fix conflict

Author: sevenan2

paddleformers/cli/hparams/model_args.py

@@ -85,6 +85,14 @@ class ModelArguments:
         default=False,
         metadata={"help": "GPT3 model, use fast layernorm"},
     )
+    fuse_attention_qkv: bool = field(
+        default=None,
+        metadata={"help": "whether to fuse attention qkv"},
+    )
+    fuse_attention_ffn: bool = field(
+        default=None,
+        metadata={"help": "whether to fuse first up and gate proj in mlp block"},
+    )
     attn_impl: str = field(default="flashmask", metadata={"help": "Attention implementation"})
     fuse_gate_detach_matmul: bool = field(
         default=True,

paddleformers/cli/hparams/parser.py

@@ -142,7 +142,7 @@ def _parse_train_args(args: Optional[Union[dict[str, Any], list[str]]] = None) -
     Returns:
         _TRAIN_CLS: _description_
     """
-    parser = PdArgumentParser(_TRAIN_ARGS)
+    parser = PdArgumentParser(_TRAIN_ARGS, conflict_handler="resolve")
     allow_extra_keys = is_env_enabled("ALLOW_EXTRA_ARGS")
     return _parse_args(parser, args, allow_extra_keys=allow_extra_keys)
 

paddleformers/cli/train/auto_parallel/workflow.py

@@ -33,8 +33,6 @@
     AutoTokenizer,
     CosineAnnealingWithWarmupDecay,
     LinearAnnealingWithWarmupDecay,
-    LlamaConfig,
-    LlamaForCausalLM,
 )
 from paddleformers.transformers.configuration_utils import LlmMetaConfig
 from paddleformers.utils.log import logger
@@ -147,6 +145,7 @@ def __init__(self, *args, **kwargs):
 
 
 def run_auto_parallel(model_args, data_args, generating_args, training_args):
+
     do_enable_linear_fused_grad_add = training_args.enable_linear_fused_grad_add
     # do_enable_mp_async_allreduce = (
     #     training_args.enable_auto_parallel
@@ -311,9 +310,6 @@ def run_auto_parallel(model_args, data_args, generating_args, training_args):
                 model = model_class.from_config(config, dtype=dtype)
         else:
             model = model_class.from_config(config, dtype=dtype)
-    
-    criterion = model.criterion
-
 
     if training_args.recompute:
 

paddleformers/trainer/argparser.py

@@ -188,7 +188,6 @@ def _add_dataclass_arguments(self, dtype: DataClassType):
                 f"removing line of `from __future__ import annotations` which opts in Postponed "
                 f"Evaluation of Annotations (PEP 563)"
             )
-
         for field in dataclasses.fields(dtype):
             if not field.init:
                 continue

paddleformers/transformers/__init__.py

@@ -193,7 +193,7 @@
     "llama.configuration": [
         "LlamaConfig",
     ],
-    "llama.modeling": ["LlamaForCausalLM", "LlamaModel", "LlamaForCausalLMPipe", "LlamaRotaryEmbedding"],
+    "llama.modeling": ["LlamaForCausalLM", "LlamaModel", "LlamaForCausalLMPipe"],
     "llama.tokenizer": ["LlamaTokenizer", "Llama3Tokenizer"],
     "llama.tokenizer_fast": ["LlamaTokenizerFast"],
     "optimization": [

paddleformers/transformers/llama/modeling.py

@@ -20,6 +20,11 @@
 from paddle.distributed.fleet.utils import recompute
 from paddle.distributed.fleet.utils.sequence_parallel_utils import ScatterOp
 
+from paddleformers.transformers.conversion_utils import (
+    StateDictNameMapping,
+    init_name_mappings,
+)
+
 from ...nn.attention.interface import ALL_ATTENTION_FUNCTIONS
 from ...nn.criterion.interface import CriterionLayer
 from ...nn.embedding import Embedding as GeneralEmbedding
@@ -28,14 +33,13 @@
 from ...nn.mlp import MLP
 from ...nn.norm import Norm as GeneralNorm
 from ...nn.pp_model import GeneralModelForCausalLMPipe
-from .auto_dist_config import get_dist_config
-
 from ...utils.log import logger
 from ..cache_utils import Cache, DynamicCache
 from ..masking_utils import create_causal_mask_and_row_indices
 from ..model_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
 from ..model_utils import PretrainedModel, register_base_model
 from ..modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from .auto_dist_config import get_dist_config
 from .configuration import LlamaConfig
 
 
@@ -162,9 +166,9 @@ def forward(
         q_shape = (batch_size, seq_len, self.num_heads, self.head_dim)
         kv_shape = (batch_size, seq_len, self.num_key_value_heads, self.head_dim)
 
-        query_states = self.q_proj(hidden_states).view(q_shape).transpose(1, 2)
-        key_states = self.k_proj(hidden_states).view(kv_shape).transpose(1, 2)
-        value_states = self.v_proj(hidden_states).view(kv_shape).transpose(1, 2)
+        query_states = self.q_proj(hidden_states).reshape(q_shape).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).reshape(kv_shape).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).reshape(kv_shape).transpose(1, 2)
 
         cos, sin = position_embeddings
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
@@ -327,8 +331,41 @@ class LlamaPretrainedModel(PretrainedModel):
     ]
 
     @classmethod
-    def _get_tensor_parallel_mappings(cls, config: LlamaConfig, is_split=True):
+    def _get_name_mappings(cls, config: LlamaConfig) -> list[StateDictNameMapping]:
+        mappings: list[StateDictNameMapping] = []
+        model_mappings = [
+            ["embed_tokens.weight"],
+            ["norm.weight"],
+        ]
+        for layer_index in range(config.num_hidden_layers):
+            layer_mappings = [
+                [f"layers.{layer_index}.self_attn.q_proj.weight", None, "transpose"],
+                [f"layers.{layer_index}.self_attn.k_proj.weight", None, "transpose"],
+                [f"layers.{layer_index}.self_attn.v_proj.weight", None, "transpose"],
+                [f"layers.{layer_index}.self_attn.o_proj.weight", None, "transpose"],
+                [f"layers.{layer_index}.self_attn.rotary_emb.inv_freq"],
+                [f"layers.{layer_index}.mlp.gate_proj.weight", None, "transpose"],
+                [f"layers.{layer_index}.mlp.down_proj.weight", None, "transpose"],
+                [f"layers.{layer_index}.mlp.up_proj.weight", None, "transpose"],
+                [f"layers.{layer_index}.input_layernorm.weight"],
+                [f"layers.{layer_index}.post_attention_layernorm.weight"],
+            ]
+            model_mappings.extend(layer_mappings)
+
+        init_name_mappings(mappings=model_mappings)
+        # base-model prefix "LlamaModel"
+        if "LlamaModel" not in config.architectures:
+            for mapping in model_mappings:
+                mapping[0] = "model." + mapping[0]
+                mapping[1] = "llama." + mapping[1]
+            if not config.tie_word_embeddings:
+                model_mappings.append(["lm_head.weight", "lm_head.weight", "transpose"])
+
+        mappings = [StateDictNameMapping(*mapping, index=index) for index, mapping in enumerate(model_mappings)]
+        return mappings
 
+    @classmethod
+    def _get_tensor_parallel_mappings(cls, config: LlamaConfig, is_split=True):
         from ..conversion_utils import split_or_merge_func
 
         fn = split_or_merge_func(