Local AI Registry

Recipe

glm-5-3-flash-ud-q2-k-xl-radeon-ai-pro-r9700-32gb-llama-cpp-tp1

glm-5-3-flash-ud-q2-k-xl-radeon-ai-pro-r9700-32gb-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
2e0e57f10
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/GLM-5.3-Flash-GGUF
Repository
unsloth/GLM-5.3-Flash-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/cmtcu7175000arw01u33ofyw2

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. -hf
  3. unsloth/GLM-5.3-Flash-GGUF:UD-Q2_K_XL
  4. --device
  5. ROCm0,ROCm1
  6. --split-mode
  7. layer
  8. --tensor-split
  9. 35,82
  10. --ctx-size
  11. 32768
  12. --parallel
  13. 1
  14. --reasoning
  15. off
  16. --reasoning-budget
  17. 0
  18. --override-tensor
  19. blk.*.ffn_gate_shexp.weight=ROCm0,blk.*.ffn_up_shexp.weight=ROCm0,blk.*.ffn_down_shexp.weight=ROCm0
  20. --metrics
  21. --jinja
FlagValue
-hfunsloth/GLM-5.3-Flash-GGUF:UD-Q2_K_XL
--deviceROCm0,ROCm1
--split-modelayer
--tensor-split35,82
--ctx-size32768
--parallel1
--reasoningoff
--reasoning-budget0
--override-tensorblk.*.ffn_gate_shexp.weight=ROCm0,blk.*.ffn_up_shexp.weight=ROCm0,blk.*.ffn_down_shexp.weight=ROCm0

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
132,76836.616.6785.7observedglm-5-3-flash-ud-q2-k-xl-radeon-ai-pro-r9700-32gb-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
radeon-ai-pro-r9700-32gb
id
glm-5-3-flash-ud-q2-k-xl-radeon-ai-pro-r9700-32gb-llama-cpp-tp1
model instance id
unsloth-glm-5-3-flash-gguf--ud-q2-k-xl
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
glm-5-3-flash-ud-q2-k-xl-radeon-ai-pro-r9700-32gb-llama-cpp-tp1-sweep
status
candidate

capabilities

engine

name
llama.cpp
version
2e0e57f10

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

metadata

localmaxxing · backend
rocm
localmaxxing · hardware label
Radeon AI Pro R9700
localmaxxing · notes
Tuned GLM-5.3-Flash UD-Q2_K_XL benchmark on Lucebox. Exact live configuration: --device ROCm0,ROCm1 --split-mode layer --tensor-split 35,82 --ctx-size 32768 --parallel 1 --reasoning off --reasoning-budget 0 --override-tensor blk.*.ffn_gate_shexp.weight=ROCm0,blk.*.ffn_up_shexp.weight=ROCm0,blk.*.ffn_down_shexp.weight=ROCm0 --metrics --jinja. The structured hardware record uses the AMD Radeon AI PRO R9700 (32 GB); the model was layer-split across the R9700 (ROCm0) and AMD Radeon 8060S unified-memory GPU (ROCm1) over M.2 OCuLink PCIe 4.0 x4. One slot, 32K context, temperature 0, reasoning disabled, cache_prompt=false. Three measured streamed requests after one warmup using the fixed 26-token count prompt and 64-token output cap; all three outputs matched. Peak workload allocation is measured from amdgpu DRM per-process VRAM+GTT counters across both GPUs and reported in GiB (101.919 GiB), not device capacity. Build/runtime: llama.cpp GLM-5-Next support PR commit 2e0e57f10. This is a short fixed-prompt comparison measurement; TTFT and total throughput are client-derived from monotonic stream timing, while prefill/decode rates are server-authoritative.
localmaxxing · observed command
llama-server -hf unsloth/GLM-5.3-Flash-GGUF:UD-Q2_K_XL --device ROCm0,ROCm1 --split-mode layer --tensor-split 35,82 --ctx-size 32768 --parallel 1 --reasoning off --reasoning-budget 0 --override-tensor blk.*.ffn_gate_shexp.weight=ROCm0,blk.*.ffn_up_shexp.weight=ROCm0,blk.*.ffn_down_shexp.weight=ROCm0 --metrics --jinja
localmaxxing · run id
cmtcu7175000arw01u33ofyw2
localmaxxing · tokenized · arguments
llama-server, -hf, unsloth/GLM-5.3-Flash-GGUF:UD-Q2_K_XL, --device, ROCm0,ROCm1, --split-mode, layer, --tensor-split, 35,82, --ctx-size, 32768, --parallel, 1, --reasoning, off, --reasoning-budget, 0, --override-tensor, blk.*.ffn_gate_shexp.weight=ROCm0,blk.*.ffn_up_shexp.weight=ROCm0,blk.*.ffn_down_shexp.weight=ROCm0, --metrics, --jinja
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/cmtcu7175000arw01u33ofyw2 ↗