Update modeling_gemmoe.py

modeling_gemmoe.py

@@ -28,6 +28,7 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 from transformers.activations import ACT2FN
 from transformers.cache_utils import Cache, DynamicCache, StaticCache
 from transformers.modeling_attn_mask_utils import (
+    AttentionMaskConverter,
     _prepare_4d_causal_attention_mask,
 )
 from transformers.modeling_outputs import SequenceClassifierOutputWithPast, MoeModelOutputWithPast, MoeCausalLMOutputWithPast
@@ -176,52 +177,71 @@ ALL_LAYERNORM_LAYERS.append(GemmoeRMSNorm)
 class GemmoeRotaryEmbedding(nn.Module):
     def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
         super().__init__()
+
         self.dim = dim
         self.max_position_embeddings = max_position_embeddings
         self.base = base
-        self._set_cos_sin_cache(seq_len=max_position_embeddings, device=device, dtype=torch.get_default_dtype())
-
-    def _set_cos_sin_cache(self, seq_len, device, dtype):
-        self.max_seq_len_cached = seq_len
-        freq_exponents = (2.0 / self.dim) * (
-            torch.arange(self.dim // 2, dtype=torch.int64, device="cpu").float()
-        )
-        timescale = self.base ** freq_exponents
-        positions = torch.arange(self.max_seq_len_cached, device="cpu", dtype=torch.int64).float()
-        radians_new = positions[..., None] / timescale[None, None, :]
-        radians_new = radians_new.squeeze(0)
-        emb = torch.cat((radians_new, radians_new), dim=-1)
-        cos = emb.cos().to(device=device, non_blocking=True)
-        sin = emb.sin().to(device=device, non_blocking=True)
-        self.register_buffer("cos_cached", cos, persistent=False)
-        self.register_buffer("sin_cached", sin, persistent=False)
-
-    def forward(self, x, position_ids=None, seq_len=None):
-        if seq_len is None:
-            seq_len = x.size(2)
-        if seq_len > self.max_seq_len_cached:
-            self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
-        return (
-            self.cos_cached[:seq_len],
-            self.sin_cached[:seq_len],
-        )
-
+        self.register_buffer("inv_freq", None, persistent=False)
+
+    @torch.no_grad()
+    def forward(self, x, position_ids, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        if self.inv_freq is None:
+            self.inv_freq = 1.0 / (
+                self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim)
+            )
+        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)
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
 def rotate_half(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(q, k, cos, sin, position_ids=None):
-    """Applies Rotary Position Embedding to the query and key tensors."""
-    seq_len, dim = q.shape[-2], q.shape[-1]
-    cos = cos[:seq_len].view(1, 1, seq_len, dim)
-    sin = sin[:seq_len].view(1, 1, seq_len, dim)
+
+# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """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.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        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) + (rotate_half(q) * sin)
     k_embed = (k * cos) + (rotate_half(k) * sin)
     return q_embed, k_embed
 
-def repeat_kv(self, hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+
+    
+# Copied from transformers.models.llama.modeling_llama.repeat_kv
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
     """
     This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
     num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
@@ -233,14 +253,9 @@ def repeat_kv(self, hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
     return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
 
 class GemmoeAttention(nn.Module):
-    """
-    Multi-headed attention module for Gemmoe model.
-
-    Args:
-        config (GemmoeConfig): The configuration object for the Gemmoe model.
-        layer_idx (Optional[int]): The index of the layer. Default is None.
-    """
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
 
+    # Ignore copy
     def __init__(self, config: GemmoeConfig, layer_idx: Optional[int] = None):
         super().__init__()
         self.config = config
@@ -251,6 +266,7 @@ class GemmoeAttention(nn.Module):
                 "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
                 "when creating this class."
             )
+
         self.attention_dropout = config.attention_dropout
         self.hidden_size = config.hidden_size
         self.num_heads = config.num_attention_heads
@@ -266,15 +282,16 @@ class GemmoeAttention(nn.Module):
                 f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
                 f" and `num_heads`: {self.num_heads})."
             )
+
         self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
         self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
         self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
-        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.attention_bias)
         self.rotary_emb = GemmoeRotaryEmbedding(
-				self.head_dim,
-				max_position_embeddings=self.max_position_embeddings,
-				base=self.rope_theta,
-			)
+            self.head_dim,
+            max_position_embeddings=self.max_position_embeddings,
+            base=self.rope_theta,
+        )
 
     def forward(
         self,
@@ -287,25 +304,6 @@ class GemmoeAttention(nn.Module):
         cache_position: Optional[torch.LongTensor] = None,
         **kwargs,
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-        """
-        Forward pass of the attention module.
-
-        Args:
-            hidden_states (torch.Tensor): The input hidden states.
-            attention_mask (Optional[torch.Tensor]): The attention mask. Default is None.
-            position_ids (Optional[torch.LongTensor]): The position IDs. Default is None.
-            past_key_value (Optional[Cache]): The past key-value cache. Default is None.
-            output_attentions (bool): Whether to output the attention weights. Default is False.
-            use_cache (bool): Whether to use caching. Default is False.
-            cache_position (Optional[torch.LongTensor]): The cache position. Default is None.
-            **kwargs: Additional keyword arguments.
-
-        Returns:
-            Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-                - The output hidden states.
-                - The attention weights (if `output_attentions=True`).
-                - The past key-value cache (if `use_cache=True`).
-        """
         bsz, q_len, _ = hidden_states.size()
 
         query_states = self.q_proj(hidden_states)
@@ -317,17 +315,16 @@ class GemmoeAttention(nn.Module):
         value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
 
         past_key_value = getattr(self, "past_key_value", past_key_value)
-
         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 past_key_value is not None:
-            # sin and cos are specific to RoPE models; position_ids needed for the static cache
+            # sin and cos are specific to RoPE models; cache_position 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)
 
-            key_states = self.repeat_kv(key_states, self.num_key_value_groups)
-            value_states = self.repeat_kv(value_states, self.num_key_value_groups)
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
 
         attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
 
@@ -341,7 +338,6 @@ class GemmoeAttention(nn.Module):
         # upcast attention to fp32
         attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
         attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
-
         attn_output = torch.matmul(attn_weights, value_states)
 
         if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
@@ -351,8 +347,8 @@ class GemmoeAttention(nn.Module):
             )
 
         attn_output = attn_output.transpose(1, 2).contiguous()
-        attn_output = attn_output.view(bsz, q_len, -1)
 
