Fix attention mask handling in the Hybrid Engine Bloom flow

deepspeed/module_inject/containers/bloom.py

@@ -23,12 +23,14 @@ def __init__(self, **kwargs):
 
         # All model specific things should be defined here instead of the base class.
         self.bigscience_bloom = True
+        self.triangular_masking = False
 
     def create_module(self, config=None):
         _config = config if config is not None else self.ds_model_config
 
         self.module = DeepSpeedBloomInference(_config, mp_group=self.mp_group)
         self.module.config.scale_attention = self.scale_attention
+        self.module.config.invert_mask = False
         return self.module
 
     def attention_qkv_mp(self, mp_replace, reversed_dim=False):

deepspeed/ops/transformer/inference/config.py

@@ -43,6 +43,7 @@ class DeepSpeedInferenceConfig(TransformerConfig):
             return_tuple: if True, returns the transformer output as a tuple, otherwise returns as a tensor
             bigscience_bloom: This flag is added temporarily for supporting the BLOOM-176B model architecture.
             use_triton: This flag is to enable triton kernels in inference or not.
+            invert_mask: If True, the attention mask is inverted when passed to attention block.
     """
 
     def __init__(self,
@@ -80,7 +81,8 @@ def __init__(self,
                  use_triton=False,
                  triton_autotune=False,
                  num_kv=-1,
-                 rope_theta=10000):
+                 rope_theta=10000,
+                 invert_mask=True):
         super(DeepSpeedInferenceConfig,
               self).__init__(hidden_size, (intermediate_size if intermediate_size > 0 else 4 * hidden_size), heads,
                              num_hidden_layers)
@@ -116,6 +118,7 @@ def __init__(self,
         self.triton_autotune = triton_autotune
         self.num_kv = num_kv
         self.rope_theta = rope_theta
+        self.invert_mask = invert_mask
 
     @classmethod
     def from_dict(cls, json_object):

deepspeed/ops/transformer/inference/ds_attention.py

@@ -254,8 +254,12 @@ def compute_attention(self, qkv_out, input_mask, layer_past, alibi):
         if input_mask.dtype == torch.bool:
             input_mask = input_mask.long()
 
+        # Invert input_mask per transformer implementation (eg, in BLOOM, it's already inverted)
+        if self.config.invert_mask:
+            input_mask = 1 - input_mask
+
         attention_probs = self.softmax_func(attn_scores=attention_scores,
-                                            attn_mask=((1 - input_mask).to(target_dtype) * minus_inf),
+                                            attn_mask=input_mask.to(target_dtype) * minus_inf,
                                             alibi=alibi,
                                             triangular=(self.config.triangular_masking
                                                         and (attention_scores.shape[-2] > 1)),