Qwen3.8-Flash-Dual-Spark-Recipe

KO | EN

gitlite — search

#pytorch #openai #sql #golang #deep-learning #java #shell #windows #macos #gemini

Qwen3.8-Flash-Dual-Spark-Recipe

(★ 9)

Solid performing deployment since day 0 for running the new qwen 3.8 flash next on 2 x gx10 / spark on VLLM for agentic workloads.

파일 탐색기

ZIP 다운로드

- tool-eval-bench.md
.dockerignore
.gitignore
Dockerfile
download-model.sh
LICENSE
ple_layer.py
README.md
run.sh
serve.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)

Serve [RadixArk/Qwen3.8-Flash-Next-NVFP4](https://huggingface.co/RadixArk/Qwen3.8-Flash-Next-NVFP4)
with **vLLM** across **2× NVIDIA DGX Spark** (GB10 / aarch64 / sm_121a), TP=2 over the RoCE fabric.
No Ray — just vLLM's `mp` distributed executor, one container per node.

## Overview

This repository is a turn-key, production-style deployment recipe for running the
**Qwen3.8 Flash Next NVFP4** checkpoint on a pair of NVIDIA DGX Spark boxes. It ships a
patched, self-contained vLLM image and a one-command launcher that handles model download,
node-to-node sync, and dual-node bring-up. Sharing this as the SGlang ones have not worked consistently for my agentic work, VLLM has proven much faster and more stable!

| | |
|---|---|
| **Model** | [Qwen3.8 Flash Next NVFP4](https://huggingface.co/RadixArk/Qwen3.8-Flash-Next-NVFP4) (hybrid ModelOpt NVFP4 + FP8 PLE) |
| **Serving stack** | vLLM (patched `ple_layer.py`) |
| **Topology** | 2× DGX Spark (GB10), TP=2, `--nnodes 2`, mp executor |
| **Context window** | 262,144 tokens |
| **Concurrency** | 8 sequences |
| **Speculative decoding** | MTP, 4 tokens |
| **Execution mode** | `--enforce-eager` (CUDA graphs / torch.compile disabled) |
| **Tool calling** | `qwen3_coder` parser + auto tool choice + `qwen3` reasoning parser |
| **Image** | [`oxbyte/qwen3.8-flash-next-dual-spark`](https://hub.docker.com/r/oxbyte/qwen3.8-flash-next-dual-spark) |

## Quick start (dual Spark)

Run this **on the head node** (rank 0), with passwordless SSH to the worker node (rank 1).
Set `WORKER`, `SSH_KEY`, and `MASTER_ADDR` for your fabric (see [Configuration](#configuration)):

```bash
git clone https://github.com/x00byte/Qwen3.8-Flash-Dual-Spark-Recipe.git
cd Qwen3.8-Flash-Dual-Spark-Recipe
./run.sh
```

`run.sh` does everything:

1. `docker pull` the published image on both nodes
2. download the 126 GiB checkpoint into `~/models/RadixArk/Qwen3.8-Flash-Next-NVFP4` (head)
3. rsync it to the worker node
4. drop page cache on both nodes (GB10 unified-memory load safety)
5. launch the worker (`--headless`, rank 1) then the head (rank 0, API on `:8888`)

Watch it boot:

```bash
docker logs -f qwen38-flash-next-nvfp4-vllm                     # head
ssh "$WORKER" docker logs -f qwen38-flash-next-nvfp4-vllm # worker
```

Check readiness:

```bash
curl http://127.0.0.1:8888/v1/models
# -> {"id":"qwen38-flash-next-nvfp4","max_model_len":262144,...}
```

## Benchmarks

# Qwen3.8-Flash-Next-NVFP4 — vLLM Benchmark (2× DGX Spark, TP2)

**Config:** vLLM `local/vllm-qwen38:ple-fp8-fix` · MTP3 speculative (0.72 acceptance) · fp8 KV · `--enforce-eager` · `--gpu-memory-utilization 0.80` · 262,144
 context · thinking off · warmed, on-box

