Local AI Registry

Recipe

gemma-4-26b-a4b-q8-k-xl-radeon-ai-pro-r9700-32gb-llama-cpp-tp2

gemma-4-26b-a4b-q8-k-xl-radeon-ai-pro-r9700-32gb-llama-cpp-tp2
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
5ea1b124e + stew675 01/10/02
Accelerators
2
Tensor parallel
2
Context tokens
262,144
Max concurrency
1
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/google/gemma-4-26B-A4B-it
Repository
google/gemma-4-26B-A4B-it
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/cmtewcr4300b2n001qknqnn05

Observed source tokens

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

  1. llama-server
  2. -m
  3. gemma-4-26B-A4B-it-UD-Q8_K_XL.gguf
  4. --host
  5. 127.0.0.1
  6. --port
  7. 18080
  8. --device
  9. ROCm0,ROCm1
  10. --split-mode
  11. tensor
  12. --tensor-split
  13. 1,1
  14. -ngl
  15. 99
  16. -c
  17. 262144
  18. -fit
  19. off
  20. --cache-ram
  21. 0
  22. --mmap
  23. --no-warmup
  24. --flash-attn
  25. on
  26. --cache-type-k
  27. q8_0
  28. --cache-type-v
  29. q8_0
  30. --parallel
  31. 1
  32. --n-predict
  33. 4096
  34. --spec-type
  35. draft-mtp
  36. --spec-draft-model
  37. gemma-4-26B-A4B-it-Q8_0-MTP.gguf
  38. --spec-draft-device
  39. ROCm0,ROCm1
  40. --spec-draft-ngl
  41. 99
  42. --spec-draft-type-k
  43. q8_0
  44. --spec-draft-type-v
  45. q8_0
  46. --spec-draft-n-max
  47. 3
  48. --spec-draft-p-min
  49. 0.0
FlagValue
-mgemma-4-26B-A4B-it-UD-Q8_K_XL.gguf
--host127.0.0.1
--port18080
--deviceROCm0,ROCm1
--split-modetensor
--tensor-split1,1
-ngl99
-c262144
-fitoff
--cache-ram0
--flash-attnon
--cache-type-kq8_0
--cache-type-vq8_0
--parallel1
--n-predict4096
--spec-typedraft-mtp
--spec-draft-modelgemma-4-26B-A4B-it-Q8_0-MTP.gguf
--spec-draft-deviceROCm0,ROCm1
--spec-draft-ngl99
--spec-draft-type-kq8_0
--spec-draft-type-vq8_0
--spec-draft-n-max3
--spec-draft-p-min0.0

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
1262,1441,974.7131.4—observedgemma-4-26b-a4b-q8-k-xl-radeon-ai-pro-r9700-32gb-llama-cpp-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
radeon-ai-pro-r9700-32gb
id
gemma-4-26b-a4b-q8-k-xl-radeon-ai-pro-r9700-32gb-llama-cpp-tp2
model instance id
google-gemma-4-26b-a4b-it--q8-k-xl
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
gemma-4-26b-a4b-q8-k-xl-radeon-ai-pro-r9700-32gb-llama-cpp-tp2-sweep
status
candidate

capabilities

engine

name
llama.cpp
version
5ea1b124e + stew675 01/10/02

serving

max concurrency
1
max context tokens
262,144
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
serving.max concurrency · provenance · captured at
2026-08-31T23:03:15Z
serving.max concurrency · reason server-capacity-derived-from-source-evidence
serving.max concurrency · state known

metadata

localmaxxing · backend
rocm
localmaxxing · batch size
1
localmaxxing · hardware label
Radeon AI Pro R9700
localmaxxing · notes
Unsloth UD-Q8_K_XL target + sidecar Q8_0-MTP draft (no in-file nextn). NOT llama-benchy. tokSOut = median of 3 essay decode-512 legs (131.39/91.94/148.31; accept 0.81/0.46/0.94). Code-prompt median was 121.95 (not submitted as tokSOut). tokSPrefill = ~10k prompt (10270 tok, 1974.69 t/s); 8k=1568, 16k=2184, 40k=2097, 80k=1678. VOID=false 512/512; draft_n present. Dual R9700 tensor-split 1,1 (Raphael iGPU unused). HIP overlay libamdhip64.so.7.2.53211. llama 5ea1b124e build-rocm-rdna-02, FA on, KV q8_0, mmap, draft-mtp n-max3 p-min0 draft on ROCm0,ROCm1 (ROCm0-only Meta() abort). HIP PR: ROCm/rocm-systems#10940. Recipe: github.com/onchainapps/rocm-llama.cpp
localmaxxing · observed command
llama-server -m gemma-4-26B-A4B-it-UD-Q8_K_XL.gguf --host 127.0.0.1 --port 18080 --device ROCm0,ROCm1 --split-mode tensor --tensor-split 1,1 -ngl 99 -c 262144 -fit off --cache-ram 0 --mmap --no-warmup --flash-attn on --cache-type-k q8_0 --cache-type-v q8_0 --parallel 1 --n-predict 4096 --spec-type draft-mtp --spec-draft-model gemma-4-26B-A4B-it-Q8_0-MTP.gguf --spec-draft-device ROCm0,ROCm1 --spec-draft-ngl 99 --spec-draft-type-k q8_0 --spec-draft-type-v q8_0 --spec-draft-n-max 3 --spec-draft-p-min 0.0
localmaxxing · run id
cmtewcr4300b2n001qknqnn05
localmaxxing · tokenized · arguments
llama-server, -m, gemma-4-26B-A4B-it-UD-Q8_K_XL.gguf, --host, 127.0.0.1, --port, 18080, --device, ROCm0,ROCm1, --split-mode, tensor, --tensor-split, 1,1, -ngl, 99, -c, 262144, -fit, off, --cache-ram, 0, --mmap, --no-warmup, --flash-attn, on, --cache-type-k, q8_0, --cache-type-v, q8_0, --parallel, 1, --n-predict, 4096, --spec-type, draft-mtp, --spec-draft-model, gemma-4-26B-A4B-it-Q8_0-MTP.gguf, --spec-draft-device, ROCm0,ROCm1, --spec-draft-ngl, 99, --spec-draft-type-k, q8_0, --spec-draft-type-v, q8_0, --spec-draft-n-max, 3, --spec-draft-p-min, 0.0
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/cmtewcr4300b2n001qknqnn05 ↗