Local AI Registry

Recipe

qwen3-8-27b-rocmfp4-ryzen-ai-max-plus-395-128gb-llama-cpp-tp1

qwen3-8-27b-rocmfp4-ryzen-ai-max-plus-395-128gb-llama-cpp-tp1
JSON APIFind compatible recipes

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

Record

Status
candidate
Source
localmaxxing
Engine
llama.cpp
Engine version
246 (0a59add)
Accelerators
1
Tensor parallel
1
Context tokens
32,768
Max concurrency
1
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/unsloth/Qwen3.8-27B-GGUF
Repository
unsloth/Qwen3.8-27B-GGUF
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

Source
https://www.localmaxxing.com/en/runs/cmt0vip5k0fvgms01179emtv0

Observed source tokens

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

    1. export
    2. HSA_ENABLE_SDMA=0
    1. ./llama-server
    2. -m
    3. Qwen3.8-27B-ROCmFP4-FAST.gguf
    4. -dev
    5. ROCm0
    6. -ngl
    7. 99
    8. -fa
    9. on
    10. -np
    11. 1
    12. -c
    13. 32768
    14. -b
    15. 2048
    16. -ub
    17. 1024
    18. -t
    19. 16
    20. -ctk
    21. q8_0
    22. -ctv
    23. q8_0
    24. --spec-type
    25. draft-mtp
    26. --spec-draft-n-max
    27. 6
    28. --spec-draft-p-min
    29. 0.60
    30. --spec-mtp-strict-qwen
    31. --host
    32. 127.0.0.1
    33. --port
    34. 18090
    35. --jinja
    36. --metrics
    37. --reasoning
    38. off
FlagValue
-mQwen3.8-27B-ROCmFP4-FAST.gguf
-devROCm0
-ngl99
-faon
-np1
-c32768
-b2048
-ub1024
-t16
-ctkq8_0
-ctvq8_0
--spec-typedraft-mtp
--spec-draft-n-max6
--spec-draft-p-min0.60
--host127.0.0.1
--port18090
--reasoningoff

Source notes

ROCmFPX llama-server built for gfx1151 (Strix Halo iGPU)

source: https://github.com/charlie12345/ROCmFPX commit 0a59add

cmake -S . -B build-gfx1151 -G Ninja -DCMAKE_BUILD_TYPE=Release # -DCMAKE_HIP_COMPILER=/opt/rocm/core-7.14/lib/llvm/bin/clang++ # -DAMDGPU_TARGETS=gfx1151 -DGGML_HIP=ON -DGGML_HIP_MMQ_MFMA=ON # -DGGML_HIP_NO_VMM=ON -DGGML_HIP_GRAPHS=ON -DLLAMA_BUILD_WEBUI=OFF -DLLAMA_CURL=OFF

request: 56-token prompt, max_tokens=1024, temperature=0, top_p=1, stream=false

1 warmup discarded, then 3 timed runs

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
132,768169.750.1—observedqwen3-8-27b-rocmfp4-ryzen-ai-max-plus-395-128gb-llama-cpp-tp1-sweep

Remaining fields

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

hardware count
1
hardware id
ryzen-ai-max-plus-395-128gb
id
qwen3-8-27b-rocmfp4-ryzen-ai-max-plus-395-128gb-llama-cpp-tp1
model instance id
unsloth-qwen3-8-27b-gguf--rocmfp4
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
qwen3-8-27b-rocmfp4-ryzen-ai-max-plus-395-128gb-llama-cpp-tp1-sweep
status
candidate

capabilities

engine

name
llama.cpp
version
246 (0a59add)

serving

max concurrency
1
max context tokens
32,768
tensor parallel
1
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
2026-09-01T10:20:39Z
serving.max concurrency · reason explicit-source-server-capacity
serving.max concurrency · state known
serving.max context tokens · provenance · captured at
2026-09-01T10:20:39Z
serving.max context tokens · reason explicit-source-context-limit
serving.max context tokens · state known