+        attn_output = attn_output.view(bsz, q_len, -1)
         attn_output = self.o_proj(attn_output)
 
         if not output_attentions:
@@ -360,17 +356,24 @@ class GemmoeAttention(nn.Module):
 
         return attn_output, attn_weights, past_key_value
 
+
+# Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2 with Llama->Gemmoe
 class GemmoeFlashAttention2(GemmoeAttention):
     """
     Gemmoe flash attention module. This module inherits from `GemmoeAttention` as the weights of the module stays
     untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
     flash attention and deal with padding tokens in case the input contains any of them.
     """
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        # TODO: Remove this attribute once Flash Attention for RoCm is bumped to 2.1.
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
         self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
 
+    # Ignore copy
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -401,8 +404,9 @@ class GemmoeFlashAttention2(GemmoeAttention):
         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:
-            # sin and cos are specific to RoPE models; position_ids needed for the static cache
+            # sin and cos are specific to RoPE models; cache_position 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)
 
@@ -419,6 +423,7 @@ class GemmoeFlashAttention2(GemmoeAttention):
         # 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():
@@ -434,6 +439,7 @@ class GemmoeFlashAttention2(GemmoeAttention):
                 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)
@@ -467,7 +473,7 @@ class GemmoeFlashAttention2(GemmoeAttention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
@@ -484,6 +490,7 @@ class GemmoeFlashAttention2(GemmoeAttention):
             query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
                 query_states, key_states, value_states, attention_mask, query_length
             )
+
             cu_seqlens_q, cu_seqlens_k = cu_seq_lens
             max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
 
@@ -499,6 +506,7 @@ class GemmoeFlashAttention2(GemmoeAttention):
                 softmax_scale=softmax_scale,
                 causal=causal,
             )
+
             attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
         else:
             attn_output = flash_attn_func(
@@ -509,15 +517,14 @@ class GemmoeFlashAttention2(GemmoeAttention):
 
     def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
         indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
-
         batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
         key_layer = index_first_axis(
             key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
         )
         value_layer = index_first_axis(
             value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
         )
-
         if query_length == kv_seq_len:
             query_layer = index_first_axis(
                 query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
@@ -546,24 +553,16 @@ class GemmoeFlashAttention2(GemmoeAttention):
             (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
         )
 
+
+# Copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Gemmoe
 class GemmoeSdpaAttention(GemmoeAttention):
     """
     Gemmoe attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
-    GemmoeAttention as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    `GemmoeAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
     SDPA API.
     """
 
-    def repeat_kv(self, x, n_rep):
-        """
-        This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
-        num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
-        """
-        batch, num_key_value_heads, slen, head_dim = x.shape
-        if n_rep == 1:
-            return x
-        x = x[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
-        return x.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
-
+    # Ignore copy
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -576,11 +575,10 @@ class GemmoeSdpaAttention(GemmoeAttention):
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
         if output_attentions:
             # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
-            # logger.warning_once(
-            "GemmoeModel is using GemmoeSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
-            'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
-            # )
-            
+            logger.warning_once(
+                "GemmoeModel is using GemmoeSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
             return super().forward(
                 hidden_states=hidden_states,
                 attention_mask=attention_mask,
@@ -590,7 +588,7 @@ class GemmoeSdpaAttention(GemmoeAttention):
                 use_cache=use_cache,
                 cache_position=cache_position,
             )
-            
+
         bsz, q_len, _ = hidden_states.size()
 
         query_states = self.q_proj(hidden_states)
@@ -605,23 +603,19 @@ class GemmoeSdpaAttention(GemmoeAttention):
         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:
-            # sin and cos are specific to RoPE models; position_ids needed for the static cache
+            # sin and cos are specific to RoPE models; cache_position 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)
 
-        key_states = self.repeat_kv(key_states, self.num_key_value_groups)
-        value_states = self.repeat_kv(value_states, self.num_key_value_groups)
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
 
         causal_mask = attention_mask
         if attention_mask is not None and cache_position is not None:
             causal_mask = causal_mask[:, :, cache_position, : key_states.shape[-2]]
 
-        # Ensure query, key, and value states have the same dtype
-        common_dtype = query_states.dtype
-        key_states = key_states.to(dtype=common_dtype)
-        value_states = value_states.to(dtype=common_dtype)
-
         # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
         # Reference: https://github.com/pytorch/pytorch/issues/112577.
         if query_states.device.type == "cuda" and causal_mask is not None:
@@ -629,10 +623,6 @@ class GemmoeSdpaAttention(GemmoeAttention):
             key_states = key_states.contiguous()
             value_states = value_states.contiguous()
 
-        # Cast causal_mask to the same dtype as query_states
-        if causal_mask is not None:
-            causal_mask = causal_mask.to(dtype=query_states.dtype)
-        
         attn_output = torch.nn.functional.scaled_dot_product_attention(
             query_states,
             key_states,
@@ -643,167 +633,113 @@ class GemmoeSdpaAttention(GemmoeAttention):
 
         attn_output = attn_output.transpose(1, 2).contiguous()
         attn_output = attn_output.view(bsz, q_len, -1)
+
         attn_output = self.o_proj(attn_output)
 
         return attn_output, None, past_key_value
 
-GEMMOE_ATTENTION_CLASSES = {
-	"eager": GemmoeAttention,
-	"flash_attention_2": GemmoeFlashAttention2,
-	"sdpa": GemmoeSdpaAttention,
-	}
-
-class GemmoeMLP(nn.Module):
-    def __init__(self, config, hidden_size=None, intermediate_size=None):
-        super().__init__()
-        self.config = config
-        self.hidden_size = config.hidden_size if hidden_size is None else hidden_size
-        self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size
 
-        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
-        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
-        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
-        self.act_fn = ACT2FN[config.hidden_act]
-
-    def forward(self, x):
-        if self.config.pretraining_tp > 1:
-            slice = self.intermediate_size // self.config.pretraining_tp
-            gate_proj_slices = self.gate_proj.weight.split(slice, dim=0)
-            up_proj_slices = self.up_proj.weight.split(slice, dim=0)
-            down_proj_slices = self.down_proj.weight.split(slice, dim=1)
-
-            gate_proj = torch.cat(
-                [F.linear(x, gate_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1
-            )
-            up_proj = torch.cat([F.linear(x, up_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1)
-
-            intermediate_states = (self.act_fn(gate_proj) * up_proj).split(slice, dim=2)
-            down_proj = [
-                F.linear(intermediate_states[i], down_proj_slices[i]) for i in range(self.config.pretraining_tp)
-            ]
-            down_proj = sum(down_proj)
-        else:
-            down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+GEMMOE_ATTENTION_CLASSES = {
+    "eager": GemmoeAttention,
+    "flash_attention_2": GemmoeFlashAttention2,
+    "sdpa": GemmoeSdpaAttention,
+}
 
-        return down_proj
-    
-class MoEGate(nn.Module):
-    def __init__(self, config):
+class GemmoeBlockSparseTop2MLP(nn.Module):
+    def __init__(self, config: GemmoeConfig):
         super().__init__()
-        self.config = config
-        self.top_k = config.num_experts_per_tok
-        self.n_routed_experts = config.n_routed_experts
-
-        self.scoring_func = config.scoring_func
-        self.alpha = config.aux_loss_alpha
-        self.seq_aux = config.seq_aux
+        self.ffn_dim = config.intermediate_size
+        self.hidden_dim = config.hidden_size
 
-        # topk selection algorithm
-        self.norm_topk_prob = config.norm_topk_prob
-        self.gating_dim = config.hidden_size
-        self.weight = nn.Parameter(torch.empty((self.n_routed_experts, self.gating_dim)))
-        self.reset_parameters()
+        self.w1 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
+        self.w2 = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False)
+        self.w3 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
 
-    def reset_parameters(self) -> None:
-        import torch.nn.init as init
-        init.kaiming_uniform_(self.weight, a=math.sqrt(5))
+        self.act_fn = approx_gelu
 
     def forward(self, hidden_states):
-        bsz, seq_len, h = hidden_states.shape
-        ### compute gating score
-        hidden_states = hidden_states.view(-1, h)
-        logits = F.linear(hidden_states, self.weight, None)
-        if self.scoring_func == 'softmax':
-            scores = logits.softmax(dim=-1)
-        else:
-            raise NotImplementedError(f'insupportable scoring function for MoE gating: {self.scoring_func}')
-
-        ### select top-k experts
-        topk_weight, topk_idx = torch.topk(scores, k=self.top_k, dim=-1, sorted=False)
-
-        ### norm gate to sum 1
-        if self.top_k > 1 and self.norm_topk_prob:
-            denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20
-            topk_weight = topk_weight / denominator
-
-        ### expert-level computation auxiliary loss
-        if self.training and self.alpha > 0.0:
-            scores_for_aux = scores
-            aux_topk = self.top_k
-            # always compute aux loss based on the naive greedy topk method
-            topk_idx_for_aux_loss = topk_idx.view(bsz, -1)
-            if self.seq_aux:
-                scores_for_seq_aux = scores_for_aux.view(bsz, seq_len, -1)
-                ce = torch.zeros(bsz, self.n_routed_experts, device=hidden_states.device)
-                ce.scatter_add_(1, topk_idx_for_aux_loss, torch.ones(bsz, seq_len * aux_topk, device=hidden_states.device)).div_(seq_len * aux_topk / self.n_routed_experts)
-                aux_loss = (ce * scores_for_seq_aux.mean(dim=1)).sum(dim=1).mean() * self.alpha
-            else:
-                mask_ce = F.one_hot(topk_idx_for_aux_loss.view(-1), num_classes=self.n_routed_experts)
-                ce = mask_ce.float().mean(0)
-                Pi = scores_for_aux.mean(0)
-                fi = ce * self.n_routed_experts
-                aux_loss = (Pi * fi).sum() * self.alpha
-        else:
-            aux_loss = None
-        return topk_idx, topk_weight, aux_loss
+        current_hidden_states = self.act_fn(self.w1(hidden_states)) * self.w3(hidden_states)
+        current_hidden_states = self.w2(current_hidden_states)
+        return current_hidden_states
 
+class GemmoeBlockSparseTop2MLP(GemmoeBlockSparseTop2MLP):
+    def __init__(self, *args, **kwargs):
+        logger.warning_once(
+            "GemmoeBLockSparseTop2MLP is deprecated by GemmoeBlockSparseTop2MLP and will be removed in v4.40."
+        )
+        super().__init__(*args, **kwargs)
 
 class GemmoeSparseMoeBlock(nn.Module):
+    """
+    This implementation is
+    strictly equivalent to standard MoE with full capacity (no
+    dropped tokens). It's faster since it formulates MoE operations
+    in terms of block-sparse operations to accomodate imbalanced
+    assignments of tokens to experts, whereas standard MoE either
+    (1) drop tokens at the cost of reduced performance or (2) set
+    capacity factor to number of experts and thus waste computation
+    and memory on padding.
+    """
+
     def __init__(self, config):
         super().__init__()
         self.hidden_dim = config.hidden_size
         self.ffn_dim = config.intermediate_size
         self.num_experts = config.num_local_experts
-        self.top_k = 2
+        self.top_k = config.num_experts_per_tok
 
-        self.gate = MoEGate(config)
+        # gating
+        self.gate = nn.Linear(self.hidden_dim, self.num_experts, bias=False)
 
-        self.experts = nn.ModuleList([GemmoeMLP(config) for _ in range(self.num_experts)])
+        self.experts = nn.ModuleList([GemmoeBlockSparseTop2MLP(config) for _ in range(self.num_experts)])
 
-    def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        """ """
         batch_size, sequence_length, hidden_dim = hidden_states.shape
         hidden_states = hidden_states.view(-1, hidden_dim)
+        # router_logits: (batch * sequence_length, n_experts)
+        router_logits = self.gate(hidden_states)
 
-        topk_idx, topk_weight, aux_loss = self.gate(hidden_states)
-
+        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
         # we cast back to the input dtype
-        topk_weight = topk_weight.to(hidden_states.dtype)
+        routing_weights = routing_weights.to(hidden_states.dtype)
 
