From da02a5dc115c04254dfb16b03aaf59dfd83788dc Mon Sep 17 00:00:00 2001 From: biondizzle Date: Sun, 17 May 2026 08:16:09 +0000 Subject: [PATCH] debug: assert token indices are correct after allocation --- vllm/nvfp4_cutedsl.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/vllm/nvfp4_cutedsl.py b/vllm/nvfp4_cutedsl.py index 4e669395..2ca3505a 100644 --- a/vllm/nvfp4_cutedsl.py +++ b/vllm/nvfp4_cutedsl.py @@ -84,7 +84,11 @@ class CuTeDSLMoERunner: # Slot -> token mapping: [0,0,...,0, 1,1,...,1, ...] (top_k repeats) self._token_indices = torch.arange( self.max_num_tokens, device=self.device - ).unsqueeze(1).expand(-1, self.top_k).reshape(-1) + ).unsqueeze(1).expand(-1, self.top_k).reshape(-1).clone() + + # Debug: verify token indices are correct + assert self._token_indices[:8].tolist() == [0, 0, 1, 1, 2, 2, 3, 3], \ + f"Token indices corrupted: {self._token_indices[:8].tolist()}" self._expert_id_range = torch.arange(self.num_experts, device=self.device)