Merge pull request #214 from zzhhjjj/match-transformers

Match Transformers RoPE implementation

src/nanotron/config/models_config.py

@@ -48,6 +48,9 @@ class LlamaConfig:
     rms_norm_eps: float = 1e-6
     rope_scaling: Optional[dict] = None
     rope_theta: float = 10000.0
+    rope_interleaved: bool = (
+        False  # The default value has been True, but for loading Llama3 checkpoints you have to set it to False
+    )
     tie_word_embeddings: bool = False
     use_cache: bool = True
     vocab_size: int = 32000

src/nanotron/models/llama.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 """PyTorch LLaMa model."""
 
-from typing import Dict, Optional, Union, List
+from typing import Dict, List, Optional, Union
 
 import torch
 from torch import nn
@@ -74,9 +74,10 @@ def init_rotary_embeddings(self):
             self.freqs_cis = self.freqs_cis.to(torch.float)
         assert self.freqs_cis.dtype == torch.float
         freqs = 1.0 / (
-            self.theta
-            ** (torch.arange(0, self.dim, 2, dtype=torch.float, device="cuda")[: (self.dim // 2)] / self.dim)
-        )
+            self.theta ** (torch.arange(0, self.dim, 2, dtype=torch.float, device="cpu")[: (self.dim // 2)] / self.dim)
+        ).to(
+            "cuda"
+        )  # should be computed on CPU, otherwise different results with Transformers.
         t = torch.arange(self.end, device="cuda")
         freqs = torch.outer(t, freqs).float()
         complex_freqs = torch.polar(torch.ones_like(freqs), freqs)
@@ -118,6 +119,78 @@ def forward(
         return x_out.type(dtype)
 
 
+## Copy from transformers. Non interleaved version of RoPE. Will be refactored later
+class LlamaRotaryEmbedding(nn.Module):
+    def __init__(self, dim: int, end: int, theta: float = 500000.0):
+        super().__init__()
+        self.dim = dim
+        self.end = end
+        self.theta = theta
+        self.init_rotary_embeddings()
+
+    def init_rotary_embeddings(self):
+        inv_freq = 1.0 / (
+            self.theta ** (torch.arange(0, self.dim, 2, dtype=torch.float, device="cpu") / self.dim)
+        )  # important to compute on CPU
+        self.register_buffer(
+            "inv_freq", torch.empty(self.dim // 2, dtype=torch.float, device="cuda"), persistent=False
+        )
+        self.inv_freq = self.inv_freq.to(
+            torch.float
+        )  # make it float32 before copy to avoid precision loss during copy_
+        self.inv_freq.copy_(inv_freq)
+
+    @torch.no_grad()
+    def forward(
+        self,
+        x: torch.Tensor,  # [batch_size, seq_length, num_heads, d_qk]
+        position_ids: Optional[torch.LongTensor],  # [batch_size, seq_length]
+    ):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 since bfloat16 loses precision on long contexts
+        # See https://github.com/huggingface/transformers/pull/29285
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+    def rotate_half(self, x):
+        """Rotates half the hidden dims of the input."""
+        x1 = x[..., : x.shape[-1] // 2]
+        x2 = x[..., x.shape[-1] // 2 :]
+        return torch.cat((-x2, x1), dim=-1)
+
+    def apply_rotary_pos_emb(self, q, k, cos, sin, unsqueeze_dim=2):
+        """Applies Rotary Position Embedding to the query and key tensors.
+
+        Args:
+            q (`torch.Tensor`): The query tensor.
+            k (`torch.Tensor`): The key tensor.
+            cos (`torch.Tensor`): The cosine part of the rotary embedding.
+            sin (`torch.Tensor`): The sine part of the rotary embedding.
+            unsqueeze_dim (`int`, *optional*, defaults to 1):
+                The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+                sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+                that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+                k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+                cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+                the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+        Returns:
+            `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+        """
+        cos = cos.unsqueeze(unsqueeze_dim)
+        sin = sin.unsqueeze(unsqueeze_dim)
+        q_embed = (q * cos) + (self.rotate_half(q) * sin)
+        k_embed = (k * cos) + (self.rotate_half(k) * sin)
+        return q_embed, k_embed
+
+
 class GLUActivation(nn.Module):
     def __init__(self, act_fn_name: str):
         super().__init__()
@@ -319,14 +392,24 @@ def __init__(
             tp_recompute_allgather=parallel_config.tp_recompute_allgather,
         )
         # TODO(kunhao): We want to have only one version per device and not one version per layer.
-        self.rotary_embedding = RotaryEmbedding(
-            dim=self.d_qk,
-            end=config.max_position_embeddings,
-            theta=config.rope_theta,
-        )
+        if config.rope_interleaved:
+            self.rotary_embedding = RotaryEmbedding(
+                dim=self.d_qk,
+                end=config.max_position_embeddings,
+                theta=config.rope_theta,
+            )
+        else:
+            self.rotary_embedding = LlamaRotaryEmbedding(
+                dim=self.d_qk,
+                end=config.max_position_embeddings,
+                theta=config.rope_theta,
+            )
+        self.rope_interleaved = config.rope_interleaved
 
