1. 基于v0.14.0rc1-a3-openeuler.tar.gz
2. 重装Vllm-ascend：
pip uninstall vllm-ascend
git clone https://github.com/vllm-project/vllm-ascend.git
cd vllm-ascend
git checkout d1dcdfc4084825d2d8f6ff39f1e69767e5f88c40
pip install -v -e .
cd ..

pkill -9 python
pkill -9 VLLM
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
export VLLM_USE_V1=1
export VLLM_VERSION=0.14.1

python -m vllm.entrypoints.openai.api_server  \
       --model=XXX/GLM 4.7 \
       --served-model-name GLM 4.7 \
       --trust-remote-code \
       --max-model-len 32768 \
       --max-num-batched-tokens 122880 \
       -tp 16 \
       --port 9998 \
       --block-size 128 \
       --no-enable-prefix-caching \
       --gpu-memory-utilization 0.95 

pkill -9 python
pkill -9 VLLM
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
export VLLM_USE_V1=1
export VLLM_VERSION=0.14.1

python -m vllm.entrypoints.openai.api_server  \
       --model=XXX/GLM 4.7 \
       --served-model-name GLM 4.7 \
       --trust-remote-code \
       --max-model-len 32768 \
       --max-num-batched-tokens 122880 \
       -tp 8 \
       --port 9998 \
       --block-size 128 \
       --no-enable-prefix-caching \
       --compilation-config '{"cudagraph_capture_sizes":[1,2,4,8,16,24,48]}' \
       --gpu-memory-utilization 0.95 

pkill -9 python
pkill -9 VLLM
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
export VLLM_USE_V1=1
export VLLM_VERSION=0.14.1

python -m vllm.entrypoints.openai.api_server  \
       --model=XXX/GLM 4.7 \
       --served-model-name GLM 4.7 \
       --trust-remote-code \
       --max-model-len 32768 \
       --max-num-batched-tokens 122880 \
       -tp 8 \
       --port 9998 \
       --block-size 128 \
       --no-enable-prefix-caching \
       --async-scheduling \
       --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY", "cudagraph_capture_sizes":[1,2,4,8,16,24,48]}' \
       --gpu-memory-utilization 0.95 

export TASK_QUEUE_ENABLE=1

export LD_PRELOAD=/usr/local/Ascend/ascend-toolkit/latest/lib64/libjemalloc.so

export HCCL_OP_EXPANSION_MODE="AIV"

export HCCL_BUFFSIZE=1024
export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15
export OMP_NUM_THREADS=64
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export VLLM_USE_V1=1
export HCCL_DETERMINISTIC="true"
export HCCL_OP_EXPANSION_MODE=AIV
export OMP_NUM_THREADS=1
export VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE=1
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl kernel.sched_migration_cost_ns=50000
export TASK_QUEUE_ENABLE=1
export VLLM_ASCEND_ENABLE_NZ=2
vllm serve /disk1/xxx/GLM-4.7-W8A8 \
        --max-model-len 131072 \
        --quantization ascend \
        --enable-expert-parallel \
        --port 8262 \
        --served-model-name GLM-4.7-w8a8 \
        --reasoning-parser glm45 \
        --trust-remote-code \
        --gpu_memory_utilization 0.9 \
        --async-scheduling \
        --max-num-seqs 64 \
        --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY", "cudagraph_capture_sizes": [1,4,8,16,32,48,64]}' \
        --data-parallel-size 2 \
        --tensor-parallel-size 8 \
        --enable-expert-parallel  \
        --additional-config '{"cudagraph_mode":"FULL_DECODE_ONLY","ascend_scheduler_config":{"enabled":false},"enable_multistream_moe":false,"chunked_prefill_for_mla":true,"enable_weight_nz_layout":true}'

git clone https://gitcode.com/Ascend/msit.git
cd msit/msmodelslim
bash install.sh

from add_safetensors import add_safetensors
add_safetensors(org_paths="/disk1/models/GLM-4.7", target_dir="/disk1/models/GLM-4.7-W8A8/", safetensors_prefix="mtp_float",
					max_file_size_gb=5, prefix="model.layers.92.")

