Add YaRN and Dynamic-YaRN RoPE Scaling Methods (#30910)

* Add YaRN and Dynamic-YaRN RoPE Scaling Methods

YaRN (Yet another RoPE extension method) combines the NTK-By-Parts
Interpolation and Attention Scaling methods, improving upon existing
RoPE interpolation methods for longer context window sizes.

Fine-tuned models maintain their original performance across benchmarks
while enabling efficient extrapolation and transfer learning for
quicker convergence, especially in compute-limited environments.

We implement YaRN and Dynamic-YaRN for the following list of models:

 - LLaMA
 - Falcon
 - GPT-NeoX
 - Olmo
 - Persimmon
 - Phi
 - StableLM
 - OpenLLaMA

New unit tests are added to assert YaRN's correct behavior on both
short and long sequence inputs.

For more details, please refer to https://arxiv.org/abs/2309.00071.

Co-authored-by: Miguel Almeida <miguel.pessanha.almeida@tecnico.ulisboa.pt>

* Refactor YaRN implementation for LLaMA

Iterate on YaRN implementation for LLaMA and remove diff from remaining
models for increased PR modularity.

This commit includes the following changes:
- Merge 'yarn_rope_scaling' and 'rope_scaling' dictionaries
- Remove unnecessary attributes ('extrapolation_factor' and 'finetuned')
  from YaRN classes
- Inherit 'forward' method in YaRN classes from superclass
- Rename 'yarn' method to 'compute_yarn_scaling'
- Extend YaRN tests with further assertions
- Fix style inconsistencies

Co-authored-by: Miguel Monte e Freitas <miguelmontefreitas@tecnico.ulisboa.pt>

* Refactor Tensor Building Logic for YaRN

- Comply with the the tensor building logic introduced in #30743
- Add referencing to the optimized Attention Factor equation
- Remove Dynamic YaRN for a more agile deployment

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

* remove unwanted file

---------

Co-authored-by: Miguel Almeida <miguel.pessanha.almeida@tecnico.ulisboa.pt>
Co-authored-by: mig-mfreitas <mig-mfreitas@users.noreply.github.com>
Co-authored-by: Joao Gante <joao@huggingface.co>

src/transformers/models/falcon/modeling_falcon.py

@@ -283,7 +283,6 @@ class FalconAttention(nn.Module):
         self.attention_dropout = nn.Dropout(config.attention_dropout)
         self.num_kv_heads = config.num_kv_heads if (self.new_decoder_architecture or not self.multi_query) else 1
 
-    # Copied from transformers.models.llama.modeling_llama.LlamaAttention._init_rope with Llama->Falcon
     def _init_rope(self):
         if self.config.rope_scaling is None:
             self.rotary_emb = FalconRotaryEmbedding(

src/transformers/models/fuyu/configuration_fuyu.py

@@ -188,7 +188,6 @@ class FuyuConfig(PretrainedConfig):
             **kwargs,
         )
 
-    # Copied from transformers.models.llama.configuration_llama.LlamaConfig._rope_scaling_validation
     def _rope_scaling_validation(self):
         """
         Validate the `rope_scaling` configuration.

src/transformers/models/gpt_neox/configuration_gpt_neox.py

@@ -154,7 +154,6 @@ class GPTNeoXConfig(PretrainedConfig):
                 "The hidden size is not divisble by the number of attention heads! Make sure to update them!"
             )
 
-    # Copied from transformers.models.llama.configuration_llama.LlamaConfig._rope_scaling_validation
     def _rope_scaling_validation(self):
         """
         Validate the `rope_scaling` configuration.

src/transformers/models/llama/configuration_llama.py

@@ -84,13 +84,22 @@ class LlamaConfig(PretrainedConfig):
         rope_theta (`float`, *optional*, defaults to 10000.0):
             The base period of the RoPE embeddings.
         rope_scaling (`Dict`, *optional*):
-            Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling
+            Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports three scaling
-            strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is
+            strategies: linear, dynamic and yarn. Their scaling factor must be a float greater than 1. The expected format is
             `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
             `max_position_embeddings` to the expected new maximum. See the following thread for more information on how
             these scaling strategies behave:
             https://www.reddit.com/r/LocalLLaMA/comments/14mrgpr/dynamically_scaled_rope_further_increases/. This is an
             experimental feature, subject to breaking API changes in future versions.
+            For the `yarn` strategy, the dictionary may also contain the following fields:
+                `original_max_position_embeddings` (`int`, *optional*):
+                    The original maximum sequence length. This is used to scale the RoPE embeddings.
+                `attention_factor` (`float`, *optional*):
+                    The attention scaling factor. If unspecified, it defaults to `0.1 ln(s) + 1`, where `s` is the                    `original_max_position_embeddings/max_position_embeddings` ratio.
+                `beta_fast` (`float`, *optional*):
+                    Parameter to set the boundary for extrapolation (only) in the linear ramp function.
+                `beta_slow` (`float`, *optional*):
+                    Parameter to set the boundary for interpolation (only) in the linear ramp function.
         attention_bias (`bool`, *optional*, defaults to `False`):
             Whether to use a bias in the query, key, value and output projection layers during self-attention.
         attention_dropout (`float`, *optional*, defaults to 0.0):
@@ -178,15 +187,52 @@ class LlamaConfig(PretrainedConfig):
         if self.rope_scaling is None:
             return
 
-        if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 2:
+        if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) < 2:
             raise ValueError(
-                "`rope_scaling` must be a dictionary with two fields, `type` and `factor`, " f"got {self.rope_scaling}"
+                "`rope_scaling` must be a dictionary with a minimum of two fields, `type` and `factor`, "
+                f"got {self.rope_scaling}"
             )
         rope_scaling_type = self.rope_scaling.get("type", None)
         rope_scaling_factor = self.rope_scaling.get("factor", None)
-        if rope_scaling_type is None or rope_scaling_type not in ["linear", "dynamic"]:
+        if rope_scaling_type is None or rope_scaling_type not in ["linear", "dynamic", "yarn"]:
             raise ValueError(
-                f"`rope_scaling`'s type field must be one of ['linear', 'dynamic'], got {rope_scaling_type}"
+                f"`rope_scaling`'s type field must be one of ['linear', 'dynamic', 'yarn'], got {rope_scaling_type}"
             )
         if rope_scaling_factor is None or not isinstance(rope_scaling_factor, float) or rope_scaling_factor <= 1.0:
             raise ValueError(f"`rope_scaling`'s factor field must be a float > 1, got {rope_scaling_factor}")
+
+        if rope_scaling_type != "yarn":
+            return
+
+        if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) > 6:
+            raise ValueError(
+                "`rope_scaling` with type "
+                f"{rope_scaling_type}"
+                " must be a dictionary with a maximum of six fields, `type`, `factor`,"
+                "`original_max_position_embeddings`, `attention_factor`, `beta_fast`, `beta_slow`, "
+                f"got {self.rope_scaling}"
+            )
+        original_max_position_embeddings = self.rope_scaling.get("original_max_position_embeddings", None)
+        attention_factor = self.rope_scaling.get("attention_factor", None)
+        beta_fast = self.rope_scaling.get("beta_fast", None)
+        beta_slow = self.rope_scaling.get("beta_slow", None)
+
+        if original_max_position_embeddings is not None and not isinstance(original_max_position_embeddings, int):
+            raise ValueError(
+                f"`rope_scaling`'s original_max_position_embeddings field must be an int, got {original_max_position_embeddings}"
+            )
+        if attention_factor is not None and not isinstance(attention_factor, float) or attention_factor < 0:
+            raise ValueError(
+                f"`rope_scaling`'s attention_factor field must be a float greater than 0, got {attention_factor}"
+            )
+        if beta_fast is not None and not isinstance(beta_fast, float):
+            raise ValueError(f"`rope_scaling`'s beta_fast field must be a float, got {beta_fast}")
+        if beta_slow is not None and not isinstance(beta_slow, float):
+            raise ValueError(f"`rope_scaling`'s beta_slow field must be a float, got {beta_slow}")
+
+        b_fast = beta_fast if beta_fast is not None else 32
+        b_slow = beta_slow if beta_slow is not None else 1
+        if b_fast < b_slow:
+            raise ValueError(
+                f"`rope_scaling`'s beta_fast field must be greater than beta_slow, got beta_fast={b_fast} and beta_slow={b_slow}"
+            )

src/transformers/models/llama/modeling_llama.py

@@ -132,6 +132,77 @@ class LlamaDynamicNTKScalingRotaryEmbedding(LlamaRotaryEmbedding):
         return cos, sin
 
