diff --git a/vllm/patches/deepseek_v4.py b/vllm/patches/deepseek_v4.py index bf7eb1a6..348bb041 100644 --- a/vllm/patches/deepseek_v4.py +++ b/vllm/patches/deepseek_v4.py @@ -1728,10 +1728,16 @@ class DeepseekV4Model(nn.Module): # BF16 dequantization: attention projections (except wo_a) for proj_name in bf16_proj_names: if not hasattr(attn, proj_name): + if layer_idx == 0: + print(f"[CLAWMINE] Layer 0: {proj_name} NOT FOUND on attn (type={type(attn).__name__})") continue mod = getattr(attn, proj_name) if not hasattr(mod, "weight"): + if layer_idx == 0: + print(f"[CLAWMINE] Layer 0: {proj_name} has no weight attr") continue + if layer_idx == 0: + print(f"[CLAWMINE] Layer 0: {proj_name} weight dtype={mod.weight.dtype}") if mod.weight.dtype in (torch.uint8, torch.int8): E2M1_LUT = torch.tensor([0, 0.5, 1, 1.5, 2, 3, 4, 6], dtype=torch.bfloat16) self._dequant_nvfp4_to_bf16(mod, E2M1_LUT)