From 48e4cb625d7762c51d5a2b81bd8a126f23e6faac Mon Sep 17 00:00:00 2001 From: biondizzle Date: Sun, 17 May 2026 06:24:15 +0000 Subject: [PATCH] fix: default activation global_scale so runner works without finalize_weights --- vllm/nvfp4_cutedsl.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/vllm/nvfp4_cutedsl.py b/vllm/nvfp4_cutedsl.py index c9542e3a..31b72dc8 100644 --- a/vllm/nvfp4_cutedsl.py +++ b/vllm/nvfp4_cutedsl.py @@ -60,8 +60,10 @@ class CuTeDSLMoERunner: self._l1_gsb = None self._l2_gsb = None - self._l1_activation_global_scale = None # set from checkpoint input_scale - self._l2_activation_global_scale = None + # Default: 1/2688 ≈ 0.000372 (amax=1 → gs=1/2688) + # Overridden in finalize_weights with checkpoint input_scale or warmup value + self._l1_activation_global_scale = 1.0 / (6.0 * 448.0) + self._l2_activation_global_scale = 1.0 / (6.0 * 448.0) # Pre-allocated cudagraph buffers (set in _allocate_buffers) self._token_indices = None