Recipe

deepseek-v4-flash-0731-fp8-rtxpro6000-vllm-tp4

deepseek-v4-flash-0731-fp8-rtxpro6000-vllm-tp4

Controller-backed candidate with recovered immutable model, image, vLLM, B12X, and FlashInfer identities. The saved server configuration reports full/piecewise CUDA graphs and 1M context, but no raw client C1/C2/C4 sweep survives, so performance remains unpromoted.

Record

Status: candidate
Source: 0xsero
Engine: vllm
Engine version: vLLM e72ad0057b5f382abb40bec1f9d731f6a22600d9; B12X 57422ad9042773165fa0d4e71ae6442fc26b48c9; FlashInfer 25dd814e03791e370f96c3148242f0dc8de504ac
Graph: full-and-piecewise
Accelerators: 4
Tensor parallel: 4
Context tokens: 1,048,576
Max concurrency: 5
KV cache tokens: 5,345,692
chat: yes
reasoning: yes
tools: yes
vision: no

Hugging Face model card

Identity

https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731

Repository: deepseek-ai/DeepSeek-V4-Flash-0731
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.

Controller configuration

vllm

No container · candidate · controller

Candidate evidence — not a Run contract

Environment

Variable	Value
`B12X_DENSE_SPLITK_TURBO`	`1`
`B12X_MLA_SM120_PREFILL_F16_ROPE`	`1`
`CUDA_VISIBLE_DEVICES`	`0,1,2,3`
`CUTE_DSL_ARCH`	`sm_120a`
`HF_HUB_OFFLINE`	`1`
`NCCL_IB_DISABLE`	`1`
`NCCL_P2P_DISABLE`	`1`
`NCCL_P2P_LEVEL`	`SYS`
`NCCL_PROTO`	`LL,LL128,Simple`
`NVIDIA_VISIBLE_DEVICES`	`GPU-c6ac75f2-cadf-6ff3-4cab-76c6033c1006,GPU-3cece4bc-432e-705e-7324-3f441d9cb4cc,GPU-7b5db8b3-3a49-c0a7-b8e4-80dc1bd3e853,GPU-fa982c97-64af-db6a-2ffb-08380e1f9375`
`TRANSFORMERS_OFFLINE`	`1`
`VLLM_CACHE_ROOT`	`/cache`
`VLLM_ENABLE_PCIE_ALLREDUCE`	`0`
`VLLM_USE_B12X_FP8_GEMM`	`0`
`VLLM_USE_B12X_MOE`	`1`
`VLLM_USE_B12X_SPARSE_INDEXER`	`1`
`VLLM_USE_V2_MODEL_RUNNER`	`1`

Remaining fields

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

hardware count: 4
hardware id: rtx-pro-6000-blackwell-96gb
id: deepseek-v4-flash-0731-fp8-rtxpro6000-vllm-tp4
model instance id: deepseek-ai-deepseek-v4-flash-0731--fp8
recipe source: 0xsero
schema version: local-ai-registry/v1
status: candidate

capabilities

chat: Yes
reasoning: Yes
tools: Yes
vision: No

engine

graph mode: full-and-piecewise
name: vllm
version: vLLM e72ad0057b5f382abb40bec1f9d731f6a22600d9; B12X 57422ad9042773165fa0d4e71ae6442fc26b48c9; FlashInfer 25dd814e03791e370f96c3148242f0dc8de504ac

serving

kv cache tokens: 5,345,692
max concurrency: 5
max context tokens: 1,048,576
tensor parallel: 4

Provenance & metadata (3)

facts

speed sweep ids · provenance · captured at: 2026-08-27T06:04:13.773Z

speed sweep ids · reason not-observed

speed sweep ids · state unknown

metadata

provenance

captured at: 2026-08-27T06:04:13.773Z

sources

captured at	kind	url
2026-08-27T06:04:13.773Z	normalized-recipe	github.com/0xSero/local-ai-registry ↗