-        hidden_states = hidden_states.repeat_interleave(self.top_k, dim=0)
+        final_hidden_states = torch.zeros(
+            (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
+        )
 
-        y = torch.empty_like(hidden_states)
+        # One hot encode the selected experts to create an expert mask
+        # this will be used to easily index which expert is going to be sollicitated
+        expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
 
-        flat_topk_idx = topk_idx.view(-1)
-        for i in range(self.num_experts):
-            expert = self.experts[i]
-            expert_output = expert(hidden_states[flat_topk_idx == i])
-            y[flat_topk_idx == i] = expert_output.to(y.dtype)  # Cast expert_output to the same dtype as y
+        # Loop over all available experts in the model and perform the computation on each expert
+        for expert_idx in range(self.num_experts):
+            expert_layer = self.experts[expert_idx]
+            idx, top_x = torch.where(expert_mask[expert_idx])
 
-        y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1)
+            if top_x.shape[0] == 0:
+                continue
 
-        final_hidden_states = y.reshape(batch_size, sequence_length, hidden_dim)
-        return final_hidden_states, aux_loss
-    
-class AddAuxiliaryLoss(torch.autograd.Function):
-    """
-    The trick function of adding auxiliary (aux) loss,
-    which includes the gradient of the aux loss during backpropagation.
-    """
-    @staticmethod
-    def forward(ctx, x, loss):
-        assert loss.numel() == 1
-        ctx.dtype = loss.dtype
-        ctx.required_aux_loss = loss.requires_grad
-        return x
+            # in torch it is faster to index using lists than torch tensors
+            top_x_list = top_x.tolist()
+            idx_list = idx.tolist()
 
-    @staticmethod
-    def backward(ctx, grad_output):
-        grad_loss = None
-        if ctx.required_aux_loss:
-            grad_loss = torch.ones(1, dtype=ctx.dtype, device=grad_output.device)
-        return grad_output, grad_loss
+            # Index the correct hidden states and compute the expert hidden state for
+            # the current expert. We need to make sure to multiply the output hidden
+            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+            current_state = hidden_states[None, top_x_list].reshape(-1, hidden_dim)
+            current_hidden_states = expert_layer(current_state) * routing_weights[top_x_list, idx_list, None]
+
+            # However `index_add_` only support torch tensors for indexing so we'll use
+            # the `top_x` tensor here.
+            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
+        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+        return final_hidden_states, router_logits
 
     
+# Copied from transformers.models.llama.modeling_llama.LlamaDecoderLayer with LLAMA->GEMMOE,Llama->Gemmoe
 class GemmoeDecoderLayer(nn.Module):
     def __init__(self, config: GemmoeConfig, layer_idx: int):
         super().__init__()
@@ -824,10 +760,31 @@ class GemmoeDecoderLayer(nn.Module):
         output_attentions: Optional[bool] = False,
         output_router_logits: Optional[bool] = False,
         use_cache: Optional[bool] = False,
-        cache_position: Optional[torch.LongTensor] = None,
         **kwargs,
     ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        if "padding_mask" in kwargs:
+            warnings.warn(
+                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+            )
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, sequence_length)` where padding elements are indicated by 0.
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_router_logits (`bool`, *optional*):
+                Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
+                should not be returned during inference.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+        """
+
         residual = hidden_states
+
         hidden_states = self.input_layernorm(hidden_states)
 
         # Self Attention
@@ -838,20 +795,15 @@ class GemmoeDecoderLayer(nn.Module):
             past_key_value=past_key_value,
             output_attentions=output_attentions,
             use_cache=use_cache,
-            cache_position=cache_position,
-            **kwargs,
         )
         hidden_states = residual + hidden_states
 
         # Fully Connected
         residual = hidden_states
         hidden_states = self.post_attention_layernorm(hidden_states)
-        hidden_states, aux_loss = self.block_sparse_moe(hidden_states)
+        hidden_states, router_logits = self.block_sparse_moe(hidden_states)
         hidden_states = residual + hidden_states
 
-        if aux_loss is not None:
-            hidden_states = AddAuxiliaryLoss.apply(hidden_states, aux_loss)
-
         outputs = (hidden_states,)
 
         if output_attentions:
@@ -860,331 +812,364 @@ class GemmoeDecoderLayer(nn.Module):
         if use_cache:
             outputs += (present_key_value,)
 
+        if output_router_logits:
+            outputs += (router_logits,)
+
         return outputs
 
+
 GEMMOE_START_DOCSTRING = r"""
-This model inherits from [PreTrainedModel]. Check the superclass documentation for the generic methods the
-library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
-etc.)
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`GemmoeConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
 """
 
+
 @add_start_docstrings(
-"The bare Gemmoe Model outputting raw hidden-states without any specific head on top.",
-GEMMOE_START_DOCSTRING,
+    "The bare Gemmoe Model outputting raw hidden-states without any specific head on top.",
+    GEMMOE_START_DOCSTRING,
 )
 
