What

Make the radixark/Megatron-Bridge GLM-5 / 5.1 / 5.2 (glm_moe_dsa) bridge fully build on the radixark/miles image's megatron-core (0.16.0rc0), for both LoRA and full-FT paths, with no megatron-core source change. Covers GLM-5.1 (DSA) and GLM-5.2 (DSA + cross-layer index sharing + rope 8e6).

Files: src/megatron/bridge/models/glm_moe_dsa/cross_layer_dsa.py (new), src/megatron/bridge/models/glm_moe_dsa/glm5_bridge.py.

Commits

e4cd5f37 — GLM-5.1: feature-detect the DSA experimental-attention spec

glm5_bridge.py sets experimental_attention_variant="dsa". On older megatron-core the dispatcher (get_experimental_attention_variant_module_spec) only wires "gated_delta_net" and raises ValueError for "dsa", and get_dsa_module_spec_for_backend omits the metainfo the variant layer-builder reads → the model can't build. transformer_layer_spec now points at _build_glm5_dsa_block_spec, whose wrapped dispatcher prefers megatron-core's native handling and only on the "dsa" ValueError (old core) back-fills via the shipped builder + sets metainfo["fuse_input_layernorm"]=False (MLA-based DSA keeps a separate, non-fused input layernorm, like the deepseek_v4 dsv4 spec). ⇒ transparent no-op on newer megatron-core; deletable after a core bump. (Previously a caller-side monkey-patch in miles bridge_lora_helpers.py, now consolidated here.)

2bdfa05d — GLM-5.2: parse the MLA rope dims (rotary_base + qk_pos_emb_head_dim)

GLM-5.2 sets rope_theta=8e6, and under transformers>=5.12 the parsed GlmMoeDsaConfig reports qk_rope_head_dim as head_dim (192) instead of the config.json value (64). The base config-mapping then sizes MLA linear_kv_down_proj as kv_lora_rank + 192 = 704, contradicting the checkpoint (kv_a_proj_with_mqa = kv_lora_rank + qk_rope_head_dim = 576 = 512 + 64).

• rotary_base: read rope_theta whether nested in rope_parameters or flat.
• qk_pos_emb_head_dim: re-read qk_rope_head_dim straight from config.json so the MLA rope/kv dims match the weights. No-op when the parse is already correct; GLM-5.1 unaffected (its head_dim already equals qk_rope_head_dim = 64).

74dade06 — GLM-5.2: DSA cross-layer index sharing (CrossLayerDSAttention)

GLM-5.2 keeps GLM-5.1's glm_moe_dsa arch, but only "computing" layers carry the lightning indexer and compute the sparse top-k; "skip" layers reuse the most recent computing layer's top-k (HF config index_topk_freq=4, index_skip_topk_offset=3 → computing Megatron-layers 1,2,3,7,11,…). megatron-core's DSA is per-layer only, so this adds a Bridge-owned CrossLayerDSAttention(DSAttention):

• anchor layers compute + publish topk_indices to a per-microbatch holder (packed_seq_params for thd, thread-local for bshd);
• skip layers del self.indexer (so the param set matches the subset checkpoint, which only stores indexer weights on computing layers) and reuse the source anchor's top-k via unfused_dsa_fn.

get_glm5_crosslayer_dsa_spec calls megatron-core's exact get_dsa_module_spec_for_backend and swaps only submodules.core_attention.module to the subclass (so the MLA structure — fused qk-layernorm, indexer submodules — is inherited verbatim). Feature-gated in _build_glm5_dsa_block_spec on dsa_index_topk_freq > 1, so GLM-5.1 (no freq → 1) keeps the existing per-layer path byte-for-byte.

Verified (via radixark/miles, 4×H200, bridge mode, TP4/EP4, bshd, mbs1)

• GLM-5.2 7-layer train-only (jybsuper/GLM-5.2-7layer): build CrossLayerDSAttention → load the 98 GB subset checkpoint cleanly (proves the indexer is built only on computing layers 1,2,3,7) → cross-layer fwd/bwd (skip layers reuse anchor top-k, no holder assert) → save LoRA adapter → TRAIN EXIT 0.
• GLM-5.1 6-layer full e2e (rollout → train → save): TRAIN EXIT 0 — the feature-gate leaves the GLM-5.1 path unchanged (regression check).

Notes

• No megatron-core source change; everything is on the Bridge side and self-disables on a newer megatron-core.
• sglang does not yet serve the GLM-5.2 cross-layer rollout, so the GLM-5.2 validation is training-side (train-only); GLM-5.1 runs the full rollout→train loop.

🤖 Generated with Claude Code