Local AI Registry

Recipe

qwen3-8-27b-autoround-int4-w4a16-intel-arc-pro-b70-32gb-vllm-tp2

qwen3-8-27b-autoround-int4-w4a16-intel-arc-pro-b70-32gb-vllm-tp2
JSON APIFind compatible recipes

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

Record

Status
candidate
Source
localmaxxing
Engine
vllm
Engine version
v0.20.2rc1.dev13+g9557d9108 local XPU stack; vllm 44fc8fde09; vllm-xpu-kernels 2dd55f380d; oneCCL 4ceafd1
Accelerators
2
Tensor parallel
2
Context tokens
2,048
Max concurrency
1
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/devan-carlin/Qwen3.8-27B-int4-AutoRound
Repository
devan-carlin/Qwen3.8-27B-int4-AutoRound
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/cmszbkxco0e11ms01l2rixxbt

Observed source tokens

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

  1. --dtype
  2. float16
  3. --tensor-parallel-size
  4. 2
  5. --max-model-len
  6. 2048
  7. --max-num-seqs
  8. 1
FlagValue
--dtypefloat16
--tensor-parallel-size2
--max-model-len2048
--max-num-seqs1

Environment

VariableValue
VALIDATION_DETERMINISTIC_GREEDY_MARGIN0.03125
VALIDATION_GDN_NATIVE_SPEC_RECURRENT_SERIAL_EXACT0
VALIDATION_LM_HEAD_INT81
VALIDATION_NUM_SPECULATIVE_TOKENS5
VLLM_XPU_GDN_SPEC_PERSISTENT_SCRATCH0
cudagraph_capture_sizes[6]

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
12,048—101.91,162.2observedqwen3-8-27b-autoround-int4-w4a16-intel-arc-pro-b70-32gb-vllm-tp2-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-autoround-int4-w4a16-intel-arc-pro-b70-32gb-vllm-tp2
model instance id
devan-carlin-qwen3-8-27b-int4-autoround--autoround-int4-w4a16
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
qwen3-8-27b-autoround-int4-w4a16-intel-arc-pro-b70-32gb-vllm-tp2-sweep
status
candidate

capabilities

engine

name
vllm
version
v0.20.2rc1.dev13+g9557d9108 local XPU stack; vllm 44fc8fde09; vllm-xpu-kernels 2dd55f380d; oneCCL 4ceafd1

serving

max concurrency
1
max context tokens
2,048
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
Supersedes cmszarna10e0nms0103hv0tve (MTP4, 100.497) on the same 25-prompt suite. Median of three fresh cold arms 100.896 / 102.042 / 101.922; every measured row cached_tokens=0. Self-determinism: all three pairwise comparisons 25/25 token-identical under a PINNED torch.compile cache shared by the arms - fresh compilations emit different-but-internally-deterministic code, so the compile cache is part of the run identity and a fresh-compile arm is outstanding. Quality: pass_all and baseline_match_all against this model's own baseline (15 comparisons). IMPORTANT: on the 12 historical selection prompts the median is 95.167, which is LOWER than the MTP4 submission's 96.627 - depth helps the newer holdout prompts and hurts the historical ones, so use 95.167 against any record set on a 12-prompt suite. Depth was swept: MTP3 96.6, MTP4 100.5, MTP5 101.9, MTP6 99.5, so 5 is the optimum. VLLM_XPU_GDN_SPEC_PERSISTENT_SCRATCH=0 is required and is a workaround for an uninitialized read in the persistent GDN scratch, not a tuning choice.
localmaxxing · observed command
VLLM_XPU_GDN_SPEC_PERSISTENT_SCRATCH=0 VALIDATION_NUM_SPECULATIVE_TOKENS=5 cudagraph_capture_sizes=[6] VALIDATION_GDN_NATIVE_SPEC_RECURRENT_SERIAL_EXACT=0 VALIDATION_LM_HEAD_INT8=1 VALIDATION_DETERMINISTIC_GREEDY_MARGIN=0.03125 --dtype float16 --tensor-parallel-size 2 --max-model-len 2048 --max-num-seqs 1
localmaxxing · run id
cmszbkxco0e11ms01l2rixxbt
localmaxxing · tokenized · arguments
--dtype, float16, --tensor-parallel-size, 2, --max-model-len, 2048, --max-num-seqs, 1
localmaxxing · tokenized · environment · VALIDATION DETERMINISTIC GREEDY MARGIN
0.03125
localmaxxing · tokenized · environment · VALIDATION GDN NATIVE SPEC RECURRENT SERIAL EXACT
0
localmaxxing · tokenized · environment · VALIDATION LM HEAD INT8
1
localmaxxing · tokenized · environment · VALIDATION NUM SPECULATIVE TOKENS
5
localmaxxing · tokenized · environment · VLLM XPU GDN SPEC PERSISTENT SCRATCH
0
localmaxxing · tokenized · environment · cudagraph capture sizes
[6]
localmaxxing · tokenized · fidelity
faithful

provenance

captured at
2026-08-30T09:10:02Z

sources

captured atkindurl
2026-08-30T09:10:02Znormalized-recipewww.localmaxxing.com/en/runs/cmszbkxco0e11ms01l2rixxbt ↗