chore(deps): upgrade mbridge from 0.15.1 to 310e8fb

[Adiactive]

Description

Upgrade mbridge from PyPI release 0.15.1 to commit 310e8fb on main (~120 upstream commits). Picks up new model architectures (Qwen3-VL dense/MoE w/ CP, Qwen3.5 + MTP, Qwen3 Omni MoE, InternVL 3.5, DeepSeek NPU), compatibility fixes for megatron-core 0.14+/0.15/0.16 and transformers v5, and several bug fixes around RoPE, weight export, and distributed checkpointing. See the linked issue for the full rationale.

Changes:

• Update mbridge requirement in pyproject.toml and pyproject.vllm.toml
• Regenerate uv.lock and uv.vllm.lock via scripts/uv_lock.sh (ran with uv 0.10.9)

Related Issue

Fixes #1257

Type of Change

• ♻️ Refactoring

Checklist

• I have read the Contributing Guide
• Pre-commit hooks pass (pre-commit run --all-files)
• Relevant tests pass; new tests added for new functionality
• Documentation updated (if applicable; built with ./docs/build_all.sh)
• Branch is up to date with main
• Self-reviewed via /review-pr command
• This PR was created by a coding agent via /create-pr
• This PR is a breaking change

Additional Context

Test Environment

• Image: ghcr.io/inclusionai/areal-runtime:v1.0.3-vllm
• Mbridge installed via uv pip install --no-deps git+https://github.com/ISEEKYAN/mbridge.git@310e8fb to keep the existing torch/TE pin intact

Tests Run

Test	Result	Notes
tests/test_estimate_num_params.py	✅ 3/3	Direct mbridge.AutoBridge API
tests/test_megatron_engine.py	✅ 3/4	DCP test pre-existing fail (see below)
tests/test_megatron_engine_distributed.py	⚠️ 4/7	TP / PP / VPP / MoE-DCP pass; 3 pre-existing fails
tests/test_tree_training.py	✅ 12/12	flex/triton × {fsdp, megatron, archon}
tests/test_offload.py	✅ 2/2	FSDP and Megatron offload paths
tests/fp8/test_fp8_rmsnorm.py + test_fp8_bf16_comparison.py	⚠️ 2/4	2 pre-existing fails (see below)

Pre-existing failures unrelated to this upgrade

1. test_dcp_save_load_weights (test_megatron_engine.py) — CheckpointingException: ShardedTensor.flattened_range is not supported. Known incompatibility between Megatron's distributed optimizer and dist-checkpointing format. Documented at cli_args.py:2177; workaround added in 7ca1fea0 for recovery. The test is @pytest.mark.slow and excluded from CI.

2. test_qwen3_dcp_save_load (test_megatron_engine_distributed.py) — same flattened_range issue (uses with_optim=True).

3. test_qwen3_context_parallel & test_qwen3moe_expert_parallel (test_megatron_engine_distributed.py) — KeyError: 'loss_mask' regression introduced by PR feat: CP-local loss and configurable CUDA memory profiling #1223 (d58cca56, merged 2026-04-23). The new CP-local loss path requires loss_mask in padded_mb, but the test's mock_input only provides input_ids / attention_mask. Adding loss_mask to the mock further uncovers a downstream unpack_sequence shape mismatch in the same CP-local path — both belong to PR feat: CP-local loss and configurable CUDA memory profiling #1223 follow-up work, out of scope for this PR.

4. test_fp8_bf16_gradient_comparison & test_fp8_bf16_partial_layers_comparison (tests/fp8/test_fp8_bf16_comparison.py) — TypeError: sft_loss_fn() got an unexpected keyword argument 'vocab_min_logits'. Caused by PR fix(tree_attn): Fix some bugs in tree training for FSDP and Megatron engines #889 (127b0264) adding vocab_min_logits=... to the engine's loss_fn call, but the fp8 test fixture's sft_loss_fn at tests/fp8/model_hooks.py:127 doesn't accept **kwargs.

Lock file diff

The lock files lose a number of wheel entries for non-x86_64 platforms (ppc64le, s390x, armv7l, riscv64, musllinux). Switching mbridge from a PyPI wheel to a git source — combined with the existing platform_machine == 'x86_64' marker on mbridge in our pyproject — lets uv prove those wheels are unreachable through the mbridge subtree and prune them. AReaL targets x86_64 Linux clusters, so there is no real platform coverage loss.

[gemini-code-assist]

Code Review

This pull request updates the mbridge dependency in pyproject.toml and pyproject.vllm.toml to a specific git commit. The review feedback recommends using a full 40-character commit hash instead of a short SHA to ensure better security and reproducibility, and also suggests removing an extra space for consistency with other entries.

[garrett4wade]

LGTM!

FYI in the future we will totally "mbridge" to "megatron-bridge" since it does not support transformers>=5.3.