From 26ecf96328a6928ad312590f1aa8caa2253442d0 Mon Sep 17 00:00:00 2001 From: biondizzle Date: Thu, 4 Jun 2026 01:30:29 +0000 Subject: [PATCH] DEBUG: check intermediate magnitude before SE L2 --- dsv4/layers/shared_expert.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/dsv4/layers/shared_expert.py b/dsv4/layers/shared_expert.py index 15267541..4330ae92 100644 --- a/dsv4/layers/shared_expert.py +++ b/dsv4/layers/shared_expert.py @@ -443,6 +443,8 @@ class Nvfp4SharedExpert: gate = gate.clamp(max=self.swiglu_limit) up = up.clamp(min=-self.swiglu_limit, max=self.swiglu_limit) intermediate = torch.nn.functional.silu(gate) * up + # DEBUG: check intermediate before L2 + print(f" SE intermediate: shape={tuple(intermediate.shape)} |max|={intermediate.abs().max().item():.6f} nan={torch.isnan(intermediate).any().item()}", flush=True) output = self._run_l2(intermediate) return output