Take input attention masks to support left-padded sequences

The previous implementation does not accept attention masks as inputs, so it will cause some unexpected behaviours at batched inference (commonly using left-padding). So I reimplemented the alibi encodings to take attention masks in user inputs. Note that this implementation largely depends on [1].

[1] https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py

modeling_baichuan.py

@@ -5,81 +5,127 @@ from typing import List, Optional, Tuple, Union
 
 import torch
 import torch.utils.checkpoint
+import torch.nn.functional as F
+from torch import nn
 from torch.nn import CrossEntropyLoss
 from transformers import PreTrainedModel
 from transformers.activations import ACT2FN
 from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
 from transformers.utils import logging
-from transformers.generation.utils import GenerationConfig
 
 from .configuration_baichuan import BaichuanConfig
 
+
 logger = logging.get_logger(__name__)
 
-def _get_interleave(n):
-    def _get_interleave_power_of_2(n):
-        start = (2 ** (-2 ** -(math.log2(n) - 3)))
-        ratio = start
-        return [start * ratio ** i for i in range(n)]
-
-    if math.log2(n).is_integer():
-        return _get_interleave_power_of_2(n)
-    else:
-        closest_power_of_2 = 2 ** math.floor(math.log2(n))
-        return _get_interleave_power_of_2(closest_power_of_2) + \
-               _get_interleave(2 * closest_power_of_2)[0::2][:n - closest_power_of_2]
-
-def _fill_with_neg_inf(t):
-    """FP16-compatible function that fills a tensor with -inf."""
-    return t.float().fill_(float("-inf")).type_as(t)
-
-def _gen_alibi_mask(n_head, max_pos):
-    slopes = torch.Tensor(_get_interleave(n_head))
-    alibi = slopes.unsqueeze(1).unsqueeze(1) * torch.arange(max_pos).unsqueeze(0).unsqueeze(0).expand(
-        n_head, -1, -1)
-    alibi = alibi.view(n_head, 1, max_pos)
-    alibi_mask = torch.triu(
-        _fill_with_neg_inf(torch.zeros([max_pos, max_pos])), 1
+
+# Copied from transformers.models.bloom.modeling_bloom._make_causal_mask
+def _make_causal_mask(
+    input_ids_shape: torch.Size, device: torch.device, past_key_values_length: int
+) -> torch.BoolTensor:
+    """
+    Make causal mask used for self-attention.
+    """
+    batch_size, target_length = input_ids_shape
+    mask = torch.empty((target_length, target_length + past_key_values_length), dtype=torch.bool, device=device)
+    # ONNX doesn't support `torch.Tensor.triu` properly, thus we use this workaround
+    seq_ids = torch.arange(target_length, device=device)
+    mask[:, past_key_values_length:] = seq_ids[:, None] < seq_ids[None, :]
+
+    if past_key_values_length > 0:
+        mask[:, :past_key_values_length] = False
+
+    expanded_mask = mask[None, None, :, :].expand(batch_size, 1, target_length, target_length + past_key_values_length)
+    return expanded_mask
+
+
+# Copied from transformers.models.bloom.modeling_bloom._expand_mask
+def _expand_mask(mask: torch.Tensor, tgt_length: int) -> torch.BoolTensor:
+    """
+    Expands attention_mask from `[batch_size, src_length]` to `[batch_size, 1, tgt_length, src_length]`.
+    """
+    batch_size, src_length = mask.shape
+    tgt_length = tgt_length if tgt_length is not None else src_length
+
+    expanded_mask = ~(mask[:, None, None, :].to(torch.bool))
+    return expanded_mask.expand(batch_size, 1, tgt_length, src_length)
+
+
+# Copied from transformers.models.bloom.modeling_bloom.build_alibi_tensor
+def build_alibi_tensor(attention_mask: torch.Tensor, num_heads: int, dtype: torch.dtype) -> torch.Tensor:
+    """
+    Link to paper: https://arxiv.org/abs/2108.12409 Alibi tensor is not causal as the original paper mentions, it
+    relies on a translation invariance of softmax for quick implementation: with l being a tensor, and a fixed value
+    `softmax(l+a) = softmax(l)`.
+
+    Args:
+    Returns tensor shaped (batch_size * num_heads, 1, max_seq_len)
+        attention_mask (`torch.Tensor`):
+            Token-wise attention mask, this should be of shape (batch_size, max_seq_len).
+        num_heads (`int`, *required*):
+            number of heads
+        dtype (`torch.dtype`, *optional*, default=`torch.bfloat16`):
+            dtype of the output tensor
+    """
+    batch_size, seq_length = attention_mask.shape
+    closest_power_of_2 = 2 ** math.floor(math.log2(num_heads))
+    base = torch.tensor(
+        2 ** (-(2 ** -(math.log2(closest_power_of_2) - 3))), device=attention_mask.device, dtype=torch.float32
     )
-    alibi_mask = alibi_mask.unsqueeze(0) + alibi
-    return alibi_mask
+    powers = torch.arange(1, 1 + closest_power_of_2, device=attention_mask.device, dtype=torch.int32)
+    slopes = torch.pow(base, powers)
+
+    if closest_power_of_2 != num_heads:
+        extra_base = torch.tensor(
+            2 ** (-(2 ** -(math.log2(2 * closest_power_of_2) - 3))), device=attention_mask.device, dtype=torch.float32
+        )
+        num_remaining_heads = min(closest_power_of_2, num_heads - closest_power_of_2)
+        extra_powers = torch.arange(1, 1 + 2 * num_remaining_heads, 2, device=attention_mask.device, dtype=torch.int32)
+        slopes = torch.cat([slopes, torch.pow(extra_base, extra_powers)], dim=0)
+
+    # Note: alibi will added to the attention bias that will be applied to the query, key product of attention
+    # => therefore alibi will have to be of shape (batch_size, num_heads, query_length, key_length)
+    # => here we set (batch_size=1, num_heads=num_heads, query_length=1, key_length=max_length)
+    # => the query_length dimension will then be broadcasted correctly
+    arange_tensor = ((attention_mask.cumsum(dim=-1) - 1) * attention_mask)[:, None, :]
+    alibi = slopes[..., None] * arange_tensor
+    return alibi.reshape(batch_size * num_heads, 1, seq_length).to(dtype)
+
 
