refactor(update-weight): LoRA adapter tensors ride the base weight stream

[yueming-yuan]

ci-sglang-pr: #36384

Part of #1360 Step 6 (design: lora_weight_sync_cleanup.md). Stacked on #2754. Engine counterpart: sgl-project/sglang#36384 — miles tracks the sglang-miles tip continuously, so this PR must merge only after that change is in sglang-miles (and sglang-miles must keep the legacy routes until this merges).

What

The second LoRA send path is deleted; adapter tensors are ordinary entries in the bucketed stream, and adapter identity moves to a control plane.

• Iterator: iter_hf_base_weights → iter_hf_weights(weights, *, include_base=True, adapters=(), materialize=True). Adapter units join the same unit stream under {lora_name}:{hf_key} names via the new _iter_hf_adapter_units hook. The megatron base implements it as the existing PP-local export hook (_export_pp_local_lora, unchanged in bridge/direct) + the PP gather now placement-gated (runs for gather_pp placements — bridge, colocate; disappears for direct+dist) instead of unconditional. The zero-export / no-lora-names guards move into the hook.
• Updater: one loop. _adapters_to_push() (single: miles_lora; multi: __miles_slot_{i} sorted) + eager register_lora_adapter for new tenants before begin (engine validates r ≤ max_lora_rank before any bytes move); the frame is unconditional again with the scope passed as begin_weight_update(..., sync_base=...) — reverting the refactor(update-weight): colocated transfer as a protocol on the unified updater #2753 call gating whose substance (no quant round-trip on adapter-only rounds) now lives in the session scope. --check-lora-weight-equal becomes a driver-side sha256 manifest accumulated during the stream and verified engine-side at end_weight_update (now on both transports; previously colocate-only).
• Deletions (net −326): send_adapter from the ABC + both implementations, the is_lora_sender role from all five protocols, get_hf_lora_weights / _export_lora_named_tensors, _send_lora_adapter / _send_multi_lora_adapters, unload-before-reload (+ session.unload_lora_adapter), the colocate LoRA IPC branch and the _repack_onto_fresh_storage cuMem workaround (adapter tensors now travel in sync-time-allocated FlattenedTensorBuckets, which is exactly why the base path never had the problem).

Behavior changes

1. Single-LoRA becomes upsert-in-place (no unload/reload window; engine-side rollback on failure).
2. Colocate multi-LoRA works (previously NotImplementedError).
3. LoRA runs stamp weight_version (frame is unconditional); the actor-side CI version assertion can be re-enabled in a follow-up.
4. Adapter placement follows base placement (PP-sharded senders under direct+dist; bridge unchanged).
5. supports_lora remains the gate for p2p/rdt/disk-delta (unchanged: unsupported).

Tests: iterator template tests retargeted to adapter units in the stream; transport-era send_adapter/orchestration test classes removed (their engine-side successors live in sglang's test_lora_stream_update.py).

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}