         # NOTE: Only supported for training (TODO(fmom): position_ids not supported yet)
-        self.flash_rotary_embedding = FlashRotaryEmbedding(dim=self.d_qk, base=config.rope_theta, interleaved=True)
+        self.flash_rotary_embedding = FlashRotaryEmbedding(
+            dim=self.d_qk, base=config.rope_theta, interleaved=config.rope_interleaved
+        )
 
         self.o_proj = TensorParallelRowLinear(
             config.num_attention_heads * self.d_qk,
@@ -405,8 +488,16 @@ def forward(
             # Compute rotary embeddings
             # Note: keep track of old rotary embedding end to check if we need to enlarge k_cache and v_cache
             old_rotary_embed_end = self.rotary_embedding.end
-            query_states = self.rotary_embedding(query_states, position_ids=position_ids)
-            key_states = self.rotary_embedding(key_states, position_ids=position_ids)
+            # interleaved version.
+            if self.rope_interleaved:
+                query_states = self.rotary_embedding(query_states, position_ids=position_ids)
+                key_states = self.rotary_embedding(key_states, position_ids=position_ids)
+            # non interleaved version.
+            else:
+                cos, sin = self.rotary_embedding(value_states, position_ids)
+                query_states, key_states = self.rotary_embedding.apply_rotary_pos_emb(
+                    query_states, key_states, cos, sin
+                )
 
             if "key" not in store:
                 # First inference iteration (Prefill)
@@ -545,7 +636,7 @@ def forward(
                     cache_seqlens=position_offsets.contiguous(),
                     softmax_scale=softmax_scale,
                     causal=True,
-                    rotary_interleaved=False,  # GPT-NeoX style
+                    rotary_interleaved=False,  # the value is not used unless rotary_cos/sin is provided. https://github.com/Dao-AILab/flash-attention
                 )
 
             store.update(
@@ -620,9 +711,9 @@ def __init__(
 
         self.post_attention_layernorm = TritonRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
         self.mlp = MLP(config=config, parallel_config=parallel_config, tp_pg=tp_pg)
-        
+
         self.recompute_layer = parallel_config.recompute_layer
-        
+
     def _core_forward(
         self,
         hidden_states: Union[torch.Tensor, TensorPointer],
@@ -641,20 +732,20 @@ def _core_forward(
         hidden_states = hidden_states + residual
 
         return hidden_states, output["sequence_mask"]
-        
+
     def _checkpointed_forward(
         self,
         hidden_states: torch.Tensor,
         sequence_mask: torch.Tensor,
-        ) -> List[torch.Tensor]:
+    ) -> List[torch.Tensor]:
         return CheckpointFunction.apply(self._core_forward, True, hidden_states, sequence_mask)
 
     def forward(
         self,
         hidden_states: Union[torch.Tensor, TensorPointer],
         sequence_mask: Union[torch.Tensor, TensorPointer],
     ) -> Dict[str, Union[torch.Tensor, TensorPointer]]:
-        
+
         if self.recompute_layer and not isinstance(hidden_states, TensorPointer):
             hidden_states, sequence_mask = self._checkpointed_forward(hidden_states, sequence_mask)
         else:
@@ -665,6 +756,7 @@ def forward(
             "sequence_mask": sequence_mask,
         }
 
+
 class Embedding(nn.Module, AttachableStore):
     def __init__(self, tp_pg: dist.ProcessGroup, config: LlamaConfig, parallel_config: Optional[ParallelismArgs]):
         super().__init__()
@@ -727,7 +819,14 @@ def __init__(
             module_input_keys={"input_ids", "input_mask"},
             module_output_keys={"input_embeds"},
         )
-
+        log_rank(f"Initialize RoPE Theta = {config.rope_theta}", logger=logger, level=logging.INFO, rank=0)
+        if config.rope_interleaved:
+            log_rank(
+                "The RoPE interleaved version differs from the Transformers implementation. It's better to set rope_interleaved=False if you need to convert the weights to Transformers",
+                logger=logger,
+                level=logging.INFO,
+                rank=0,
+            )
         self.decoder = nn.ModuleList(
             [
                 PipelineBlock(

src/nanotron/nn/layer_norm.py

@@ -22,6 +22,8 @@ def forward(
         )
 
 
+# This is equivalent to LLaMA RMSNorm
+# https://github.com/huggingface/transformers/blob/28952248b19db29ca25ccf34a5eec413376494a9/src/transformers/models/llama/modeling_llama.py#L112
 class TritonRMSNorm(nn.Module):
     def __init__(self, hidden_size, eps=1e-5, device=None, dtype=None):
         factory_kwargs = {"device": device, "dtype": dtype}