Recipe
qwen3-8-27b-fp8-intel-arc-pro-b70-32gb-vllm-tp2-n13vly2h
qwen3-8-27b-fp8-intel-arc-pro-b70-32gb-vllm-tp2-n13vly2hObserved LocalMaxxing leaderboard run. Evidence for compatibility, not an executable launch contract.
Record
- Status
- candidate
- Source
- localmaxxing
- Engine
- vllm
- Engine version
- 0.27.2rc1.dev77+gac7509e2b; XPU kernels 1e90ffa672; neural.download W8A16+active-width+active-Mamba overlay 2b79af686423
- Accelerators
- 2
- Tensor parallel
- 2
- Context tokens
- 256
- Max concurrency
- 1
- chat
- unknown
- reasoning
- unknown
- tools
- unknown
- vision
- unknown
Hugging Face model card
Identity
https://huggingface.co/Qwen/Qwen3.8-27B-FP8- Repository
- Qwen/Qwen3.8-27B-FP8
- 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
Observed source tokens
Mechanical split of the source command. Unverified against the engine CLI. The registry does not offer Run for this recipe.
repro/qwen38-27b-fp8-vllm-tp2-asrock-b70/run-w8a16-dynamic-mtp-server.sh
Environment
| Variable | Value |
|---|---|
MODEL_DIR | /path/to/qwen3.8-27b-fp8 |
VLLM_CACHE_DIR | /path/to/new-dynamic-mtp-cache |
Measured speed
| Concurrency | Context | Prefill | Decode | TTFT ms | Status | Sweep |
|---|---|---|---|---|---|---|
| 1 | 256 | — | 58.5 | 108.1 | observed | qwen3-8-27b-fp8-intel-arc-pro-b70-32gb-vllm-tp2-n13vly2h-sweep |
Remaining fields
Identity, launch, related records, and measured speed are shown above. This is the rest of the normalized record.
- hardware count
- 2
- hardware id
- intel-arc-pro-b70-32gb
- id
- qwen3-8-27b-fp8-intel-arc-pro-b70-32gb-vllm-tp2-n13vly2h
- model instance id
- qwen-qwen3-8-27b-fp8--fp8
- recipe source
- localmaxxing
- schema version
- local-ai-registry/v1
- speed sweep ids
- qwen3-8-27b-fp8-intel-arc-pro-b70-32gb-vllm-tp2-n13vly2h-sweep
- status
- candidate
capabilities
engine
- name
- vllm
- version
- 0.27.2rc1.dev77+gac7509e2b; XPU kernels 1e90ffa672; neural.download W8A16+active-width+active-Mamba overlay 2b79af686423
serving
- max concurrency
- 1
- max context tokens
- 256
- tensor parallel
- 2
Provenance & metadata (3)
facts
- capabilities.chat · provenance · captured at
- 2026-08-30T09:10:02Z
capabilities.chat · reason capability-not-verified
capabilities.chat · state unknown
- capabilities.reasoning · provenance · captured at
- 2026-08-30T09:10:02Z
capabilities.reasoning · reason capability-not-verified
capabilities.reasoning · state unknown
- capabilities.tools · provenance · captured at
- 2026-08-30T09:10:02Z
capabilities.tools · reason capability-not-verified
capabilities.tools · state unknown
- capabilities.vision · provenance · captured at
- 2026-08-30T09:10:02Z
capabilities.vision · reason capability-not-verified
capabilities.vision · state unknown
- engine.graph mode · provenance · captured at
- 2026-08-30T09:10:02Z
engine.graph mode · reason runtime-detail-not-published
engine.graph mode · state unknown
- serving.kv cache tokens · provenance · captured at
- 2026-08-30T09:10:02Z
serving.kv cache tokens · reason kv-cache-capacity-not-published
serving.kv cache tokens · state unknown
metadata
- localmaxxing · backend
- xpu
- localmaxxing · hardware label
- Intel Arc Pro B70
- localmaxxing · notes
- Qwen3.8 27B official FP8 target on two Intel Arc Pro B70 cards with the neural.download block-W8A16, active-width GDN, and active-Mamba allocation patches. Dynamic service uses the publisher MTP layer serially to depth 8 for one active request and depth 1 under load; every accepted draft token is verified by the unchanged official FP8 target. Fixed 12-prompt cold realistic suite, one unique request each, 128 outputs, streamed token-ID timing, prefix cache disabled, cached_tokens=0 for all requests. Conventional 99-interval median was 58.537756 tok/s on this submitted cold server; an independent second cold server measured 58.244309 tok/s. Exact evidence, patches, commands, cold-start timing, and the separately scoped 146.814418 tok/s high-acceptance fixture are in the linked result packet.
- localmaxxing · observed command
- MODEL_DIR=/path/to/qwen3.8-27b-fp8 VLLM_CACHE_DIR=/path/to/new-dynamic-mtp-cache repro/qwen38-27b-fp8-vllm-tp2-asrock-b70/run-w8a16-dynamic-mtp-server.sh
- localmaxxing · run id
- cmtb5n45n0021qq01n13vly2h
- localmaxxing · tokenized · arguments
- repro/qwen38-27b-fp8-vllm-tp2-asrock-b70/run-w8a16-dynamic-mtp-server.sh
- localmaxxing · tokenized · environment · MODEL DIR
- /path/to/qwen3.8-27b-fp8
- localmaxxing · tokenized · environment · VLLM CACHE DIR
- /path/to/new-dynamic-mtp-cache
- localmaxxing · tokenized · fidelity
- faithful
provenance
- captured at
- 2026-08-30T09:10:02Z
sources
| captured at | kind | url |
|---|---|---|
| 2026-08-30T09:10:02Z | normalized-recipe | www.localmaxxing.com/en/runs/cmtb5n45n0021qq01n13vly2h ↗ |