新增配置文件合并脚本configmerge.py
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
import os
import json
import sys
import argparse

# 直接指定路径
INPUT_DIR = "/disk1/models/GLM-4.7-W8A8"  # 修改为你的模型目录
OUTPUT_FILE = "/disk1/models/GLM-4.7-W8A8/config.json"  # 修改为你想要的输出文件路径

def merge_configs():
    """
    合并配置文件:
    1. 从quant_model_description.json读取内容
    2. 将内容合并到config.json的quantization_config部分
    3. 保存文件config.json
    """
    # 构建文件路径
    config_path = os.path.join(INPUT_DIR, "config.json")
    quant_desc_path = os.path.join(INPUT_DIR, "quant_model_description.json")

    # 检查文件存在
    if not os.path.exists(config_path):
        print(f"错误: 配置文件不存在: {config_path}")
        return False

    if not os.path.exists(quant_desc_path):
        print(f"错误: 量化描述文件不存在: {quant_desc_path}")
        return False

    try:
        # 读取config.json
        with open(config_path, 'r', encoding='utf-8') as f:
            config_data = json.load(f)

        # 读取quant_model_description_w8a8.json
        with open(quant_desc_path, 'r', encoding='utf-8') as f:
            quant_desc_data = json.load(f)

        # 确保config.json有quantization_config字段
        if "quantization_config" not in config_data:
            config_data["quantization_config"] = {}

        # 合并配置
        config_data["quantization_config"].update(quant_desc_data)

        # 确保有必要的字段
        if "moe_quantize" not in config_data:
            config_data["moe_quantize"] = "w8a8_dynamic"

        # 保存新配置文件
        with open(OUTPUT_FILE, 'w', encoding='utf-8') as f:
            json.dump(config_data, f, indent=4)

        print(f"成功: 配置已合并并保存到 {OUTPUT_FILE}")
        return True

    except Exception as e:
        print(f"错误: 处理配置文件时发生异常: {str(e)}")
        return False

def main(): 
    # 执行合并
    success = merge_configs()
    if not success:
        sys.exit(1)

if __name__ == "__main__":
    main()

"ascend_fusion_config": {"fusion_ops_gmmswigluquant": false}

mv /usr/local/Ascend/cann-8.5.0/opp/built-in/op_impl/ai_core/tbe/op_tiling/lib/linux/aarch64/libopmaster_ct.so /usr/local/Ascend/cann-8.5.0/opp/built-in/op_impl/ai_core/tbe/op_tiling/lib/linux/aarch64/libopmaster_ct.so_bak

mv /usr/local/Ascend/cann-8.5.0/opp/built-in/op_impl/ai_core/tbe/op_tiling/lib/linux/aarch64/liboptiling.so /usr/local/Ascend/cann-8.5.0/opp/built-in/op_impl/ai_core/tbe/op_tiling/lib/linux/aarch64/liboptiling.so_bak

mkdir -p /usr/local/Ascend/cann-8.5.0/opp/built-in/op_impl/ai_core/tbe/kernel/ascend910b/ops_transformer/fused_infer_attention_score_bak
mv /usr/local/Ascend/cann-8.5.0/opp/built-in/op_impl/ai_core/tbe/kernel/ascend910b/ops_transformer/fused_infer_attention_score /usr/local/Ascend/cann-8.5.0/opp/built-in/op_impl/ai_core/tbe/kernel/ascend910b/ops_transformer/fused_infer_attention_score_bak

cp /mnt/project/g9/glm4.7/FIA/op_tiling_aarch64/libopmaster_ct.so /usr/local/Ascend/cann-8.5.0/opp/built-in/op_impl/ai_core/tbe/op_tiling/lib/linux/aarch64/
cp /mnt/project/g9/glm4.7/FIA/op_tiling_aarch64/libopmaster_rt2.0.so /usr/local/Ascend/cann-8.5.0/opp/built-in/op_impl/ai_core/tbe/op_tiling/lib/linux/aarch64/
cp /mnt/project/g9/glm4.7/FIA/op_tiling_aarch64/liboptiling.so /usr/local/Ascend/cann-8.5.0/opp/built-in/op_impl/ai_core/tbe/op_tiling/lib/linux/aarch64/

cp -r /mnt/project/g9/glm4.7/FIA/910B_opp_kernel_aarch64/fused_infer_attention_score /usr/local/Ascend/cann-8.5.0/opp/built-in/op_impl/ai_core/tbe/kernel/ascend910b/ops_transformer/

ll /usr/local/Ascend/cann-8.5.0/opp/built-in/op_impl/ai_core/tbe/kernel/ascend910b/ops_transformer/fused_infer_attention_score/

