aliyun / aliyun/SimAI

仿真异常结束,weight grad事件无法触发

Open
#101 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
1.2k
Forks
184
Avg merge
13h 4m
Merged PRs (30d)
1

Description

请教下各位老师,是否遇到过仿真异常结束的问题:
1、修改: astra-sim-alibabacloud/inputs/config/SimAI.conf中CC_MODE 3
2、正常仿真:
命令:
AS_SEND_LAT=3 AS_NVLS_ENABLE=1 ./bin/SimAI_simulator -t 80 -w ./aicb/results/workload/None-gpt_7B-world_size8-tp4-pp1-ep1-gbs2-mbs1-seq2048-MOE-False-GEMM-False-flash_attn-True.txt -n ./HPN_7_0_8_gpus_8_in_one_server_with_single_plane_100Gbps_A100 -c astra-sim-alibabacloud/inputs/config/SimAI.conf
输出:

***** info: input gradient collective for layer: layernorm is finished************
DEBUG - Processing event: 16 in layer: layernorm
DEBUG - Processing event: 16 in layer: layernorm
DEBUG - Processing event: 16 in layer: layernorm
DEBUG - Processing event: 16 in layer: layernorm
DEBUG - Processing event: 16 in layer: layernorm
DEBUG - Processing event: 16 in layer: layernorm
info: no weight grad collective for layer: layernorm
info: no input grad collective for layer: grad_param_compute
info: no weight grad collective for layer: grad_param_compute
info: no input grad collective for layer: grad_param_comm
chunk size is: 6488588288 , size is: 6488588288 , layer_num is: 1 , node: 0
info: reduce-scatter weight grad collective issued for layer: grad_param_comm, involved dimensions:  0, 1, 1, 1, 1, 1, 1, 1, 1, 1,
info: no input grad collective for layer: grad_gather
Layer issue wg all gather at tick: 540210436
chunk size is: 3244294144 , size is: 3244294144 , layer_num is: 0 , node: 0
info: all-gather weight grad collective issued for layer: grad_gather, involved dimensions:  0, 1, 1, 1, 1, 1, 1, 1, 1, 1,
pass: 0 finished at time: 540210436
DEBUG - Processing event: 12 in layer: grad_param_comm
DEBUG - Processing event: 12 in layer: grad_param_comm
DEBUG - Processing event: 12 in layer: grad_param_comm
DEBUG - Processing event: 12 in layer: grad_param_comm
DEBUG - Processing event: 12 in layer: grad_param_comm
DEBUG - Processing event: 12 in layer: grad_param_comm
DEBUG - Processing event: 12 in layer: grad_param_comm
DEBUG - Processing event: 12 in layer: grad_param_comm
DEBUG - Processing event: 15 in layer: grad_param_comm
DEBUG - PHY_MTP defined: NO
DEBUG - Current generator ID: 7 (0 is required for print)
DEBUG - Processing event: 15 in layer: grad_param_comm

3、异常仿真:
命令:
AS_LOG_LEVEL=INFO AS_SEND_LAT=3 AS_NVLS_ENABLE=1 ./bin/SimAI_simulator -t 80 -w ./aicb/results/workload/None-gpt_7B-world_size1024-tp4-pp1-ep1-gbs256-mbs1-seq4096-MOE-False-GEMM-False-flash_attn-True.txt -n ./HPN_7_0_1024_gpus_8_in_one_server_with_single_plane_100Gbps_A100 -c astra-sim-alibabacloud/inputs/config/SimAI.conf
输出:

DEBUG - Processing event: 16 in layer: layernorm
DEBUG - Processing event: 16 in layer: layernorm
info: no weight grad collective for layer: layernorm
DEBUG - Processing event: 16 in layer: layernorm
DEBUG - Processing event: 16 in layer: layernorm
DEBUG - Processing event: 16 in layer: layernorm
DEBUG - Processing event: 16 in layer: layernorm
info: no input grad collective for layer: grad_param_compute
info: no weight grad collective for layer: grad_param_compute
info: no input grad collective for layer: grad_param_comm
chunk size is: 6522142720 , size is: 6522142720 , layer_num is: 1 , node: 0
info: reduce-scatter weight grad collective issued for layer: grad_param_comm, involved dimensions:  0, 1, 1, 1, 1, 1, 1, 1, 1, 1,
info: no input grad collective for layer: grad_gather
Layer issue wg all gather at tick: 546518210
chunk size is: 3261071360 , size is: 3261071360 , layer_num is: 0 , node: 0
info: all-gather weight grad collective issued for layer: grad_gather, involved dimensions:  0, 1, 1, 1, 1, 1, 1, 1, 1, 1,
pass: 0 finished at time: 546518210
root@05-06-78-999:/data/SimAI# 

4、根据打印分析,千卡7B CC_MODE 10时候无法触发weight grad事件
5、另外不修改astra-sim-alibabacloud/inputs/config/SimAI.conf配置,在千卡70B大模型仿真也遇到同样无法触发weight grad事件的问题

期待各位老师指点,万分感谢!

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start with astra-sim-alibabacloud/inputs/config/SimAI.conf and reproduce the two commands using the referenced workload and network files. Compare the normal and abnormal logs around the weight-grad event, including CC_MODE and the 70B case; done means identifying why the event is not triggered and documenting a reproducible resolution.

Written by the indexing model from the issue text.

Assessment

Domain
tooling
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.