 class GemmoePreTrainedModel(PreTrainedModel):
-	config_class = GemmoeConfig
-	base_model_prefix = "model"
-	supports_gradient_checkpointing = True
-	_keep_in_fp32_modules = ["inv_freq", "rotary_emb", "cos_cached", "sin_cached"]
-	_no_split_modules = ["GemmoeDecoderLayer"]
-	_skip_keys_device_placement = ["past_key_values", "causal_mask"]
-	_supports_flash_attn_2 = True
-	_supports_sdpa = True
-	_supports_cache_class = True
- 
-	def _init_weights(self, module):
-		std = self.config.initializer_range
-		if isinstance(module, nn.Linear):
-			module.weight.data.normal_(mean=0.0, std=std)
-			if module.bias is not None:
-				module.bias.data.zero_()
-		elif isinstance(module, nn.Embedding):
-			module.weight.data.normal_(mean=0.0, std=std)
-			if module.padding_idx is not None:
-				module.weight.data[module.padding_idx].zero_()
-
-	def _setup_cache(self, cache_cls, max_batch_size, max_cache_len: Optional[int] = None):
-		if self.config._attn_implementation == "flash_attention_2" and cache_cls == StaticCache:
-			raise ValueError(
-				"`static` cache implementation is not compatible with `attn_implementation==flash_attention_2` "
-				"make sure to use `sdpa` in the mean time, and open an issue at https://github.com/huggingface/transformers"
-			)
-		if max_cache_len > self.model.causal_mask.shape[-1] or self.device != self.model.causal_mask.device:
-			causal_mask = torch.full((max_cache_len, max_cache_len), fill_value=1, device=self.device)
-			self.register_buffer("causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False)
-
-		for layer in self.model.layers:
-			weights = layer.self_attn.o_proj.weight
-			layer.self_attn.past_key_value = cache_cls(
-				self.config, max_batch_size, max_cache_len, device=weights.device, dtype=weights.dtype
-			)
-
-	def _reset_cache(self):
-		for layer in self.model.layers:
-			layer.self_attn.past_key_value = None
+    config_class = GemmoeConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _keep_in_fp32_modules = ["inv_freq", "rotary_emb", "cos_cached", "sin_cached"]
+    _no_split_modules = ["GemmoeDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values", "causal_mask"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_cache_class = True
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    def _setup_cache(self, cache_cls, max_batch_size, max_cache_len: Optional[int] = None):
+        if self.config._attn_implementation == "flash_attention_2" and cache_cls == StaticCache:
+            raise ValueError(
+                "`static` cache implementation is not compatible with `attn_implementation==flash_attention_2` "
+                "make sure to use `sdpa` in the mean time, and open an issue at https://github.com/huggingface/transformers"
+            )
+
+        if max_cache_len > self.model.causal_mask.shape[-1] or self.device != self.model.causal_mask.device:
+            causal_mask = torch.full((max_cache_len, max_cache_len), fill_value=1, device=self.device)
+            self.register_buffer("causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False)
+
+        for layer in self.model.layers:
+            weights = layer.self_attn.o_proj.weight
+            layer.self_attn.past_key_value = cache_cls(
+                self.config, max_batch_size, max_cache_len, device=weights.device, dtype=weights.dtype
+            )
+
+    def _reset_cache(self):
+        for layer in self.model.layers:
+            layer.self_attn.past_key_value = None
+
 
 GEMMOE_INPUTS_DOCSTRING = r"""
-Args:
-input_ids (torch.LongTensor of shape (batch_size, sequence_length)):
-Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
-it.
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+
+            Two formats are allowed:
+            - a [`~cache_utils.Cache`] instance;
+            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
+            cache format.
+
+            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+            legacy cache format will be returned.
+
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+            the complete sequence length.
 """
 
+
 @add_start_docstrings(
-"The bare Gemmoe Model outputting raw hidden-states without any specific head on top.",
-GEMMOE_START_DOCSTRING,
+    "The bare Gemmoe Model outputting raw hidden-states without any specific head on top.",
+    GEMMOE_START_DOCSTRING,
 )
 
