Datasets:

---

Etched

/

forward_passes

like 0

Follow

Etched 11.8k

Deepseek R1 in FP8 with (PP, EP): (2,1), (1,2), (1,4), (1,8), (1,16)

Dense run of identical users for separating prefill from decode with:

• Prefill: Batch_tokens per user = X and KV-cache per user = X/2
• Decode: Batch_tokens per user = 1 and KV-cache per user = scaled in increments of 128 up to maximum per system

In addition to:

• 100k samples from log-normal distributions for each unique number of users on each system configuration (based on expert and pipeline parallelism): ~5.1M configurations total
• Num_users and maximum kv-cache scale with the total number of systems (EP * PP) to account for greater total capacity of the configuration