Recipe

deepseek-v4-flash-fp8-dgx-spark-gb10-128gb-vllm-tp1

deepseek-v4-flash-fp8-dgx-spark-gb10-128gb-vllm-tp1

Observed LocalMaxxing leaderboard run. Evidence for compatibility, not an executable launch contract.

Record

Status
candidate
Source
localmaxxing
Engine
vllm
Engine version
0.21.1rc1.dev339+g1967a5627bc3-tp2
Accelerators
1
Tensor parallel
1
Context tokens
200,000
Max concurrency
1
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash
Repository
deepseek-ai/DeepSeek-V4-Flash
Status
known
Link type
Exact Hub repository

Public Hugging Face repository confirmed by the Hub API.

Candidate: useful compatibility or speed evidence. The registry does not offer Run until promotion requirements are met.

Observed configuration

vllm

Evidence only · candidate · reference

Candidate evidence — not a Run contract

Source
https://www.localmaxxing.com/en/runs/cmpw3cr6l003tr101wblsaxvj

Observed source tokens

Mechanical split of the source command. Unverified against the engine CLI. The registry does not offer Run for this recipe.

  1. vllm
  2. serve
  3. deepseek-ai/DeepSeek-V4-Flash
  4. --tensor-parallel-size
  5. 2
  6. --max-model-len
  7. 200000
  8. --kv-cache-dtype
  9. fp8
  10. --gpu-memory-utilization
  11. 0.8
FlagValue
--tensor-parallel-size2
--max-model-len200000
--kv-cache-dtypefp8
--gpu-memory-utilization0.8

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
1200,0001,238.545.7152observeddeepseek-v4-flash-fp8-dgx-spark-gb10-128gb-vllm-tp1-sweep

Remaining fields

Identity, launch, related records, and measured speed are shown above. This is the rest of the normalized record.

hardware count
1
hardware id
dgx-spark-gb10-128gb
id
deepseek-v4-flash-fp8-dgx-spark-gb10-128gb-vllm-tp1
model instance id
deepseek-ai-deepseek-v4-flash--fp8
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
deepseek-v4-flash-fp8-dgx-spark-gb10-128gb-vllm-tp1-sweep
status
candidate

capabilities

draft launch

accelerator backend
nvidia
arguments
--model, deepseek-ai/DeepSeek-V4-Flash, --tensor-parallel-size, 1, --host, 0.0.0.0, --port, 8000, --max-model-len, 200000
container port
8,000
host port
8,000
image
vllm/vllm-openai@sha256:0a51ea5b4ae2dc5d81890e5173f54203d2a3ae0cfffe51b8fd2afd4391bfd967
ipc
host
kind
docker
shm size
16g
synthesized · generated at
2026-08-31T22:12:17Z
synthesized · image provenance
deepseek-fp8-rtx-pro-6000-blackwell-96gb-vllm-tp1
synthesized · template
vllm-openai-v1

mounts

read onlytarget
No/root/.cache/huggingface

engine

name
vllm
version
0.21.1rc1.dev339+g1967a5627bc3-tp2

serving

max concurrency
1
max context tokens
200,000
tensor parallel
1
Provenance & metadata (3)

facts

capabilities.chat · provenance · captured at
2026-08-30T09:26:09Z
capabilities.chat · reason capability-not-verified
capabilities.chat · state unknown
capabilities.reasoning · provenance · captured at
2026-08-30T09:26:09Z
capabilities.reasoning · reason capability-not-verified
capabilities.reasoning · state unknown
capabilities.tools · provenance · captured at
2026-08-30T09:26:09Z
capabilities.tools · reason capability-not-verified
capabilities.tools · state unknown
capabilities.vision · provenance · captured at
2026-08-30T09:26:09Z
capabilities.vision · reason capability-not-verified
capabilities.vision · state unknown
engine.graph mode · provenance · captured at
2026-08-30T09:26:09Z
engine.graph mode · reason runtime-detail-not-published
engine.graph mode · state unknown
serving.kv cache tokens · provenance · captured at
2026-08-30T09:26:09Z
serving.kv cache tokens · reason kv-cache-capacity-not-published
serving.kv cache tokens · state unknown

metadata

localmaxxing · backend
cuda
localmaxxing · hardware label
GB10 Grace Blackwell
localmaxxing · notes
2x NVIDIA DGX Spark (GB10) cluster, TP=2 over RoCE. FP8 KV cache + MXFP4 MoE, MTP=2, 200K ctx (aidendle94/sparkrun-vllm-ds4-gb10). Warm single-stream, temp=0. ttftMs=median of 3 (short prompt, streaming). tokSOut=median of 3 (code prompt, max_tokens=256, completion_tokens/decode_seconds). tokSPrefill=median of 3 (~2000-token unique prompts, prefix-cache defeated). Unified memory 128GB/node (242GB combined pool); peakVramGb=209 = 105GB(node1)+104GB(node2) actually used.
localmaxxing · observed command
vllm serve deepseek-ai/DeepSeek-V4-Flash --tensor-parallel-size 2 --max-model-len 200000 --kv-cache-dtype fp8 --gpu-memory-utilization 0.8
localmaxxing · run id
cmpw3cr6l003tr101wblsaxvj
localmaxxing · tokenized · arguments
vllm, serve, deepseek-ai/DeepSeek-V4-Flash, --tensor-parallel-size, 2, --max-model-len, 200000, --kv-cache-dtype, fp8, --gpu-memory-utilization, 0.8
localmaxxing · tokenized · fidelity
faithful

provenance

captured at
2026-08-30T09:26:09Z

sources

captured atkindurl
2026-08-30T09:26:09Znormalized-recipewww.localmaxxing.com/en/runs/cmpw3cr6l003tr101wblsaxvj