fix mixtral moe

llm/data.py

@@ -44,11 +44,11 @@ def get_convert_example(model):
 
     if base_model_prefix == "chatglm":
         return convert_example_chatglm
-    elif base_model_prefix in ["chatglm_v2", "llama", "bloom", "opt", "qwen"]:
+    elif base_model_prefix in ["chatglm_v2", "llama", "bloom", "opt", "qwen", "mixtral"]:
         return convert_example_common
     else:
         raise ValueError(
-            f"Unknown base_model_prefix: {model.base_model_prefix}. Supported base_model_prefix list: chatglm, bloom, llama."
+            f"Unknown base_model_prefix: {model.base_model_prefix}. Supported base_model_prefix list: chatglm, bloom, llama, qwen, mixtral"
         )
 
 
@@ -107,7 +107,7 @@ def tokenize_rounds_example(tokenizer, example, data_args):
     # 0. prepare data
     context_data = example.get("context", {})
     context_data["is_training"] = True
-    
+
     example["src"] = example["src"] if isinstance(example["src"], list) else [example["src"]]
     example["tgt"] = example["tgt"] if isinstance(example["tgt"], list) else [example["tgt"]]
 

paddlenlp/transformers/mixtral/configuration.py

@@ -12,25 +12,28 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+""" Mixtral model configuration"""
 
-from paddlenlp.transformers import PretrainedConfig
+from paddlenlp.transformers.configuration_utils import PretrainedConfig
 
-__all__ = ["MixtralConfig"]
+__all__ = [
+    "MixtralConfig",
+]
 
 
 class MixtralConfig(PretrainedConfig):
     r"""
-    This is the configuration class to store the configuration of a [`MixtralModel`]. It is used to instantiate an
-    Mixtral model according to the specified arguments, defining the model architecture. Instantiating a configuration
-    with the defaults will yield a similar configuration to that of the Mixtral-7B-v0.1 or Mixtral-7B-Instruct-v0.1.
+    This is the configuration class to store the configuration of a [`~MixtralModel`]. It is used to instantiate an Mixtral
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the Mixtral-7B-v0.1 or Mixtral-7B-Instruct-v0.1.
 
     Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
     documentation from [`PretrainedConfig`] for more information.
 