## Single-Stream Decode (temp 0, 512-token gens, median of 3)

| Content type | tok/s |
   |---|---:|
   | Counting / structured | 63.7 |
   | Code | 53.6 |
   | Mixed (reasoning + code) | 47.4 |
   | Freeform prose | 39.3 |

## Prefill / TTFT (cold, unique prompts, 1 output token)

| Prompt length | TTFT | Prefill throughput |
   |---|---:|---:|
   | 1K | 0.71 s | 1,452 tok/s |
   | 32K | 11.2 s | 2,932 tok/s |
   | 128K | 49.5 s | 2,652 tok/s |

## Concurrency (code prompts, 400-token generations)

| Streams | Aggregate tok/s | Per-stream tok/s |
   |---|---:|---:|
   | ×1 | 47.4 | 48.3 |
   | ×2 | 86.0 | 47.9 |
   | ×4 | 118.4 | 39.7 |
   | ×8 | 196.3 | 32.5 |

## Agentic Long-Context — 4 × 128K Concurrent

| Metric | Value |
   |---|---:|
   | Wall time (4 × 128,032-token prompts, 512-tok gens) | 68.9 s |
   | Prefill aggregate | 7,433 tok/s |
   | TTFT | ~50–54 s |
   | Per-stream decode (post-prefill) | 25–36 tok/s |

## Resources & Stability

| Metric | Value |
   |---|---:|
   | KV cache pool | 1,905,857 tokens (7.27× 262K) |
   | Memory headroom (idle) | ~9 GB |
   | 16-way quality stress | 48/48 clean, 0 NaN |
   | MTP acceptance | 0.72 |

Full tool-calling evaluation with
[tool-eval-bench](https://github.com/SeraphimSerapis/tool-eval-bench) — the standard **69-scenario**
suite (15 categories), run against the served model at 262K context with thinking disabled
(`seed=42`, `temperature=0`, sequential).

> **Score: 91 / 100 — ★★★★★ Excellent** · 59 passed · 8 partial · 2 failed · 126/138 points

| Metric | Value |
|---|---|
| **Final score** | **91 / 100** |
| **Rating** | ★★★★★ Excellent |
| Passed / Partial / Failed | 59 / 8 / 2 |
| Points | 126 / 138 |
| Quality | 91 / 100 |
| Responsiveness | 62 / 100 (median turn 2.2s) |
| Deployability | 82 / 100 (α=0.7) |
| Total wall time | 562 s (69 scenarios) |
| Engine | vLLM `0.1.dev20073+g8e685d198` |

### Category scores

| Category | Score |
|---|:---:|
| Tool Selection | **100%** (6/6) |
| Parameter Precision | **100%** (6/6) |
| Error Recovery | **100%** (6/6) |
| Localization | **100%** (6/6) |
| Structured Reasoning | **100%** (6/6) |
| Safety & Boundaries | **100%** (26/26) |
| Creative Composition | **100%** (6/6) |
| Multi-Step Chains | 88% (7/8) |
| Toolset Scale | 88% (7/8) |
| Context & State | 85% (17/20) |
| Restraint & Refusal | 83% (5/6) |
| Code Patterns | 83% (5/6) |
| Autonomous Planning | 83% (5/6) |
| Structured Output | 83% (10/12) |
| Instruction Following | 80% (8/10) |

### Difficulty tier breakdown

| Tier | Scenarios | Passed | Pass rate |
|---|:---:|:---:|:---:|
| Trivial (★) | 4 | 3 | 75% |
| Easy (★★) | 17 | 16 | 94% |
| Moderate (★★★) | 31 | 27 | 87% |
| Hard (★★★★) | 17 | 13 | 76% |

### Notable results

- **Safety & Boundaries is perfect (26/26)** — all prompt-injection, authority-escalation,
  and parameter-validation scenarios passed, so the ★★★★★ rating is not safety-capped.
- **100% on six categories**, including tool selection, localization, and structured reasoning.
- **2 failures** (both infrastructure-independent, model-level):
  - `TC-45` tool_choice=required compliance — no tool call emitted despite a required tool.
  - `TC-68` schema violation resistance — called tools when none were needed.
- **8 partials** are mostly "correct but suboptimal" — e.g. reaching for a calculator on trivial
  math (TC-11, TC-39) or an unnecessary extra tool call (TC-28).

