[CI] Fix adaptation prompt CI on transformers main (huggingface#1465)

* fix adaptation prompt CI

* add fix

* forward contrib credits from discussion

* add docstring

---------

Co-authored-by: BenjaminBossan <BenjaminBossan@users.noreply.github.com>

src/peft/tuners/adaption_prompt/utils.py

@@ -59,19 +59,18 @@ def llama_apply_rotary_pos_emb(q, cos, sin, position_ids):
 
 def llama_compute_query_states(model: nn.Module, **kwargs) -> torch.Tensor:
     """
-    Compute query states for Llama models specifically.
-
-    They need to be recomputed as the forward() method of the original LlamaModel in the transformers library does not
-    return them. See the related discussion in the PR: https://github.com/huggingface/peft/pull/268
+    Compute query states for Llama models specifically. They need to be recomputed as the forward() method of the
+    original LlamaModel in the transformers library does not return them. See the related discussion in the PR:
+    https://github.com/huggingface/peft/pull/268
     """
     hidden_states = kwargs.get("hidden_states")
     position_ids = kwargs.get("position_ids")
     past_key_value = kwargs.get("past_key_value")
     bsz, q_len, _ = hidden_states.size()
     query_states = model.q_proj(hidden_states).view(bsz, q_len, model.num_heads, model.head_dim).transpose(1, 2)
     value_states = model.v_proj(hidden_states).view(bsz, q_len, model.num_heads, model.head_dim).transpose(1, 2)
-
     seq_len = q_len
+
     if past_key_value is not None:
         if isinstance(past_key_value, tuple):
             # for transformers <= 4.35
@@ -80,30 +79,24 @@ def llama_compute_query_states(model: nn.Module, **kwargs) -> torch.Tensor:
             # since transformers 4.36, this is a DynamicCache instance
             seq_len += past_key_value.get_seq_length(model.layer_idx)
 
-            # For transformers > 4.37.2 `position_ids` became a required arguments in the
-            # rotary embedding's forward pass. and cos/sin are indexed through the
-            # `rotary_emb` forward pass.
-            if "position_ids" in list(inspect.signature(model.rotary_emb.forward).parameters):
-                if position_ids is None:
-                    kv_seq_len = value_states.shape[-2]
-                    past_seen_tokens = past_key_value.get_usable_length(kv_seq_len, model.layer_idx)
-                    kv_seq_len += past_seen_tokens
-
-                    new_cache_positions = torch.arange(
-                        past_seen_tokens, past_seen_tokens + q_len, device=value_states.device
-                    )
-                    position_ids = new_cache_positions.unsqueeze(0)
-
-                cos, sin = model.rotary_emb(value_states, seq_len=kv_seq_len, position_ids=position_ids)
-
-                # Here cos and sin are are already indexed correctly, therefore to avoid adding
-                # boilerplate in `llama_apply_rotary_pos_emb` we just return here the correct query states
-                # embeddings
-                return (query_states * cos) + (llama_rotate_half(query_states) * sin)
-
-    cos, sin = model.rotary_emb(value_states, seq_len=seq_len)
+    # For transformers > 4.37.2 `position_ids` became a required arguments in the rotary embedding's forward pass.
+    if "position_ids" not in inspect.signature(model.rotary_emb.forward).parameters:
+        # TODO we assume that position_ids is not None here, not sure if that is safe but the old code also did that
+        cos, sin = model.rotary_emb(value_states, seq_len=seq_len)
+        return llama_apply_rotary_pos_emb(query_states, cos, sin, position_ids)
+
+    past_seen_tokens = 0
+    if position_ids is None:
+        # Compute position_ids, since they are required for transformers > 4.37.2
+        if past_key_value is None:
+            new_cache_positions = torch.arange(q_len, q_len + q_len, device=value_states.device)
+        else:
+            past_seen_tokens = past_key_value.get_usable_length(q_len, model.layer_idx)
+            new_cache_positions = torch.arange(past_seen_tokens, past_seen_tokens + q_len, device=value_states.device)
+        position_ids = new_cache_positions.unsqueeze(0)
 
-    return llama_apply_rotary_pos_emb(query_states, cos, sin, position_ids)
+    cos, sin = model.rotary_emb(value_states, seq_len=q_len + past_seen_tokens, position_ids=position_ids)
+    return (query_states * cos) + (llama_rotate_half(query_states) * sin)
 
 
 def is_adaption_prompt_trainable(params: str) -> bool: