[GDN] Add configurable output gate activation

[alifurkanstahl]

Summary

Add configurable output-gate activation to GatedDeltaNet.

GatedDeltaNet previously used the default SiLU/Swish activation of FusedRMSNormGated. This PR exposes that choice through a new output_gate_activation option supporting swish, silu, and sigmoid.

The option is wired through GatedDeltaNetConfig → GatedDeltaNetBlock → GatedDeltaNet.

This removes the hard-coded output-gate activation from native GatedDeltaNet and allows callers to select sigmoid gating when required by a model architecture.

The default remains swish, so existing callers and checkpoints are unaffected. No Gated Delta Rule kernel or recurrence is changed.

Test plan

Hardware: NVIDIA RTX 5060 Ti

Dependent test files identified with:

python scripts/find_dependent_tests.py fla/layers/gated_deltanet.py
python scripts/find_dependent_tests.py fla/models/gated_deltanet

Full dependent suite:

python -m pytest \
  tests/layers/test_attn_varlen_pack_layout.py \
  tests/layers/test_gated_deltanet.py \
  tests/layers/test_gated_deltanet_output_gate.py \
  tests/layers/test_layer_cache_layer_idx.py \
  tests/models/test_hybrid_attention.py \
  tests/models/test_modeling_gated_deltanet.py \
  tests/models/test_modeling_yoco.py \
  -q

• 284 passed in 1100.08s

Additional targeted validation:

• python -m pytest tests/layers/test_gated_deltanet_output_gate.py -v — 8 passed
• python -m pytest tests/layers/test_gated_deltanet.py -v — 5 passed
• python -m pytest tests/layers -k "gated_deltanet or gdn" -v — 14 passed, 42 deselected
• python -m pytest tests/modules/test_layernorm_gated.py -k "not large_batch" -v — 12 passed, 2 deselected
• pre-commit run --all-files — passed
• git diff --check — clean

The new tests cover the default activation, sigmoid/SILU wiring, config serialization, state-dict compatibility, validation, positional config compatibility, and sigmoid forward/backward execution through the native GatedDeltaNet layer.

Benchmark / NCU (kernel changes only)

N/A — no kernel code changed. This only selects an activation already supported by FusedRMSNormGated.

Breaking changes

None. New optional configuration/constructor argument; the default remains swish and existing state dicts are unchanged.

Checklist

• I have read CONTRIBUTING.md and follow its conventions (code style, docstrings, commit prefixes).
• I have read AGENTS.md and, where my change matches its scope, the relevant skill under .agents/skills.
• Dependent tests pass locally or in CI, and new behavior is covered by tests where applicable (tick as N/A for changes with no testable code, e.g. docs-only).
• Kernel changes include same-hardware before/after benchmark numbers, dense + varlen where applicable (N/A — no kernel code changed).
• This PR is minor/cosmetic-only (typo, formatting, style-only tweaks) — tick only if it is, and justify below.