Merge branch 'v0.14.1-dev_yql_1.28' into 'v0.14.1-dev'

修复了awq的shape的bug，以及兼容了lmslim注册导入的的情况

See merge request dcutoolkit/deeplearing/vllm!394

vllm/model_executor/layers/quantization/awq.py

@@ -33,7 +33,8 @@ if TYPE_CHECKING:
  
 from vllm.model_executor.layers.quantization.awq_triton import awq_gemm_triton
 triton_configs_dict={}
-
+def is_layer_skipped_awq(prefix: str, modules_to_not_convert: list[str]):
+    return any(module_name in prefix for module_name in modules_to_not_convert)
 def get_triton_cache(file_path):
         #会将所报错的json文件以字典的形式return出来
      
@@ -376,7 +377,7 @@ class AWQLinearMethod(LinearMethodBase):
         qzeros = layer.qzeros
         scales = layer.scales
         pack_factor = self.quant_config.pack_factor
-        out_shape = x.shape[:-1] + (qweight.shape[-1] * pack_factor,)
+        out_shape = (x.shape[:-1] + (qweight.shape[0] * 1, ))
         reshaped_x = x.reshape(-1, x.shape[-1])
         
         m = reshaped_x.shape[0]

vllm/platforms/rocm.py

@@ -530,7 +530,7 @@ class RocmPlatform(Platform):
                 " is not set, enabling VLLM_USE_TRITON_AWQ."
             )
             envs.VLLM_USE_TRITON_AWQ = False
-        os.environ["VLLM_USE_TRITON_AWQ"] = "1"
+        # os.environ["VLLM_USE_TRITON_AWQ"] = "1"
 
     @classmethod
     def get_punica_wrapper(cls) -> str:

vllm/utils/__init__.py

@@ -6,7 +6,7 @@ import json
 import uuid
 
 import torch
-
+from vllm.utils.torch_utils import direct_register_custom_op
 MASK_64_BITS = (1 << 64) - 1