export VLLM_ASCEND_BALANCE_SCHEDULING=1 
注意vllm版本对齐： vllm_ascend\patch\platform\__init__.py和脚本设置的vllm版本

export VLLM_ASCEND_ENABLE_FUSED_MC2=1

"multistream_overlap_shared_expert": true
"enable_shared_expert_dp": true

export VLLM_ASCEND_ENABLE_FLASHCOMM1=1

export HCCL_BUFFSIZE=512
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export VLLM_USE_V1=1
export HCCL_OP_EXPANSION_MODE=AIV
export VLLM_RPC_TIMEOUT=3600000
export VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=30000

export VLLM_ASCEND_BALANCE_SCHEDULING=1

export VLLM_VERSION=0.14.1
export LD_PRELOAD=/usr/lib64/libjemalloc.so.2:$LD_PRELOAD

export VLLM_TORCH_PROFILER_WITH_STACK=0

export VLLM_ASCEND_ENABLE_TOPK_OPTIMIZE=1
export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
export VLLM_ASCEND_ENABLE_FUSED_MC2=2 # 1: dispatchFFNcombine;  2: dispatchgmmcombinedecode
rm -rf /disk1/lcb/profiling/glm_4.7_in4k_out_10_bs_16_no_stack
### 注意，下面的环境变量是在拉起服务端设置，在发送请求的地方设置不会生效。
export VLLM_TORCH_PROFILER_DIR="/disk1/lcb/profiling/glm_4.7_in4k_out_10_bs_16_no_stack" 

rm -rf GLM4.7.log
nohup vllm serve  /disk1/models/GLM-4.7-W8A8 \
  --data-parallel-size 2 \
  --tensor-parallel-size 8 \
  --enable-expert-parallel \
  --seed 1024 \
  --served-model-name dsv3 \
  --max-model-len 140000 \
  --no-enable-prefix-caching \
  --max-num-batched-tokens 8192 \
  --max-num-seqs 16 \
  --async-scheduling \
  --quantization ascend \
  --trust-remote-code \
  --gpu-memory-utilization 0.9 \
  --speculative-config '{"num_speculative_tokens": 3, "model":"/disk1/models/GLM-4.7-W8A8", "method":"mtp"}' \
  --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
  --additional-config '{"enable_shared_expert_dp": true, "ascend_compilation_config": {"fuse_qknorm": true}, "ascend_fusion_config": {"fusion_ops_gmmswigluquant": false}, "multistream_overlap_shared_expert":"true"}' \
>>GLM4.7.log &

任务安排：
    mooncake kv cache性能验证，你需要先如下几组性能数据：
    1. 基线性能，即不添加mooncake kv cache的性能
    2. 基线性能 + 添加 mooncake kv cache 的性能
	4. 通过修改服务启动命令中的TP个数，分为TP2、TP4、TP8、TP16，需要和ASCEND_RT_VISIBLE_DEVICES对应，比如TP2，则ASCEND_RT_VISIBLE_DEVICES=0,1，TP4则ASCEND_RT_VISIBLE_DEVICES=0,1,2,3
	5. 一共2*4组数据，给出最终的数据结果。
注意点：
    1. 同时需要保存服务日志和压测日志，方便留存。
    2. 日志保存地址：放在本路径下logs文件夹即可，没有的话则新建，若文件夹不存在则需要新创建。
    3. 只看第一遍的性能数据，给出分析文档MD，记录保存的性能测试、服务启动命令、压测命令、mooncake接受率，用作归档，并附上压测结论。
	
启动服务：
1. 容器：qwen3.5_kv_cache
不存在的话就先创建容器：

2. 基线初始启动命令：
确保export PYTHONPATH=/vllm-workspace/vllm:/vllm-workspace/vllm-ascend:${PYTHONPATH}是正确的安装地址。
3. 代码路径：/vllm-workspace/

