use apply_rotary_emb_torch for z100l&k100

vllm/envs.py

@@ -1116,7 +1116,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
                  ("true", "1")),  
     # vLLM will use lightop moe_sum_mul_add
     "VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD":
-        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD", "True").lower() in
+        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD", "False").lower() in
                  ("true", "1")),  
     # vLLM will use lightop moe_sum 
     "VLLM_USE_LIGHTOP_MOE_SUM":

vllm/model_executor/layers/rotary_embedding.py

@@ -39,10 +39,11 @@ from vllm.model_executor.custom_op import CustomOp
 from vllm.platforms import current_platform
 import vllm.envs as envs
 from vllm.utils import direct_register_custom_op
+from vllm.utils import SUPPORT_TC
 
 if current_platform.is_cuda():
     from vllm.vllm_flash_attn.layers.rotary import apply_rotary_emb
-if current_platform.is_rocm():
+if current_platform.is_rocm() and SUPPORT_TC:
     from flash_attn.layers.rotary import apply_rotary_emb
 
 
@@ -91,10 +92,13 @@ def _apply_rotary_emb(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor,
             positional embeddings.
     """
     if current_platform.is_cuda():
+        if SUPPORT_TC:
             return apply_rotary_emb(x.unsqueeze(0), cos, sin,
                                     not is_neox_style).squeeze(0)
         else:
             return _apply_rotary_emb_torch(x, cos, sin, is_neox_style)
+    else:
+        return _apply_rotary_emb_torch(x, cos, sin, is_neox_style)
 
 
 @CustomOp.register("rotary_embedding")

vllm/model_executor/models/keye.py

@@ -55,6 +55,7 @@ from .utils import (AutoWeightsLoader, WeightsMapper,
                     maybe_prefix, merge_multimodal_embeddings)
 from .vision import get_vit_attn_backend
 from vllm.platforms import current_platform
+from vllm.utils import SUPPORT_TC
 
 logger = init_logger(__name__)
 
@@ -331,6 +332,7 @@ def apply_rotary_pos_emb_flashatt(
     cos = cos.chunk(2, dim=-1)[0].contiguous()
     sin = sin.chunk(2, dim=-1)[0].contiguous()
 
+    if SUPPORT_TC:
         if not current_platform.is_rocm():
             from vllm.vllm_flash_attn.layers.rotary import apply_rotary_emb
         else:

vllm/model_executor/models/qwen2_vl.py

@@ -85,6 +85,7 @@ import re
 from vllm import _custom_ops as ops
 from vllm.model_executor.utils import pad_weight, gemm_bank_conf
 from vllm.platforms import current_platform
+from vllm.utils import SUPPORT_TC
 
 logger = init_logger(__name__)
 
@@ -246,7 +247,7 @@ def apply_rotary_pos_emb_vision(t: torch.Tensor,
     apply_rotary_emb = apply_rotary_emb_torch
     if current_platform.is_cuda():
         from vllm.vllm_flash_attn.layers.rotary import apply_rotary_emb
-    if current_platform.is_rocm():
+    if current_platform.is_rocm() and SUPPORT_TC:
         from flash_attn.layers.rotary import apply_rotary_emb
     output = apply_rotary_emb(t_, cos, sin).type_as(t)
     return output