From 1e103a3e231f4d884699e1b05bfb3dc009d52666 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Fri, 21 Mar 2025 16:42:32 -0700 Subject: [PATCH] Update llama.py --- unsloth/models/llama.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py index 88494df3d2..c2c278ec22 100644 --- a/unsloth/models/llama.py +++ b/unsloth/models/llama.py @@ -929,6 +929,8 @@ def LlamaModel_fast_forward_inference( temp_mlp = torch.empty((2, bsz, 1, mlp_size), dtype = X.dtype, device = "cuda:0") temp_gate, temp_up = temp_mlp[0], temp_mlp[1] + print(type(past_key_values), len(past_key_values)) + seq_len = past_key_values[0][0].shape[-2] if bsz != 1: attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(