From 361b25272de9667649ed717b2ddda03d863122bc Mon Sep 17 00:00:00 2001 From: "Wang, Yi A" Date: Thu, 2 Nov 2023 20:55:51 -0700 Subject: [PATCH] fix backward error in DDP when running reward model finetune in RLHF Signed-off-by: Wang, Yi A --- optimum/habana/transformers/models/llama/modeling_llama.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/optimum/habana/transformers/models/llama/modeling_llama.py b/optimum/habana/transformers/models/llama/modeling_llama.py index d9f3a7cf7a..60c9962ec1 100755 --- a/optimum/habana/transformers/models/llama/modeling_llama.py +++ b/optimum/habana/transformers/models/llama/modeling_llama.py @@ -599,6 +599,9 @@ def prepare_inputs_for_generation( def apply_customized_rope(q, k, cos, sin, position_ids): if q.device.type == "hpu" and FusedRoPE: - return FusedRoPE.apply(q, cos, sin, position_ids), FusedRoPE.apply(k, cos, sin, position_ids) + # TODO: remove `.clone()` when SynapseAI v1.15 is released + return FusedRoPE.apply(q, cos.clone(), sin.clone(), position_ids), FusedRoPE.apply( + k, cos.clone(), sin.clone(), position_ids + ) else: return apply_rotary_pos_emb(q, k, cos, sin, position_ids)