chore(beep boop 🤖): Bump uv.lock (main, mcore-dev) (2026-05-20)

[svcnvidia-nemo-ci]

🚀 PR to bump uv.lock in main.

🤖 This PR will be merged automatically once CI passes.

[svcnvidia-nemo-ci]

/ok to test 8e84cec

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[yaoyu-33]

MCore bump auto-fix status for dev:

Classification: MCore broke Bridge
Evidence: The current main/mcore-dev bump #3903 is still failing on 2026-05-20. gh run view 26156612801 --log-failed returned an empty log, so I used the Actions job-log API. Failed job L0_Launch_models_nemotron_omni / 76952734284 and L0_Launch_recipes_nemotron_omni / 76952734293 fail at src/megatron/bridge/models/nemotron_omni/nemotron_omni_provider.py:257 with TypeError: LLaVAModel.__init__() got an unexpected keyword argument 'dynamic_resolution'. In the checked-out bump tree, MCore dev commit bd5c98f7796259362a2e8c99b5c7b0d92b4e9d84 lacks the LLaVA constructor parameters present on MCore main commit 38986a98aae6a0cc4c8ae7b435db3288a890b0cb: dynamic_resolution, RADIO controls, sound model/projection fields, and temporal video fields. L0_Launch_models_wan / 76952734267 fails because MCore dev now calls DiTSelfAttention.get_query_key_value_tensors(..., head_wise_gate=...). gb200_L0_Launch_recipes_deepseek_fsdp / 76952734283 reaches validation but fails convergence with correlation=0.510130, final_loss_current=7.510771, final_loss_golden=8.127562, and failed metrics correlation, final_loss.
Fix PR: not opened
Guards: none added, none removed. I audited existing guard patterns and confirmed the relevant current guards use local inspect.signature(...) checks with TODO removal comments, but I did not add another guard because the Nemotron Omni LLaVA gap is broader than one kwarg and the DeepSeek convergence change is not safe to patch blindly.
Validation: Read Linear MB-401 first, re-read #3903 metadata/diff/checks/run state, inspected MCore compare cf081d5df0b34b665d214ff936f27489d7396876...bd5c98f7796259362a2e8c99b5c7b0d92b4e9d84 (15 commits ahead), checked for existing open mcore-dev-autofix / #3903 / dynamic_resolution / head_wise_gate / DiTSelfAttention / deepseek_fsdp fix PRs and found none, then checked out #3903 at 8e84cecca293d9dc428ccf6e80c85288fe04be16 locally. No code was changed, so no local or CW interactive tests were run. Handoff time: 2026-05-20 07:06 PDT.
Next action: Maintainer decision needed. A narrow Bridge PR can update WAN DiTSelfAttention.get_query_key_value_tensors to accept the new head_wise_gate kwarg, but that will not green #3903 by itself. MCore dev owners should align the LLaVA API/behavior with main for Nemotron Omni or confirm a Bridge-side compatibility wrapper strategy. DeepSeek/Megatron-FSDP owners should validate the GB200 convergence drift before any golden-value update.