+class RMSNorm(nn.Module):
 
-class RMSNorm(torch.nn.Module):
     def __init__(self, hidden_size, epsilon=1e-6):
         super().__init__()
-        self.weight = torch.nn.Parameter(torch.empty(hidden_size))
+        self.weight = nn.Parameter(torch.ones(hidden_size))
         self.epsilon = epsilon
 
-    def forward(self, hidden_states):
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        input_dtype = hidden_states.dtype
         variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
         hidden_states = hidden_states * torch.rsqrt(variance + self.epsilon)
 
-        # convert into half-precision
-        if self.weight.dtype in [torch.float16, torch.bfloat16]:
-            hidden_states = hidden_states.to(self.weight.dtype)
+        return (self.weight * hidden_states).to(input_dtype)
 
-        return self.weight * hidden_states
 
+class MLP(nn.Module):
 
-class MLP(torch.nn.Module):
     def __init__(
-            self,
-            hidden_size: int,
-            intermediate_size: int,
-            hidden_act: str,
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
     ):
         super().__init__()
-        self.gate_proj = torch.nn.Linear(hidden_size, intermediate_size, bias=False)
-        self.down_proj = torch.nn.Linear(intermediate_size, hidden_size, bias=False)
-        self.up_proj = torch.nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
+        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
         self.act_fn = ACT2FN[hidden_act]
 
     def forward(self, x):
         return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
 
 
-class BaichuanAttention(torch.nn.Module):
+class BaichuanAttention(nn.Module):
 
     def __init__(self, config: BaichuanConfig):
         super().__init__()
@@ -93,62 +139,89 @@ class BaichuanAttention(torch.nn.Module):
             raise ValueError(
                 f"hidden_size {self.hidden_size} is not divisible by num_heads {self.num_heads}"
             )
-        self.W_pack = torch.nn.Linear(self.hidden_size, 3 * self.hidden_size, bias=False)
-        self.o_proj = torch.nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+
+        # Layer-wise attention scaling
+        self.inv_norm_factor = 1.0 / math.sqrt(self.head_dim)
+        self.beta = 1.0
+
+        self.W_pack = nn.Linear(self.hidden_size, 3 * self.hidden_size, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
 
     def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
         return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
 
     def forward(
-            self,
-            hidden_states: torch.Tensor,
-            attention_mask: Optional[torch.Tensor] = None,
-            past_key_value: Optional[Tuple[torch.Tensor]] = None,
-            output_attentions: bool = False,
-            use_cache: bool = False,
+        self,
+        hidden_states: torch.Tensor,
+        alibi: torch.Tensor,
+        attention_mask: torch.Tensor,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
 
         bsz, q_len, _ = hidden_states.size()
 
