From 8546ed725f9ef321fe3323df71edecb17b750cf0 Mon Sep 17 00:00:00 2001 From: biondizzle Date: Thu, 4 Jun 2026 01:38:24 +0000 Subject: [PATCH] DEBUG: check SE input magnitude --- dsv4/layers/shared_expert.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/dsv4/layers/shared_expert.py b/dsv4/layers/shared_expert.py index 4330ae92..72ca19ba 100644 --- a/dsv4/layers/shared_expert.py +++ b/dsv4/layers/shared_expert.py @@ -425,6 +425,9 @@ class Nvfp4SharedExpert: """Actual implementation — called via custom autograd to be torch.compile-safe.""" self._ensure_initialized() + # DEBUG: check input + print(f" SE input: shape={tuple(hidden_states.shape)} |max|={hidden_states.abs().max().item():.6f} nan={torch.isnan(hidden_states).any().item()}", flush=True) + if self._fused_swiglu: # P1: Fused L1 GEMM + SwiGLU + clamp in one kernel launch intermediate = self._run_l1_fused(hidden_states)