[Refactor] MLP weight prefetch to consistency with MoE Model's prefetching in terms of code and usage (#6442)

### What this PR does / why we need it?
Refactor MLP weight prefetch to consistency with MoE Model's prefetching
in terms of code and usage.
Environments VLLM_ASCEND_ENABLE_PREFETCH_MLP,
VLLM_ASCEND_MLP_DOWN_PREFETCH_SIZE and
VLLM_ASCEND_MLP_GATE_UP_PREFETCH_SIZE is removed, usage as following:

--additional-config '{"weight_prefetch_config": { "enabled": true,
"prefetch_ratio": {"mlp": { "gate_up": 1.0, "down": 1.0} }}}'

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.14.1
- vLLM main:
dc917cceb8

---------

Signed-off-by: leo-pony <nengjunma@outlook.com>
This commit is contained in:
Nengjun Ma
2026-02-04 09:08:18 +08:00
committed by GitHub
parent fa56abea9f
commit 78fad4e348
18 changed files with 250 additions and 171 deletions

View File

@@ -17,7 +17,7 @@
import torch
from vllm.model_executor.layers.activation import QuickGELU, SiluAndMul
from vllm_ascend.utils import get_weight_prefetch_method
class AscendQuickGELU(QuickGELU):
@@ -33,7 +33,10 @@ class AscendSiluAndMul(SiluAndMul):
def forward_oot(self, x: torch.Tensor) -> torch.Tensor:
import torch_npu
torch.ops.vllm.maybe_prefetch_mlp_down_proj(x)
weight_prefetch_method = get_weight_prefetch_method()
if weight_prefetch_method:
weight_prefetch_method.maybe_prefetch_mlp_weight_preprocess(weight_prefetch_method.MLP_DOWN, x)
out = torch_npu.npu_swiglu(x)
torch.ops.vllm.maybe_wait_prefetch_done(out)
if weight_prefetch_method:
weight_prefetch_method.maybe_prefetch_mlp_weight_postprocess(out)
return out