intentional main-over-dev follow-up

• LatentMoE theoretical memory estimate (Fix LatentMoE theoretical memory estimate #5145) — favored main's fix in megatron/training/theoretical_memory_usage.py and restored its focused unit test in tests/unit_tests/training/test_weight_and_optimizer_memory.py. This intentionally replaces the old dev routed-expert formulas; main merge commit: 1bcb3b9e0b029c6109d8d32872645bd540052963.
• Fused group MLP activation offload ([feat] Support fine-grained activation offloading in fused group mlp #5082, restored by fix: restore fused group MLP offload in main2dev sync #5493) — restored main's fused group MLP offload support across fine_grained_activation_offload.py, moe/experts.py, transformer_config.py, docs, and the matching offload tests. This intentionally replaces the older dev-only documentation/config/code paths for expert_fc1 / moe_act-only offload; main merge commit: 5c660c3b4a9d91e7ed32997b56cd6d333ed04bc5.

New main features brought in additively

• megatron/core/inference/shards.py, shards_spec.py, tests/unit_tests/inference/test_inference_shards.py — the inference shard-spec API + --inference-shards arg (Disag MR1). Self-contained (only depend on mpu / HyperCommGrid / ProcessGroupCollection), wired through arguments.py → shards_spec.py.
• Knowledge-distillation modules (megatron/training/distillation/*), megatron/rl/rl_profiling.py, FSDP experimental/*, MIMO runtime, and all cleanly-merged main changes.

Deferred (kept dev; needs human reconciliation)

The runtime integration from the 3 newest commits modifies dev files that have diverged reimplementations, so taking main would drop dev lines (guard violation). Deferred to keep dev's stack intact and CI green:

• Mamba prefix matching (Fix Mamba prefix match for chunked prefill #4758) — dev keeps its own _find_mamba_match_count / prefix-match implementation in dynamic_engine.py / dynamic_context.py.
• Disag runtime refit (Disag MR1: Add inference shard specs and pg-collection building #5186/Disag MR2: Refit into multiple destination pools and tied-embedding + UVM fixes #5187) — refit.py, dynamic_engine.py, dynamic_context.py, benchmark_refit.py, and the CUDA-graph admission-gating path kept at dev. (planner.py / unified_memory.py came in as main's purely-additive versions; cross-boundary imports verified.)

Deletions / restorations

• .github/workflows/sync-skills.yml — removed, matching main's intentional deletion (41dbab44a, ci: Remove sync skills workflow #5091). Dev's removal of multi-approval-bot.yml (ci: Remove multi-approval action from dev branch #3576) and the h100/{bert,t5}.yaml recipes is preserved.
• megatron/rl/parallel_utils.py — restored to dev. Main moved build_inference_pg_collection into shards.py; since the Disag refit refactor is deferred, dev's benchmark_refit.py / test_model_swap.py callers still need it. Both modules coexist (no broken imports).
• tests/unit_tests/inference/engines/test_cg_admission_gating.py — not included (main-new test for the admission-gating path, which is deferred with dev's dynamic_engine.py).

Local validation

• Pre-push guard: passed for CODEOWNERS and dep-triple. The dev-feature audit reports the expected intentional replacements documented above for Fix LatentMoE theoretical memory estimate #5145 (1bcb3b9e0b029c6109d8d32872645bd540052963) and [feat] Support fine-grained activation offloading in fused group mlp #5082/fix: restore fused group MLP offload in main2dev sync #5493 (5c660c3b4a9d91e7ed32997b56cd6d333ed04bc5).
• Lint (CI parity over changed megatron/core + tests/): black, isort, ruff clean; pylint 10.00/10.
• All changed Python parses; cross-boundary inference imports resolve.

CI to follow via /ok to test.

[FDecaYed]

/ok to test 213314b

[FDecaYed]

I see @Phlip79 fixed #5145 issue and #5493 is merged to fix missing #5082

approving

[svcnvidia-nemo-ci]

Superseded by today's nightly sync.

[Phlip79]

/ok to test f6c77e7