From 4a573995925b0682a37524924690492c5f253fe7 Mon Sep 17 00:00:00 2001 From: biondizzle Date: Mon, 18 May 2026 15:59:59 +0000 Subject: [PATCH] Add debug prints for input_global_scale_inv check --- vllm/patches/deepseek_v4.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/vllm/patches/deepseek_v4.py b/vllm/patches/deepseek_v4.py index c1b7cf01..481b67fc 100644 --- a/vllm/patches/deepseek_v4.py +++ b/vllm/patches/deepseek_v4.py @@ -1723,8 +1723,12 @@ class DeepseekV4Model(nn.Module): # by dequantizing to BF16 temporarily and running a warmup. for proj_name in ["fused_wqa_wkv", "wq_b", "wo_b"]: if not hasattr(attn, proj_name): + if layer_idx == 0: + print(f"[CLAWMINE] Layer 0: {proj_name} NOT on attn") continue mod = getattr(attn, proj_name) + if layer_idx == 0: + print(f"[CLAWMINE] Layer 0: {proj_name} dtype={mod.weight.dtype} has_input_global_scale_inv={hasattr(mod, 'input_global_scale_inv')} has_input_scale={hasattr(mod, 'input_scale')}") if not hasattr(mod, "input_global_scale_inv"): continue if mod.weight.dtype not in (torch.uint8, torch.int8):