Qwen3.8-27B DFlash2 on 2× DGX Spark (GB10) — 135 tok/s single-stream

Mirrors: canonical repo is on GitHub at fattchris/qwen-dflash2-spark (identical content).

One-stop, copy-paste reproducible stack: run Qwen3.8-27B NVFP4 with DFlash2 block speculative decoding on two NVIDIA DGX Spark (GB10) nodes over RoCE, hitting 124–135 tok/s single-stream (C1) — a +43% speedup over DSpark (MTP-style) speculative decoding at equal quality, verified end-to-end.

Everything needed is in this repo: build the vLLM image, fix the checkpoints, launch the 2-node pair, and verify. No external state assumed beyond two DGX Sparks with Docker + SSH, internet to pull the base image and checkpoints.

TL;DR quickstart

git clone https://github.com/fattchris/qwen-dflash2-spark.git
cd qwen-dflash2-spark
# on BOTH nodes:                     ~30 min (bandwidth-bound)
bash scripts/00-download-models.sh
# on node 1 only:                    ~1–2 h (compiles vLLM fork)
bash scripts/01-build-image.sh
# on node 1 only:                    ~10 min
bash scripts/02-make-bf16head.sh
# from node 1:                       ~5 min  (rsync ~23 GB)
bash scripts/03-replicate.sh
# from anywhere with SSH to both:    ~4–5 min to health 200
bash scripts/04-launch-pair.sh
# verify generation + throughput:
bash scripts/06-verify.sh && bash scripts/07-quality-gate.sh

Result: OpenAI-compatible endpoint at http://NODE1:8004/v1, model name qwen3.8-27b-dflash2, 131–135 tok/s typical on the edit-heavy C1 benchmark.

What this is

Target model: unsloth/Qwen3.8-27B-NVFP4 (FP4 weights, FP8 attention) — the bandwidth-floor baseline on GB10.
Draft model: z-lab/Qwen3.8-27B-DFlash2 (1.92B params, block_size 8, local convolution + candidate selector head).
Serving stack: vLLM from PR vllm-project/vllm#52816 (z-lab/vllm-fork@19c9351, branch subsir/upstream-dflash2), compiled aarch64 in the official vllm/vllm-openai:0.27.1-aarch64 image, TP=2 across two Sparks via RoCE v2.
The one real fix that makes the combo work: dequantize lm_head to BF16 in a copy of the NVFP4 checkpoint. Every public NVFP4 Qwen3.8 checkpoint quantizes lm_head, but DFlash2's candidate selector performs TopK over the target model's lm_head logits and hard-rejects quantized heads — so nothing out of the box works. 02-make-bf16head.sh does the surgery + config fix atomically and hardlink-safely.

Results (GB10 ×2, TP=2, NVFP4 target + 1.92B draft)

Config	tok/s (C1 edit-heavy)	accept	mean tok/pass
DSpark (MTP) K=14	91–95	68.6%	~9.7
DFlash2 K=7	73–74	99.6%	7.0
DFlash2 K=16	124–135	93.1%	15.9
DFlash2 K=24	121–129	66.4%	16.9

K sweep on GB10: sweet spot is K=16 (2 draft blocks). K=24 drops acceptance (66%) and net throughput. K must be a multiple of block_size 8.
×5 repeat variance: 131.5–134.8 tok/s, accept 93.1% every run.
Quality: 6/6 greedy correct vs baseline (math/Canberra/syllogism/bugfix/code) — BF16 lm_head dequant introduces zero measurable drift.
Aggregate/concurrent throughput scales differently — C1 numbers are single-stream only, and only comparable to other C1 numbers.

Repo layout

scripts/00-download-models.sh   # hf download target + draft
scripts/01-build-image.sh       # compile z-lab vllm-fork inside official image
scripts/Dockerfile.df2          # overlay built vllm/ onto clean official image
scripts/02-make-bf16head.sh + make_bf16head.py   # the lm_head surgery
scripts/03-replicate.sh         # rsync checkpoints node1 → node2
scripts/04-launch-pair.sh       # TP2 launch, health-wait
scripts/06-verify.sh            # health + generation + bench
scripts/07-quality-gate.sh      # greedy math gate (17*23+45 == 436)
scripts/qwen-dflash2-rank.sh    # the actual docker run (single rank)
bench/edit_bench.py             # C1 edit-heavy/fresh benchmark (from engine counters)
bench/quality_capture.py        # 6-prompt greedy capture
bench/prod_tps.py               # simple client-side tps probe
docs/architecture.md            # how the pieces fit, RoCE env details
docs/README-agent.md            # runbook for AI agents operating this stack
docs/results.md                 # full benchmark tables + methodology

Step-by-step (with what each does)

0. Prereqs

Two DGX Spark nodes with: Docker, passwordless SSH from your workstation and node1→node2, internet. Defaults assume admin@192.168.1.205 (node1/rank0) and admin@192.168.1.206 (node2/rank1); override with NODE1= NODE2= env vars. Note the image only exists for aarch64 — this stack is GB10-specific.

