fix bug

python/sglang/srt/layers/attention/vision.py

@@ -297,8 +297,10 @@ class VisionFlash3Attention(nn.Module):
         self,
         **kwargs,
     ):
-        if not _is_cuda:
-            raise Exception("VisionFlash3Attention is only available for cuda")
+        # if not _is_cuda:
+        #     raise Exception("VisionFlash3Attention is only available for cuda")
+        if _is_hip:
+            from sglang.srt.layers.attention.flashattention_interface import flash_attn_varlen_func
         super().__init__()
 
     def forward(

python/sglang/srt/layers/linear.py

@@ -1389,16 +1389,16 @@ class RowParallelLinear(LinearBase):
         if use_fused_silu_mul_quant:
             xq, xs = lm_fuse_silu_mul_quant(input_parallel)
             silu_quant_args = [xq, xs]
-            with use_symmetric_memory(parallel_state.get_tp_group()) as sm:
+            with use_symmetric_memory(get_tp_group()) as sm:
                 output_parallel = self.quant_method.apply(self, input_parallel,
                                                           bias=bias_,
                                                           silu_quant_args=silu_quant_args
                 )
-                sm.tag(output_parallel)
+                # sm.tag(output_parallel)
         else:
-            with use_symmetric_memory(parallel_state.get_tp_group()) as sm:
+            with use_symmetric_memory(get_tp_group()) as sm:
                 output_parallel = self.quant_method.apply(self, input_parallel, bias=bias_)
-                sm.tag(output_parallel)
+                # sm.tag(output_parallel)
 
         if self.reduce_results and self.tp_size > 1 and not skip_all_reduce:
             output = tensor_model_parallel_all_reduce(output_parallel)

python/sglang/srt/layers/moe/ep_moe/layer.py

@@ -13,7 +13,7 @@ from sglang.srt.layers.moe import (
     get_deepep_mode,
     get_moe_a2a_backend,
     get_moe_runner_backend,
-    should_use_flashinfer_trtllm_moe,
+    # should_use_flashinfer_trtllm_moe,
 )
 from sglang.srt.layers.moe.ep_moe.kernels import (
     ep_gather,

python/sglang/srt/model_executor/forward_batch_info.py

@@ -362,7 +362,7 @@ class ForwardBatch:
             input_embeds=batch.input_embeds,
             token_type_ids=batch.token_type_ids,
             tbo_split_seq_index=batch.tbo_split_seq_index,
-            dimensions=batch.dimensions,
+            # dimensions=batch.dimensions,
         )
         device = model_runner.device
 

python/sglang/srt/models/deepseek_v2.py

@@ -868,11 +868,12 @@ class DeepseekV2MoE(nn.Module):
                 # fused in biased_grouped_topk so we can skip here
                 final_hidden_states *= self.routed_scaling_factor
             if shared_output is not None:
-                with use_symmetric_memory(parallel_state.get_tp_group()) as sm:
-                    final_hidden_states_out = torch.empty_like(final_hidden_states)
-                torch.add(final_hidden_states, shared_output, out=final_hidden_states_out)
-                final_hidden_states = final_hidden_states_out
-                sm.tag(final_hidden_states)
+                # with use_symmetric_memory(parallel_state.get_tp_group()) as sm:
+                #     final_hidden_states_out = torch.empty_like(final_hidden_states)
+                # torch.add(final_hidden_states, shared_output, out=final_hidden_states_out)
+                # final_hidden_states = final_hidden_states_out
+                # sm.tag(final_hidden_states)
+                final_hidden_states += shared_output
         if (
             self.tp_size > 1
             and not should_allreduce_fusion