 
+class LlamaYarnScalingRotaryEmbedding(LlamaRotaryEmbedding):
+    def __init__(
+        self,
+        dim,
+        max_position_embeddings=2048,
+        base=10000,
+        scaling_factor=1,
+        original_max_position_embeddings=2048,
+        attention_factor=None,
+        beta_fast=32,
+        beta_slow=1,
+        device=None,
+    ):
+        super().__init__(dim, max_position_embeddings, base, device, scaling_factor)
+
+        self.original_max_position_embeddings = original_max_position_embeddings
+        self.attention_factor = attention_factor
+        self.beta_fast = beta_fast
+        self.beta_slow = beta_slow
+
+        if self.attention_factor is None:
+            # Recommended attention factor for LLaMA models.
+            # For more details please refer to https://arxiv.org/pdf/2309.00071, Eq. 22.
+            self.attention_factor = 0.1 * math.log(scaling_factor) + 1.0
+
+        self.compute_yarn_scaling(device)
+
+    # Inverse dimension formula to find the dimension based on the number of rotations
+    def find_correction_dim(self, num_rotations, dim, base=10000, max_position_embeddings=2048):
+        return (dim * math.log(max_position_embeddings / (num_rotations * 2 * math.pi))) / (2 * math.log(base))
+
+    # Find dimension range bounds based on rotations
+    def find_correction_range(self, low_rot, high_rot, dim, base=10000, max_position_embeddings=2048):
+        low = math.floor(self.find_correction_dim(low_rot, dim, base, max_position_embeddings))
+        high = math.ceil(self.find_correction_dim(high_rot, dim, base, max_position_embeddings))
+        return max(low, 0), min(high, dim - 1)
+
+    def linear_ramp_mask(self, min, max, dim):
+        if min == max:
+            max += 0.001  # Prevent singularity
+
+        linear_func = (torch.arange(dim, dtype=torch.float32) - min) / (max - min)
+        ramp_func = torch.clamp(linear_func, 0, 1)
+        return ramp_func
+
+    def forward(self, x, position_ids=None):
+        # Difference to the original RoPE: applies a scaling factor computed with
+        # the YaRN method (NTK-by-Parts + Attn Scaling)
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        cos, sin = super().forward(x, position_ids)
+        cos = cos * self.mscale
+        sin = sin * self.mscale
+        return cos, sin
+
+    def compute_yarn_scaling(self, device):
+        pos_freqs = self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim)
+        inv_freq_extrapolation = 1.0 / pos_freqs
+        inv_freq_interpolation = 1.0 / (self.scaling_factor * pos_freqs)
+
+        low, high = self.find_correction_range(
+            self.beta_fast, self.beta_slow, self.dim, self.base, self.original_max_position_embeddings
+        )
+        # Get n-dimensional rotational scaling corrected for extrapolation
+        inv_freq_mask = 1 - self.linear_ramp_mask(low, high, self.dim // 2).float().to(device)
+        inv_freq = inv_freq_interpolation * (1 - inv_freq_mask) + inv_freq_extrapolation * inv_freq_mask
+
+        self.register_buffer("inv_freq", inv_freq)
+        # Get n-dimensional magnitude scaling corrected for interpolation
+        self.mscale = self.attention_factor
+
+
 def rotate_half(x):
     """Rotates half the hidden dims of the input."""
     x1 = x[..., : x.shape[-1] // 2]
@@ -258,6 +329,15 @@ class LlamaAttention(nn.Module):
         else:
             scaling_type = self.config.rope_scaling["type"]
             scaling_factor = self.config.rope_scaling["factor"]
+            # Yarn parameters
+            kwargs = {
+                "dim": self.config.rope_scaling.get("original_max_position_embeddings", None),
+                "max_position_embeddings": self.config.rope_scaling.get("attention_factor", None),
+                "base": self.config.rope_scaling.get("beta_fast", None),
+                "scaling_factor": self.config.rope_scaling.get("beta_slow", None),
+            }
+            kwargs = {k: v for k, v in kwargs.items() if v is not None}
+
             if scaling_type == "linear":
                 self.rotary_emb = LlamaLinearScalingRotaryEmbedding(
                     self.head_dim,
@@ -272,6 +352,14 @@ class LlamaAttention(nn.Module):
                     scaling_factor=scaling_factor,
                     base=self.rope_theta,
                 )
+            elif scaling_type == "yarn":
+                self.rotary_emb = LlamaYarnScalingRotaryEmbedding(
+                    self.head_dim,
+                    max_position_embeddings=self.max_position_embeddings,
+                    scaling_factor=scaling_factor,
+                    base=self.rope_theta,
+                    **kwargs,
+                )
             else:
                 raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
 

src/transformers/models/olmo/configuration_olmo.py

@@ -160,7 +160,6 @@ class OlmoConfig(PretrainedConfig):
             **kwargs,
         )
 