-        proj = self.W_pack(hidden_states)
+        proj = self.W_pack(hidden_states) # [batch_size, seq_length, 3 x hidden_size]
         proj = proj.unflatten(-1, (3, self.hidden_size)).unsqueeze(0).transpose(0, -2).squeeze(-2)
-        query_states = proj[0].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
-        key_states = proj[1].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
-        value_states = proj[2].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        query_states = proj[0].view(bsz, q_len, self.num_heads, self.head_dim)
+        key_states = proj[1].view(bsz, q_len, self.num_heads, self.head_dim)
+        value_states = proj[2].view(bsz, q_len, self.num_heads, self.head_dim)
 
-        kv_seq_len = key_states.shape[-2]
-        if past_key_value is not None:
-            kv_seq_len += past_key_value[0].shape[-2]
+        query_states = query_states.transpose(1, 2).reshape(bsz * self.num_heads, q_len, self.head_dim)
+        key_states = key_states.permute(0, 2, 3, 1).reshape(bsz * self.num_heads, self.head_dim, q_len)
+        value_states = value_states.transpose(1, 2).reshape(bsz * self.num_heads, q_len, self.head_dim)
 
         if past_key_value is not None:
             # reuse k, v, self_attention
-            key_states = torch.cat([past_key_value[0], key_states], dim=2)
-            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+            past_key, past_value = past_key_value
+            key_states = torch.cat([past_key, key_states], dim=2)
+            value_states = torch.cat([past_value, value_states], dim=1)
+
+        _, _, kv_seq_len = key_states.shape
 
         past_key_value = (key_states, value_states) if use_cache else None
 
-        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+        # [batch_size * num_heads, q_length, kv_length]
+        # we use `torch.Tensor.baddbmm` instead of `torch.baddbmm` as the latter isn't supported by TorchScript v1.11
+        matmul_result = alibi.baddbmm(
+            batch1=query_states,
+            batch2=key_states,
+            beta=self.beta,
+            alpha=self.inv_norm_factor,
+        )
 
-        if attention_mask is not None:
-            if attn_weights.size(-2) == 1:
-                attention_mask = attention_mask[:, -1:, :]
-            attn_weights = attn_weights + attention_mask.unsqueeze(0)
-            attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
+        # change view to [batch_size, num_heads, q_length, kv_length]
+        attention_scores = matmul_result.view(bsz, self.num_heads, q_len, kv_seq_len)
 
-        attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1)
-        attn_output = torch.matmul(attn_weights, value_states)
+        # cast attention scores to fp32, compute scaled softmax and cast back to initial dtype
+        # [batch_size, num_heads, q_length, kv_length]
+        input_dtype = attention_scores.dtype
+        # `float16` has a minimum value of -65504.0, whereas `bfloat16` and `float32` have a minimum value of `-3.4e+38`
+        if input_dtype == torch.float16:
+            attention_scores = attention_scores.to(torch.float)
+        attn_weights = torch.masked_fill(attention_scores, attention_mask, torch.finfo(attention_scores.dtype).min)
+        attention_probs = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(input_dtype)
 
-        attn_output = attn_output.transpose(1, 2)
-        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+        # change view [batch_size x num_heads, q_length, kv_length]
+        attention_probs_reshaped = attention_probs.view(bsz * self.num_heads, q_len, kv_seq_len)
+
+        # matmul: [batch_size * num_heads, q_length, head_dim]
+        attn_output = torch.bmm(attention_probs_reshaped, value_states)
+
+        attn_output = attn_output.view(bsz, self.num_heads, q_len, self.head_dim)
+
+        attn_output = attn_output.transpose(1, 2).reshape(bsz, q_len, self.hidden_size)
         attn_output = self.o_proj(attn_output)
 
         if not output_attentions:
-            attn_weights = None
+            attention_probs = None
 
-        return attn_output, attn_weights, past_key_value
+        return attn_output, attention_probs, past_key_value
 
