remove internal ticket #

vllm/hpu/ops.py

@@ -133,11 +133,11 @@ def static_fused_moe(hidden_states, w1, w2, score, topk):
     return final_hidden_states.view(-1, D)
 
 
-#TODO: remove after SW-195415 fix
+#TODO: remove after fusedsdpa fix for query_head != kv_head
 def repeat_kv(kv: torch.Tensor, n_rep: int) -> torch.Tensor:
     """
     This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep).
-    The hidden states go from (batch, num_key_value_heads, seqlen, head_dim) to
+    The kv go from (batch, num_key_value_heads, seqlen, head_dim) to
     (batch, num_attention_heads, seqlen, head_dim)
     """
     batch, num_key_value_heads, slen, head_dim = kv.shape
@@ -180,7 +180,7 @@ def prompt_attention(
         if query_heads != kv_heads:
             attn_weights = attn_weights.flatten(1, 2)
     else:
-        #TODO: remove after SW-195415 fix
+        #TODO: remove after fusedsdpa fix for query_heads != kv_heads
         if query_heads != kv_heads:
             key = repeat_kv(key, int(query_heads // kv_heads))
             value = repeat_kv(value, int(query_heads // kv_heads))