From 18f3310aadc1c981510d14bb193d70ef3bca835f Mon Sep 17 00:00:00 2001
From: amote-i <49533125+amote-i@users.noreply.github.com>
Date: Wed, 22 Apr 2026 17:51:28 +0800
Subject: [PATCH] [NPU] [DOC] Update Ascend NPU best practice (#23459)
---
.../ascend-npus/ascend_npu_best_practice.mdx | 17 ++++++++---------
1 file changed, 8 insertions(+), 9 deletions(-)
diff --git a/docs_new/docs/hardware-platforms/ascend-npus/ascend_npu_best_practice.mdx b/docs_new/docs/hardware-platforms/ascend-npus/ascend_npu_best_practice.mdx
index ced8ef268..e57d590d1 100644
--- a/docs_new/docs/hardware-platforms/ascend-npus/ascend_npu_best_practice.mdx
+++ b/docs_new/docs/hardware-platforms/ascend-npus/ascend_npu_best_practice.mdx
@@ -357,12 +357,12 @@ you encounter issues or have any questions, please [open an issue](https://githu
| Qwen3-Next-A3B-Instruct |
Atlas 800I A3 |
- 2 |
+ 1 |
PD Mixed |
3.5K+1.5K |
20ms |
W8A8 INT8 |
- Optimal Configuration |
+ Optimal Configuration |
| Qwen3-14B |
@@ -689,7 +689,7 @@ python -m sglang_router.launch_router \
We tested it based on the `RANDOM` dataset.
```shell Command
-python -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 6688 --max-concurrency 768 --random-input-len 3500 --random-output-len 1500 --num-prompts 3072 --random-range-ratio 1 --request-rate 16
+python -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 6688 --max-concurrency 1024 --random-input-len 3584 --random-output-len 1536 --num-prompts 7168 --random-range-ratio 1 --request-rate 40
```
### DeepSeek-R1 2K-2K 50ms on A3 24 Cards Disaggregation Mode
@@ -4163,11 +4163,11 @@ We tested it based on the `RANDOM` dataset.
python3 -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 7239 --max-concurrency 1 --random-output-len 1500 --random-input-len 3500 --num-prompts 4 --random-range-ratio 1
```
-### Qwen3-Next 3_5K-1_5K 20ms on A3 2 Cards Mixed Mode
+### Qwen3-Next 3_5K-1_5K 20ms on A3 1 Cards Mixed Mode
Model: Qwen3-Next-80B-A3B-Instruct
-Hardware: Atlas 800I A3 2Card
+Hardware: Atlas 800I A3 1Card
DeployMode: PD Mixed
@@ -4231,7 +4231,7 @@ echo "${LOCAL_HOST2}"
python3 -m sglang.launch_server --model-path ${MODEL_PATH} \
--page-size 128 \
- --tp-size 4 --dp-size 2 \
+ --tp-size 2 \
--trust-remote-code \
--attention-backend ascend \
--device npu \
@@ -4240,10 +4240,9 @@ python3 -m sglang.launch_server --model-path ${MODEL_PATH} \
--host 127.0.0.1 --port 6699 \
--mem-fraction-static 0.85 \
--disable-radix-cache --max-prefill-tokens 28672 --context-length 26384 --max-total-tokens 122304 \
- --enable-dp-attention --enable-dp-lm-head \
--speculative-algorithm NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 --speculative-draft-model-quantization unquant \
- --chunked-prefill-size -1 --max-running-requests 16 \
- --cuda-graph-bs 2 4 8 \
+ --chunked-prefill-size -1 --max-running-requests 2 \
+ --cuda-graph-bs 2 \
--mamba-ssm-dtype bfloat16 \
--speculative-draft-model-path /path/to/Qwen3-Next-80B-A3B-Instruct
```