 
-class BaichuanLayer(torch.nn.Module):
+class BaichuanLayer(nn.Module):
+
     def __init__(self, config: BaichuanConfig):
         super().__init__()
         self.hidden_size = config.hidden_size
@@ -162,12 +235,13 @@ class BaichuanLayer(torch.nn.Module):
         self.post_attention_layernorm = RMSNorm(config.hidden_size, epsilon=config.rms_norm_eps)
 
     def forward(
-            self,
-            hidden_states: torch.Tensor,
-            attention_mask: Optional[torch.Tensor] = None,
-            past_key_value: Optional[Tuple[torch.Tensor]] = None,
-            output_attentions: Optional[bool] = False,
-            use_cache: Optional[bool] = False,
+        self,
+        hidden_states: torch.Tensor,
+        alibi: torch.Tensor,
+        attention_mask: torch.Tensor,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
     ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
 
         residual = hidden_states
@@ -177,6 +251,7 @@ class BaichuanLayer(torch.nn.Module):
         # Self Attention
         hidden_states, self_attn_weights, present_key_value = self.self_attn(
             hidden_states=hidden_states,
+            alibi=alibi,
             attention_mask=attention_mask,
             past_key_value=past_key_value,
             output_attentions=output_attentions,
@@ -192,6 +267,9 @@ class BaichuanLayer(torch.nn.Module):
 
         outputs = (hidden_states,)
 
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
         if use_cache:
             outputs += (present_key_value,)
 
@@ -203,15 +281,16 @@ class BaichuanPreTrainedModel(PreTrainedModel):
     base_model_prefix = "model"
     supports_gradient_checkpointing = True
     _no_split_modules = ["BaichuanLayer"]
+    _skip_keys_device_placement = "past_key_values"
     _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
 
     def _init_weights(self, module):
         std = self.config.initializer_range
-        if isinstance(module, torch.nn.Linear):
+        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, torch.nn.Embedding):
+        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_()
@@ -221,49 +300,69 @@ class BaichuanPreTrainedModel(PreTrainedModel):
             module.gradient_checkpointing = value
 
 
-
 class BaichuanModel(BaichuanPreTrainedModel):
+
     def __init__(self, config: BaichuanConfig):
         super().__init__(config)
         self.padding_idx = config.pad_token_id
         self.vocab_size = config.vocab_size
         self.n_head = config.num_attention_heads
-        self.embed_tokens = torch.nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
-        self.layers = torch.nn.ModuleList([BaichuanLayer(config) for _ in range(config.num_hidden_layers)])
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([BaichuanLayer(config) for _ in range(config.num_hidden_layers)])
         self.norm = RMSNorm(config.hidden_size, epsilon=config.rms_norm_eps)
 
         self.gradient_checkpointing = config.gradient_checkpointing
         self.post_init()
-        self.max_cache_pos = config.model_max_length
-        self.first_run = True    
 
     def get_input_embeddings(self):
         return self.embed_tokens
         
     def set_input_embeddings(self, value):
         self.embed_tokens = value  
-        
-    def get_alibi_mask(self, tensor, seq_length_with_past):
-        if self.first_run:
-            self.first_run = False
-            self.register_buffer("future_mask", _gen_alibi_mask(self.n_head, self.max_cache_pos).to(tensor), persistent=False)
-        if seq_length_with_past > self.max_cache_pos:
-            self.max_cache_pos = seq_length_with_past
-            self.register_buffer("future_mask", _gen_alibi_mask(self.n_head, self.max_cache_pos).to(tensor), persistent=False)
-        mask = self.future_mask[:self.n_head, :seq_length_with_past, :seq_length_with_past] 
-        return mask
+
+    def build_alibi_tensor(self, attention_mask: torch.Tensor, num_heads: int, dtype: torch.dtype) -> torch.Tensor:
+        return build_alibi_tensor(attention_mask, num_heads, dtype)
+
+    def _prepare_attn_mask(
+        self, attention_mask: torch.Tensor, input_shape: Tuple[int, int], past_key_values_length: int
+    ) -> torch.BoolTensor:
+        # create causal mask
+        # [batch_size, seq_length] -> [batch_size, 1, tgt_length, src_length]
+        combined_attention_mask = None
+        device = attention_mask.device
+        _, src_length = input_shape
+
+        if src_length > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape, device=device, past_key_values_length=past_key_values_length
+            )
+
+        # [batch_size, seq_length] -> [batch_size, 1, tgt_length, src_length]
+        expanded_attn_mask = _expand_mask(attention_mask, tgt_length=src_length)
+        combined_attention_mask = (
+            expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask | combined_attention_mask
+        )
+
+        return combined_attention_mask
 
     def forward(
-            self,
-            input_ids: torch.LongTensor = None,
-            past_key_values: Optional[List[torch.FloatTensor]] = None,
-            inputs_embeds: Optional[torch.FloatTensor] = None,
-            use_cache: Optional[bool] = False,
-            output_attentions: Optional[bool] = False,
-            output_hidden_states: Optional[bool] = False,
-            return_dict: Optional[bool] = True,
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = 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,
     ) -> Union[Tuple, BaseModelOutputWithPast]:
-
+        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 not None and inputs_embeds is not None:
             raise ValueError("You cannot provide both input_ids and inputs_embeds simultaneously")
@@ -275,19 +374,21 @@ class BaichuanModel(BaichuanPreTrainedModel):
             raise ValueError("You need to provide input_ids or inputs_embeds")
 