 class GemmoeModel(GemmoePreTrainedModel):
-	"""
-	Transformer decoder consisting of config.num_hidden_layers layers. Each layer is a [GemmoeDecoderLayer]Args:
-		config: GemmoeConfig
-	"""
-
-
-	def __init__(self, config: GemmoeConfig):
-		super().__init__(config)
-		self.padding_idx = config.pad_token_id
-		self.vocab_size = config.vocab_size
-
-		self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
-		self.layers = nn.ModuleList(
-			[GemmoeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
-		)
-          
-		self.norm = GemmoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-
-		self.gradient_checkpointing = False
-
-		# Register a causal mask to separate causal and padding mask creation. Merging happens in the attention class.
-		# NOTE: This is not friendly with TorchScript, ONNX, ExportedProgram serialization for very large `max_position_embeddings`.
-		causal_mask = torch.full(
-			(config.max_position_embeddings, config.max_position_embeddings), fill_value=True, dtype=torch.bool
-		)
-		self.register_buffer("causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False)
-
-		# Initialize weights and apply final processing
-		self.post_init()
-
-	def get_input_embeddings(self):
-		return self.embed_tokens
-
-	def set_input_embeddings(self, value):
-		self.embed_tokens = value
-
-	@add_start_docstrings_to_model_forward(GEMMOE_INPUTS_DOCSTRING)
-	@replace_return_docstrings(output_type=MoeModelOutputWithPast, config_class=_CONFIG_FOR_DOC)
-	def forward(
-		self,
-		input_ids: torch.LongTensor = None,
-		attention_mask: Optional[torch.Tensor] = None,
-		position_ids: Optional[torch.LongTensor] = None,
-		past_key_values: Optional[List[torch.FloatTensor]] = None,
-		inputs_embeds: Optional[torch.FloatTensor] = None,
-		use_cache: Optional[bool] = None,
-		output_attentions: Optional[bool] = None,
-		output_hidden_states: Optional[bool] = None,
-		output_router_logits: Optional[bool] = None,
-		return_dict: Optional[bool] = None,
-		cache_position: Optional[torch.LongTensor] = None,
-	) -> Union[Tuple, MoeModelOutputWithPast]:
-		"""
-		Forward pass of the sequence classification model.
-
-		Args:
-			input_ids: Input token IDs.
-			attention_mask: Attention mask.
-			position_ids: Position IDs.
-			past_key_values: Past key-value pairs.
-			inputs_embeds: Input embeddings.
-			labels: Labels for sequence classification.
-			use_cache: Whether to use cache.
-			output_attentions: Whether to output attentions.
-			output_hidden_states: Whether to output hidden states.
-			return_dict: Whether to return a dictionary or tuple.
-
-		Returns:
-			Output of the sequence classification model.
-		"""
-		output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-		output_hidden_states = (
-			output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-		)
-		output_router_logits = (
-			output_router_logits if output_router_logits is not None else self.config.output_router_logits
-		)
-		use_cache = use_cache if use_cache is not None else self.config.use_cache
-		return_dict = return_dict if return_dict is not None else self.config.use_return_dict
-
-		if (input_ids is None) ^ (inputs_embeds is not None):
-			raise ValueError(
-				"You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
-			)
-
-		if self.gradient_checkpointing and self.training and use_cache:
-			logger.warning_once(
-				"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
-			)
-			use_cache = False
-
-		if inputs_embeds is None:
-			inputs_embeds = self.embed_tokens(input_ids)
-
-		past_seen_tokens = 0
-		if use_cache:  # kept for BC (cache positions)
-			if not isinstance(past_key_values, StaticCache):
-				past_key_values = DynamicCache.from_legacy_cache(past_key_values)
-			past_seen_tokens = past_key_values.get_seq_length()
-
-		if cache_position is None:
-			cache_position = torch.arange(
-				past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
-			)
-
-		if position_ids is None:
-			position_ids = cache_position.unsqueeze(0)
-
-		causal_mask = self._update_causal_mask(attention_mask, inputs_embeds)
-
-		hidden_states = inputs_embeds
-
-		# Normalize
-		scale_factor = torch.tensor(math_sqrt(self.config.hidden_size), dtype=hidden_states.dtype)
-		hidden_states = hidden_states * scale_factor
-		# Decoder layers
-		all_hidden_states = () if output_hidden_states else None
-		all_self_attns = () if output_attentions else None
-		all_router_logits = () if output_router_logits else None
-		next_decoder_cache = None
-
-		for decoder_layer in self.layers:
-			if output_hidden_states:
-				all_hidden_states += (hidden_states,)
-
-			if self.gradient_checkpointing and self.training:
-				layer_outputs = self._gradient_checkpointing_func(
-					decoder_layer.__call__,
-					hidden_states,
-					causal_mask,
-					position_ids,
-					past_key_values,
-					output_attentions,
-					output_router_logits,
-					use_cache,
-					cache_position,
-				)
-			else:
-				layer_outputs = decoder_layer(
-					hidden_states,
-					attention_mask=causal_mask,
-					position_ids=position_ids,
-					past_key_value=past_key_values,
-					output_attentions=output_attentions,
-					output_router_logits=output_router_logits,
-					use_cache=use_cache,
-					cache_position=cache_position,
-				)
-
-			hidden_states = layer_outputs[0]
-			if use_cache:
-				next_decoder_cache = layer_outputs[2 if output_attentions else 1]
-			if output_attentions:
-				all_self_attns += (layer_outputs[1],)
-			if output_router_logits:
-				all_router_logits += (layer_outputs[-1],)
-
-		hidden_states = self.norm(hidden_states)
-
-		# Add hidden states from the last decoder layer
-		if output_hidden_states:
-			all_hidden_states += (hidden_states,)
-
-		next_cache = None
-		if use_cache:
-			next_cache = (
-				next_decoder_cache.to_legacy_cache() if isinstance(next_decoder_cache, Cache) else next_decoder_cache
-			)
-
-		if not return_dict:
-			return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_router_logits] if v is not None)
-
-		return MoeModelOutputWithPast(
-			last_hidden_state=hidden_states,
-			past_key_values=next_cache,
-			hidden_states=all_hidden_states,
-			attentions=all_self_attns,
-			router_logits=all_router_logits
-		)
-
-	def _update_causal_mask(self, attention_mask, input_tensor):
-		"""
-		Update the causal mask based on the attention mask and input tensor.
-
-		Args:
-			attention_mask (torch.Tensor): The attention mask.
-			input_tensor (torch.Tensor): The input tensor.
-
-		Returns:
-			torch.Tensor: The updated causal mask.
-		"""
-
-		if self.config._attn_implementation == "flash_attention_2":
-			if attention_mask is not None and 0.0 in attention_mask:
-				return attention_mask
-			return None
-
-		batch_size, seq_length = input_tensor.shape[:2]
-		dtype = input_tensor.dtype
-		device = input_tensor.device
-
-		# support going beyond cached `max_position_embedding`
-		if seq_length > self.causal_mask.shape[-1]:
-			logger.info(f"Resizing causal mask buffer from {self.causal_mask.shape[-1]} to {2 * self.causal_mask.shape[-1]}")
-			causal_mask = torch.full((2 * self.causal_mask.shape[-1], 2 * self.causal_mask.shape[-1]), fill_value=1)
-			self.register_buffer("causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False)
-
-		# We use the current dtype to avoid any overflows
-		min_dtype = torch.finfo(dtype).min
-		causal_mask = self.causal_mask[None, None, :, :].repeat(batch_size, 1, 1, 1).to(dtype) * min_dtype
-		causal_mask = causal_mask.to(dtype=dtype, device=device)
-		
-		if attention_mask is not None and attention_mask.dim() == 2:
-			mask_length = attention_mask.shape[-1]
-			padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
-			causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype)
-		
-		if self.config._attn_implementation == "sdpa" and attention_mask is not None:
-			# TODO: For dynamo, rather use a check on fullgraph=True once this is possible (https://github.com/pytorch/pytorch/pull/120400).
-			is_tracing = (
-				torch.jit.is_tracing()
-				or isinstance(input_tensor, torch.fx.Proxy)
-				or (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling())
-			)
-			
-			if not is_tracing and torch.any(attention_mask != 1):
-				# Attend to all tokens in masked rows from the causal_mask, for example the relevant first rows when
-				# using left padding. This is required by
-				# F.scaled_dot_product_attention memory-efficient attention path.
-				# Details: https://github.com/pytorch/pytorch/issues/110213
-				causal_mask = causal_mask.mul(~torch.all(causal_mask == min_dtype, dim=-1, keepdim=True)).to(dtype)
-
-		return causal_mask
-
-class GemmoeForCausalLM(GemmoePreTrainedModel):
-    r"""
-    The Gemmoe Model transformer with a language modeling head on top for causal language modeling (CLM).
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`GemmoeDecoderLayer`]
 
     Args:
-        config (GemmoeConfig): The configuration object for the Gemmoe model.
+        config: GemmoeConfig
+    """
+
+    def __init__(self, config: GemmoeConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
 