The complete per-scenario trace is in [`benchmarks/tool-eval-bench.md`](benchmarks/tool-eval-bench.md).

## Configuration

Everything is env-overridable. `run.sh` handles orchestration; `serve.sh` (baked into the image)
holds the vLLM flags.

| Env | Default | Meaning |
|---|---|---|
| `IMAGE` | `oxbyte/qwen3.8-flash-next-dual-spark:latest` | image to pull/run |
| `WORKER` | `user@worker-ip` | worker node `user@fabric-ip` |
| `SSH_KEY` | `~/.ssh/id_ed25519` | key for the worker node |
| `MODEL_REPO` | `RadixArk/Qwen3.8-Flash-Next-NVFP4` | HuggingFace repo |
| `MODEL_DIR` | `~/models/RadixArk/Qwen3.8-Flash-Next-NVFP4` | local checkpoint path |
| `MASTER_ADDR` | `192.168.1.1` | head node's fabric IP (rank 0) |
| `MASTER_PORT` | `29501` | torch.distributed port |
| `SPEC_TOKENS` | `4` | MTP speculative tokens |
| `MAX_MODEL_LEN` | `262144` | context length |
| `MAX_NUM_SEQS` | `8` | concurrent sequences |
| `GPU_MEMORY_UTILIZATION` | `0.85` | vLLM gpu-memory-utilization |

## Exact `vllm serve` command

This is the full command the published image runs (see `serve.sh`). It's shown with the model
mounted at `/model` — swap in your own path if you mount it elsewhere.

**Head node (rank 0) — API server + engine core:**

```bash
vllm serve /model \
  --served-model-name qwen38-flash-next-nvfp4 \
  --quantization modelopt_fp4 \
  --tensor-parallel-size 2 \
  --pipeline-parallel-size 1 \
  --nnodes 2 \
  --master-addr "$MASTER_ADDR" \
  --master-port 29501 \
  --node-rank 0 \
  --distributed-executor-backend mp \
  --enforce-eager \
  --speculative-config '{"method":"mtp","num_speculative_tokens":4}' \
  --max-model-len 262144 \
  --max-num-seqs 8 \
  --max-num-batched-tokens 8192 \
  --enable-chunked-prefill \
  --enable-prefix-caching \
  --gpu-memory-utilization 0.85 \
  --tool-call-parser qwen3_coder \
  --enable-auto-tool-choice \
  --reasoning-parser qwen3 \
  --host 0.0.0.0 \
  --port 8888
```

**Worker node (rank 1) — workers only, no API server:**

```bash
vllm serve /model \
  --served-model-name qwen38-flash-next-nvfp4 \
  --quantization modelopt_fp4 \
  --tensor-parallel-size 2 \
  --pipeline-parallel-size 1 \
  --nnodes 2 \
  --master-addr "$MASTER_ADDR" \
  --master-port 29501 \
  --node-rank 1 \
  --distributed-executor-backend mp \
  --enforce-eager \
  --speculative-config '{"method":"mtp","num_speculative_tokens":4}' \
  --max-model-len 262144 \
  --max-num-seqs 8 \
  --max-num-batched-tokens 8192 \
  --enable-chunked-prefill \
  --enable-prefix-caching \
  --gpu-memory-utilization 0.85 \
  --tool-call-parser qwen3_coder \
  --enable-auto-tool-choice \
  --reasoning-parser qwen3 \
  --host 0.0.0.0 \
  --port 8888 \
  --headless
```

## Problems we hit — and how we fixed them

### 1. Dropping `--enforce-eager` locked up the whole node

We tried removing `--enforce-eager` (and/or adding `--optimization-level 1`) to see whether CUDA
graphs would help. Instead, after weights finished loading, vLLM entered a `torch.compile` /
inductor compile of the backbone, logged `No available shared memory broadcast block found in 60
seconds` every minute for ~20 minutes, then exited with code 255 — and took the DGX Spark down
with it (the unified-memory box became unresponsive).

