[shardformer]fix flash attention, when mask is casual, just don't unpad it

colossalai/shardformer/modeling/chatglm2.py

@@ -51,7 +51,8 @@ def forward(self: CoreAttention, query_layer, key_layer, value_layer, attention_
                 attn_mask_type = AttnMaskType.causal
             else:
                 flash_attention_mask = ~(attention_mask[:, :, -1].squeeze(1).to(torch.bool)).contiguous()
-                attn_mask_type = AttnMaskType.paddedcausal
+                if not torch.all(flash_attention_mask):
+                    attn_mask_type = AttnMaskType.paddedcausal
 
             attention = ColoAttention(
                 embed_dim=self.hidden_size_per_partition,

colossalai/shardformer/modeling/gpt2.py

@@ -771,11 +771,12 @@ def forward(
             attn_mask_type = AttnMaskType.causal
             flash_attention_mask = None
         if attention_mask != None:
-            if attn_mask_type == AttnMaskType.causal:
-                attn_mask_type == AttnMaskType.paddedcausal
-            else:
-                attn_mask_type = AttnMaskType.padding
             flash_attention_mask = ~(attention_mask[:, :, -1].squeeze(1).to(torch.bool)).contiguous()
+            if not torch.all(flash_attention_mask):
+                if attn_mask_type == AttnMaskType.causal:
+                    attn_mask_type == AttnMaskType.paddedcausal
+                else:
+                    attn_mask_type = AttnMaskType.padding
 
         scale = value.size(-1) ** -0.5
         if self.scale_attn_by_inverse_layer_idx:

colossalai/shardformer/modeling/llama.py

@@ -465,7 +465,8 @@ def forward(
                     f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
                 )
             flash_attention_mask = ~(attention_mask[:, :, -1].squeeze(1).to(torch.bool)).contiguous()
-            attn_mask_type = AttnMaskType.paddedcausal
+            if not torch.all(flash_attention_mask):
+                attn_mask_type = AttnMaskType.paddedcausal
 
         attention = ColoAttention(embed_dim=self.hidden_size, num_heads=self.num_heads)
         attn_output = attention(

colossalai/shardformer/modeling/opt.py

@@ -581,7 +581,8 @@ def forward(
                     f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
                 )
             flash_attention_mask = ~(attention_mask[:, :, -1].squeeze(1).to(torch.bool)).contiguous()
-            attn_mask_type = AttnMaskType.paddedcausal
+            if not torch.all(flash_attention_mask):
+                attn_mask_type = AttnMaskType.paddedcausal
 
         attention = ColoAttention(
             embed_dim=self.embed_dim, num_heads=self.num_heads, dropout=self.dropout, scale=self.scaling

colossalai/shardformer/modeling/whisper.py

@@ -106,7 +106,10 @@ def forward(
                         f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
                     )
                 flash_attention_mask = ~(attention_mask[:, :, -1].squeeze(1).to(torch.bool).contiguous())
-                attn_type = AttnMaskType.paddedcausal
+                if not torch.all(flash_attention_mask):
+                    attn_type = AttnMaskType.paddedcausal
+                else:
+                    attn_type = AttnMaskType.causal
 
         attention = ColoAttention(
             embed_dim=self.embed_dim, num_heads=self.num_heads, dropout=self.dropout, scale=self.scaling

examples/language/llama2/pretrain.py

@@ -76,6 +76,7 @@ def tokenize_batch_for_pretrain(batch, tokenizer: Optional[LlamaTokenizer] = Non
 
 def all_reduce_mean(tensor: torch.Tensor) -> torch.Tensor:
     dist.all_reduce(tensor, op=dist.ReduceOp.SUM)
+    tensor = tensor.data
     tensor.div_(dist.get_world_size())
     return tensor