From faf7c8cc51c16b30a1f22789cd3e6778c5a2164b Mon Sep 17 00:00:00 2001 From: biondizzle Date: Sun, 17 May 2026 15:18:07 +0000 Subject: [PATCH] Debug: print runner max_num_tokens and max_chunks --- vllm/nvfp4_cutedsl.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/vllm/nvfp4_cutedsl.py b/vllm/nvfp4_cutedsl.py index e39cf8d4..a440a42f 100644 --- a/vllm/nvfp4_cutedsl.py +++ b/vllm/nvfp4_cutedsl.py @@ -86,6 +86,10 @@ class CuTeDSLMoERunner: self._max_chunks_per_expert = cutedsl_ceil_div( self.max_num_tokens * self.top_k, self.num_experts * 128 ) + import os + print(f"[CLAWMINE] Runner init: max_num_tokens={self.max_num_tokens} top_k={self.top_k} " + f"num_experts={self.num_experts} max_chunks={self._max_chunks_per_expert} " + f"pid={os.getpid()}", flush=True) self._buffers_allocated = False def _fill_token_indices(self):