-    Example usage:
-    ```python
-    >>> from transformers import AutoTokenizer, GemmoeForCausalLM
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [GemmoeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = GemmoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.gradient_checkpointing = False
 
-    >>> model = GemmoeForCausalLM.from_pretrained("google/gemmoe-7b")
-    >>> tokenizer = AutoTokenizer.from_pretrained("google/gemmoe-7b")
+        # Register a causal mask to separate causal and padding mask creation. Merging happens in the attention class.
+        # NOTE: This is not friendly with TorchScript, ONNX, ExportedProgram serialization for very large `max_position_embeddings`.
+        causal_mask = torch.full(
+            (config.max_position_embeddings, config.max_position_embeddings), fill_value=True, dtype=torch.bool
+        )
+        self.register_buffer("causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False)
+        # Initialize weights and apply final processing
+        self.post_init()
 
-    >>> prompt = "What is your favorite condiment?"
-    >>> inputs = tokenizer(prompt, return_tensors="pt")
+    def get_input_embeddings(self):
+        return self.embed_tokens
 
-    >>> # Generate
-    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
-    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
-    "What is your favorite condiment?"
-    ```
-    """
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(GEMMOE_INPUTS_DOCSTRING)
+    # Ignore copy
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, MoeModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
+            )
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        past_seen_tokens = 0
+        if use_cache:  # kept for BC (cache positions)
+            if not isinstance(past_key_values, StaticCache):
+                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+            past_seen_tokens = past_key_values.get_seq_length()
+
+        if cache_position is None:
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds)
+
+        # embed positions
+        hidden_states = inputs_embeds
+
+        # normalized
+        hidden_states = hidden_states * (self.config.hidden_size**0.5)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = None
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = None
+        if use_cache:
+            next_cache = (
+                next_decoder_cache.to_legacy_cache() if isinstance(next_decoder_cache, Cache) else next_decoder_cache
+            )
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return MoeModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+    # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
+    # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
+    # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
+    # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
+    def _update_causal_mask(self, attention_mask, input_tensor):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        batch_size, seq_length = input_tensor.shape[:2]
+        dtype = input_tensor.dtype
+        device = input_tensor.device
+
+        # support going beyond cached `max_position_embedding`
+        if seq_length > self.causal_mask.shape[-1]:
+            causal_mask = torch.full((2 * self.causal_mask.shape[-1], 2 * self.causal_mask.shape[-1]), fill_value=1)
+            self.register_buffer("causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False)
+
+        # We use the current dtype to avoid any overflows
+        min_dtype = torch.finfo(dtype).min
+
+        causal_mask = self.causal_mask[None, None, :, :].to(dtype=dtype, device=device) * min_dtype
+        causal_mask = causal_mask.expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            if attention_mask.dim() == 2:
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
+                causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype)
+            elif attention_mask.dim() == 4:
+                mask_shape = attention_mask.shape
+                mask_slice = (attention_mask.eq(0.0)).to(dtype=dtype) * min_dtype
+                causal_mask[: mask_shape[0], : mask_shape[1], : mask_shape[2], : mask_shape[3]] = mask_slice
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type == "cuda"
+        ):
+            # TODO: For dynamo, rather use a check on fullgraph=True once this is possible (https://github.com/pytorch/pytorch/pull/120400).
+            is_tracing = (
+                torch.jit.is_tracing()
+                or isinstance(input_tensor, torch.fx.Proxy)
+                or (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling())
+            )
+            if not is_tracing and torch.any(attention_mask != 1):
+                # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+                # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+                # Details: https://github.com/pytorch/pytorch/issues/110213
+                causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+class GemmoeForCausalLM(GemmoePreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
@@ -1193,9 +1178,8 @@ class GemmoeForCausalLM(GemmoePreTrainedModel):
         self.vocab_size = config.vocab_size
         self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
         self.router_aux_loss_coef = config.router_aux_loss_coef
-        self.num_experts = 8
+        self.num_experts = config.num_local_experts
         self.num_experts_per_tok = config.num_experts_per_tok
-
         # Initialize weights and apply final processing
         self.post_init()
 
@@ -1219,6 +1203,7 @@ class GemmoeForCausalLM(GemmoePreTrainedModel):
 
     @add_start_docstrings_to_model_forward(GEMMOE_INPUTS_DOCSTRING)
     @replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    # Ignore copy
     def forward(
         self,
         input_ids: torch.LongTensor = None,
@@ -1232,7 +1217,6 @@ class GemmoeForCausalLM(GemmoePreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         output_router_logits: Optional[bool] = None,
         return_dict: Optional[bool] = None,
-        cache_position: Optional[torch.LongTensor] = None,
     ) -> Union[Tuple, MoeCausalLMOutputWithPast]:
         r"""
         Args:
@@ -1248,26 +1232,29 @@ class GemmoeForCausalLM(GemmoePreTrainedModel):
         ```python
         >>> from transformers import AutoTokenizer, GemmoeForCausalLM
 
-        >>> model = GemmoeForCausalLM.from_pretrained("google/gemmoe-7b")
-        >>> tokenizer = AutoTokenizer.from_pretrained("google/gemmoe-7b")
+        >>> model = GemmoeForCausalLM.from_pretrained("mistralai/Gemmoe-8x7B-v0.1")
+        >>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Gemmoe-8x7B-v0.1")
 
-        >>> prompt = "What is your favorite condiment?"
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
         >>> inputs = tokenizer(prompt, return_tensors="pt")
 
         >>> # Generate
         >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
         >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
-        "What is your favorite condiment?"
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
         ```"""
+
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_router_logits = (
-            output_router_logits if output_router_logits is not None else getattr(self.config, "output_router_logits", False)
+            output_router_logits if output_router_logits is not None else self.config.output_router_logits
         )
+
         output_hidden_states = (
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
         outputs = self.model(
             input_ids=input_ids,
             attention_mask=attention_mask,
@@ -1279,42 +1266,39 @@ class GemmoeForCausalLM(GemmoePreTrainedModel):
             output_hidden_states=output_hidden_states,
             output_router_logits=output_router_logits,
             return_dict=return_dict,
-            cache_position=cache_position,
         )
 
         hidden_states = outputs[0]
-
-        # Ensure hidden_states and lm_head have compatible dtypes
-        hidden_states = hidden_states.to(dtype=self.lm_head.weight.dtype)
-
         logits = self.lm_head(hidden_states)
+        logits = logits.float()
 
         loss = None
         if labels is not None:
+            # Shift so that tokens < n predict n
             shift_logits = logits[..., :-1, :].contiguous()
             shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
             loss_fct = CrossEntropyLoss()
             shift_logits = shift_logits.view(-1, self.config.vocab_size)
             shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
             shift_labels = shift_labels.to(shift_logits.device)
             loss = loss_fct(shift_logits, shift_labels)
 
