[port] Megatron-Core APIs needed by Megatron-Bridge main (#4452 MLA resolver, gdn alias, #5865 FSDP names)

[yushengsu-thu]

Megatron-Bridge main (radixark bridge-new @ 2e09c234a) needs three things from Megatron-Core that miles-main does not have. Required by radixark/miles#3291.

Commits

1. megatron.core.transformer.mla_qk_norm_config + megatron.core.models.backends.get_backend: the MLA half of Port Multi-Latent Attention to HybridModel NVIDIA/Megatron-LM#4452 (HybridModel half not ported; absorbed_mla.py keeps miles-main's combined/split K/V up-projection structure).
2. experimental_attention_variant="gdn" accepted as an alias of gated_delta_net (upstream Bump FLA to 0.5.1 and integrate GDN2 kernels NVIDIA/Megatron-LM#5765 rename; Megatron-Bridge selects Qwen3.5 linear attention with the new name).
3. FullyShardedDataParallelV1 (= FullyShardedDataParallel) and a non-constructible FullyShardedDataParallelV2 placeholder in mcore_fsdp_adapter.py: the names Add MFSDP v2 MCore integration with FullyShardedOptimizer NVIDIA/Megatron-LM#5865 introduced and Megatron-Bridge's unwrap_model imports since fix(fsdp): drop stale MFSDP v2 guard rails for clip_grad, precision-aware optimizer, HSDP, and optimizer CUDA graphs NVIDIA-NeMo/Megatron-Bridge#5933. MFSDP v2 itself is not ported.

Validation

• Commits 1–2: 2× RTX 4090, bridge-new unit tests (deepseek, glm_moe_dsa, kimi, model_bridge, peft) 695 passed / 14 skipped, identical to NVIDIA main @ f6c33bde4; Megatron-LM MLA tests unchanged (8 passed / 1 pre-existing failure). ci(docker): bump Megatron-Bridge to bridge-new (CI probe) and testing miles#3291 rounds 2–7 run against this PR via ci-megatron-pr: #102; bridge-mode LoRA / multi-LoRA / GLM / Nemotron suites green.
• Commit 3: tests/unit_tests/distributed/mfsdp_v1/test_mcore_fsdp_adapter_names.py (alias + placeholder class; not yet run on GPU).

[Zhichenzzz]

LGTHuman, lets wait for the ci pass of radixark/miles#3291

[Zhichenzzz]

LGTHuman, please check the cc comments 👇

[Zhichenzzz]

Reviewed against upstream NVIDIA/Megatron-LM@main and Megatron-Bridge bridge-new @ 2e09c234a (the consumer this port targets). Overall the port is faithful; a few things worth addressing or knowing about before merge.

1. Behavior change for MLA with q_lora_rank=None (not hit by any shipped miles recipe).
QKNormConfigResolver._mla_fused_linear_or_default returns the backend's fused norm+linear whenever the spec's linear_q_proj is not already fused. miles-main's gpt_layer_specs MLA spec passes linear_q_proj=backend.column_parallel_linear(), so with TE + qk_layernorm=True + no query LoRA, linear_q_proj silently changes from TEColumnParallelLinear to TELayerNormColumnParallelLinear and gains a layer_norm_weight with no HF counterpart. The path is _resolve_mla_qk_layernorm -> else branch (no q-LoRA) -> _mla_fused_linear_or_default(submodules.linear_q_proj) -> _require_linear(self.fused_norm_linear_impl), which ignores the non-fused class the spec asked for.
All 12 MLA recipes in scripts/models/ set --q-lora-rank, so none are affected today. The exposed path is miles' megatron_to_hf/kimi_vl.py (Moonlight-style, no q-LoRA), which only maps linear_q_proj.weight. This is upstream's chosen behavior (Megatron-Bridge added MLASelfAttentionWithoutQueryNorm to work around it); worth a line in the PR body and, ideally, a guard or converter update on the miles side.

2. attention_latent_norm_epsilon is not ported.
Upstream MLATransformerConfig has attention_latent_norm_epsilon: float | None = None, defaulted to layernorm_epsilon in __post_init__. Megatron-Bridge's MLASelfAttentionWithoutQueryNorm._resolve_qk_norm_config does hasattr(self.config, "attention_latent_norm_epsilon") and, when missing, replaces self.config with a SimpleNamespace(**vars(self.config), ...) for the rest of the module's life. That fallback only runs on the no-q-LoRA DeepSeek path, but it is a 4-line port to make it never run. Suggest adding the field + the __post_init__ default.

3. get_backend is ported with the older single-argument signature.
Upstream main now takes use_kitchen, use_te_op_fuser, cross_entropy_*, cuda_graph_impl and has get_backend_from_config. Bridge at 2e09c234a only calls get_backend(transformer_impl), so this is fine for the pinned commit; the next bridge-new bump may need the wider signature. Not a blocker.

4. Minor consistency nit.
megatron/training/training.py:1030 uses a literal in ("gated_delta_net", "gdn") while every other site goes through is_gated_delta_net_variant. Harmless, but the alias now has two sources of truth.

Checked and clean

• mla_qk_norm_config.py is byte-identical to upstream main; MLA / FusedMLA / AbsorbedMLA wiring matches upstream line for line, and the _resolve_qk_norm_config hook has the exact signature Bridge's DeepSeek subclass overrides.
• The backend API the resolver needs (layer_norm(rms_norm, for_qk), column_parallel_layer_norm_linear, column_parallel_linear) exists on all three miles-main providers.
• Existing miles MLA specs (DeepSeek-V3/V3.2, GLM-5.x, Kimi-K2, JoyAI) pass the resolver's new validation: fused up-projections with IdentityOp norms is exactly the accepted MLA form. glm5 (DSAMLASelfAttention) and deepseek_v4 build their own attention and never enter the resolver.
• gdn alias: __post_init__ canonicalizes to gated_delta_net, Bridge's finalize() re-runs MCoreTransformerConfig.__post_init__ after provider.experimental_attention_variant = "gdn", and Bridge has no == "gdn" comparisons on the variant afterwards. No literal gated_delta_net comparisons remain outside transformer_config.py/training.py.
• FSDP names: V1 is FullyShardedDataParallel, V2 importable and non-constructible; matches what Bridge's unwrap_model / setup.py / train.py import and isinstance against. Test is CPU-runnable.
• Test edits to test_schedule_layer_1f1b.py / test_finalize_model_grads.py / test_submodule_callables.py only parametrize qk_layernorm; nothing previously valid was removed.