biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: default activation global_scale so runner works without finalize_weights

Browse Source
This commit is contained in:
biondizzle
2026-05-17 06:24:15 +00:00
parent d2965b432d
commit 48e4cb625d
1 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/nvfp4_cutedsl.py
View File

@@ -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