From f27fa0da93d47f028910ca5d78b427573cd34e97 Mon Sep 17 00:00:00 2001
From: Kurkur <102506892+litmei@users.noreply.github.com>
Date: Tue, 2 Jun 2026 13:14:14 +0800
Subject: [PATCH] [NPU][Docs] Kimi-K2.5 best practice (#26774)
---
.../ascend-npus/ascend_npu_best_practice.mdx | 230 ++++++++++++++++++
.../ascend_npu_kimi_k2.5_examples.mdx | 8 +-
2 files changed, 234 insertions(+), 4 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 b39552a39..1dd298559 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
@@ -786,6 +786,86 @@ you encounter issues or have any questions, please [open an issue](https://githu
+## Kimi Series Models
+
+### Low Latency
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+ | Model |
+ Hardware |
+ Cards |
+ Deploy Mode |
+ Dataset |
+ TPOT |
+ Quantization |
+ Configuration |
+
+
+
+
+ | Kimi-K2.5-w4a8 |
+ Atlas 800I A3 |
+ 8 |
+ PD Mixed |
+ 3.5K+1.5K |
+ 20ms |
+ W4A8 INT8 |
+ Optimal Configuration |
+
+
+
+
+### High Throughput
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+ | Model |
+ Hardware |
+ Cards |
+ Deploy Mode |
+ Dataset |
+ TPOT |
+ Quantization |
+ Configuration |
+
+
+
+
+ | Kimi-K2.5-w4a8 |
+ Atlas 800I A3 |
+ 8 |
+ PD Mixed |
+ 3.5K+1.5K |
+ 50ms |
+ W4A8 INT8 |
+ Optimal Configuration |
+
+
+
+
## Optimal Configuration
### DeepSeek-R1 3_5K-1_5K 50ms on A3 32 Cards Disaggregation Mode
@@ -5775,3 +5855,153 @@ 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 --random-input-len 131072 --random-output-len 1024 --num-prompts 192 --random-range-ratio 1 --max-concurrency 48
```
+
+
+### Kimi K2.5 w4a8 3_5K-1_5K 20ms on A3 8 Cards Mixed Mode
+
+Model: Kimi-K2.5-w4a8
+
+Hardware: Atlas 800I A3 8Card
+
+DeployMode: PD Mixed
+
+Dataset: random
+
+Input Output Length: 3.5K+1.5K
+
+TPOT: 20ms
+
+#### Model Deployment
+
+```bash Command
+echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
+sysctl -w vm.swappiness=0
+sysctl -w kernel.numa_balancing=0
+sysctl -w kernel.sched_migration_cost_ns=50000
+
+unset https_proxy
+unset http_proxy
+unset HTTPS_PROXY
+unset HTTP_PROXY
+unset ASCEND_LAUNCH_BLOCKING
+source /usr/local/Ascend/ascend-toolkit/set_env.sh
+source /usr/local/Ascend/nnal/atb/set_env.sh
+source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
+
+export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600
+export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
+export HCCL_SOCKET_IFNAME=lo
+export GLOO_SOCKET_IFNAME=lo
+export STREAMS_PER_DEVICE=32
+export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
+export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=48
+export HCCL_BUFFSIZE=1200
+export SGLANG_ENABLE_SPEC_V2=1
+export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
+export SGLANG_NPU_USE_MLAPO=1
+export SGLANG_NPU_USE_MULTI_STREAM=1
+export SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE=1
+export SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES=200
+
+MODEL_PATH=xxx
+DRAFT_PATH=xxx
+
+python3 -m sglang.launch_server \
+ --model-path $MODEL_PATH --quantization modelslim --dtype bfloat16 \
+ --model-loader-extra-config '{"enable_multithread_load": true}' \
+ --host 0.0.0.0 --port 6699 \
+ --trust-remote-code --device npu --attention-backend ascend \
+ --tp-size 16 --base-gpu-id 0 --mem-fraction-static 0.78 --max-running-requests 64 \
+ --chunked-prefill-size 32768 --context-length 8192 --max-prefill-tokens 16384 \
+ --enable-multimodal --mm-attention-backend ascend_attn --sampling-backend ascend \
+ --enable-dp-attention --dp-size 16 \
+ --moe-a2a-backend deepep --deepep-mode auto \
+ --cuda-graph-bs 1 2 3 4 --disable-radix-cache \
+ --speculative-algorithm EAGLE3 \
+ --speculative-draft-model-path $DRAFT_PATH \
+ --speculative-num-steps 4 --speculative-eagle-topk 1 \
+ --speculative-num-draft-tokens 5 \
+ --speculative-draft-model-quantization unquant
+```
+
+#### Benchmark
+
+We tested it based on the `RANDOM` dataset.
