support dynamic ntk in transformers

tools/transformers/internlm_model/modeling_internlm.py

@@ -124,6 +124,65 @@ class InternLMRotaryEmbedding(torch.nn.Module):
         )
 
 
+class InternLMDynamicNTKScalingRotaryEmbedding(torch.nn.Module):
+    """实现dynamic ntk rope；
+
+    需要保证：
+    1. 长度小于 seq len 时能够不断地复用；
+    2. 长度超过 seq len 时，每一个 新的token，都需要一个新的base；
+
+    Args:
+        InternLMRotaryEmbedding (_type_): _description_
+    """
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None, scaling_factor=1.0):
+        super().__init__()
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
+        self.register_buffer("inv_freq", inv_freq)
+        self.dim = dim
+        self.base = base
+        self.scaling_factor = scaling_factor
+
+        # Build here to make `torch.jit.trace` work.
+        self.max_position_embeddings = max_position_embeddings
+        self.max_seq_len_cached = max_position_embeddings
+        t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
+        freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+        # Different from paper, but it uses a different permutation in order to obtain the same calculation
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
+        self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
+    
+    def _update_cached(self, x, seq_len=None):
+        self.max_seq_len_cached = max(seq_len, self.max_position_embeddings)
+        if seq_len > self.max_position_embeddings:
+            base = self.base * (
+                (self.scaling_factor * seq_len / self.max_position_embeddings) - (self.scaling_factor - 1)
+            ) ** (self.dim / (self.dim - 2))
+            inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2).float().to(x.device) / self.dim))
+        else:
+            inv_freq = self.inv_freq
+        t = torch.arange(self.max_seq_len_cached, device=inv_freq.device, dtype=inv_freq.dtype)
+        freqs = torch.einsum("i,j->ij", t, inv_freq)
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
+        self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
+    
+
+    def forward(self, x, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        # This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
+        if seq_len <= self.max_position_embeddings:
+            # Reset the tables if the sequence length has changed,
+            if self.max_seq_len_cached > self.max_position_embeddings:
+                self._update_cached(x, seq_len)
+        else:
+            self._update_cached(x, seq_len)
+    
+        return (
+            self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
+            self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
+        )
+
 def rotate_half(x):
     """Rotates half the hidden dims of the input."""
     x1 = x[..., : x.shape[-1] // 2]
@@ -135,10 +194,18 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
     # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
     cos = cos.squeeze(1).squeeze(0)  # [seq_len, dim]
     sin = sin.squeeze(1).squeeze(0)  # [seq_len, dim]
-    cos = cos[position_ids].unsqueeze(1)  # [bs, 1, seq_len, dim]
+    cos = cos.unsqueeze(0).unsqueeze(0).expand(len(position_ids), -1, -1, -1)
-    sin = sin[position_ids].unsqueeze(1)  # [bs, 1, seq_len, dim]
+    sin = sin.unsqueeze(0).unsqueeze(0).expand(len(position_ids), -1, -1, -1)
-    q_embed = (q * cos) + (rotate_half(q) * sin)
+    if q.size(2) == 1:
-    k_embed = (k * cos) + (rotate_half(k) * sin)
+        q_embed = (q * cos[:, :, -1, :]) + (rotate_half(q) * sin[:, :, -1, :])
+    else:
+        q_embed = (q * cos) + (rotate_half(q) * sin)
+
+    if k.size(2) == 1:
+        k_embed = (k * cos[:, :, -1, :]) + (rotate_half(k) * sin[:, :, -1, :])
+    else:
+        k_embed = (k * cos) + (rotate_half(k) * sin)
+
     return q_embed, k_embed
 
 
@@ -179,7 +246,26 @@ class InternLMAttention(nn.Module):
         self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.bias)
         self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.bias)
         self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.bias)
-        self.rotary_emb = InternLMRotaryEmbedding(self.head_dim, max_position_embeddings=self.max_position_embeddings)
+        self.rotary_emb = self._init_rope()
+
+    def _init_rope(self):
+        if self.config.rotary["type"] == "origin":
+            self.rotary_emb = InternLMRotaryEmbedding(
+                self.head_dim, 
+                max_position_embeddings=self.max_position_embeddings,
+                base=self.config.rotary["base"],
+            )
+        elif self.config.rotary["type"] == "dynamic":
+            self.rotary_emb = InternLMDynamicNTKScalingRotaryEmbedding(
+                self.head_dim, 
+                max_position_embeddings=self.max_position_embeddings,
+                base=self.config.rotary["base"],
+                scaling_factor=self.config.rotary.get("scaling_factor", 1.0)
+            )
+        else:
+            raise ValueError("Currently we only support rotary embedding's type being one of ('origin', 'dynamic').")
+
+        return self.rotary_emb
 
     def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
         return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
@@ -199,20 +285,18 @@ class InternLMAttention(nn.Module):
         key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
         value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
 
-        kv_seq_len = key_states.shape[-2]
-        if past_key_value is not None:
-            kv_seq_len += past_key_value[0].shape[-2]
-        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
-        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
-        # [bsz, nh, t, hd]
-
         if past_key_value is not None:
             # reuse k, v, self_attention
             key_states = torch.cat([past_key_value[0], key_states], dim=2)
             value_states = torch.cat([past_key_value[1], value_states], dim=2)
 
+        # print(use_cache)
         past_key_value = (key_states, value_states) if use_cache else None
 
+        kv_seq_len = key_states.shape[-2]
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
         attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
 
         if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):