Upload merged_deepseek_R1_all_configs_100000_samples.json.gz (1.35 GB)
#6
by
mfein
- opened
No description provided.
This compressed data set contains the following:
- 100k samples for each unique number of users on each system configuration (based on expert and pipeline parallelism): ~5.1M configurations total
- EP1 and run with PP2 since Deepseek will not properly fit onto 1 single system with 671B in FP8
- EP 2, 4, 8, and 16 all run with PP1, which is optimal and strictly better than running with PP>1 for an equivalent number of systems
- Num_users and maximum kv-cache scale with the total number of systems (EP * PP) to account for greater total capacity of the configuration
- Batch tokens per user and kv-cache per user are sampled from log-normal distributions
mfein
changed pull request status to
merged