4. 服务启动命令：
pkill -9 python
pkill -9 VLLM
export TASK_QUEUE_ENABLE=1
export VLLM_USE_V1=1
export ASCEND_RT_VISIBLE_DEVICES=9,10,11,12
# AIV
#export PYTHONPATH=/vllm-workspace/vllm:/disk1/lcb/vllm-ascend/vllm_ascend_main_1227_30B/vllm-ascend:${PYTHONPATH}
export HCCL_OP_EXPANSION_MODE="AIV"
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export LD_PRELOAD=/disk1/lcb/libjemalloc.so:$LD_PRELOAD

export OMP_NUM_THREADS=1
export VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE=1
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl kernel.sched_migration_cost_ns=50000
export VLLM_ASCEND_ENABLE_NZ=2
vllm serve /mnt/weight/Qwen3-30B-A3B_w8a8 \
 --served-model-name qwen3 \
 --dtype bfloat16 \
 --max_model_len 23000 \
 --quantization ascend \
 --max-num-batched-tokens 40960 \
 --tensor-parallel-size 4 \
 --enable-auto-tool-choice \
 --tool-call-parser hermes \
 --no-enable_expert_parallel \
 --gpu-memory-utilization 0.90 \
 --profiler-config '{"profiler": "torch", "torch_profiler_dir": "/data/lcb/profiling/qwen3-30B","torch_profiler_with_stack": false}' \
 --no-enable-prefix-caching \
 --host 0.0.0.0 \
 --port 8889 \
 --block-size 128 \
 --async-scheduling \
 --distributed_executor_backend "mp" \
 --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY","cudagraph_capture_sizes":[4,8,16,32,64,96,128]}'

3.  拉起mooncake kv cache服务端:
3.1 先新建mooncake服务端文件mooncake.json，可以放在/data1路径下：
    local_hostname需要修改成本机host的IP。
    master_server_address需要修改成本机host的IP+新建一个端口号，你可以设置成50088。
    {
    "local_hostname": "xxx",
    "metadata_server": "P2PHANDSHAKE",
    "protocol": "ascend",
    "device_name": "",
    "use_ascend_direct": true,
    "alloc_in_same_node": true,
    "master_server_address": "xxx:50088",
    "global_segment_size": 103079215104
    }
3.2 启动mooncake kv cache 服务端：
    mooncake_master --port 50088
3.3 修改原始配置文件：
3.3.1 添加环境变量：
#指定mooncake配置文件
export LD_LIBRARY_PATH=/usr/local/lib/:/usr/local/Ascend/ascend-toolkit/latest/python/site-packages/mooncake/:$LD_LIBRARY_PATH
export MOONCAKE_CONFIG_PATH="xxx/mooncake.json"
export ACL_OP_INIT_MODE=1
export ASCEND_BUFFER_POOL=4:8
export ASCEND_CONNECT_TIMEOUT=10000
export ASCEND_TRANSFER_TIMEOUT=10000
3.3.2 启动命令里面添加：
--kv-transfer-config \
  '{
      "kv_connector": "AscendStoreConnector",
      "kv_role": "kv_both",
      "kv_connector_extra_config": {
        "backend": "mooncake",
        "kvpool_rpc_port":"0"
      }
  }'

4.  压测工具，新创建一个test容器，用刚刚的创建容器命令：
     python /xxx/vllm/benchmarks/multi_turn/benchmark_serving_multi_turn.py --url http://127.0.0.1:10113 --model xxx --served-model-name xxx --seed 1234 --input-file xxx/vllm/benchmarks/multi_turn/generate_multi_turn.json  --num-clients 16 --max-active-conversations 64 --no-early-stop
    参考服务启动命令，帮我修改--model、--served-model-name、--url、--input-file xxx/vllm/benchmarks/multi_turn/generate_multi_turn.json，若保持一致则无需修改
    数据集：wget https://www.gutenberg.org/ebooks/1184.txt.utf-8  mv 1184.txt.utf-8 pg1184.txt

    代理配置：
5. 修改点： 
    （1） generate_multi_turn.json和/xxx/vllm/benchmarks/multi_turn/benchmark_serving_multi_turn.py在vllm的安装路径下面，其中的num_conversations要和max-active-conversations保持一致。
	（2） num-clients需要配置为TP的16倍，max-active-conversations为num-clients的4倍，比如TP2时，num-clients为32. max-active-conversations为128.


