poolside

/

Laguna-S-2.1-NVFP4

like 88

Follow

Poolside 825

If you disable DFlash and use sglang image with tag dev-cu13 (released on 30.06.2026) - it works perfectly. My settings to try/run it on 1 B200 are:

sglang serve
--model-path /models/Laguna-S-2.1-NVFP4
--served-model-name Laguna-S-2.1
--host 0.0.0.0
--port 8000
--allow-auto-truncate
--attention-backend trtllm_mha
--mm-attention-backend fa4
--disable-overlap-schedule
--mem-fraction-static 0.8
--cuda-graph-max-bs-decode 32
--max-running-requests 32
--context-length 175000
--tool-call-parser poolside_v1
--reasoning-parser poolside_v1
--chunked-prefill-size 8192
--trust-remote-code
--enable-metrics

Looping on my DGX Spark

Hi guys! Thanks for the feedback.

We've just uploaded a new checkpoint at the RC1 revision poolside/Laguna-S-2.1-NVFP4@RC1. In our internal testing, it loops far less commonly in W4A4 & not at all in W4A16.

Because it loops in W4A4 still, we'll follow up with another checkpoint that'll be the official patch. Expect this in the next few days, at the latest.

Hi guys! Thanks for the feedback.

We've just uploaded a new checkpoint at the RC1 revision poolside/Laguna-S-2.1-NVFP4@RC1. In our internal testing, it loops far less commonly in W4A4 & not at all in W4A16.

Because it loops in W4A4 still, we'll follow up with another checkpoint that'll be the official patch. Expect this in the next few days, at the latest.

When will RC1 be pushed to main?

Hi guys! Thanks for the feedback.

We've just uploaded a new checkpoint at the RC1 revision poolside/Laguna-S-2.1-NVFP4@RC1. In our internal testing, it loops far less commonly in W4A4 & not at all in W4A16.

Because it loops in W4A4 still, we'll follow up with another checkpoint that'll be the official patch. Expect this in the next few days, at the latest.

When will RC1 be pushed to main?

In the next ~10 minutes or so.

It's now on main. As I've previously mentioned, we intend to provide another, final checkpoint in the next little while.

thanks for the fix. Should i disable Dflash to use this new main and wait for the new DFlash fix you guys are working on or is it OK to use this new main with the current poolside/Laguna-S-2.1-DFlash-NVFP4@main

It's now on main. As I've previously mentioned, we intend to provide another, final checkpoint in the next little while.

Thank you for the quick turnaround!

thanks for the fix. Should i disable Dflash to use this new main and wait for the new DFlash fix you guys are working on or is it OK to use this new main with the current poolside/Laguna-S-2.1-DFlash-NVFP4@main

You can use the Dflash that exists now, but the acceptance rates may be a little bit lower than you were getting before.

Hi guys! Thanks for the feedback.

We've just uploaded a new checkpoint at the RC1 revision poolside/Laguna-S-2.1-NVFP4@RC1. In our internal testing, it loops far less commonly in W4A4 & not at all in W4A16.

Because it loops in W4A4 still, we'll follow up with another checkpoint that'll be the official patch. Expect this in the next few days, at the latest.

When using vllm, does this mean I should add --quantization w4a16 when running on RTX 6000 Pro?