Skip to content

[Bug]: PD 分离部署乱码 #2359

Open

Description

@renren123

Your environment

Ascend Atlas A3 * 2

🐛 Describe the bug

xLLM PD 分离双机验证矩阵(a3_1 master + a3_2 P + a3_3 D)

环境:xllm 291b79f | 驱动 26.1.1 | CANN 9.0.0 | 模型 GLM-5.2-W8A8(EP16/DP2/world16/block128)
固定参数:--npu_kernel_backend ATB|XLITE --etcd_addr=192.168.0.1:2379 --enable_disagg_pd=true --host=<本机IP> --transfer_listen_port=26000+i(P)/26100+i(D) --disagg_pd_port=7777+i(P)/7787+i(D) --master_node_addr=<rank0:18888(P)/18898(D)>

时间 后端 P graph D graph D util 结果 报错/现象
09-21 20:17 ATB+python true true 0.86 ⚠️ 链路通但输出错乱 PD KV transfer 后 decode 内容退化(Let let's "1...Alpha:1),transfer 注册正常(mooncake 177 buffers/78层),非 </think> 乱码
09-21 20:21 ATB+python true true 0.75 ❌ kv_cache_estimation.cpp:659 Check failed: cache_size_in_bytes()>0(公式:KV预算=空闲−(1−util)×total,0.75 时为负)
09-21 20:21 ATB+python true true 0.86 ❌ D 崩 llm_master.cpp:62 engine_->init failed;Init_model_async 部分成功部分 failed(同一线程先 succeed 后 failed)
09-21 20:27 ATB+python true true 0.86 ⚠️ 端到端通 master 28888 HTTP server started;curl 返回但内容错乱(1{1- for180[ 1sumat1、1ning answer a11)
09-21 20:34 ATB+python true false(改) 0.86 ❌ D 大面积崩 OOM NPU 8/9: 17.4GB已分配/355MB free,386MB分配失败;16 实例只活 1 个实体(50850)占 8 卡
09-21 21:5x ATB+python true false 0.86 ⏳ 待重试(D 空窗) —

A3 双机PD分离有合适的参数吗?上面验证矩阵中对应参数输出乱码

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions