Merge branch 'v0.9.2-dev' into v0.9.2-dev-ds

vllm/model_executor/layers/rotary_embedding.py

@@ -916,7 +916,7 @@ class DeepseekScalingRotaryEmbedding(RotaryEmbedding):
     direct_register_custom_op(
         op_name="rotary_embedding_deepseek_fuse",
         op_func=rotary_embedding_deepseek_fuse,
-        mutates_args=[], 
+        mutates_args=["query", "key"], 
         fake_impl=rotary_embedding_deepseek_fuse_fake,
     )
     
@@ -958,8 +958,7 @@ class DeepseekScalingRotaryEmbedding(RotaryEmbedding):
                     BLOCK_SIZE=BLOCK_SIZE,
                     num_warps=1)
 
-            # if envs.VLLM_USE_LIGHTOP:
-            if False:
+            if envs.VLLM_USE_LIGHTOP:
                 torch.ops.vllm.rotary_embedding_deepseek_fuse(positions, query, key, self.head_size, self.cos_sin_cache, self.is_neox_style)
             else:
                 call(query)

vllm/platforms/rocm.py

@@ -16,6 +16,14 @@ from vllm.utils import cuda_device_count_stateless
 
 from .interface import DeviceCapability, Platform, PlatformEnum, _Backend
 
+from vllm.utils import SUPPORT_TC
+
+if not SUPPORT_TC:
+    os.environ['VLLM_USE_V1'] = '0'
+    os.environ['VLLM_USE_FLASH_ATTN_PA'] = '0'
+    os.environ['VLLM_USE_FLASH_MLA'] = '0'
+    
+
 if TYPE_CHECKING:
     from vllm.config import ModelConfig, VllmConfig