Stories

Comment by GreenGames

Comment

Comment by GreenGames

We built a standalone C++/ggml speculative decoder for Qwen3.5-27B Q4_K_M with a DFlash block-diffusion draft.

207.6 tok/s peak (5.46x over AR); HE 10-prompt bench averages 129.5 tok/s at DDTree budget=22, single RTX 3090, 24 GB. 3.43x over autoregressive and 2.8x over the best public SGLang AWQ number.

TL;DR - Peak 207.6 tok/s DFlash vs 38.0 tok/s AR (5.46x). HE bench: 129.5 tok/s mean at DDTree budget=22. - 3.43x over autoregressive Q4_K_M baseline (37.78 tok/s). - 2.8x vs SGLang AWQ reference (46.6 tok/s) on the same RTX 3090. - 128K context fits on 24 GB. Q4_0 KV + rolling 4096-slot target feature buffer. 134.78 tok/s at ctx=131072. - Only ggml. Never link libllama. ~2000 LOC C++/CUDA in libdflash27b.a around ggml_gated_delta_net.

Why the experiment exists Qwen3.5-27B is a hybrid model: every 4th layer is full softmax attention, the rest (48 of 64) are Gated DeltaNet. SSM state cache alongside the KV cache. That combo doesn't have a good single-3090 decode path today: llama.cpp has the GGUF loader and ggml_gated_delta_net, but no DFlash speculative decoding. vLLM / SGLang ship z-lab's DFlash integration, but only on BF16 (54 GB, doesn't fit on 24 GB). AWQ target on SGLang runs plain AR at 46.6 tok/s but can't host a BF16 draft + DDTree state in 24 GB. z-lab's reference benchmarks run BF16 on B200, 54+ GB class. We wanted the fastest single-3090 decode on a 24 GB card. The answer: port only the graph glue to ggml, keep the existing DeltaNet kernel, run DFlash block-diffusion draft with a DDTree verifier, compress KV to Q4_0 for long context.

From autoregressive to DDTree Same 10-prompt HE bench, n_gen=256, Q4_K_M target, BF16 draft. AL = average accept length. DDTree paper reports +35-42% over chain DFlash on pure-attention Qwen3 variants. On our hybrid Q4_K_M/RTX 3090 combo we see +15% over chain. The gap comes from Q4 quantization flattening the draft softmax, partially patched with a chain pre-seed in build_ddtree. Draft-ceiling bound, not verify-memory bound: a bigger tree won't help, only a better draft will.

Key wins - f16 intermediate cache: half the bandwidth, +5% at the same tree budget. Bit-identical to AR at 40 tokens. - Persist-write kernel (ggml_gated_delta_net_tree_persist): skips a 9 ms ggml_cpy per step, +11%. - target_feat compaction after sibling accept: unlocked real tree rescue on 9/10 prompts. - extract_draft_topk reverse bug: sort_heap + cmp_greater already produces descending order; an extra std::reverse was sending the worst candidate to the tree root. One-line fix. - verify_logits_buf overflow: sized vocabq_len but DDTree reads vocab(budget+1) past budget 15. Silent memory corruption. One-line size fix.

128K context on 24 GB Flash-attention in ggml-cuda supports Q4_0 K+V natively, so KV compression is just ggml_cpy with the F32->Q4_0 quantizer on write. 8x over f16. Combined with a rolling 4096-slot target_feat ring, target_feat shrinks from 6.6 GB to 0.2 GB at 128K. Tradeoffs: Q4_0 KV costs ~3% quality on HE (AL 8.56 -> 8.33) at short context, dramatically better at long ones. Only thing that lets 128K fit on 24 GB.

Prefill Short prompts (<=2048 tok): PREFILL_UBATCH=16. Matches DFlash block size. Long prompts (>2048 tok): auto-bump to PREFILL_UBATCH=192. 13K prefill: 40.9 s -> 15.07 s (2.7x, ~913 tok/s).

What comes next - Daemon mode: keep the model resident, first-token latency 10 s -> ms. - Temperature / top-k sampling in verify. Currently greedy-only. - Q5_K_M / Q6_K: better quants should recover most of the ~30-point accept gap vs BF16. - Full llama.cpp integration: qwen35 arch, llama-speculative-dflash.cpp wiring. - Metal/Vulkan: not planned. CUDA only, anyone who wants Metal can fork.

As soon as Qwen3.6-27B comes out, we'll do the same for it. Repo in the first comment (open source, MIT).

