[5/N] pass the whole config to model (#9983)

Signed-off-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
youkaichao
2024-11-09 14:17:28 +08:00
committed by GitHub
parent 49d2a41a86
commit 1a95f10ee7
75 changed files with 583 additions and 654 deletions
+6 -12
View File
@@ -22,13 +22,11 @@
# limitations under the License.
"""Inference-only DeciLM model compatible with HuggingFace weights."""
from typing import Iterable, Optional, Tuple
from typing import Iterable, Tuple
import torch
from transformers import LlamaConfig
from vllm.config import CacheConfig, LoRAConfig
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.config import VllmConfig
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.models.llama import LlamaForCausalLM
@@ -55,17 +53,13 @@ class DeciLMForCausalLM(LlamaForCausalLM):
def __init__(
self,
config: LlamaConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
lora_config: Optional[LoRAConfig] = None,
vllm_config: VllmConfig,
prefix: str = "",
) -> None:
config = vllm_config.model_config.hf_config
config.num_key_value_heads = max(config.num_key_value_heads_per_layer)
delattr(config, "num_key_value_heads_per_layer")
super().__init__(config=config,
cache_config=cache_config,
quant_config=quant_config,
lora_config=lora_config)
super().__init__(vllm_config=vllm_config)
def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
stacked_params_mapping = [