Repository navigation
blog: understanding PR #6516 — SenseNova-U1.5-8B-MoT and its distilled 8-step LoRA - #52
Open
hsliuustc0106 wants to merge 2 commits into
Open
hsliuustc0106 wants to merge 2 commits into
hsliuustc0106 wants to merge 2 commits into
Conversation
added 2 commits
September 15, 2026 17:38
4 tasks
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
New PR-analysis post (EN + ZH), first
feature: lorapost in-tree, with 3 new SVG figures and its xiaohongshu note set.Upstream PR: vllm-project/vllm-omni#6516 — Support SenseNova-U1.5-8B-MoT and its distilled 8-step LoRA (merged 2026-09-01, first in v0.29.0rc1; closes #6471).
The story the post tells: a "model support" PR whose model already ran — the real work was (1) the official 8-step distilled LoRA that loaded successfully and silently changed nothing (a
.qkv_proj⊂.qkv_proj_mot_gensubstring bug in_prepare_lora_deltaproducing a double-height delta), and (2) the review escalation into an Nsight-profiled rebuild of the AR think loop: paged KV cache (seqused_k+block_table) + CUDA graphs, think decode −78%, GPU busy 97.7%, kernel launches −67.9%, plus the kept/dropped experiment table and the H200 validation.Included:
blog/_posts/2026-09-15-understanding-pr-6516-sensenova-u15-distilled-lora.md+.zh.md(paired, plain-first, three reader tiers)blog/xhs/note set (6 cards + note.txt; title 15/20, body 693/1000)Verification:
jekyll buildclean (both URLs render, figures resolve);ruby scripts/check-blog-summaries.rb→ 23 posts OK.All numbers cite the PR review thread with comment permalinks (author on 1× A800, reviewer on L20X/H200); the post states explicitly that the cookbook has not yet run its own SenseNova traces.