Add RoPE scaling for Llama3.1

Author: turboderp

exllamav2/config.py

@@ -10,7 +10,9 @@
 T = TypeVar('T')
 no_default = object()
 
-def read(input_dict: dict[str, Any], expected_type: type, keys: str | list[str], default = no_default) -> T:
+def read(input_dict: dict[str, Any], expected_type: type | list[type], keys: str | list[str], default = no_default) -> T:
+
+    expected_types = expected_type if isinstance(expected_type, list) else [expected_type]
 
     if isinstance(keys, str): keys = [keys]
 
@@ -34,10 +36,10 @@ def read(input_dict: dict[str, Any], expected_type: type, keys: str | list[str],
             if expected_type == int and isinstance(x, float) and x == int(x):
                 x = int(x)
 
-            if isinstance(x, expected_type):
-                return cast(T, x)
-            else:
-                raise TypeError(f"Value for {key} is not of expected type {expected_type}")
+            for t in expected_types:
+                if isinstance(x, t):
+                    return cast(T, x)
+            raise TypeError(f"Value for {key} is not of expected type {expected_type}")
 
     if default != no_default: return default
     raise ValueError(f"Missing any of the following keys: {keys}")
@@ -105,7 +107,10 @@ class ExLlamaV2Config:
     attn_logit_softcapping: float | None
     sliding_window: int
     norm_head: int | None
-
+    l3_rope_factor: float | None
+    l3_rope_low_freq_factor: float | None
+    l3_rope_high_freq_factor: float | None
+    l3_rope_original_max_position_embeddings: int | None
     checkpoint_fused_mlp: bool
     checkpoint_offset_qzeros: bool
 
@@ -191,10 +196,13 @@ def prepare(self, no_tensors: bool = False):
         # Vocab params
 
         self.bos_token_id = read(read_config, int, "bos_token_id", None)  # 1
-        self.eos_token_id = read(read_config, int, "eos_token_id", None)  # 2
+        self.eos_token_id = read(read_config, [int, list], "eos_token_id", None)  # 2
         self.pad_token_id = read(read_config, int, "pad_token_id", None)  # 0
         self.vocab_size = read(read_config, int, "vocab_size")
 
+        if isinstance(self.eos_token_id, list):
+            self.eos_token_id = self.eos_token_id[0]  # TODO: Figure out a way to maybe use all the EOS tokens somehow
+
         # Standard params
 
         self.initializer_range = read(read_config, float, ["initializer_range"])
@@ -287,6 +295,13 @@ def prepare(self, no_tensors: bool = False):
                 self.alt_rope_method = "su"
             # if scaling_type == "yarn":
             #     self.scale_alpha_value = factor
+            rope_type = rs.get("rope_type", None)
+            if rope_type == "llama3":
+                self.alt_rope_method = "llama3"
+                self.l3_rope_factor = rs["factor"]
+                self.l3_rope_low_freq_factor = rs["low_freq_factor"]
+                self.l3_rope_high_freq_factor = rs["high_freq_factor"]
+                self.l3_rope_original_max_position_embeddings = rs["original_max_position_embeddings"]
 
         # Checkpoint format (for GPTQ models)
 

exllamav2/model.py

@@ -129,6 +129,31 @@ def get_scratch_slice(self, size_bytes):
         return scratch_slice
 
 
+    @staticmethod
+    def _apply_scaling(
+        freqs: torch.Tensor,
+        scale_factor: float = 8,
+        low_freq_factor: float = 1,
+        high_freq_factor: float = 4,
+        old_context_len: int = 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 prepare_sincos(self):
 
         device = _torch_device(self.device_idx)
@@ -163,6 +188,19 @@ def prepare_sincos(self):
 
             inv_freq = 1.0 / (ext_factors * base ** (torch.arange(0, head_dim, 2, device = device).float() / head_dim))
 
+        # Llama 3.1
+
+        elif cfg.alt_rope_method == "llama3":
+
+            inv_freq = 1.0 / (base ** (torch.arange(0, head_dim, 2, device = device).float() / head_dim))
+            inv_freq = self._apply_scaling(
+                inv_freq,
+                cfg.l3_rope_factor,
+                cfg.l3_rope_low_freq_factor,
+                cfg.l3_rope_high_freq_factor,
+                cfg.l3_rope_original_max_position_embeddings,
+            )
+
         # Regular
 
         else: