Recipe
qwen3-6-27b-dflash-q4-0-intel-arc-pro-b70-32gb-llama-cpp-tp3
qwen3-6-27b-dflash-q4-0-intel-arc-pro-b70-32gb-llama-cpp-tp3Observed LocalMaxxing leaderboard run. Evidence for compatibility, not an executable launch contract.
Record
- Status
- candidate
- Source
- localmaxxing
- Engine
- llama.cpp
- Engine version
- db44417+sycl-async-cpy+qwen-split+single-kernel-allreduce
- Accelerators
- 3
- Tensor parallel
- 3
- Context tokens
- 512
- chat
- unknown
- reasoning
- unknown
- tools
- unknown
- vision
- unknown
Hugging Face model card
Identity
https://huggingface.co/z-lab/Qwen3.6-27B-DFlash- Repository
- z-lab/Qwen3.6-27B-DFlash
- 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
llama.cpp
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.
/home/steve/src/llama.cpp-q4-b70/build-sycl-2026-bmg-g31/bin/llama-bench-m/home/steve/models/qwen3.6-27b-q4_0-gguf/Qwen3.6-27B-Q4_0.gguf--prio0-devSYCL0/SYCL1/SYCL2-ngl99-p0-n512-smtensor-ts1/1/1-b512-ub32-ctkf16-ctvf16-t8--poll50-fa1-r3-ojsonl
| Flag | Value |
|---|---|
-m | /home/steve/models/qwen3.6-27b-q4_0-gguf/Qwen3.6-27B-Q4_0.gguf |
--prio | 0 |
-dev | SYCL0/SYCL1/SYCL2 |
-ngl | 99 |
-p | 0 |
-n | 512 |
-sm | tensor |
-ts | 1/1/1 |
-b | 512 |
-ub | 32 |
-ctk | f16 |
-ctv | f16 |
-t | 8 |
--poll | 50 |
-fa | 1 |
-r | 3 |
-o | jsonl |
Environment
| Variable | Value |
|---|---|
GGML_SYCL_ASYNC_CPY_TENSOR | 1 |
GGML_SYCL_ASYNC_PEER_COPY | 0 |
GGML_SYCL_COMM_ALLREDUCE | 1 |
GGML_SYCL_COMM_SINGLE_KERNEL | 1 |
GGML_SYCL_DISABLE_DNN | 1 |
GGML_SYCL_DISABLE_GRAPH | 0 |
GGML_SYCL_DISABLE_OPT | 0 |
GGML_SYCL_PRIORITIZE_DMMV | 0 |
ONEAPI_DEVICE_SELECTOR | level_zero:2,1,3 |
Measured speed
| Concurrency | Context | Prefill | Decode | TTFT ms | Status | Sweep |
|---|---|---|---|---|---|---|
| 1 | 512 | — | 41.7 | — | observed | qwen3-6-27b-dflash-q4-0-intel-arc-pro-b70-32gb-llama-cpp-tp3-sweep |
Remaining fields
Identity, launch, related records, and measured speed are shown above. This is the rest of the normalized record.
- hardware count
- 3
- hardware id
- intel-arc-pro-b70-32gb
- id
- qwen3-6-27b-dflash-q4-0-intel-arc-pro-b70-32gb-llama-cpp-tp3
- model instance id
- z-lab-qwen3-6-27b-dflash--q4-0
- recipe source
- localmaxxing
- schema version
- local-ai-registry/v1
- speed sweep ids
- qwen3-6-27b-dflash-q4-0-intel-arc-pro-b70-32gb-llama-cpp-tp3-sweep
- status
- candidate
capabilities
engine
- name
- llama.cpp
- version
- db44417+sycl-async-cpy+qwen-split+single-kernel-allreduce
serving
- max context tokens
- 512
- tensor parallel
- 3
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
- serving.max concurrency · provenance · captured at
- 2026-08-31T23:03:15Z
serving.max concurrency · reason server-capacity-not-evidenced
serving.max concurrency · state unknown
metadata
- localmaxxing · backend
- xpu
- localmaxxing · batch size
- 1
- localmaxxing · hardware label
- Intel Arc Pro B70
- localmaxxing · notes
- Quality-preserving Qwen3.6 27B Q4_0 GGUF tensor-split run on three Intel Arc Pro B70 32GB GPUs. Same Q4_0 weights and f16 KV cache, greedy decode, no speculative decode, no power-limit changes. Uses selector order 2,1,3 so physical B70 2 is the root for the experimental direct 2-4 GPU single-kernel F32 allreduce. Local patches add SYCL async cpy_tensor, Qwen recurrent split anchoring, and peer-read/peer-write allreduce. Low-level four-device SYCL peer stress test passed. Validated over 512 generated tokens, 3 repeats: samples 41.6977, 41.6966, 41.8155 tok/s. Previous root 0,1,2 TP3 result was 41.367 tok/s; best TP2 result was 39.849 tok/s; single-B70 baseline was 24.723 tok/s.
- localmaxxing · observed command
- ONEAPI_DEVICE_SELECTOR=level_zero:2,1,3 GGML_SYCL_DISABLE_GRAPH=0 GGML_SYCL_DISABLE_DNN=1 GGML_SYCL_DISABLE_OPT=0 GGML_SYCL_PRIORITIZE_DMMV=0 GGML_SYCL_ASYNC_PEER_COPY=0 GGML_SYCL_ASYNC_CPY_TENSOR=1 GGML_SYCL_COMM_ALLREDUCE=1 GGML_SYCL_COMM_SINGLE_KERNEL=1 /home/steve/src/llama.cpp-q4-b70/build-sycl-2026-bmg-g31/bin/llama-bench -m /home/steve/models/qwen3.6-27b-q4_0-gguf/Qwen3.6-27B-Q4_0.gguf --prio 0 -dev SYCL0/SYCL1/SYCL2 -ngl 99 -p 0 -n 512 -sm tensor -ts 1/1/1 -b 512 -ub 32 -ctk f16 -ctv f16 -t 8 --poll 50 -fa 1 -r 3 -o jsonl
- localmaxxing · run id
- cmoqqed6s0007jv049wnizwle
- localmaxxing · tokenized · arguments
- /home/steve/src/llama.cpp-q4-b70/build-sycl-2026-bmg-g31/bin/llama-bench, -m, /home/steve/models/qwen3.6-27b-q4_0-gguf/Qwen3.6-27B-Q4_0.gguf, --prio, 0, -dev, SYCL0/SYCL1/SYCL2, -ngl, 99, -p, 0, -n, 512, -sm, tensor, -ts, 1/1/1, -b, 512, -ub, 32, -ctk, f16, -ctv, f16, -t, 8, --poll, 50, -fa, 1, -r, 3, -o, jsonl
- localmaxxing · tokenized · environment · GGML SYCL ASYNC CPY TENSOR
- 1
- localmaxxing · tokenized · environment · GGML SYCL ASYNC PEER COPY
- 0
- localmaxxing · tokenized · environment · GGML SYCL COMM ALLREDUCE
- 1
- localmaxxing · tokenized · environment · GGML SYCL COMM SINGLE KERNEL
- 1
- localmaxxing · tokenized · environment · GGML SYCL DISABLE DNN
- 1
- localmaxxing · tokenized · environment · GGML SYCL DISABLE GRAPH
- 0
- localmaxxing · tokenized · environment · GGML SYCL DISABLE OPT
- 0
- localmaxxing · tokenized · environment · GGML SYCL PRIORITIZE DMMV
- 0
- localmaxxing · tokenized · environment · ONEAPI DEVICE SELECTOR
- level_zero:2,1,3
- 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/cmoqqed6s0007jv049wnizwle ↗ |