Recipe

glm-5-3-flash-unsloth-dynamic-iq4-xs-rtx-3090-24gb-llama-cpp-tp8

glm-5-3-flash-unsloth-dynamic-iq4-xs-rtx-3090-24gb-llama-cpp-tp8

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

Record

Status
candidate
Source
localmaxxing
Engine
llama.cpp
Engine version
b10667-2e0e57f1
Accelerators
8
Tensor parallel
8
Context tokens
131,072
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/cmtc0g6680041qq01j6ruzjdg

Observed source tokens

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

  1. /mnt/nvme/src/llama.cpp-glm53-mtp-candidate/build-cuda-nccl/bin/llama-server
  2. --model
  3. /mnt/nvme/models/unsloth/GLM-5.3-Flash-GGUF/UD-IQ4_XS/GLM-5.3-Flash-UD-IQ4_XS-00001-of-00005.gguf
  4. --alias
  5. glm-5.3-flash-ud-iq4-xs
  6. --device
  7. CUDA5,CUDA6,CUDA7,CUDA4,CUDA3,CUDA2,CUDA1,CUDA0
  8. --split-mode
  9. layer
  10. --tensor-split
  11. 8,5,6,5,5,6,6,6
  12. --n-gpu-layers
  13. all
  14. --n-cpu-moe
  15. 0
  16. --fit
  17. off
  18. --load-mode
  19. mmap
  20. --ctx-size
  21. 131072
  22. --parallel
  23. 1
  24. --no-kv-unified
  25. --batch-size
  26. 2048
  27. --ubatch-size
  28. 32
  29. --cache-type-k
  30. f16
  31. --cache-type-v
  32. f16
  33. --flash-attn
  34. off
  35. --threads
  36. 8
  37. --threads-batch
  38. 24
  39. --cache-prompt
  40. --cache-ram
  41. 0
  42. --no-cache-idle-slots
  43. --ctx-checkpoints
  44. 4
  45. --checkpoint-min-step
  46. 8192
  47. --metrics
  48. --slots
  49. --host
  50. 192.168.178.61
  51. --port
  52. 18082
  53. --temp
  54. 0
  55. --top-p
  56. 1
  57. --top-k
  58. 0
  59. --min-p
  60. 0
  61. --reasoning
  62. on
  63. --reasoning-effort
  64. low
  65. --spec-type
  66. draft-mtp
  67. --spec-draft-n-max
  68. 2
  69. --spec-draft-n-min
  70. 1
  71. --no-spec-draft-backend-sampling
FlagValue
--model/mnt/nvme/models/unsloth/GLM-5.3-Flash-GGUF/UD-IQ4_XS/GLM-5.3-Flash-UD-IQ4_XS-00001-of-00005.gguf
--aliasglm-5.3-flash-ud-iq4-xs
--deviceCUDA5,CUDA6,CUDA7,CUDA4,CUDA3,CUDA2,CUDA1,CUDA0
--split-modelayer
--tensor-split8,5,6,5,5,6,6,6
--n-gpu-layersall
--n-cpu-moe0
--fitoff
--load-modemmap
--ctx-size131072
--parallel1
--batch-size2048
--ubatch-size32
--cache-type-kf16
--cache-type-vf16
--flash-attnoff
--threads8
--threads-batch24
--cache-ram0
--ctx-checkpoints4
--checkpoint-min-step8192
--host192.168.178.61
--port18082
--temp0
--top-p1
--top-k0
--min-p0
--reasoningon
--reasoning-effortlow
--spec-typedraft-mtp
--spec-draft-n-max2
--spec-draft-n-min1

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
1131,072148.442.12,460observedglm-5-3-flash-unsloth-dynamic-iq4-xs-rtx-3090-24gb-llama-cpp-tp8-sweep

