Support GLM5.2 and recipe for GLM 5.1 & 5.2 on GB200 and H100

[HollowMan6]

What does this PR do ?

Support GLM5.2 and use correct GLM5.x related flags, as introduced in NVIDIA/Megatron-LM#5246 and related PRs.

Changelog

• Pass related flags correctly to mcore

GitHub Actions CI

See the CI section in the Contributing doc for how to trigger the CI. A Nvidia developer will need to approve and trigger the CI for external contributors.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

If you haven't finished some of the above items you can still open "Draft" PR.

Additional Information

• Related to # (issue)

[HollowMan6]

/claude review

[claude]

Review - GLM-5.2 support + GLM-5.1/5.2 perf recipes

Overall clean and well-tested. One blocking bug in the tests.

[High] Arithmetic assertion is wrong and will fail CI - tests/unit_tests/recipes/test_glm5_perf_recipes.py lines 127-135. The GLM-5.2 H100 recipe sets PP=13, EP=16, ETP=1, so 416 // (13161) = 416 // 208 = 2, but the test asserts == 4. The recipe is self-consistent (dense DP = 416 // (TPPPCP) = 416 // 208 = 2, expert DP = 416 // (PPEPETP) = 2), so the expected value should be 2. Inline suggestion posted. As written, this test fails and the whole file errors out.

Notes (non-blocking):

• glm5_bridge.py:115 reads hf_config.indexer_rope_interleave unconditionally (no getattr). It exists on GLM-5/5.1/5.2 HF configs and the toy config and new unit test both define it, so this is fine - just flagging that a config missing this key raises AttributeError rather than falling back.
• Dispatcher selection matches guidance: hybridep on GB200, deepep on H100.

Suggested test cases:

• glm51_sft_192gpu_gb200_bf16_config
• glm52_sft_192gpu_gb200_bf16_config
• glm51_sft_416gpu_h100_bf16_config
• glm52_sft_416gpu_h100_bf16_config
• test_glm5_perf_recipes_are_flat_and_preserve_bridge_dsa_fields[glm51_sft_192gpu_gb200_bf16_config]
• test_glm5_perf_recipes_are_flat_and_preserve_bridge_dsa_fields[glm52_sft_192gpu_gb200_bf16_config]
• test_glm5_perf_recipes_are_flat_and_preserve_bridge_dsa_fields[glm51_sft_416gpu_h100_bf16_config]
• test_glm5_perf_recipes_are_flat_and_preserve_bridge_dsa_fields[glm52_sft_416gpu_h100_bf16_config]
• test_glm52_h100_pipeline_layout_keeps_dsa_index_sharing_within_each_stage
• test_provider_bridge_maps_dsa_architecture_from_hf_config