diff --git a/vllm_ascend/patch/platform/patch_kv_cache_utils.py b/vllm_ascend/patch/platform/patch_kv_cache_utils.py
index 777d29c0..6ad63b96 100644
--- a/vllm_ascend/patch/platform/patch_kv_cache_utils.py
+++ b/vllm_ascend/patch/platform/patch_kv_cache_utils.py
@@ -372,7 +372,14 @@ def _get_kv_cache_groups_uniform_page_size_with_multi_groups(
     # group_size = 22
     # TODO(lxs): generalize the logic for determining group size.
     # Now, we use num_hidden_layers // 2 as the group size for DSV4.
-    group_size = cdiv(len(kv_cache_spec_list), 2)
+    # MTP layers are extra dense draft layers; keep them on the non-compress
+    # side instead of letting them shrink the number of KV tensors and drop the
+    # tail main-model layer from allocation.
+    num_mtp_layers = sum(
+        1 for layer_name in kv_cache_spec_list
+        if ".mtp." in f".{layer_name}."
+    )
+    group_size = cdiv(len(kv_cache_spec_list) - num_mtp_layers, 2) + num_mtp_layers
     grouped_layers = []
     group_layer_specs = []
     for layer_spec, layers in same_type_layers.items():
@@ -636,6 +643,21 @@ def  get_kv_cache_config_from_groups_multispec(
             # full.0, sw.0, sw.1: share a Tensor with size=available_memory//2
             # full.1, sw.2: share another Tensor with size=available_memory//2
             group_size = max(len(group.layer_names) for group in kv_cache_groups)
+            unique_layer_names = {
+                layer_name
+                for group in kv_cache_groups
+                for layer_name in group.layer_names
+            }
+            num_mtp_layers = sum(
+                1 for layer_name in unique_layer_names
+                if ".mtp." in f".{layer_name}."
+            )
+            if num_mtp_layers:
+                group_size = max(
+                    group_size,
+                    cdiv(len(unique_layer_names) - num_mtp_layers, 2) +
+                    num_mtp_layers,
+                )
 
             page_size = get_uniform_page_size(
                 [group.kv_cache_spec for group in kv_cache_groups])
@@ -657,18 +679,13 @@ def  get_kv_cache_config_from_groups_multispec(
                     for layer_name in kv_cache_groups[j].layer_names:
                         if layer_name in allocate_complete_layers:
                             continue
-                        group_used = False
-                        for gid in layer_kv_cache_group_idx[layer_name]:
-                            if gid in used_group_idx_set:
-                                group_used = True
-                                break
-                            else:
-                                used_layer_kv_cache_group_idx[layer_name].add(gid)
-                        if group_used is True:
+                        group_idxs = layer_kv_cache_group_idx[layer_name]
+                        if any(gid in used_group_idx_set for gid in group_idxs):
                             continue
                         shared_by.append(layer_name)
-                        used_group_idx_set.extend(layer_kv_cache_group_idx[layer_name])
-                        if len(used_layer_kv_cache_group_idx[layer_name]) == len(layer_kv_cache_group_idx[layer_name]):
+                        used_group_idx_set.extend(group_idxs)
+                        used_layer_kv_cache_group_idx[layer_name].update(group_idxs)
+                        if len(used_layer_kv_cache_group_idx[layer_name]) == len(group_idxs):
                             allocate_complete_layers.append(layer_name)
                 kv_cache_tensors.append(
                     KVCacheTensor(size=page_size * num_blocks,
diff --git a/vllm_ascend/worker/model_runner_v1.py b/vllm_ascend/worker/model_runner_v1.py
index 8296dc47..dcf2ab40 100644
--- a/vllm_ascend/worker/model_runner_v1.py
+++ b/vllm_ascend/worker/model_runner_v1.py
@@ -2918,7 +2918,24 @@ class NPUModelRunner(GPUModelRunner):
                 if layer_name in self.runner_only_attn_layers:
                     continue
                 layer_names.add(layer_name)
-        assert layer_names == set(kv_cache_raw_tensors.keys()), "Some layers are not correctly initialized"
+        initialized_layer_names = set(kv_cache_raw_tensors.keys())
+        if layer_names != initialized_layer_names:
+            missing_layers = sorted(layer_names - initialized_layer_names)
+            unexpected_layers = sorted(initialized_layer_names - layer_names)
+            logger.error(
+                "KV cache initialization mismatch: expected=%d initialized=%d "
+                "missing=%s unexpected=%s kv_cache_groups=%d kv_cache_tensors=%d "
+                "tensor_shared_by_sample=%s runner_only_attn_layers=%s",
+                len(layer_names),
+                len(initialized_layer_names),
+                missing_layers[:64],
+                unexpected_layers[:64],
+                len(kv_cache_config.kv_cache_groups),
+                len(kv_cache_config.kv_cache_tensors),
+                [list(t.shared_by) for t in kv_cache_config.kv_cache_tensors[:8]],
+                sorted(self.runner_only_attn_layers)[:64],
+            )
+        assert layer_names == initialized_layer_names, "Some layers are not correctly initialized"
         return kv_cache_raw_tensors