Remaining fields

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

hardware count
8
hardware id
rtx-3090-24gb
id
glm-5-3-flash-unsloth-dynamic-iq4-xs-rtx-3090-24gb-llama-cpp-tp8
model instance id
unsloth-glm-5-3-flash-gguf--unsloth-dynamic-iq4-xs
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
glm-5-3-flash-unsloth-dynamic-iq4-xs-rtx-3090-24gb-llama-cpp-tp8-sweep
status
candidate

capabilities

engine

name
llama.cpp
version
b10667-2e0e57f1

serving

max concurrency
1
max context tokens
131,072
tensor parallel
8
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
cuda
localmaxxing · hardware label
RTX 3090
localmaxxing · notes
Exact Unsloth UD-IQ4_XS GGUF; custom MTP candidate binary; quant repo revision ac47690c15c8. Two warmups and six sequential 512-token measured runs at temperature 0. Server --cache-prompt was enabled, but cache_prompt=false was sent per request; cached_tokens=0 and timings.cache_n=0 in every measured run. tokSTotal follows the current API wording: (prompt+output tokens)/client end-to-end seconds. Temperature-zero output hashes: 1 unique across 6 runs. MTP draft acceptance: 1752/2616 (0.6697). Service PID/config stable=True; benchmark incidents=0; systemd restarts=0; suspicious benchmark-window log lines=0. Per-GPU power values are medians of each run's mean active draw; peak VRAM is summed across GPUs.
localmaxxing · observed command
/mnt/nvme/src/llama.cpp-glm53-mtp-candidate/build-cuda-nccl/bin/llama-server --model /mnt/nvme/models/unsloth/GLM-5.3-Flash-GGUF/UD-IQ4_XS/GLM-5.3-Flash-UD-IQ4_XS-00001-of-00005.gguf --alias glm-5.3-flash-ud-iq4-xs --device CUDA5,CUDA6,CUDA7,CUDA4,CUDA3,CUDA2,CUDA1,CUDA0 --split-mode layer --tensor-split 8,5,6,5,5,6,6,6 --n-gpu-layers all --n-cpu-moe 0 --fit off --load-mode mmap --ctx-size 131072 --parallel 1 --no-kv-unified --batch-size 2048 --ubatch-size 32 --cache-type-k f16 --cache-type-v f16 --flash-attn off --threads 8 --threads-batch 24 --cache-prompt --cache-ram 0 --no-cache-idle-slots --ctx-checkpoints 4 --checkpoint-min-step 8192 --metrics --slots --host 192.168.178.61 --port 18082 --temp 0 --top-p 1 --top-k 0 --min-p 0 --reasoning on --reasoning-effort low --spec-type draft-mtp --spec-draft-n-max 2 --spec-draft-n-min 1 --no-spec-draft-backend-sampling
localmaxxing · run id
cmtc0g6680041qq01j6ruzjdg
localmaxxing · tokenized · arguments
/mnt/nvme/src/llama.cpp-glm53-mtp-candidate/build-cuda-nccl/bin/llama-server, --model, /mnt/nvme/models/unsloth/GLM-5.3-Flash-GGUF/UD-IQ4_XS/GLM-5.3-Flash-UD-IQ4_XS-00001-of-00005.gguf, --alias, glm-5.3-flash-ud-iq4-xs, --device, CUDA5,CUDA6,CUDA7,CUDA4,CUDA3,CUDA2,CUDA1,CUDA0, --split-mode, layer, --tensor-split, 8,5,6,5,5,6,6,6, --n-gpu-layers, all, --n-cpu-moe, 0, --fit, off, --load-mode, mmap, --ctx-size, 131072, --parallel, 1, --no-kv-unified, --batch-size, 2048, --ubatch-size, 32, --cache-type-k, f16, --cache-type-v, f16, --flash-attn, off, --threads, 8, --threads-batch, 24, --cache-prompt, --cache-ram, 0, --no-cache-idle-slots, --ctx-checkpoints, 4, --checkpoint-min-step, 8192, --metrics, --slots, --host, 192.168.178.61, --port, 18082, --temp, 0, --top-p, 1, --top-k, 0, --min-p, 0, --reasoning, on, --reasoning-effort, low, --spec-type, draft-mtp, --spec-draft-n-max, 2, --spec-draft-n-min, 1, --no-spec-draft-backend-sampling
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/cmtc0g6680041qq01j6ruzjdg