From 455ecb563151473efbe485b6d16a6cfb41c16cad Mon Sep 17 00:00:00 2001
From: biondizzle <biondizzle@gmail.com>
Date: Sun, 17 May 2026 15:47:38 +0000
Subject: [PATCH] Fix: define padded_max_slots before using it in shared buffer
 allocation

---
 vllm/nvfp4_cutedsl.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/vllm/nvfp4_cutedsl.py b/vllm/nvfp4_cutedsl.py
index 5e234263..a57dc111 100644
--- a/vllm/nvfp4_cutedsl.py
+++ b/vllm/nvfp4_cutedsl.py
@@ -145,6 +145,8 @@ class CuTeDSLMoERunner:
         # Padded hidden/activated: SHARED across all runners (not per-layer)
         # These are only used during run() which is sequential across layers.
         # Per-layer allocation would be 72 MB × 60 layers = 4.3 GB → OOM.
+        max_rows_per_expert = self._max_chunks_per_expert * 128
+        padded_max_slots = self.num_experts * max_rows_per_expert
         device_key = str(self.device)
         if not hasattr(CuTeDSLMoERunner, '_shared_padded_bufs'):
             CuTeDSLMoERunner._shared_padded_bufs = {}