metadata

localmaxxing · backend
rocm
localmaxxing · batch size
1
localmaxxing · hardware label
Ryzen AI Max 395
localmaxxing · notes
Radeon 8060S integrated GPU (gfx1151) on BOSGAME M5 Ryzen AI MAX+ 395, 128GB unified memory, Ubuntu 26.04, ROCm 7.14. iGPU only — the discrete R9700 was present in the system but not used (server bound to ROCm0 = 8060S; verified in the engine device log). Engine is a ROCmFPX llama.cpp fork (github.com/charlie12345/ROCmFPX, commit 0a59add, version 246) built specifically for gfx1151, with Qwen MTP self-speculative decoding. Steady-state decode from 3 consecutive timed runs after 1 discarded warmup: 51.39 / 50.17 / 50.11 tok/s; submitting the slowest (50.11) as the conservative figure. Engine-reported prompt eval across the timed runs: 167.9 / 170.4 / 169.7 tok/s. MTP draft acceptance rate 0.994 with mean accepted length 6.87 of 6 max draft tokens. Weights are Qwen3.8-27B-ROCmFP4-FAST.gguf, a ROCmFP4 repack of the unsloth Qwen3.8-27B GGUF release; the quantization label 'ROCmFP4' is the engine-native format name and has no exact entry in the accepted quantization list. Note that -ctk/-ctv q8_0 means the KV cache is 8-bit, not fp16.
localmaxxing · observed command
# ROCmFPX llama-server built for gfx1151 (Strix Halo iGPU) # source: https://github.com/charlie12345/ROCmFPX commit 0a59add # cmake -S . -B build-gfx1151 -G Ninja -DCMAKE_BUILD_TYPE=Release \ # -DCMAKE_HIP_COMPILER=/opt/rocm/core-7.14/lib/llvm/bin/clang++ \ # -DAMDGPU_TARGETS=gfx1151 -DGGML_HIP=ON -DGGML_HIP_MMQ_MFMA=ON \ # -DGGML_HIP_NO_VMM=ON -DGGML_HIP_GRAPHS=ON -DLLAMA_BUILD_WEBUI=OFF -DLLAMA_CURL=OFF export HSA_ENABLE_SDMA=0 ./llama-server \ -m Qwen3.8-27B-ROCmFP4-FAST.gguf \ -dev ROCm0 \ -ngl 99 -fa on -np 1 \ -c 32768 -b 2048 -ub 1024 -t 16 \ -ctk q8_0 -ctv q8_0 \ --spec-type draft-mtp --spec-draft-n-max 6 --spec-draft-p-min 0.60 \ --spec-mtp-strict-qwen \ --host 127.0.0.1 --port 18090 --jinja --metrics --reasoning off # request: 56-token prompt, max_tokens=1024, temperature=0, top_p=1, stream=false # 1 warmup discarded, then 3 timed runs
localmaxxing · run id
cmt0vip5k0fvgms01179emtv0
localmaxxing · tokenized · arguments
./llama-server, -m, Qwen3.8-27B-ROCmFP4-FAST.gguf, -dev, ROCm0, -ngl, 99, -fa, on, -np, 1, -c, 32768, -b, 2048, -ub, 1024, -t, 16, -ctk, q8_0, -ctv, q8_0, --spec-type, draft-mtp, --spec-draft-n-max, 6, --spec-draft-p-min, 0.60, --spec-mtp-strict-qwen, --host, 127.0.0.1, --port, 18090, --jinja, --metrics, --reasoning, off
localmaxxing · tokenized · fidelity
faithful
localmaxxing · tokenized · notes
ROCmFPX llama-server built for gfx1151 (Strix Halo iGPU), source: https://github.com/charlie12345/ROCmFPX commit 0a59add, cmake -S . -B build-gfx1151 -G Ninja -DCMAKE_BUILD_TYPE=Release # -DCMAKE_HIP_COMPILER=/opt/rocm/core-7.14/lib/llvm/bin/clang++ # -DAMDGPU_TARGETS=gfx1151 -DGGML_HIP=ON -DGGML_HIP_MMQ_MFMA=ON # -DGGML_HIP_NO_VMM=ON -DGGML_HIP_GRAPHS=ON -DLLAMA_BUILD_WEBUI=OFF -DLLAMA_CURL=OFF, request: 56-token prompt, max_tokens=1024, temperature=0, top_p=1, stream=false, 1 warmup discarded, then 3 timed runs

localmaxxing · tokenized · steps

01234567
exportHSA_ENABLE_SDMA=0UnknownUnknownUnknownUnknownUnknownUnknown
./llama-server-mQwen3.8-27B-ROCmFP4-FAST.gguf-devROCm0-ngl99-fa

provenance

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

sources

captured atkindurl
2026-08-30T09:10:02Znormalized-recipewww.localmaxxing.com/en/runs/cmt0vip5k0fvgms01179emtv0 ↗