From 993bb345d11ef60b08fe8ff4cac2c46274934711 Mon Sep 17 00:00:00 2001 From: biondizzle Date: Thu, 4 Jun 2026 00:15:38 +0000 Subject: [PATCH] DEBUG: fix VERBOSE reference in shared_expert, always print L2 gsa debug --- dsv4/layers/shared_expert.py | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/dsv4/layers/shared_expert.py b/dsv4/layers/shared_expert.py index c5ca22dc..e07e921a 100644 --- a/dsv4/layers/shared_expert.py +++ b/dsv4/layers/shared_expert.py @@ -368,8 +368,7 @@ class Nvfp4SharedExpert: x_fp4, x_sf, gsa_l2_gpu = quantize_nvfp4_gpu_fused(intermediate) # DEBUG: verify no CUDA errors from quantize kernel torch.cuda.synchronize() # DEBUG: catch async errors - if VERBOSE >= 1: - print(f" SE L2 gsa: gsa_gpu shape={tuple(gsa_l2_gpu.shape)} dtype={gsa_l2_gpu.dtype} dev={gsa_l2_gpu.device} _l2_gsa_buf shape={tuple(self._l2_gsa_buf.shape)} dev={self._l2_gsa_buf.device}", flush=True) + print(f" SE L2 gsa: gsa_gpu shape={tuple(gsa_l2_gpu.shape)} dtype={gsa_l2_gpu.dtype} dev={gsa_l2_gpu.device} _l2_gsa_buf shape={tuple(self._l2_gsa_buf.shape)} dev={self._l2_gsa_buf.device}", flush=True) self._l2_gsa_buf.copy_(gsa_l2_gpu[:1].reshape(1)) # GPU → GPU, no sync else: x_fp4, x_sf = quantize_activation_nvfp4(