-    # Copied from transformers.models.llama.configuration_llama.LlamaConfig._rope_scaling_validation
     def _rope_scaling_validation(self):
         """
         Validate the `rope_scaling` configuration.

src/transformers/models/olmo/modeling_olmo.py

@@ -236,7 +236,6 @@ class OlmoAttention(nn.Module):
         self.o_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=config.attention_bias)
         self._init_rope()
 
-    # Copied from transformers.models.llama.modeling_llama.LlamaAttention._init_rope with Llama->Olmo
     def _init_rope(self):
         if self.config.rope_scaling is None:
             self.rotary_emb = OlmoRotaryEmbedding(

src/transformers/models/persimmon/configuration_persimmon.py

@@ -138,7 +138,6 @@ class PersimmonConfig(PretrainedConfig):
             **kwargs,
         )
 
-    # Copied from transformers.models.llama.configuration_llama.LlamaConfig._rope_scaling_validation
     def _rope_scaling_validation(self):
         """
         Validate the `rope_scaling` configuration.

src/transformers/models/phi/configuration_phi.py

@@ -165,7 +165,6 @@ class PhiConfig(PretrainedConfig):
             **kwargs,
         )
 
-    # Copied from transformers.models.llama.configuration_llama.LlamaConfig._rope_scaling_validation
     def _rope_scaling_validation(self):
         """
         Validate the `rope_scaling` configuration.

src/transformers/models/stablelm/configuration_stablelm.py

@@ -164,7 +164,6 @@ class StableLmConfig(PretrainedConfig):
             **kwargs,
         )
 
-    # Copied from transformers.models.llama.configuration_llama.LlamaConfig._rope_scaling_validation
     def _rope_scaling_validation(self):
         """
         Validate the `rope_scaling` configuration.

tests/models/llama/test_modeling_llama.py

@@ -55,6 +55,7 @@ if is_torch_available():
         LlamaDynamicNTKScalingRotaryEmbedding,
         LlamaLinearScalingRotaryEmbedding,
         LlamaRotaryEmbedding,
+        LlamaYarnScalingRotaryEmbedding,
     )
 
 
@@ -397,7 +398,7 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
     def test_save_load_fast_init_from_base(self):
         pass
 
-    @parameterized.expand([("linear",), ("dynamic",)])
+    @parameterized.expand([("linear",), ("dynamic",), ("yarn",)])
     def test_model_rope_scaling_from_config(self, scaling_type):
         config, _ = self.model_tester.prepare_config_and_inputs_for_common()
         short_input = ids_tensor([1, 10], config.vocab_size)
@@ -491,6 +492,26 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
             torch.testing.assert_close(ntk_sin_long, original_sin_long)
         self.assertTrue((ntk_scaling_rope.inv_freq <= original_rope.inv_freq).all())
 
+        # Sanity check Yarn RoPE scaling
+        yarn_scaling_rope = LlamaYarnScalingRotaryEmbedding(
+            head_dim,
+            max_position_embeddings=config.max_position_embeddings,
+            base=config.rope_theta,
+            scaling_factor=scaling_factor,
+        ).to(torch_device)
+        yarn_cos_short, yarn_sin_short = yarn_scaling_rope(x, position_ids_short)
+        yarn_cos_long, yarn_sin_long = yarn_scaling_rope(x, position_ids_long)
+        torch.testing.assert_close(yarn_cos_short, yarn_cos_long[:, :short_input_length, :])
+        torch.testing.assert_close(yarn_sin_short, yarn_sin_long[:, :short_input_length, :])
+        with self.assertRaises(AssertionError):
+            torch.testing.assert_close(yarn_cos_short, original_cos_short)
+        with self.assertRaises(AssertionError):
+            torch.testing.assert_close(yarn_sin_short, original_sin_short)
+        with self.assertRaises(AssertionError):
+            torch.testing.assert_close(yarn_cos_long, original_cos_long)
+        with self.assertRaises(AssertionError):
+            torch.testing.assert_close(yarn_sin_long, original_sin_long)
+
     @require_flash_attn
     @require_torch_gpu
     @require_bitsandbytes