Files
nvfp4-megamoe-kernel/vllm
biondizzle a10c582cf4 Add swiglu_limit=10.0 activation clamping (was missing)
DeepSeek-V4 uses SiluAndMulWithClamp(10.0) which clamps:
- silu(gate) to max 10.0
- up to [-10.0, 10.0]

Our runner was doing plain F.silu(gate) * up without clamping.
Large gate values could produce unbounded SiLU output, causing
numerical issues in the L2 GEMM. This is likely contributing to
garbage model output.
2026-05-17 17:52:16 +00:00
..