DeepSeek-V4-Flash-DSpark-GB10-2x-DGX-Spark-1m-fp4-fp8 - GitLite

KO | EN

gitlite — search

#ui #animation #reactjs #gemini #java #openai #github #hacktoberfest #privacy #go

DeepSeek-V4-Flash-DSpark-GB10-2x-DGX-Spark-1m-fp4-fp8

(★ 9)

DeepSeek-V4-Flash DSpark speculative decoding on 2x DGX Spark (GB10/sm_121), 1M context — our fp8 recipe + independent reproduction and cross-build benchmarks of the NVFP4-KV build. Honest, apples-to-apples measurements.

파일 탐색기

ZIP 다운로드

- - bench_concurrent.py
  - correctness_test.py
  - staggered_bench.py
- - depth-curve-16k-512k.txt
  - depth-point-1M-run.txt
  - depth-point-1M-summary.txt
  - depth-point-1M.json
- 20260630-fp8-dspark-checkpoint.md
- 20260630-nvfp4-1m-context-curve-checkpoint.md
- 20260630-nvfp4-c16-reproduction-checkpoint.md
- 20260701-sampling-coherence-probe.md
- bench_vllm.py
- ARCHITECTURE.md
- GB10-PERFORMANCE.md
- LONG-CONTEXT-FINDINGS.md
- recipe.md
- - test_dspark_weight_routing.py
- PATCHES.md
.env.dspark.example
.gitignore
apply-patches.sh
CREDITS.md
launch-fp8-node.sh
LICENSE
NOTICE
prepare-model-cache.sh
README.md
run-fp8-node.sh
run-nvfp4-node.sh
smoke.sh

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

Running DeepSeek's **DSpark** speculative decoding for `DeepSeek-V4-Flash-DSpark` on a 2× DGX
Spark cluster (GB10, consumer Grace-Blackwell, `sm_121`), TP=2 over 200G RoCEv2, at 1M-token
context — plus an **honest, cross-build benchmark suite**. This repo carries our own fp8 DSpark
GB10 bring-up (patches + launcher), and it also documents our independent **reproduction and
measurement** of the community NVFP4-KV build. We author the fp8 recipe; we do **not** author the
NVFP4 recipe — that one is tonyd2wild's and is referenced by link, not vendored here. Our unique
contribution is the **extended benchmark data**: the 1M-context depth curve, the concurrency
scaling, and a long-context coherence / sampling study.

No marketing. All throughput is apples-to-apples (`completion_tokens / wall_time`, non-streaming),
and every gap to datacenter numbers is explained rather than hidden.

---

## Two recipes on GB10

### 1. fp8 DSpark — **ours**

Our GB10 / `sm_121` bring-up of an fp8 DSpark KV path (`fp8_ds_mla`): the DSpark draft head and
proposer wired into vLLM's spec-decode loop, plus the numerical bug-fixes that took the draft from
"predicts noise" to "matches the reference." Implemented as vLLM **patches** (a new `dspark`
speculative method) — it does **not** fork the base model, and the running prod image is never
touched.

- Apply the patches: [`apply-patches.sh`](apply-patches.sh) → everything under
  [`patches/`](patches/) (the `DSparkV4MTP` draft head, the `DSparkProposer`, the
  spec-config wiring, the EAGLE3 aux plumbing).
- Launch (fp8, TP=2 over RoCE, 1M context): [`run-fp8-node.sh`](run-fp8-node.sh)
  (production, systemd-managed, reboot-proof, RoCE-GID re-alignment) and
  [`launch-fp8-node.sh`](launch-fp8-node.sh) (load-test variant).
- Full recipe, anchors, and the three bug-fixes: [`docs/recipe.md`](docs/recipe.md),
  [`patches/PATCHES.md`](patches/PATCHES.md), [`docs/ARCHITECTURE.md`](docs/ARCHITECTURE.md).

### 2. NVFP4-KV DSpark — **tonyd2wild's** (we reproduced + measured, did not author)

The faster single-stream / higher-acceptance line uses NVFP4 KV. **That recipe is not ours.** It
is [tonyd2wild's NVFP4-KV recipe](https://github.com/tonyd2wild/DeepSeek-v4-Flash-DSpark-1M-NVFP4-KV-2x-DGX-Spark)
(commit `89bb82b`). We cloned it, reproduced it byte-identically on our own two-node cluster, and measured
what nobody had published. To **build** it, go to the upstream repo. In this repo the NVFP4 line is
**referenced by link, not vendored**; what we ship is our launcher for the reproduction run
([`run-nvfp4-node.sh`](run-nvfp4-node.sh)) and our measurements (below). Credit chain in
[`CREDITS.md`](CREDITS.md).

