feat(mcore): add GLM-5/DeepSeek-V3 model support (mbridge + megatron-bridge)

[dingzhiqiang]

Summary

Adds GLM-5.1 / DeepSeek-V3 / GLM-4.7-Flash model support — three architectures open-source AReaL did not previously support. Coverage spans both the mbridge path (default) and the megatron-bridge path opted into via mcore.bridge_type=megatron-bridge.

This PR stacks on top of #1372 (Bailing-MoE megatron-bridge adapter), which introduced the shared cross-cutting infrastructure (optional mbridge import wrapping, Bridge type-annotation cleanups, migration doc skeleton). The diff shown here only contains GLM-5/DSV3 additions; #1372 must merge first.

Previously, both this work and the Bailing adapter were combined in #1362 (now closed). They were split into PR-A (#1372) and this PR-B for easier review.

What's added

New model code

• areal/models/mcore/deepseek_v3.py — HF config → MLATransformerConfig conversion, homogeneous MLA layer specs, _has_dsa() helper. Handles DeepseekV3ForCausalLM, GlmMoeDsaForCausalLM (GLM-5.1), and Glm4MoeForCausalLM (GLM-4.7-Flash) — all three share the underlying MLA + MoE topology; GLM-5.1 additionally exposes a DSA indexer.
• areal/models/mcore/deepseek_v3_bridge.py — mbridge LLMBridge subclass registered as deepseek_v3 / glm_moe_dsa / glm4_moe_lite.
• areal/models/mcore/dsa_mla_attention.py — custom DSAMLASelfAttention module that inherits Attention directly (not DSAttention) so packed THD inputs work without modification. Implements the DSA indexer (wq_b, wk, k_norm, weights_proj) called by the layer spec.
• areal/models/mcore/glm5_megatron_bridge.py — megatron-bridge MegatronModelBridge subclass for GlmMoeDsaForCausalLM, with DSA indexer weight mappings and MTP layer support.
• areal/experimental/ops/dsa/ — six tilelang kernel files implementing the DSA indexer forward/backward and the sparse MLA forward/backward used by DSAMLASelfAttention. Specialized for DSV3/GLM-5.1 latent geometry (kv_lora_rank=512, qk_rope_head_dim=64).

Registry / engine wiring

• areal/models/mcore/registry.py — add _DEEPSEEK_V3_ARCHITECTURES set and _supplement_dsa_config() helper; register DSV3/GLM-5/GLM-4.7 architectures in make_hf_and_mcore_config (no-bridge fallback) and make_mcore_layer_specs; inject AReaL's DSA-aware layer spec into provider.transformer_layer_spec for the megatron-bridge path when _has_dsa(hf_config) is True.
• areal/engine/megatron_engine.py — import deepseek_v3_bridge (lazy via try/except, since it depends on mbridge) and glm5_megatron_bridge (unconditional, since it uses megatron-bridge) so their decorators fire on engine load.

Docs

• docs/en/best_practices/migrate_to_megatron_bridge.md — extend the supported-architectures table with DSV3/GLM-5/GLM-4.7 entries and describe the DSA-aware layer spec injection alongside the Bailing one.

Validation

The mbridge ↔ megatron-bridge numerical equivalence (matching starting logp / grad_norm / loss with same seed and config) has been validated on the internal branch this PR is ported from. All provider.<field> assignments in glm5_megatron_bridge.py are preserved as-is from the validated source.

Test plan

• pre-commit run --all-files — passing locally
• CI smoke tests (no regression on existing models) — to run on PR
• Manual: GLM-5.1 GRPO step with bridge_type: megatron-bridge on a real cluster
• Manual: DeepSeek-V3 GRPO with both bridge_type: mbridge and bridge_type: megatron-bridge for 5-step loss alignment

Dependencies

• Depends on feat(mcore): add Bailing-MoE V2.5 megatron-bridge adapter #1372 (PR-A). Targets the chucai.dzq/bailing-megatron-bridge branch; once feat(mcore): add Bailing-MoE V2.5 megatron-bridge adapter #1372 merges I'll retarget this PR onto main and rebase.

Known limitations / not in scope

• VPP > 1 with the DSA layer-spec injection: the lambda currently captures stage-0 specs (validated runs were VPP=1). Tracked as follow-up.
• The two tilelang kernels (tilelang_sparse_mla_{fwd,bwd}.py) are specialized for DSV3/GLM-5.1 latent geometry — D=512, dim_plus_tail_dim=576. Supporting other latent dimensions requires kernel rewrite and re-tuning.
• examples/math/glm5_grpo_megatron_bridge.yaml deferred (would require placeholder paths and large multi-node config; users should adapt existing examples/math/gsm8k_grpo_megatron.yaml with mcore.bridge_type: megatron-bridge).

[gemini-code-assist]

Code Review

This pull request introduces support for DeepSeek V3, GLM-5.1, and GLM-4.7-Flash models in megatron-core, implementing custom Multi-Latent Attention (MLA) and Dynamic Sparse Attention (DSA) operators with TileLang kernels. The review feedback identifies several important issues: a critical mismatch in the number of returned gradients in the autograd IndexerFunction.backward which will trigger a runtime error, potential AttributeErrors when directly accessing HuggingFace configuration attributes, a missing validation check when DSA is enabled but Transformer Engine is disabled, and a missing dimension assertion in the TileLang backward kernel.