         seq_length_with_past = seq_length
-
+        past_key_values_length = 0
         if past_key_values is not None:
-            past_key_values_length = past_key_values[0][0].shape[2]
+            past_key_values_length = past_key_values[0][0].shape[1]
             seq_length_with_past = seq_length_with_past + past_key_values_length
 
         if inputs_embeds is None:
             inputs_embeds = self.embed_tokens(input_ids)
 
-        # embed positions
-        attention_mask = self.get_alibi_mask(inputs_embeds, seq_length_with_past)
-
         hidden_states = inputs_embeds
 
+        if attention_mask is None:
+            attention_mask = torch.ones((batch_size, seq_length_with_past), device=hidden_states.device)
+        else:
+            attention_mask = attention_mask.to(hidden_states.device)
+
         if self.gradient_checkpointing and self.training:
             if use_cache:
                 logger.warning_once(
@@ -295,6 +396,15 @@ class BaichuanModel(BaichuanPreTrainedModel):
                 )
                 use_cache = False
 
+        # Compute alibi tensor: check build_alibi_tensor documentation
+        alibi = self.build_alibi_tensor(attention_mask, self.n_head, dtype=hidden_states.dtype)
+
+        causal_mask = self._prepare_attn_mask(
+            attention_mask,
+            input_shape=(batch_size, seq_length),
+            past_key_values_length=past_key_values_length,
+        )
+
         # decoder layers
         all_hidden_states = () if output_hidden_states else None
         all_self_attns = () if output_attentions else None
@@ -318,13 +428,15 @@ class BaichuanModel(BaichuanPreTrainedModel):
                 layer_outputs = torch.utils.checkpoint.checkpoint(
                     create_custom_forward(decoder_layer),
                     hidden_states,
-                    attention_mask,
+                    alibi,
+                    causal_mask,
                     None,
                 )
             else:
                 layer_outputs = decoder_layer(
                     hidden_states,
-                    attention_mask=attention_mask,
+                    alibi=alibi,
+                    attention_mask=causal_mask,
                     past_key_value=past_key_value,
                     output_attentions=output_attentions,
                     use_cache=use_cache,
@@ -345,8 +457,10 @@ class BaichuanModel(BaichuanPreTrainedModel):
             all_hidden_states += (hidden_states,)
 
         next_cache = next_decoder_cache if use_cache else None
+
         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 BaseModelOutputWithPast(
             last_hidden_state=hidden_states,
             past_key_values=next_cache,
@@ -356,10 +470,12 @@ class BaichuanModel(BaichuanPreTrainedModel):
     
 
 class BaichuanForCausalLM(BaichuanPreTrainedModel):
+
     def __init__(self, config):
         super().__init__(config)
         self.model = BaichuanModel(config)
-        self.lm_head = torch.nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
 
         # Initialize weights and apply final processing
         self.post_init()
@@ -383,22 +499,28 @@ class BaichuanForCausalLM(BaichuanPreTrainedModel):
         return self.model
         
     def forward(
-            self,
-            input_ids: torch.LongTensor = None,
-            past_key_values: Optional[List[torch.FloatTensor]] = None,
-            inputs_embeds: Optional[torch.FloatTensor] = None,
-            labels: Optional[torch.LongTensor] = None,
-            use_cache: Optional[bool] = None,
-            output_attentions: Optional[bool] = False,
-            output_hidden_states: Optional[bool] = False,
-            return_dict: Optional[bool] = True,
-            **kwargs
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs
     ) -> Union[Tuple, CausalLMOutputWithPast]:
-
+        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
+        )
+        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,
             past_key_values=past_key_values,
             inputs_embeds=inputs_embeds,
             use_cache=use_cache,
@@ -436,7 +558,7 @@ class BaichuanForCausalLM(BaichuanPreTrainedModel):
         )
 
     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, **kwargs
     ):  
         if past_key_values:
             input_ids = input_ids[:, -1:]
@@ -451,6 +573,7 @@ class BaichuanForCausalLM(BaichuanPreTrainedModel):
             {   
                 "past_key_values": past_key_values,
                 "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
             }   
         )   
         return model_inputs