Merge branch 'v0.14.1-dev_yql_1.28_2' into 'v0.14.1-dev'

修复vit attn的导入问题，以及w4a16的gptq的接口问题

See merge request dcutoolkit/deeplearing/vllm!395

vllm/model_executor/layers/quantization/gptq.py

@@ -385,7 +385,7 @@ class GPTQLinearMethod(LinearMethodBase):
             layer.scales,
             layer.g_idx,
             layer.exllama_state == ExllamaState.READY,
-            self.use_v2_format,
+            # self.use_v2_format,
             self.quant_config.weight_bits,
         )
         if bias is not None:

vllm/model_executor/layers/quantization/kernels/mixed_precision/exllama.py

@@ -155,12 +155,12 @@ class ExllamaLinearKernel(MPLinearKernel):
         # gptq_gemm supports GPTQv2 format by passing use_v2_format=True.
         # However, the MPLinearLayerConfig doesn't contain format info.
         # So hardcode GPTQv1 format here, to keep its behavior unchanged.
-        use_v2_format = False
+        #use_v2_format = False
 
         assert w_zp is not None, "Zero points are required by Exllama"
         assert w_g_idx is not None, "Group index is required by Exllama"
         output = ops.gptq_gemm(
-            x_2d, w_q, w_zp, w_s, w_g_idx, True, use_v2_format, c.weight_type.size_bits
+            x_2d, w_q, w_zp, w_s, w_g_idx, True, c.weight_type.size_bits
         )
 
         if bias is not None:

vllm/v1/attention/backends/fa_utils.py

@@ -23,7 +23,7 @@ elif current_platform.is_xpu():
 elif current_platform.is_rocm():
     try:
         from vllm._custom_ops import reshape_and_cache_cuda
-        from flash_attn import vllm_flash_attn_varlen_func
+        from flash_attn import flash_attn_varlen_func, vllm_flash_attn_varlen_func
     except ImportError as e:
         raise ImportError(
             "Rocm platform requires upstream flash-attn "