Recipe

deepseek-v4-flash-0731-ud-iq4-xs-rtx-3090-24gb-llama-cpp-tp7

deepseek-v4-flash-0731-ud-iq4-xs-rtx-3090-24gb-llama-cpp-tp7

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

Record

Status
candidate
Source
localmaxxing
Engine
llama.cpp
Engine version
b10270-071327508
Accelerators
7
Tensor parallel
7
Context tokens
262,144
Max concurrency
1
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/unsloth/DeepSeek-V4-Flash-0731-GGUF
Repository
unsloth/DeepSeek-V4-Flash-0731-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/cmshcu8v301fqpp01htchp8sl

Observed source tokens

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

  1. /home/thread/src/llama.cpp-dsv4-iq4-dspark-0713275/build-cuda-sm86/bin/llama-server
  2. --model
  3. /mnt/nvme/models/unsloth/DeepSeek-V4-Flash-0731-GGUF/UD-IQ4_XS/DeepSeek-V4-Flash-0731-UD-IQ4_XS-00001-of-00004.gguf
  4. --model-draft
  5. /mnt/nvme/models/unsloth/DeepSeek-V4-Flash-0731-GGUF/dspark/dspark-DeepSeek-V4-Flash-0731-Q8_0.gguf
  6. --alias
  7. deepseek-v4-flash-0731-iq4-xs-dspark
  8. --host
  9. 0.0.0.0
  10. --port
  11. 18081
  12. --device
  13. CUDA0,CUDA1,CUDA2,CUDA3,CUDA4,CUDA5,CUDA6
  14. --split-mode
  15. layer
  16. --fit
  17. on
  18. --fit-target
  19. 4000
  20. --fit-ctx
  21. 262144
  22. --spec-type
  23. draft-dspark
  24. --spec-draft-n-max
  25. 3
  26. --jinja
  27. --temp
  28. 1.0
  29. --parallel
  30. 1
  31. --no-mmap
  32. --batch-size
  33. 2048
  34. --ubatch-size
  35. 512
  36. --cache-type-k
  37. f16
  38. --cache-type-v
  39. f16
  40. --top-p
  41. 1.0
  42. --min-p
  43. 0.0
  44. --flash-attn
  45. on
  46. --no-warmup
  47. --cache-prompt
  48. --cache-ram
  49. 2048
  50. --cache-idle-slots
  51. --slots
  52. --metrics
  53. --no-webui
FlagValue
--model/mnt/nvme/models/unsloth/DeepSeek-V4-Flash-0731-GGUF/UD-IQ4_XS/DeepSeek-V4-Flash-0731-UD-IQ4_XS-00001-of-00004.gguf
--model-draft/mnt/nvme/models/unsloth/DeepSeek-V4-Flash-0731-GGUF/dspark/dspark-DeepSeek-V4-Flash-0731-Q8_0.gguf
--aliasdeepseek-v4-flash-0731-iq4-xs-dspark
--host0.0.0.0
--port18081
--deviceCUDA0,CUDA1,CUDA2,CUDA3,CUDA4,CUDA5,CUDA6
--split-modelayer
--fiton
--fit-target4000
--fit-ctx262144
--spec-typedraft-dspark
--spec-draft-n-max3
--temp1.0
--parallel1
--batch-size2048
--ubatch-size512
--cache-type-kf16
--cache-type-vf16
--top-p1.0
--min-p0.0
--flash-attnon
--cache-ram2048

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
1262,144413.150.8849.8observeddeepseek-v4-flash-0731-ud-iq4-xs-rtx-3090-24gb-llama-cpp-tp7-sweep