         aux_loss = None
         if output_router_logits:
-            router_logits = outputs.router_logits if return_dict else outputs[-1]
-            if router_logits is not None:
-                aux_loss = load_balancing_loss_func(
-                    router_logits,
-                    self.num_experts,
-                    self.num_experts_per_tok,
-                    attention_mask,
-                )
-                if labels is not None:
-                    loss += self.router_aux_loss_coef * aux_loss.to(loss.device)
+            aux_loss = load_balancing_loss_func(
+                outputs.router_logits if return_dict else outputs[-1],
+                self.num_experts,
+                self.num_experts_per_tok,
+                attention_mask,
+            )
+            if labels is not None:
+                loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device
 
         if not return_dict:
             output = (logits,) + outputs[1:]
-            if aux_loss is not None:
+            if output_router_logits:
                 output = (aux_loss,) + output
             return (loss,) + output if loss is not None else output
 
@@ -1329,9 +1313,15 @@ class GemmoeForCausalLM(GemmoePreTrainedModel):
         )
 
     def prepare_inputs_for_generation(
-        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        output_router_logits=False,
+        **kwargs,
     ):
-        past_length = 0
+        # Omit tokens covered by past_key_values
         if past_key_values is not None:
             if isinstance(past_key_values, Cache):
                 cache_length = past_key_values.get_seq_length()
@@ -1341,11 +1331,19 @@ class GemmoeForCausalLM(GemmoePreTrainedModel):
                 cache_length = past_length = past_key_values[0][0].shape[2]
                 max_cache_length = None
 
+            # Keep only the unprocessed tokens:
+            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
+            # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
+            # input)
             if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
                 input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
+            # input_ids based on the past_length.
             elif past_length < input_ids.shape[1]:
                 input_ids = input_ids[:, past_length:]
-            
+            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
+
+            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
             if (
                 max_cache_length is not None
                 and attention_mask is not None
@@ -1355,37 +1353,27 @@ class GemmoeForCausalLM(GemmoePreTrainedModel):
 
         position_ids = kwargs.get("position_ids", None)
         if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
             position_ids = attention_mask.long().cumsum(-1) - 1
             position_ids.masked_fill_(attention_mask == 0, 1)
             if past_key_values:
                 position_ids = position_ids[:, -input_ids.shape[1] :]
 
-        if self.generation_config.cache_implementation == "static":
-            cache_position = kwargs.get("cache_position", None)
-            if cache_position is None:
-                past_length = 0
-            else:
-                past_length = cache_position[-1] + 1
-            input_ids = input_ids[:, -1].unsqueeze(-1)
-            position_ids = position_ids[:, -1].unsqueeze(-1)
-
-        cache_position = torch.arange(past_length, past_length + position_ids.shape[-1], device=position_ids.device)
-
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and past_key_values is None:
             model_inputs = {"inputs_embeds": inputs_embeds}
         else:
-            model_inputs = {"input_ids": input_ids.contiguous()}
+            model_inputs = {"input_ids": input_ids}
 
         model_inputs.update(
             {
-                "position_ids": position_ids.contiguous(),
-                "cache_position": cache_position,
+                "position_ids": position_ids,
                 "past_key_values": past_key_values,
                 "use_cache": kwargs.get("use_cache"),
                 "attention_mask": attention_mask,
+                "output_router_logits": output_router_logits,
             }
         )
-
         return model_inputs
 
     @staticmethod
@@ -1418,6 +1406,7 @@ class GemmoeForSequenceClassification(GemmoePreTrainedModel):
         self.num_labels = config.num_labels
         self.model = GemmoeModel(config)
         self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
+
         # Initialize weights and apply final processing
         self.post_init()
 
@@ -1428,7 +1417,6 @@ class GemmoeForSequenceClassification(GemmoePreTrainedModel):
         self.model.embed_tokens = value
 
     @add_start_docstrings_to_model_forward(GEMMOE_INPUTS_DOCSTRING)
-    @replace_return_docstrings(output_type=SequenceClassifierOutputWithPast, config_class=_CONFIG_FOR_DOC)
     def forward(
         self,
         input_ids: torch.LongTensor = None,
@@ -1442,25 +1430,14 @@ class GemmoeForSequenceClassification(GemmoePreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
     ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
-        """
-        Forward pass of the sequence classification model.
-
-        Args:
-            input_ids (torch.LongTensor, optional): Input token IDs.
-            attention_mask (torch.Tensor, optional): Attention mask.
-            position_ids (torch.LongTensor, optional): Position IDs.
-            past_key_values (List[torch.FloatTensor], optional): Past key-value pairs.
-            inputs_embeds (torch.FloatTensor, optional): Input embeddings.
-            labels (torch.LongTensor, optional): Labels for sequence classification.
-            use_cache (bool, optional): Whether to use cache.
-            output_attentions (bool, optional): Whether to output attentions.
-            output_hidden_states (bool, optional): Whether to output hidden states.
-            return_dict (bool, optional): Whether to return a dictionary or tuple.
-
-        Returns:
-            Union[Tuple, SequenceClassifierOutputWithPast]: Output of the sequence classification model.
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
         """
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
         transformer_outputs = self.model(
             input_ids,
             attention_mask=attention_mask,
@@ -1486,8 +1463,10 @@ class GemmoeForSequenceClassification(GemmoePreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1
-                sequence_lengths = sequence_lengths.clamp(min=0).to(logits.device)
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
 
@@ -1516,7 +1495,6 @@ class GemmoeForSequenceClassification(GemmoePreTrainedModel):
             elif self.config.problem_type == "multi_label_classification":
                 loss_fct = BCEWithLogitsLoss()
                 loss = loss_fct(pooled_logits, labels)
-
         if not return_dict:
             output = (pooled_logits,) + transformer_outputs[1:]
             return ((loss,) + output) if loss is not None else output
@@ -1527,4 +1505,4 @@ class GemmoeForSequenceClassification(GemmoePreTrainedModel):
             past_key_values=transformer_outputs.past_key_values,
             hidden_states=transformer_outputs.hidden_states,
             attentions=transformer_outputs.attentions,
-        )
+        )