     Args:
         vocab_size (`int`, *optional*, defaults to 32000):
             Vocabulary size of the Mixtral model. Defines the number of different tokens that can be represented by the
-            `inputs_ids` passed when calling [`MixtralModel`]
+            `inputs_ids` passed when calling [`~MixtralModel`]
         hidden_size (`int`, *optional*, defaults to 4096):
             Dimension of the hidden representations.
         intermediate_size (`int`, *optional*, defaults to 14336):
@@ -39,13 +42,6 @@ class MixtralConfig(PretrainedConfig):
             Number of hidden layers in the Transformer encoder.
         num_attention_heads (`int`, *optional*, defaults to 32):
             Number of attention heads for each attention layer in the Transformer encoder.
-        num_key_value_heads (`int`, *optional*, defaults to 8):
-            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
-            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
-            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
-            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
-            by meanpooling all the original heads within that group. For more details checkout [this
-            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `8`.
         hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
             The non-linear activation function (function or string) in the decoder.
         max_position_embeddings (`int`, *optional*, defaults to `4096*32`):
@@ -64,25 +60,47 @@ class MixtralConfig(PretrainedConfig):
             The id of the "beginning-of-sequence" token.
         eos_token_id (`int`, *optional*, defaults to 2):
             The id of the "end-of-sequence" token.
-        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
-            Whether the model's input and output word embeddings should be tied.
+        tie_word_embeddings(`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings.
         rope_theta (`float`, *optional*, defaults to 1000000.0):
             The base period of the RoPE embeddings.
         sliding_window (`int`, *optional*):
-            Sliding window attention window size. If not specified, will default to `4096`.
+            Sliding window attention window size.
         attention_dropout (`float`, *optional*, defaults to 0.0):
             The dropout ratio for the attention probabilities.
         num_experts_per_tok (`int`, *optional*, defaults to 2):
             The number of experts to root per-token, can be also interpreted as the `top-p` routing
-            parameter.
+            parameter
         num_local_experts (`int`, *optional*, defaults to 8):
             Number of experts per Sparse MLP layer.
         output_router_logits (`bool`, *optional*, defaults to `False`):
             Whether or not the router logits should be returned by the model. Enabeling this will also
             allow the model to output the auxiliary loss. See [here]() for more details
         router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
             The aux loss factor for the total loss.
-    """
+        use_fused_rope(`bool`, *optional*, defaults to False):
+            Enable rope fusion or not.
+        num_key_value_heads (`int`, *optional*, defaults to 8):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
+            `num_attention_heads`.
+        Example:
+    ```python
+    >>> from paddlenlp.transformer import MixtralModel, MixtralConfig
+
+    >>> # Initializing a Mixtral mixtral-7b style configuration
+    >>> configuration = MixtralConfig()
+
+    >>> # Initializing a model from the mixtral-7b style configuration
+    >>> model = MixtralModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
 
     model_type = "mixtral"
     keys_to_ignore_at_inference = ["past_key_values"]
@@ -92,64 +110,82 @@ def __init__(
         vocab_size=32000,
         hidden_size=4096,
         intermediate_size=14336,
+        max_position_embeddings=4096 * 32,
+        seq_length=2048,
         num_hidden_layers=32,
         num_attention_heads=32,
         num_key_value_heads=8,
         hidden_act="silu",
-        max_position_embeddings=4096 * 32,
         initializer_range=0.02,
         rms_norm_eps=1e-5,
         use_cache=True,
+        use_recompute=False,
+        recompute_granularity="full",
+        no_recompute_layers=None,
+        use_flash_attention=False,
+        attention_dropout=0.0,
+        use_fused_rope=False,
+        rope_theta=1e6,
+        tensor_parallel_output=True,
+        sequence_parallel=False,
+        fuse_sequence_parallel_allreduce=False,
         pad_token_id=None,
         bos_token_id=1,
         eos_token_id=2,
         tie_word_embeddings=False,
-        rope_theta=1e6,
-        sliding_window=None,
-        attention_dropout=0.0,
         num_experts_per_tok=2,
         num_local_experts=8,
-        output_router_logits=False,
         router_aux_loss_coef=0.001,
-        use_flash_attention=False,
-        use_fused_rope=False,
-        use_recompute=False,
-        recompute_granularity="full",
+        output_router_logits=False,
+        sliding_window=None,
         **kwargs,
     ):
         self.vocab_size = vocab_size
         self.hidden_size = hidden_size
         self.intermediate_size = intermediate_size
+        self.max_position_embeddings = max_position_embeddings
+        self.seq_length = seq_length
         self.num_hidden_layers = num_hidden_layers
         self.num_attention_heads = num_attention_heads
+        self.attention_dropout = attention_dropout
 
-        # for backward compatibility
         if num_key_value_heads is None:
             num_key_value_heads = num_attention_heads
-
         self.num_key_value_heads = num_key_value_heads
         self.hidden_act = hidden_act
-        self.max_position_embeddings = max_position_embeddings
+
         self.initializer_range = initializer_range
         self.rms_norm_eps = rms_norm_eps
+
         self.use_cache = use_cache
+        self.use_recompute = use_recompute
+        self.recompute_granularity = recompute_granularity
+        self.no_recompute_layers = no_recompute_layers
+        self.use_flash_attention = use_flash_attention
+        self.tensor_parallel_output = tensor_parallel_output
+        self.sequence_parallel = sequence_parallel
+        self.fuse_sequence_parallel_allreduce = fuse_sequence_parallel_allreduce
+
+        self.pad_token_id = pad_token_id
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+
+        self.use_fused_rope = use_fused_rope
         self.rope_theta = rope_theta
-        self.sliding_window = sliding_window
-        self.attention_dropout = attention_dropout
+
+        # ----------------- Experts -------------------- #
         self.num_experts_per_tok = num_experts_per_tok
         self.num_local_experts = num_local_experts
-        self.output_router_logits = output_router_logits
         self.router_aux_loss_coef = router_aux_loss_coef
+        self.output_router_logits = output_router_logits
 
-        self.use_flash_attention = use_flash_attention
-        self.use_fused_rope = use_fused_rope
-        self.use_recompute = use_recompute
-        self.recompute_granularity = recompute_granularity
+        self.sliding_window = sliding_window
 
         super().__init__(
             pad_token_id=pad_token_id,
             bos_token_id=bos_token_id,
             eos_token_id=eos_token_id,
             tie_word_embeddings=tie_word_embeddings,
+            tensor_parallel_output=tensor_parallel_output,
             **kwargs,
         )