feat(deepseek-v4): add native LoRA construction and conversion bridge

[zhoutong-hai]

Depends on #2771. The grouped wo_a execution path must land before this PR enables and wraps that target.

Problem

Current Miles pins Megatron-Bridge at 7f0fb345, which does not register DeepseekV4ForCausalLM. Letting Bridge select a provider therefore cannot construct the current native V4 graph. Treating it as V3 or using incomplete inferred mappings is worse: a strict current-main load exposed thousands of unmapped or missing checkpoint parameters.

V4 also owns both a main-attention wq_b and a nested DSA indexer linear_wq_b. The generic MLA alias maps bare wq_b to the indexer, so an unqualified V4 treatment can wrap the wrong module.

Finally, a V4 portable export is valid only as an adapter against the pinned native base. The conversion-only bridge cannot safely claim to emit a complete merged base, and a failed adapter conversion must not leave a success marker or strand peer ranks in a barrier.

Fix

• Detect the exact V4 architecture and construct it through the Miles native provider/spec.
• Apply PEFT in a pre-wrap hook, before DDP construction.
• Qualify the five main V4 attention targets so wq_b cannot also match the nested indexer.
• Register a conversion-only Bridge for the current native graph, with explicit attention mappings and explicit replicated/column/automatic semantics for hyper-connections, compressors, routing tables, attention sink, and indexer weights.
• Keep portable V4 output adapter-only against the pinned base.
• Make required adapter export fail closed on empty or failed Bridge conversion.
• Collect writer/export failures across all ranks before raising, so rank-local I/O failure cannot leave peers at the trailing barrier.
• Write .complete only after a valid adapter is present.
• Reject unsupported native V4 Multi-LoRA instead of silently falling back.

This does not duplicate the MXFP4 ingestion work in #2717. It is scoped to current main and its pinned Bridge/native V4 graph. If the large Megatron bump in #2706 lands first, the mapping layer will need a focused rebase because that PR intentionally renames and relayouts the V4 graph. The general PEFT-export work in #2580 is related; this PR adds the V4 requirement that portable export cannot remain best-effort.

Validation

• 8 focused tests cover mapping ownership, native-provider dispatch, main/indexer target qualification, adapter-only layout, exact architecture selection, empty export, completion-marker failure, and two-rank writer-failure propagation.
• The writer-failure case confirms both ranks receive the same rank-0 I/O error instead of deadlocking.
• Tests run against the exact pinned Megatron-Bridge and Miles Megatron sources; unavailable GPU-only packages were stubbed only for import.
• Ruff and diff checks pass.
• The underlying current-main graph and explicit mappings were exercised on a 32-rank V4 LoRA run: strict base load reached zero unmapped/missing warnings, then completed finite updates and trainer-to-rollout publications.

Remaining qualification

A successful adapter-only artifact still needs to be loaded back against the pinned base before portable export is called complete. This PR remains draft until that round trip is recorded and #2771 lands.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.