1. Download models (~30 min, both nodes)

bash scripts/00-download-models.sh

Pulls unsloth/Qwen3.8-27B-NVFP4 (~~21 GB) and z-lab/Qwen3.8-27B-DFlash2 (~~4 GB). Idempotent — safe to re-run.

2. Build image (~1–2 h, node1)

bash scripts/01-build-image.sh

Clones the z-lab vllm fork at the pinned PR commit, builds it in-place inside the pulled official image (compiler + CUDA deps already present), tars the built vllm/ tree, and overlays it onto a fresh official image via Dockerfile.df2. Tag: local/vllm-dflash2-pr52816:v2. Then docker save it to node2 — see below.

Wait — cross-node image: 01 builds on node1 only. Move it to node2:

docker save local/vllm-dflash2-pr52816:v2 | gzip > /tmp/df2.tgz
scp /tmp/df2.tgz node2:/tmp/ && ssh node2 'gunzip -c /tmp/df2.tgz | docker load'

~22.5 GB; over a 10 GbE LAN ~30 min, over Tailscale slower.

3. lm_head surgery (~10 min, node1)

bash scripts/02-make-bf16head.sh

Creates unsloth-nvfp4-bf16head/ — a hardlink copy of the NVFP4 checkpoint with lm_head.weight dequantized (w * scale → bf16, vocab 151k × 4.1k hidden) and the config fixed (see Pitfalls #2). The original checkpoint is left pristine (verify with fix_configs.py's proof mode).

4. Replicate to node2 (~5 min)

bash scripts/03-replicate.sh

rsyncs the bf16head checkpoint (uses rsync hardlink semantics where possible)

draft dir to node2.

4b. Container image to node2 — don't skip (see step 2 note).

5. Launch pair (~4–5 min to health)

bash scripts/04-launch-pair.sh

Installs the rank script (with your K) on both nodes, removes stale containers, starts rank1 (headless) then rank0 (master/API), and health-polls up to 7.5 min.

6–7. Verify

bash scripts/06-verify.sh
bash scripts/07-quality-gate.sh

06 checks health, a real generation ("DFLASH2 OK"), and a single bench run. 07 is a greedy math gate: 17*23+45 must answer 436.

Pitfalls (learned the hard way — read before debugging)

lm_head must be BF16 — DFlash2's selector TopKs over target logits; the image asserts UnquantizedEmbeddingMethod. Dequant via 02. (This is the headline fix; every public NVFP4 Qwen3.8 checkpoint ships a quantized lm_head.)
Quant-config precedence: config_groups.group_0.targets overrides the ignore list. Your ignore: ["lm_head"] does nothing while re:.*lm_head sits in targets. 02 strips it from targets AND adds it to ignore. (Cost me three failed boots to find.)
Hardlink bleed: shutil.copytree(copy_function=os.link) keeps the original and the copy sharing inodes — a later json.dump into config.json rewrites the shared inode and corrupts the original checkpoint. Always write-temp-then-os.replace() configs (02 does).
Copy path exists: 02 runs copytree into an existing dir → symlink loops. It removes .safetensors in DST first; if DST is partial, delete it and re-run. 4b. Served model name is exactly qwen3.8-27b-dflash2 (dot-separated, set by --served-model-name in the rank script). Clients must match it.
NCCL/RoCE env: exact working set is in qwen-dflash2-rank.sh (NCCL_IB_HCA=rocep1s0f1, RoCE v2, NCCL_CUMEM_ENABLE=0, VLLM_MARLIN_USE_ATOMIC_ADD==1 etc.). Don't hand-tune without capturing a baseline first.
Health timing: boot → 200 takes ~4–5 min. My first 650 s health-poll loop assumed faster; don't panic at 200 s.
TP2 MPI-rank pairing: rank1 must be up before rank0 starts serving (rank1 joins via --headless). 04 launches r1 first for this reason. If you restart rank0 only, expect a stuck handshake → restart both, r1 first.
Stale background jobs: if you automate around this stack, ensure old launcher jobs are dead before relaunching — a zombie once docker rm -f'd a healthy pair and relaunched the wrong checkpoint.
launchd/kickstart gotcha (Mac LiteLLM gateways): launchctl kickstart -k recycles only the worker; the master keeps the old config. Kill the master PID and let launchd respawn for config changes to take.
K must be multiple of 8 (draft block_size). K=16 peak on GB10.

Credit & lineage

DFlash2: z-lab (draft weights + vLLM fork). Blog: inco.ai/blog/dflash2.
NVFP4 target: unsloth/Qwen3.8-27B-NVFP4.
vLLM PR 52816 by SubSir (z-lab), open as of 2026-08-19.
This repo = integration engineering + the lm_head BF16 fix + GB10 K-sweep + 2-node RoCE recipe. MIT licensed. Benchmarks by Chris Fontes.

Downloads last month: -; Downloads are not tracked for this model. How to track