---

## Quick start (fp8 build, 2× GB10)

Prerequisites: two DGX Spark (GB10) nodes with the RoCE fabric up and a base SM12x
DeepSeek-V4-Flash vLLM image. Run steps 1–2 **on both nodes**; step 3 with the right rank
per node.

```bash
# 1. cache the ~156 GiB checkpoint on each node (download + shard-completeness verify)
./prepare-model-cache.sh

# 2. build the fp8 DSpark image from your base image (or build once and push to a registry)
bash apply-patches.sh <base-sm12x-image> dspark-vllm:fp8

# 3. set the cluster env, then start the node  (0 = head / OpenAI API :8000, 1 = worker)
cp .env.dspark.example .env.dspark      # fill in your IPs / RoCE HCA
set -a; . ./.env.dspark; set +a
./run-fp8-node.sh 0                      # on the head node
./run-fp8-node.sh 1                      # on the worker node

# 4. verify (from anywhere that can reach the head node)
./smoke.sh                              # checks /v1/models + one chat completion
```

For the faster NVFP4-KV line, build tonyd2wild's image (link above) and launch with
`./run-nvfp4-node.sh 0|1` in place of steps 2–3.

---

## Benchmarks (our unique contribution)

The measurement suite is the point of this repo. Everything here is our data on 2× DGX Spark GB10,
TP=2 over RoCE, apples-to-apples.

- **[`benchmarks/20260630-fp8-dspark-checkpoint.md`](benchmarks/20260630-fp8-dspark-checkpoint.md)**
  — our fp8 DSpark build: single-stream +21–24% over no-spec, the `block_size 5→4` control, and
  the concurrency numbers (seqs=12 prod: ~99 tok/s at 8 streams, 141–146 tok/s at 12).
- **[`benchmarks/20260630-nvfp4-1m-context-curve-checkpoint.md`](benchmarks/20260630-nvfp4-1m-context-curve-checkpoint.md)**
  — the **1M-context single-stream depth curve** of tonyd2wild's NVFP4 build (42.7 tok/s at 16K →
  17.6 at 512K → 17.2 at a real 1.03M prompt; the curve flattens past 512K), plus cold-prefill /
  prefix-cache timing at full depth — the numbers nobody in the community had published.
- **[`benchmarks/20260630-nvfp4-c16-reproduction-checkpoint.md`](benchmarks/20260630-nvfp4-c16-reproduction-checkpoint.md)**
  — independent reproduction of tonyd2wild's C16 concurrency numbers (his 315.1 re-derives as our
  control's 319), validating his benchmark.
- **[`benchmarks/20260701-sampling-coherence-probe.md`](benchmarks/20260701-sampling-coherence-probe.md)**
  — a long-context sampling / coherence probe on the NVFP4 build: an **honest negative result** (no
  sampler config is a reliable cross-depth coherence fix), with two verifiable takeaways — the
  collapse is not a fixed depth threshold, and `min_p` is blocked under spec-decode.
- **[`docs/LONG-CONTEXT-FINDINGS.md`](docs/LONG-CONTEXT-FINDINGS.md)** — the detailed analysis
  behind the NVFP4 measurements: the per-position acceptance decay driving the depth curve, the
  acceptance-not-bandwidth mechanism, the flappy coherence-collapse characterization, and two
  documented negative results.

---

## Honest stance

- The **NVFP4 numbers are measurements of tonyd2wild's build, not our recipe.** We reproduced and
  measured it; we did not author it and claim **no speed win over it** (our C16 reproduction ties
  his within his own run-to-run variance).
- Our own recipe is the **fp8 DSpark** line (patches + launcher above).
- Our unique data is the **1M-context depth curve**, the **concurrency scaling**, and the
  **coherence / sampling study**.
- No "world-first," no "we beat," no "SOTA." Where DSpark on GB10 is a modest single-stream gain,
  we say so; where it wins (concurrency, long context), we show the numbers.

---

## Credits & license

Full attribution — our fp8 contribution, the lineage we build on, and the NVFP4 line we reproduced
but did not author — is in **[`CREDITS.md`](CREDITS.md)**. This repo is released under the **MIT
License** (see [`LICENSE`](LICENSE) and [`NOTICE`](NOTICE)). **No model weights are redistributed by
this repository.**