+
+```bash Command
+python3 -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 6699 --random-range-ratio 1 --max-concurrency 64 --random-output-len 1500 --random-input-len 3500 --num-prompts 64
+```
+
+
+### Kimi K2.5 w4a8 3_5K-1_5K 50ms on A3 8 Cards Mixed Mode
+
+Model: Kimi-K2.5-w4a8
+
+Hardware: Atlas 800I A3 8Card
+
+DeployMode: PD Mixed
+
+Dataset: random
+
+Input Output Length: 3.5K+1.5K
+
+TPOT: 50ms
+
+#### Model Deployment
+
+```bash Command
+echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
+sysctl -w vm.swappiness=0
+sysctl -w kernel.numa_balancing=0
+sysctl -w kernel.sched_migration_cost_ns=50000
+
+unset https_proxy
+unset http_proxy
+unset HTTPS_PROXY
+unset HTTP_PROXY
+unset ASCEND_LAUNCH_BLOCKING
+source /usr/local/Ascend/ascend-toolkit/set_env.sh
+source /usr/local/Ascend/nnal/atb/set_env.sh
+source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
+
+export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600
+export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
+export HCCL_SOCKET_IFNAME=lo
+export GLOO_SOCKET_IFNAME=lo
+export STREAMS_PER_DEVICE=32
+export DEEP_NORMAL_MODE_USE_INT8_QUANT=1
+export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=96
+export HCCL_BUFFSIZE=1200
+export SGLANG_ENABLE_SPEC_V2=1
+export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
+export SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE=1
+export SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES=200
+
+MODEL_PATH=xxx
+DRAFT_PATH=xxx
+
+python3 -m sglang.launch_server \
+ --model-path $MODEL_PATH --quantization modelslim --dtype bfloat16 \
+ --model-loader-extra-config '{"enable_multithread_load": true}' \
+ --host 0.0.0.0 --port 6699 \
+ --trust-remote-code --device npu --attention-backend ascend \
+ --tp-size 16 --base-gpu-id 0 --mem-fraction-static 0.7 --max-running-requests 120 \
+ --chunked-prefill-size 32768 --context-length 8192 --max-prefill-tokens 16384 \
+ --enable-multimodal --mm-attention-backend ascend_attn --sampling-backend ascend \
+ --enable-dp-attention --dp-size 16 \
+ --moe-a2a-backend deepep --deepep-mode auto \
+ --cuda-graph-bs 1 2 4 8 12 16 24 32 48 64 96 120 --disable-radix-cache \
+ --speculative-algorithm EAGLE3 \
+ --speculative-draft-model-path $DRAFT_PATH \
+ --speculative-num-steps 4 --speculative-eagle-topk 1 \
+ --speculative-num-draft-tokens 5 \
+ --speculative-draft-model-quantization unquant
+```
+
+#### Benchmark
+
+We tested it based on the `RANDOM` dataset.
+
+```bash Command
+python3 -m sglang.bench_serving --dataset-name random --backend sglang --host 127.0.0.1 --port 6699 --random-range-ratio 1 --max-concurrency 120 --random-output-len 1500 --random-input-len 3500 --num-prompts 120
+```
diff --git a/docs_new/docs/hardware-platforms/ascend-npus/ascend_npu_kimi_k2.5_examples.mdx b/docs_new/docs/hardware-platforms/ascend-npus/ascend_npu_kimi_k2.5_examples.mdx
index dac84c9b3..445f8fe36 100644
--- a/docs_new/docs/hardware-platforms/ascend-npus/ascend_npu_kimi_k2.5_examples.mdx
+++ b/docs_new/docs/hardware-platforms/ascend-npus/ascend_npu_kimi_k2.5_examples.mdx
@@ -300,10 +300,10 @@ SGLang Model Gateway (former Router)
python -m sglang_router.launch_router \
--pd-disaggregation \
--policy cache_aware \
- --prefill http://'your prefill ip1':8000 8998 \
- --prefill http://'your prefill ip2':8000 8999 \
- --prefill http://'your prefill ip3':8000 9000 \
- --decode http://'your decode ip1':8001 \
+ --prefill http://:8000 8998 \
+ --prefill http://:8000 8999 \
+ --prefill http://:8000 9000 \
+ --decode http://:8001 \
--host 127.0.0.1 \
--port 6688 \
```