[yueming-yuan]

Fixes found while verifying miles p2p (RDMA) weight updates on GLM-5.2 744B fp8 with MTP enabled. Companion miles PR: radixark/miles#2605.

Worker-keyed engine-info bootstrap registry (engine_info_bootstrap_server.py, remote_instance_weight_transporter.py, model_runner.py, http_server.py)
With speculative decoding, the draft worker's ModelRunner.initialize() also registers with the EngineInfoBootstrapServer. The registry was keyed by tp_rank alone, last write wins, and the draft registers after the target — so the p2p client received the draft model's weight table and RDMA session for every rank: ordinary target params appeared missing on the remote and every write bucket failed. The registry key is now (worker, tp_rank) (worker in {"target", "draft"}), the transporter tags its registrations, and the two query endpoints take an optional worker parameter defaulting to "target" — existing clients keep their behavior, and a p2p client can address the draft explicitly. Long-running MTP training needs exactly that: the trained nextn layer must reach the draft, or speculative acceptance decays to zero as the target trains away from the frozen draft (observed as accept rate: 0.00 followed by an out-of-range index assert in decode ~100 steps in).

Fused DSA indexer weights unmappable by ParameterMapper (deepseek_v2.py)
With use_dsa_indexer_fusion, the model has one bf16 indexer.wk_weights_proj per layer and no indexer.wk / indexer.weights_proj params. The checkpoint loader folds the unfused names via _load_fused_indexer_wk, but ParameterMapper — built from stacked_params_mapping and used by the p2p staging path — had no entry, so every indexer wk/weights_proj tensor sent by the trainer was silently dropped. Added the two stacked entries, gated on the fusion config (not a named_parameters() probe: the NextN subclass builds its mappings before its model exists), with dot-anchored patterns so already-fused checkpoint names cannot be rewritten into nonsense.

NextN name mapping for external weight mappers (deepseek_nextn.py)
mutate_weight_preload maps model.layers.{num_hidden_layers}.* to model.decoder.* so ParameterMapper.from_model on the draft model resolves converter-emitted nextn names the same way the loader does.

Registration was one-shot and swallowed failures (remote_instance_weight_transporter.py)
A transient connection failure while PUTting to the bootstrap server was logged and forgotten; the missing rank then surfaced minutes later as a 404 at the first weight update, far from the cause. Both registrations now retry (24 x 5s, timeout 10s) and raise after exhaustion.

Inverse ue8m0 scale transform for non-128-divisible row counts (fp8_utils.py)
The weights-checker comparator crashed on params like fused_qkv_a_proj_with_mqa (mn=2624): the inverse transform assumed the repeated-scale layout's row count is a multiple of the 128 block size. The main blocks and the partial tail block are now de-repeated separately.

---

CI States

Latest PR Test (Base): ❌ Run #32296566471
Latest PR Test (Extra): ❌ Run #32296566119
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.