fix(mtp): rename MTP submodule transformer_layer -> mtp_model_layer

[Zhichenzzz]

Problem

This fork names the MTP inner module self.transformer_layer, but megatron-bridge registers all MTP weight mappings under mtp_model_layer. So megatron_to_hf_lookup returns None for every MTP param → 'NoneType' object has no attribute 'megatron_module' on bridge checkpoint load, and ... has no attribute 'param_weight' on export (Qwen3.6-27B GDN). Upstream NVIDIA renamed transformer_layer → mtp_model_layer; this fork is the outlier.

Fix

Rename the MTP instance attribute self.transformer_layer → self.mtp_model_layer in MultiTokenPredictionLayer (the ModuleSpec field name is unchanged). This makes the live param names match the bridge.

⚠️ Cross-repo: merge together with the miles PR

Renaming the live param names breaks miles' raw (non-bridge) converters that key on the old name. The paired miles PR updates them (qwen3_5.py / qwen3_next.py / update_weight/common.py). These two PRs must be merged together.

Validation

Real megatron-bridge Qwen3.5 mapping_registry.megatron_to_hf_lookup: all 8 MTP param types flip from None (old transformer_layer) to a resolved mapping (QKV/Auto/GatedMLPMapping) under the new mtp_model_layer name.

Fixes radixark/miles#1289

---

Update: full upstream alignment + e2e validation

Upstream Megatron-LM renamed the whole MTP naming chain, not just the module attribute. This PR now tracks all of it:

• MultiTokenPredictionLayerSubmodules.transformer_layer → mtp_model_layer (dataclass field)
• get_mtp_layer_spec(transformer_layer_spec) / get_mtp_layer_spec_for_backend(...) → mtp_model_layer_spec (kwarg; only in-tree call site get_gpt_mtp_block_spec updated)
• self.transformer_layer → self.mtp_model_layer (module attribute — checkpoint param names derive from this)
• remaining in-tree consumers of the old attribute: pipeline_parallel/schedules.py (set_current_microbatch raised AttributeError with MTP enabled), cuda_graphs.py, fine_grained_callables.py, model_chunk_schedule_plan.py

Impact / compatibility notes for reviewers:

1. Affects every MTP model (DeepSeek V3/V4, Qwen3-Next, Qwen3.5/3.6, GLM MoE): new checkpoints save MTP params as mtp.layers.N.mtp_model_layer.*.
2. Old torch_dist checkpoints containing MTP params will not load after this rename (key mismatch). Upstream made the same break with no compat shim; non-MTP checkpoints are unaffected.
3. Bridge mappings that already expect the new name (qwen35_vl_bridge, deepseek/common.py) are fixed by this PR — they were silently broken before (load_hf_weights crashed with NoneType has no attribute megatron_module on mtp.layers.0.transformer_layer.*).
4. Companion PRs: fix(mtp): track megatron mtp_model_layer rename in raw converters miles#1307 (converters, dual-name), [model] fix: register both MTP submodule spellings in qwen3_next_bridge Megatron-Bridge#11 (qwen3_next_bridge, dual-name).

e2e validated:

• Qwen3.5-35B-A3B VL, bridge mode: load_hf_weights 0 missing / 0 "No mapping found", geo3k RL step train_rollout_logprob_abs_diff=0.0129, reward 0.46 (and a separate 200-step run with the attribute rename: reward 0.17→0.95+, abs_diff flat 0.013–0.020).
• Qwen3.5-35B-A3B, raw mode with MTP enabled: model build through the renamed spec field, torch_dist load, MTP expert weight sync clean, set_current_microbatch passes (crashed before the consumer fix).