Recipe

qwen3-fp8-intel-arc-pro-b70-32gb-vllm-tp4

qwen3-fp8-intel-arc-pro-b70-32gb-vllm-tp4

Record

Status
candidate
Source
localmaxxing
Engine
vllm
Engine version
g6607a80da (custom)
Accelerators
4
Tensor parallel
4
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/Qwen/Qwen3-Coder-Next
Repository
Qwen/Qwen3-Coder-Next
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/cmqkf5q9g00vpnv01bs4wzeu0

No tokenized launch fields. This is measured or documented compatibility, not a Docker launch.

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
1262,14471.795historicalqwen3-fp8-intel-arc-pro-b70-32gb-vllm-tp4-sweep

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
intel-arc-pro-b70-32gb
id
qwen3-fp8-intel-arc-pro-b70-32gb-vllm-tp4
model instance id
qwen-qwen3-coder-next--fp8
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
qwen3-fp8-intel-arc-pro-b70-32gb-vllm-tp4-sweep
status
candidate

capabilities

engine

name
vllm
version
g6607a80da (custom)

serving

tensor parallel
4
Provenance & metadata (3)

facts

capabilities.chat · provenance · captured at
2026-08-27T06:04:13.773Z
capabilities.chat · reason capability-not-verified
capabilities.chat · state unknown
capabilities.reasoning · provenance · captured at
2026-08-27T06:04:13.773Z
capabilities.reasoning · reason capability-not-verified
capabilities.reasoning · state unknown
capabilities.tools · provenance · captured at
2026-08-27T06:04:13.773Z
capabilities.tools · reason capability-not-verified
capabilities.tools · state unknown
capabilities.vision · provenance · captured at
2026-08-27T06:04:13.773Z
capabilities.vision · reason capability-not-verified
capabilities.vision · state unknown
description · provenance · captured at
2026-08-27T06:04:13.773Z
description · reason not-observed
description · state unknown
engine.graph mode · provenance · captured at
2026-08-27T06:04:13.773Z
engine.graph mode · reason runtime-detail-not-published
engine.graph mode · state unknown
metadata.localmaxxing.engine flags.attentionBackend · provenance · captured at
2026-08-27T06:04:13.773Z
metadata.localmaxxing.engine flags.attentionBackend · reason not-observed
metadata.localmaxxing.engine flags.attentionBackend · state unknown
metadata.localmaxxing.engine flags.flashAttn · provenance · captured at
2026-08-27T06:04:13.773Z
metadata.localmaxxing.engine flags.flashAttn · reason not-observed
metadata.localmaxxing.engine flags.flashAttn · state unknown
metadata.localmaxxing.engine flags.gpuLayers · provenance · captured at
2026-08-27T06:04:13.773Z
metadata.localmaxxing.engine flags.gpuLayers · reason not-observed
metadata.localmaxxing.engine flags.gpuLayers · state unknown
serving.kv cache tokens · provenance · captured at
2026-08-27T06:04:13.773Z
serving.kv cache tokens · reason kv-cache-capacity-not-published
serving.kv cache tokens · state unknown
serving.max concurrency · provenance · captured at
2026-08-27T06:04:13.773Z
serving.max concurrency · reason not-observed
serving.max concurrency · state unknown
serving.max context tokens · provenance · captured at
2026-08-27T06:04:13.773Z
serving.max context tokens · reason not-observed
serving.max context tokens · state unknown

metadata

localmaxxing · base lineage
derived
localmaxxing · engine flags · kvCacheDtype
bf16
localmaxxing · engine flags · mtpEnabled
No
localmaxxing · engine flags · specDecoding
No
localmaxxing · engine flags · tensorParallel
4
localmaxxing · hardware label
4x Intel Arc Pro B70 128GB
localmaxxing · notes
Big reason for the difference: stock vLLM disables graph capture on Intel XPU → eager-only (~11 tok/s decode). Un-gating XPUGraph (torch.xpu.XPUGraph, Level-Zero command-graph capture) takes single-stream ~11 → ~102 tok/s (~9.5×). The gain is command-graph replay eliminating per-kernel host-launch overhead — torch.compile/Inductor alone adds ~nothing. Custom (not stock vLLM): Un-gated XPU graph capture (cudagraph_mode=FULL_DECODE_ONLY, capture ladder [1,2,4,8,16,32,48,100]). Graph-capturable paged-decode kernel. Capture-safe all-reduce in SYCL/ESIMD (Level-Zero IPC peer buffers + custom GPU barrier): stock oneCCL all-reduce corrupts inside a captured graph, and PCIe device-to-device atomics are unreliable on Battlemage → local-write/remote-read + uncached loads + system_acquire fences. Ceiling: ~104–108 tok/s single-stream is PCIe-bound (all-reduce latency caps it). Correctness verified — GSM8K graph vs eager within noise.
localmaxxing · revision unpinned
Yes
localmaxxing · run id
cmqkf5q9g00vpnv01bs4wzeu0

provenance

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

sources

captured atkindurl
2026-08-27T06:04:13.773Znormalized-recipewww.localmaxxing.com/en/runs/cmqkf5q9g00vpnv01bs4wzeu0