Support llama3.1 (pytorch#4376)

Summary:
Pull Request resolved: pytorch#4376

Add scaled RoPE

Reviewed By: Gasoonjia

Differential Revision: D60129927

fbshipit-source-id: b8d2fadcd3e6985740965ad0185b8fb516806c22

examples/models/llama2/llama_transformer.py

@@ -8,6 +8,7 @@
 # Please refer to README.md in the same folder for more information.
 
 from dataclasses import dataclass
+from functools import partial
 from typing import Optional, Tuple
 
 import torch
@@ -101,6 +102,7 @@ class ModelArgs:
         None  # The official name to override self.rope_freq_base.
     )
     rope_freq_base: float = 10000.0  # The base frequency for RoPE. Keep it for BC.
+    use_scaled_rope: bool = False  # Use scaled RoPE, introduced in llama3.1.
     # Additional Model Metadata needed at runtime
     bos_idx: int = 1
     eos_idx: int = 3
@@ -453,7 +455,9 @@ def __init__(self, params: ModelArgs):
         if params.use_hf_rope:
             self.precompute_freqs_cis = hf_precompute_freqs_cis
         else:
-            self.precompute_freqs_cis = precompute_freqs_cis
+            self.precompute_freqs_cis = partial(
+                precompute_freqs_cis, use_scaled=params.use_scaled_rope
+            )
         freqs_cos, freqs_sin = self.precompute_freqs_cis(
             params.dim // params.n_heads,
             (

examples/models/llama2/rope.py

@@ -7,19 +7,49 @@
 
 # Different RoPE implementations
 
+import math
 from typing import Tuple
 
 import torch
 
 # ======================== Stock Implementation ========================
 
 
-def precompute_freqs_cis(dim: int, end: int, theta: float):
+def apply_scaling(freqs: torch.Tensor):
+    # Values obtained from grid search
+    scale_factor = 8
+    low_freq_factor = 1
+    high_freq_factor = 4
+    old_context_len = 8192  # original llama3 length
+
+    low_freq_wavelen = old_context_len / low_freq_factor
+    high_freq_wavelen = old_context_len / high_freq_factor
+    new_freqs = []
+    for freq in freqs:
+        wavelen = 2 * math.pi / freq
+        if wavelen < high_freq_wavelen:
+            new_freqs.append(freq)
+        elif wavelen > low_freq_wavelen:
+            new_freqs.append(freq / scale_factor)
+        else:
+            assert low_freq_wavelen != high_freq_wavelen
+            smooth = (old_context_len / wavelen - low_freq_factor) / (
+                high_freq_factor - low_freq_factor
+            )
+            new_freqs.append((1 - smooth) * freq / scale_factor + smooth * freq)
+    return torch.tensor(new_freqs, dtype=freqs.dtype, device=freqs.device)
+
+
+def precompute_freqs_cis(
+    dim: int, end: int, theta: float = 10000.0, use_scaled: bool = False
+):
     freqs = 1.0 / (
         theta ** (torch.arange(0, dim, 2, device="cpu")[: (dim // 2)].float() / dim)
     )
     t = torch.arange(end, device=freqs.device)  # pyre-ignore
-    freqs = torch.outer(t, freqs).float()  # pyre-ignore
+    if use_scaled:
+        freqs = apply_scaling(freqs)  # pyre-ignore
+    freqs = torch.outer(t, freqs).float()
     freqs_cos = torch.cos(freqs)
     freqs_sin = torch.sin(freqs)
     return freqs_cos, freqs_sin