Add comments on RoPE initialization (#1176)

Author: WoosukKwon

vllm/model_executor/layers/attention.py

@@ -264,6 +264,15 @@ def __init__(
         self.is_neox_style = is_neox_style
 
         # Create the cos and sin cache.
+        # NOTE(woosuk): The HF implementation uses `torch.arange(...).float()`.
+        # However, we use `torch.arange(..., dtype=torch.float)` instead to
+        # avoid numerical issues with large base values (e.g., 10000000).
+        # This may cause a slight numerical difference between the HF
+        # implementation and ours.
+        # NOTE(woosuk): To exactly match the HF implementation, we need to
+        # use CPU to compute the cache and then move it to GPU. However, we
+        # create the cache on GPU for faster initialization. This may cause
+        # a slight numerical difference between the HF implementation and ours.
         inv_freq = 1.0 / (base**(torch.arange(
             0, rotary_dim, 2, dtype=torch.float, device="cuda") / rotary_dim))
         t = torch.arange(max_position, dtype=torch.float, device="cuda")
@@ -274,7 +283,6 @@ def __init__(
 
         # FIXME(woosuk): This assumes that we configure the default dtype when
         # initializing the model.
-        # TODO(woosuk): Make it more robust.
         torch_dtype = torch.get_default_dtype()
         cache = cache.to(torch_dtype)
         # Embedding size: [max_position, rotary_dim]