[main] mlp weight prefetch in Qwen Dense Models (#2816)

### What this PR does / why we need it?
This PR prefetchs the weight of mlp layers in Qwen Dense Models to
optimize the performance in Decode phase mainly.

### Does this PR introduce _any_ user-facing change?
 No.

### How was this patch tested?
CI passed with new added/existing test.

- vLLM version: main
- vLLM main:
a1213fae5f

Signed-off-by: rjg-lyh <1318825571@qq.com>
Co-authored-by: Shuming19 <313093131@qq.com>
This commit is contained in:
rjg-lyh
2025-09-11 21:20:09 +08:00
committed by GitHub
parent c3c2221503
commit 0005479b9c
17 changed files with 313 additions and 24 deletions

View File

@@ -199,8 +199,12 @@ def torchair_quant_method_register():
def torchair_ops_patch():
from vllm_ascend.ops.activation import AscendSiluAndMul
from vllm_ascend.ops.layernorm import AscendRMSNorm
from vllm_ascend.ops.rotary_embedding import (
AscendDeepseekScalingRotaryEmbedding, AscendRotaryEmbedding)
from vllm_ascend.torchair.ops import (torchair_activation,
torchair_layernorm)
from vllm_ascend.torchair.ops.torchair_rotary_embedding import (
deepseek_rope_init_func, native_rope_deepseek_forward,
qwen_rope_init_func, rope_forward)
@@ -210,3 +214,6 @@ def torchair_ops_patch():
AscendDeepseekScalingRotaryEmbedding.__init__ = deepseek_rope_init_func # type: ignore[method-assign]
AscendDeepseekScalingRotaryEmbedding.forward = native_rope_deepseek_forward # type: ignore[method-assign]
AscendRMSNorm.forward_oot = torchair_layernorm.torchair_rmsnorm_forward_oot # type: ignore[method-assign]
AscendSiluAndMul.forward_oot = torchair_activation.torchair_silu_and_mul_forward_oot # type: ignore[method-assign]