From 825fdfb197fd699e459a16a6c1a4ef7e05b400a6 Mon Sep 17 00:00:00 2001
From: Ruri <33858552+zhoux77899@users.noreply.github.com>
Date: Mon, 27 Oct 2025 09:42:09 +0800
Subject: [PATCH] [v0.11.0][Feat] Prefetching Attention QKV Linear Weight With
 `AddRmsNormQuant` Custom Op (#3649)

### What this PR does / why we need it?

- `qkv_proj.weight` prefetching has been implemented with `Quant` op,
when `AddRmsNormQuant` is enabled (#3465) `qkv_proj.weight` prefetching
won't work
- Implement `qkv_proj.weight` prefetching with `AddRmsNormQuant`, which
has been merged on `main` branch (#3517)

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

None.

### How was this patch tested?

Tested on `Qwen3-235B-A22B-W8A8`
<img width="1868" height="109" alt="image"

src="https://github.com/user-attachments/assets/0bc28082-0287-4d5c-b8f6-f907c3134d36"
/>


- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

---------

<!--  Thanks for sending a pull request!

BEFORE SUBMITTING, PLEASE READ
https://docs.vllm.ai/en/latest/contributing/overview.html

-->
### What this PR does / why we need it?
<!--
- Please clarify what changes you are proposing. The purpose of this
section is to outline the changes and how this PR fixes the issue.
If possible, please consider writing useful notes for better and faster
reviews in your PR.

- Please clarify why the changes are needed. For instance, the use case
and bug description.

- Fixes #
-->

### Does this PR introduce _any_ user-facing change?
<!--
Note that it means *any* user-facing change including all aspects such
as API, interface or other behavior changes.
Documentation-only updates are not considered user-facing changes.
-->

### How was this patch tested?
<!--
CI passed with new added/existing test.
If it was tested in a way different from regular unit tests, please
clarify how you tested step by step, ideally copy and paste-able, so
that other reviewers can test and check, and descendants can verify in
the future.
If tests were not added, please describe why they were not added and/or
why it was difficult to add.
-->

Signed-off-by: zhoux77899 <zhouxiang100@huawei.com>
---
 tests/ut/ops/test_layernorm.py | 13 +++++++------
 vllm_ascend/ops/layernorm.py   | 22 ++++++++++++++++++++++
 2 files changed, 29 insertions(+), 6 deletions(-)

diff --git a/tests/ut/ops/test_layernorm.py b/tests/ut/ops/test_layernorm.py
index ed9dd44..7a829b6 100644
--- a/tests/ut/ops/test_layernorm.py
+++ b/tests/ut/ops/test_layernorm.py
@@ -117,6 +117,7 @@ class TestAscendRMSNorm(PytestBase):
         mock_forward_context.layer_idx = 0
         mock_forward_context.num_hidden_layers = num_hidden_layers
         mock_forward_context.fusion_linear = "gate_up_dense"
+        mock_forward_context.weight_prefetch_method = None
 
         # Ensure fusion and layer_idx increment are handled correctly
         x = torch.randn(4, 8, dtype=torch.float16)
@@ -125,13 +126,13 @@ class TestAscendRMSNorm(PytestBase):
 
         x_out, residual_out = layer.forward_oot(x, residual)
 
-        assert mock_get_forward_context.call_count == 1
+        assert mock_get_forward_context.call_count == 2
         assert mock_forward_context.fusion_linear == "qkv_dense"
         assert mock_forward_context.layer_idx == 1
 
         x_out, residual_out = layer.forward_oot(x, residual)
 
-        assert mock_get_forward_context.call_count == 2
+        assert mock_get_forward_context.call_count == 4
         assert mock_forward_context.fusion_linear == "gate_up_dense"
         assert mock_forward_context.layer_idx == 1
 
@@ -139,14 +140,14 @@ class TestAscendRMSNorm(PytestBase):
             mock_forward_context.fusion_linear = "gate_moe"
         x_out, residual_out = layer.forward_oot(x, residual)
 
-        assert mock_get_forward_context.call_count == 3
+        assert mock_get_forward_context.call_count == 6
         fusion_linear_expected = "qkv_moe" if torch_npu_check else "qkv_dense"
         assert mock_forward_context.fusion_linear == fusion_linear_expected
         assert mock_forward_context.layer_idx == 2
 
         x_out, residual_out = layer.forward_oot(x, residual)
 
-        assert mock_get_forward_context.call_count == 4
+        assert mock_get_forward_context.call_count == 7
         fusion_linear_expected = "gate_moe" if torch_npu_check else "qkv_dense"
         assert mock_forward_context.fusion_linear == fusion_linear_expected
         assert mock_forward_context.layer_idx == 2
@@ -156,13 +157,13 @@ class TestAscendRMSNorm(PytestBase):
         # last layer returned directly
         x_out, residual_out = layer.forward_oot(x, residual)
 
-        assert mock_get_forward_context.call_count == 5
+        assert mock_get_forward_context.call_count == 8
         assert mock_forward_context.fusion_linear == "qkv_moe"
         assert mock_forward_context.layer_idx == 3
 
         x_out, residual_out = layer.forward_oot(x, residual)
 
-        assert mock_get_forward_context.call_count == 6
+        assert mock_get_forward_context.call_count == 9
         assert mock_forward_context.fusion_linear == "qkv_moe"
         assert mock_forward_context.layer_idx == 3
 
diff --git a/vllm_ascend/ops/layernorm.py b/vllm_ascend/ops/layernorm.py
index 55eab21..239eeb0 100644
--- a/vllm_ascend/ops/layernorm.py
+++ b/vllm_ascend/ops/layernorm.py
@@ -38,6 +38,21 @@ def _addrmsnorm_forward_oot(
 
     torch_npu_check = version_check()
     if layer is not None and not is_310p():
+        layer_cls_name = layer.__class__.__name__
+        try:
+            weight_prefetch_method = get_forward_context(
+            ).weight_prefetch_method
+        except AssertionError:
+            weight_prefetch_method = None
+
+        # prefetch qkvo_proj.weight preprocess
+        if weight_prefetch_method:
+            weight_prefetch_method.maybe_prefetch_attn_weight_preprocess(
+                layer_cls_name=layer_cls_name,
+                weight=layer.weight,
+                start_flag=x,
+            )
+        # add_rms_norm_quant
         if torch_npu_check:
             x, _, residual = torch_npu.npu_add_rms_norm_quant(
                 x,
@@ -55,6 +70,13 @@ def _addrmsnorm_forward_oot(
                 layer.aclnn_input_scale,
                 layer.aclnn_input_offset,
                 epsilon=self.variance_epsilon)
+        # prefetch qkvo_proj.weight postprocess
+        if weight_prefetch_method:
+            weight_prefetch_method.maybe_prefetch_attn_weight_postprocess(
+                layer_cls_name=layer_cls_name,
+                stop_flag=x,
+            )
+
     else:
         if is_310p():
             orig_dtype = residual.dtype