update torch.float16

vllm/model_executor/layers/fused_moe/fused_moe.py

@@ -1885,7 +1885,7 @@ def fused_experts_impl(
                                 use_nn_moe=use_nn_moe)
 
         if activation == "silu":
-            if envs.VLLM_USE_FUSE_SILU_AND_MUL and intermediate_cache1.dtype == intermediate_cache2.dtype == "fp16":
+            if envs.VLLM_USE_FUSE_SILU_AND_MUL and intermediate_cache1.dtype == intermediate_cache2.dtype == torch.float16:
                 from lightop import fuse_silu_and_mul
                 fuse_silu_and_mul(intermediate_cache1.view(-1, N),intermediate_cache2)    
             else:

vllm/v1/attention/backends/flash_attn.py

@@ -566,7 +566,7 @@ class FlashAttentionImpl(AttentionImpl):
                     layer._v_scale,
                 )
             else:
-                if envs.VLLM_USE_OPT_RESHAPE_AND_CACHE and key.dtype == value.dtype == "fp16":
+                if envs.VLLM_USE_OPT_RESHAPE_AND_CACHE and key.dtype == value.dtype == torch.float16:
                     from lightop import reshape_and_cache_cuda
                     reshape_and_cache_cuda(
                         key, value,