z-lab runs BF16 on B200 (54+ GB). There is no z-lab path that fits on a 24 GB 3090. That is literally the entire point of our work, and it is stated in the second paragraph. If you had checked the HF model card you linked before posting, you would see the same thing. Before this repo, there was no path to run this... SGLang's GGUF path for this model is broken. llama.cpp doesn't have DFlash speculative decoding at all. If you wanted to run this hybrid model fast on a 24 GB consumer card, there was nothing...
That took weeks of real engineering.
Calling that "vibecoded" because we used a bit of AI in the README is clean is the laziest possible critique. An LLM reading the DFlash paper does not catch verify_logits_buf being sized vocabq_len when DDTree reads vocab(budget+1). That is hours of debugging with nvidia-smi and memory sanitizers, not prompting.
The 207 and 129.5 numbers are both in the second sentence of the post and again in the TL;DR. 207.6 is peak tok/s in the linked demo video, 129.5 is the HumanEval 10-prompt mean at DDTree budget=22. We specify both just behind the title.
On the Q4 KV cache: the tradeoff is disclosed with actual numbers. AL 8.56 -> 8.33 at short context (3% drop), dramatically better at long context. It’s the only way 128K allocates on 24 GB. The binary is env-selectable, you can run BF16 KV if you don’t need 128K. Both are benchmarked.
I was discussing details I read in your repo. How did you conclude that I didn't read the post? I'm skeptical a human is writing these comments because everything you're posting reads like LLM output
> On the Q4 KV cache: the tradeoff is disclosed with actual numbers. AL 8.56 -> 8.33 at short context (3% drop), dramatically better at long context.
I'm sorry, but you're not the first (or LLM) to think of using Q4 KV cache to fit more context in VRAM.
The degradation is far more than 3% on real evals. Q8 only recently became usable on Qwen3.5 in llama.cpp with the context rotation changes. Before that bf16 was necessary to get decent performance in real tasks.
Q4 is a non-starter for real work. The fact that you're still trying to defend it tells me you haven't used this for anything other than token/sec racing.
You wrote this reply with Claude, and it's lying about it only being README.md. OP, and I, know this because you and Claude documented it.*
I use the same tools, I'm not mad at you for using it. It's just, idk man, you want to use it tactically in ways that are a net benefit to you. Not in ways that embarrass you or lie.
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).
It didn't really add anything to the conversation, and if I wanted to know what an LLM thought, I'd ask it. The reason for the rule is people come here to interact with other people.
The crazy thing is how effort posts went from the most valuable part of this site to the most hateful part of the site, by the very people claiming to be protecting the site
Super interesting take Paul. Curious btw, how are these teams actually encoding their “institutional knowledge” into constraints? Like is it some manual config or more like natural‑language rules that evolve with the codebase?
Some teams are using Claude or similar models in GitHub Actions, which automatically review PRs. The rules are basically natural language encoded in a YAML file that's committed in the codebase. Pretty lightweight to get started.
Other teams upgrade to dedicated tools like cubic. We have a feature where you can encode your rules either in our UI, or we're releasing a feature where you can write them directly in your codebase. We'll check them on every PR and leave comments when something violates a constraint.
The in-codebase approach is nice because the rules live next to the code they're protecting, so they evolve naturally as your system changes.
The "in-codebase" approach is the right one, but a YAML file with plain text is a half-measure. The most reliable rule that "lives next to the code" is an architectural test. An ArchUnit test verifying that "all routes in /billing/* call requireAuth" is also code, it's versioned with the project, and it breaks the build deterministically
That is a more robust engineering solution, unlike semantic text interpretation, which can fail
Hi there, Alessandro and Francesco here. We just launched an experimental feature in C/ua called App-Use. It lets you create virtual desktops scoped to specific apps (e.g., "Safari and Notes only") to give your agents focused, lightweight control without full-screen access.
Use cases:
- Run multiple agents in parallel with isolated app views
- Automate your iPhone using the iPhone Mirroring app
- Improve agent task precision and reduce VLM distractions
Works only on macOS (Sequoia+) and requires experiments=["app-use"]. No extra processes, just clever compositing.
Again, so? The grocery store on my block is beneficially owned by a Belgian and Dutch corporation which feels no compunction in union-busting. What stops a Swede from similarly parasitizing my local economy?
Between steps it would preserve cookies, but atm when the playwright browser launches, it starts with a fresh browser state, so you'd have to o-auth to log in each time.
We're adding browser state persistence soon, hoping to enable it so once you sign in with google once, it can stay signed in on your local machine.
I'm friends with a few of these founders so I'm refraining from recommending one service over another. That said, the feature that I love the most that a few of them share is the PR summarization. It helps humans do their PR review job particularly well!
z-lab runs BF16 on B200 (54+ GB). There is no z-lab path that fits on a 24 GB 3090. That is literally the entire point of our work, and it is stated in the second paragraph. If you had checked the HF model card you linked before posting, you would see the same thing. Before this repo, there was no path to run this... SGLang's GGUF path for this model is broken. llama.cpp doesn't have DFlash speculative decoding at all. If you wanted to run this hybrid model fast on a 24 GB consumer card, there was nothing...
That took weeks of real engineering.
Calling that "vibecoded" because we used a bit of AI in the README is clean is the laziest possible critique. An LLM reading the DFlash paper does not catch verify_logits_buf being sized vocabq_len when DDTree reads vocab(budget+1). That is hours of debugging with nvidia-smi and memory sanitizers, not prompting.
The 207 and 129.5 numbers are both in the second sentence of the post and again in the TL;DR. 207.6 is peak tok/s in the linked demo video, 129.5 is the HumanEval 10-prompt mean at DDTree budget=22. We specify both just behind the title.
On the Q4 KV cache: the tradeoff is disclosed with actual numbers. AL 8.56 -> 8.33 at short context (3% drop), dramatically better at long context. It’s the only way 128K allocates on 24 GB. The binary is env-selectable, you can run BF16 KV if you don’t need 128K. Both are benchmarked.