Remaining fields

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

hardware count
7
hardware id
rtx-3090-24gb
id
deepseek-v4-flash-0731-ud-iq4-xs-rtx-3090-24gb-llama-cpp-tp7
model instance id
unsloth-deepseek-v4-flash-0731-gguf--ud-iq4-xs
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
deepseek-v4-flash-0731-ud-iq4-xs-rtx-3090-24gb-llama-cpp-tp7-sweep
status
candidate

capabilities

engine

name
llama.cpp
version
b10270-071327508

serving

max concurrency
1
max context tokens
262,144
tensor parallel
7
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
6 valid measured runs in two 3-run batches, each after 2 warmups; fixed standard prompt; temperature 0; streaming. Server prefix cache enabled, but cache_prompt=false per request; cached_tokens=0/cache_n=0 verified each run. tokSOut=median llama.cpp predicted tok/s; tokSPrefill=promptTokens/client TTFT; tokSTotal=outputTokens/client end-to-end time; power=median per-GPU mean during requests. DSpark Q8_0 speculative draft, n_max=3. Temp-zero output hash stable. Reliability caveat: an initial fourth consecutive 512-token run crashed llama.cpp in ggml_cuda_mul_mat_cublas with CUBLAS_STATUS_INVALID_VALUE; systemd recovered it. The final six samples used one planned clean restart between the two 3-run batches to avoid a repeat.
localmaxxing · observed command
/home/thread/src/llama.cpp-dsv4-iq4-dspark-0713275/build-cuda-sm86/bin/llama-server --model /mnt/nvme/models/unsloth/DeepSeek-V4-Flash-0731-GGUF/UD-IQ4_XS/DeepSeek-V4-Flash-0731-UD-IQ4_XS-00001-of-00004.gguf --model-draft /mnt/nvme/models/unsloth/DeepSeek-V4-Flash-0731-GGUF/dspark/dspark-DeepSeek-V4-Flash-0731-Q8_0.gguf --alias deepseek-v4-flash-0731-iq4-xs-dspark --host 0.0.0.0 --port 18081 --device CUDA0,CUDA1,CUDA2,CUDA3,CUDA4,CUDA5,CUDA6 --split-mode layer --fit on --fit-target 4000 --fit-ctx 262144 --spec-type draft-dspark --spec-draft-n-max 3 --jinja --temp 1.0 --parallel 1 --no-mmap --batch-size 2048 --ubatch-size 512 --cache-type-k f16 --cache-type-v f16 --top-p 1.0 --min-p 0.0 --flash-attn on --no-warmup --cache-prompt --cache-ram 2048 --cache-idle-slots --slots --metrics --no-webui
localmaxxing · run id
cmshcu8v301fqpp01htchp8sl
localmaxxing · tokenized · arguments
/home/thread/src/llama.cpp-dsv4-iq4-dspark-0713275/build-cuda-sm86/bin/llama-server, --model, /mnt/nvme/models/unsloth/DeepSeek-V4-Flash-0731-GGUF/UD-IQ4_XS/DeepSeek-V4-Flash-0731-UD-IQ4_XS-00001-of-00004.gguf, --model-draft, /mnt/nvme/models/unsloth/DeepSeek-V4-Flash-0731-GGUF/dspark/dspark-DeepSeek-V4-Flash-0731-Q8_0.gguf, --alias, deepseek-v4-flash-0731-iq4-xs-dspark, --host, 0.0.0.0, --port, 18081, --device, CUDA0,CUDA1,CUDA2,CUDA3,CUDA4,CUDA5,CUDA6, --split-mode, layer, --fit, on, --fit-target, 4000, --fit-ctx, 262144, --spec-type, draft-dspark, --spec-draft-n-max, 3, --jinja, --temp, 1.0, --parallel, 1, --no-mmap, --batch-size, 2048, --ubatch-size, 512, --cache-type-k, f16, --cache-type-v, f16, --top-p, 1.0, --min-p, 0.0, --flash-attn, on, --no-warmup, --cache-prompt, --cache-ram, 2048, --cache-idle-slots, --slots, --metrics, --no-webui
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/cmshcu8v301fqpp01htchp8sl