Replies GreenGames · 2026-04-20
Open on HN
Loading the discussion…

Domain filters

Stories from these domains are hidden from every list. Subdomains match too: blocking substack.com also hides danluu.substack.com.

    New collection

    Delete this collection?

    About YAVCHN

    YAVCHN is a reader for Hacker News and Lobsters, with articles and discussions in separate windows or Classic pages.

    Created by Paul Parks and built with PUDL.

    YAVCHN source code on GitHub

    Privacy policy · Terms of use

    Help

    Keyboard

    j / k
    Move down and up the story list. The arrow keys scroll whatever has focus.
    Enter
    Read the marked story in the article reader.
    ]
    Read the next story in the same article-reader applet. Back returns to the previous story.
    p
    Pin or unpin the marked story, which keeps it in Pinned.
    n / N
    Move to the next or previous top-level comment in the window in front.
    c
    Collapse or expand that comment.
    f
    Hide or show the story list.
    Esc
    Close a menu or this help.
    Access key m
    Go to the menu bar. Most browsers take it with Alt on Windows and Linux, and Safari with Control and Option.
    ?
    Show this help.

    Windows

    Each story opens in a window holding its article above its discussion; drag the bar between them to share the room differently. A window can be moved by its title bar, resized from any edge, snapped to a half or a corner by dragging it there, maximised, or minimised to the bar at the foot of the page. Use Window > New reader window to open an empty reader, or Story > Open in new reader window to open another reader for the current article. Docked readers keep their articles when you select another story from the sidebar. Minimized readers can be restored and reused for their site. A window's Next story link reads on down the list in the same window.

    A link in a comment or an article to another Hacker News or Lobsters thread opens that thread in a window too. A link to a single HN comment opens the comment above its replies.

    While a story's window is in front, the Story and Discussion menus in the menu bar hold its commands: pinning, Next story, sorting, collapsing every thread, jumping to the first new comment. Each window also remembers where you were in its article and discussion, so a reload, or Back to a story that Next took you past, finds your place again. Closing a window forgets it.

    The whole arrangement lives in the address, so a bookmark or a shared link brings it back, and Back undoes the last change. Moving between Hacker News, Lobsters, their lists, Pinned and Find changes only the list, and leaves the windows open.

    The list

    The pin at the start of a row keeps the story in Pinned, and the cross at its end hides it. Pinned can be narrowed by words in the title, site or author, by source, and to the stories you haven't opened yet, and ordered by when you pinned them, by points or by comments; the filters are part of the address, so a filtered view can be bookmarked. Scroll past the end of the list to load more. Domain filters, in the View menu, hide every story from a site.

    Collections are named lists of stories. Story > Add to collection files the story in front into one or more of them, and the Collections feed shows them all or one at a time; the menu that chooses collections also creates, renames, and deletes them. A note is your own text on a story. Choose Add note in a story's toolbar to write one; it saves as you type. Rows with a note carry the note mark, and the Notes feed lists every noted story and searches the text of your notes.

    Browsing view

    View > Windowed and View > Classic select the browsing view and save your default in this browser. Window view reuses a reader for each feed. Classic view opens stories and applets as pages. Open as a page is a one-off action that does not change your saved default. Use the Windowed selector to return an article to a window. Direct page links always open as pages.

    Applets

    The Applets menu in the menu bar holds three tools, each a window of its own. Replies to me takes your Hacker News user name and lists the replies to your last thirty comments and stories, checking again every three minutes while it is open, and marking what is new since you last marked them read. Look up a user opens a profile on Hacker News or Lobsters, with their submissions and recent comments, as a commenter's name in any discussion does; the bar at the top of a profile looks up someone else in the same window, and Back returns to the one before. Who is hiring? filters the posts of HN's monthly hiring threads by the words you type.

    They read only what the sites publish to everyone, so none of them asks for a login, and your user name stays in this browser anonymously and is included in retained applet state when signed in.

    Find

    Find takes any link and lists every time it was submitted to Hacker News and Lobsters, so you can read each discussion of it.

    About

    YAVCHN never sees your Hacker News or Lobsters login. The discussion is fetched from each site's public API; to vote or reply, follow the link above the discussion, or the arrow beside a comment, to the source's own site. Without a YAVCHN account, your data stays in this browser. When signed in, pins, collections, notes, blocked domains, and retained reading state are stored with your account and synchronized across devices. Hidden stories and layout stay in this browser. The privacy policy has the details.

    Open source: github.com/paulmooreparks/yavchn. Built with PUDL.