**Fix:** keep `--enforce-eager`. On GB10, CUDA graphs and torch.compile aren't worth the risk;
eager is the stable path. The 91/100 benchmark in this README was produced in eager mode.

### 2. PLE n-gram embedding was silently served wrong

The checkpoint is hybrid: routed experts are ModelOpt **NVFP4**, but the PLE n-gram table ships as
**FP8** shards with a single global `ngram_embedding.weight_scale`. vLLM only enabled its FP8 PLE
path when the outer quant config was an `Fp8Config`; here it's `modelopt`, so it silently upcast
the FP8 bytes to bf16 with no scale applied — no crash, just wrong embeddings.

**Fix:** patch `_get_ple_embedding_quant_method()` to use the FP8 path whenever
`ple_embedding_dtype == "float8_e4m3fn"`, regardless of the outer quant config.

### 3. `KeyError: attribute 'weight_scale' already exists`

The moment the gate above was fixed, a second bug surfaced: the FP8 PLE method registered
`weight_scale` as a parameter in `create_weights`, while the loader registered it as a buffer.

**Fix:** stop registering it as a parameter — the buffer (read by `_dequantize_embeddings`) is the
one that survives. Both fixes live in `ple_layer.py`.

### 4. Follower node crashed at KV-cache init

Without `--headless` on the worker, the follower node hit
`AssertionError: collective_rpc should not be called on follower node` during KV-cache init.

**Fix:** run the rank-1 container with `--headless`. The head runs the API server + engine core;
the worker is workers only.

### 5. Warm page cache starved the GPU allocator mid-load

On GB10's unified memory, a warm page cache from a previous load could starve the GPU allocator
halfway through the 126 GiB checkpoint load.

**Fix:** drop page cache before loading (`echo 3 > /proc/sys/vm/drop_caches`). `run.sh` does this
from a `--privileged` one-shot container on both nodes.

### 6. `tool_choice:"auto"` returned HTTP 400

Tool-calling requests with `tool_choice:"auto"` failed with 400 until we enabled auto tool choice
on the server.

**Fix:** add `--enable-auto-tool-choice` alongside `--tool-call-parser qwen3_coder`.

## Rebuilding the image

The published image is self-contained (you never need the base). To rebuild or audit:

```bash
docker build -t oxbyte/qwen3.8-flash-next-dual-spark:latest .
```

The `FROM` is NVIDIA's dev image `vllm/vllm-openai:qwen38-flash-next` (sm_121a / aarch64 / CUDA 13).
You only need it if you rebuild — `docker pull` users don't.

## Hardware notes (DGX Spark / GB10)

- RoCE interface `enp1s0f1np1`, MTU 9000, `NCCL_IB_HCA=rocep1s0f1,roceP2p1s0f1`.
- Weights load ~61.7 GiB/rank; drop page cache before load or a warm cache can starve the GPU
  allocator midway (`run.sh` does this automatically).
- Load takes ~8–10 minutes for the 126 GiB checkpoint.
- KV cache lands at ~3.16M tokens — 8 concurrent sequences at 262K fits comfortably.

## Repository layout

```
.
├── Dockerfile          # image build (patch + entrypoint)
├── serve.sh            # in-image vLLM entrypoint
├── run.sh              # one-command dual-node bring-up
├── download-model.sh   # HF checkpoint downloader (uses the image's hf CLI)
├── ple_layer.py        # PLE quant path fix
├── benchmarks/         # eval results + methodology
└── README.md
```

## License

Apache 2.0 — see [LICENSE](LICENSE). `ple_layer.py` is derived from vLLM (also Apache 2.0).

Qwen3.8-Flash-Dual-Spark-Recipe

오너의 다른 레포

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Qwen3.8-Flash-Dual-Spark-Recipe

오너의 다른 레포

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Download README