Complete clamped SwiGLU across expert paths

[FDecaYed]

Complete clamped SwiGLU across expert paths

Summary

Add clamped SwiGLU as one usable feature across the primitive, routed and
shared experts, token-weighted fallbacks, and the Transformer Engine grouped
MLP path.

Scope and non-goals

• Support bias-free, biased, and token-weighted forward/backward.
• Preserve the legacy unclamped path when the option is disabled.
• Add validation and Transformer Engine operator/constructor-signature gating,
  so optional GroupedLinear kwargs are passed only when the installed
  interface accepts them.
• Do not include unrelated activation rewrites, routing features, DSv4
  attention, or recipe changes.

Provenance

This local recut is reconstructed from the frozen main baseline
bb5647a9bdd0, not a replay. It references master #5795 and combines #5940
and #5941 into one functional unit, including the completion fixes derived
from #4481 and #5130. Original clamped-SwiGLU and DSv4 integration credit:
@hxbai in #4481, #5130, and #5795.

Dependencies

None beyond the frozen baseline used for this local recut. This is a reusable
prerequisite outside the nine-PR DSv4-specific series.

Tests

• Primitive forward/backward for biased, bias-free, and token-weighted modes.
• Routed, shared, and fallback expert paths.
• Transformer Engine grouped-MLP behavior, optional constructor kwargs, and
  unsupported-interface validation.
• Legacy behavior when clamping is disabled.

Publication

Publish in the first prerequisite round on then-current main. This is not a
stacked-review PR.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[FDecaYed]

/ok to test 33bd241

[FDecaYed]

/ok to test 388eeb4

[FDecaYed]

/ok to test ec202a4

[FDecaYed]

/ok to test 4ba9d91

[svcnvidia-nemo-ci]

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/31685310896

[svcnvidia-nemo-ci]

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/31685597753