Update modeling_gemmoe.py

modeling_gemmoe.py

@@ -24,8 +24,6 @@ import torch.nn.functional as F
 import torch.utils.checkpoint
 from torch import nn
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
-import flash_attn_cuda_utils
-
 from transformers.activations import ACT2FN
 from transformers.cache_utils import Cache, DynamicCache, StaticCache
 from transformers.modeling_attn_mask_utils import (
@@ -375,72 +373,84 @@ class GemmoeFlashAttention2(GemmoeAttention):
         # TODO: Remove this attribute once Flash Attention for RoCm is bumped to 2.1.
         self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
 
-def forward(
-    self,
-    hidden_states: torch.Tensor,
-    attention_mask: Optional[torch.LongTensor] = None,
-    position_ids: Optional[torch.LongTensor] = None,
-    past_key_value: Optional[Cache] = None,
-    output_attentions: bool = False,
-    use_cache: bool = False,
-    cache_position: Optional[torch.LongTensor] = None,
-    **kwargs,
-) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-    output_attentions = False
-
-    bsz, q_len, _ = hidden_states.size()
-
-    query_states = self.q_proj(hidden_states)
-    key_states = self.k_proj(hidden_states)
-    value_states = self.v_proj(hidden_states)
-
-    query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
-    key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
-    value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
-
-    cos, sin = self.rotary_emb(value_states, position_ids, seq_len=None)
-    query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, None)
-
-    past_key_value = getattr(self, "past_key_value", past_key_value)
-    if past_key_value is not None:
-        cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
-        key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
-
-    query_states = query_states.transpose(1, 2)
-    key_states = key_states.transpose(1, 2)
-    value_states = value_states.transpose(1, 2)
-
-    dropout_rate = self.attention_dropout if self.training else 0.0
-
-    input_dtype = query_states.dtype
-    if input_dtype == torch.float32:
-        if torch.is_autocast_enabled():
-            target_dtype = torch.get_autocast_gpu_dtype()
-        elif hasattr(self.config, "_pre_quantization_dtype"):
-            target_dtype = self.config._pre_quantization_dtype
-        else:
-            target_dtype = self.q_proj.weight.dtype
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        output_attentions = False
 
-        logger.warning_once(
-            f"The input hidden states seems to be silently casted in float32, this might be related to"
-            f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
-            f" {target_dtype}."
-        )
-        query_states = query_states.to(target_dtype)
-        key_states = key_states.to(target_dtype)
-        value_states = value_states.to(target_dtype)
+        bsz, q_len, _ = hidden_states.size()
 
-    attn_output = flash_attn_cuda_utils.pyt_flash_scaled_dot_attention(
-        query_states, key_states, value_states, attn_mask=attention_mask, dropout_prob=dropout_rate
-    )
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
 
-    attn_output = attn_output.reshape(bsz, q_len, -1).contiguous()
-    attn_output = self.o_proj(attn_output)
+        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=None)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, None)
 
-    if not output_attentions:
-        attn_weights = None
+        past_key_value = getattr(self, "past_key_value", past_key_value)
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; position_ids needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
-    return attn_output, attn_weights, past_key_value
+        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
+        # to be able to avoid many of these transpose/reshape/view.
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        dropout_rate = self.attention_dropout if self.training else 0.0
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in the correct dtype just to be sure everything works as expected.
+        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
+        # in fp32. (GemmoeRMSNorm handles it correctly)
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = self._flash_attention_forward(
+            query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
 
     def _flash_attention_forward(
         self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None