v0.7.2 Release Tracker

[simon-mo]

We will make a new release as soon as these PRs are merged.

• [Bugfix][Kernel] Fix per-token/per-channel quantization for Hopper scaled mm #12696
• [VLM] Qwen2.5-VL #12604
• [VLM] Add MLA with pure RoPE support for deepseek-vl2 models #12729
• [Perf] Mem align KV caches for CUDA devices (MLA perf improvement) #12676
• [core][distributed] exact ray placement control #12732
• [Bugfix] Better FP8 supported defaults #12796

[robertgshaw2-redhat]

Great. Should probably be a 0.7.2 if we include Qwen2.5-VL

[tlrmchlsmth]

#12704 would be good to get in

[mgoin]

I would like to get this in #12710

[whybeyoung]

can you fix #12690 ?

[tlrmchlsmth]

I think we should get #12729 in as well. It fixes DeepSeek-VL2, which is broken on 0.7.1 (see #11578 (comment))

[simon-mo]

@whybeyoung V1 is not a release blocker yet. Thank you!

[Swipe4057]

Can this PR be included in the release? #12156

[Mullestafa]

The leftover blockers #12604 and #12732 seem to be merged already.

[mgoin]

One more blocker #12796

[SugerFlat]

When is it expected to be released, as we are in great need of this version?