Remove VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE (#5272)

`VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE` is only used together with
`VLLM_ASCEND_ENABLE_PREFETCH_MLP` which is useless totally. This PR
remove it.
- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>

vllm_ascend/ascend_forward_context.py

@@ -102,8 +102,7 @@ def set_ascend_forward_context(
 
         # TODO(rjg-lyh): refactor mlp weight prefetch method
         # set for mlp weight prefetch
-        prefetch_mlp_enabled = envs_ascend.VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE and \
-            envs_ascend.VLLM_ASCEND_ENABLE_PREFETCH_MLP and \
+        prefetch_mlp_enabled = envs_ascend.VLLM_ASCEND_ENABLE_PREFETCH_MLP and \
             forward_context.layer_idx is not None and \
             num_tokens is not None and num_tokens < 500
         if prefetch_mlp_enabled:

vllm_ascend/envs.py

@@ -108,11 +108,6 @@ env_variables: Dict[str, Callable[[], Any]] = {
     "VLLM_ASCEND_MLP_DOWN_PREFETCH_SIZE":
     lambda: int(
         os.getenv("VLLM_ASCEND_MLP_DOWN_PREFETCH_SIZE", 18 * 1024 * 1024)),
-    # Whether to enable dense model and general optimizations for better performance.
-    # Since we modified the base parent class `linear`, this optimization is also applicable to other model types.
-    # However, there might be hidden issues, and it is currently recommended to prioritize its use with dense models.
-    "VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE":
-    lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE", '0'))),
     # Whether to enable msMonitor tool to monitor the performance of vllm-ascend.
     "MSMONITOR_USE_DAEMON":
     lambda: bool(int(os.getenv("MSMONITOR_USE_DAEMON", '0'))),

vllm_ascend/ops/linear_op.py

@@ -53,13 +53,13 @@ from vllm.distributed import (split_tensor_along_last_dim,
 from vllm.distributed.parallel_state import get_tp_group
 from vllm.forward_context import get_forward_context
 
+from vllm_ascend import envs as envs_ascend
 from vllm_ascend.ascend_config import get_ascend_config
 from vllm_ascend.distributed.parallel_state import (get_flashcomm2_odp_group,
                                                     get_flashcomm2_otp_group,
                                                     get_mlp_tp_group,
                                                     get_otp_group)
-from vllm_ascend.utils import (dense_optim_enable, enable_sp,
-                               flashcomm2_enable,
+from vllm_ascend.utils import (enable_sp, flashcomm2_enable,
                                get_flashcomm2_reorgnized_batch_ids,
                                matmul_allreduce_enable, mlp_tp_enable,
                                oproj_tp_enable, shared_expert_dp_enabled)
@@ -135,7 +135,7 @@ class CustomRowParallelOp(CustomLinearOp):
 
     def apply(self, input_):
         output, output_bias = self.apply_impl(input_)
-        if dense_optim_enable():
+        if envs_ascend.VLLM_ASCEND_ENABLE_PREFETCH_MLP:
             torch.ops.vllm.maybe_prefetch_mlp_gate_up_proj(output, self.prefix)
         if not self.return_bias:
             return output

vllm_ascend/utils.py

@@ -772,10 +772,6 @@ def matmul_allreduce_enable() -> bool:
     return envs_ascend.VLLM_ASCEND_ENABLE_MATMUL_ALLREDUCE
 
 
-def dense_optim_enable() -> bool:
-    return envs_ascend.VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE
-
-
 def enable_sp(vllm_config=None, enable_shared_expert_dp: bool = False) -> bool:
     global _ENABLE_SP
     if _ENABLE_SP is None: