Local AI Registry

Recipe

deepseek-v4-flash-0731-ud-iq2-m-rtx-3090-24gb-llama-cpp-tp4

deepseek-v4-flash-0731-ud-iq2-m-rtx-3090-24gb-llama-cpp-tp4
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
69 (fork ds4-longctx @ e6718011c)
Accelerators
4
Tensor parallel
4
Context tokens
131,072
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/cmsi0l2sc01lepp019uelu1ga

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. DeepSeek-V4-Flash-0731-UD-IQ2_M-00001-of-00003.gguf
  4. -ngl
  5. 999
  6. -ts
  7. 1,1,0.95,1.05
  8. --split-mode
  9. layer
  10. --flash-attn
  11. on
  12. --no-repack
  13. --ctx-size
  14. 131072
  15. --batch-size
  16. 8192
  17. --ubatch-size
  18. 384
  19. --ctx-checkpoints
  20. 1
  21. -t
  22. 8
  23. --parallel
  24. 1
  25. --jinja
  26. --reasoning-budget
  27. 1024
  28. --reasoning-budget-message
  29. Okay, that is enough thinking. Let me get to work.
FlagValue
-mDeepSeek-V4-Flash-0731-UD-IQ2_M-00001-of-00003.gguf
-ngl999
-ts1,1,0.95,1.05
--split-modelayer
--flash-attnon
--ctx-size131072
--batch-size8192
--ubatch-size384
--ctx-checkpoints1
-t8
--parallel1
--reasoning-budget1024
--reasoning-budget-messageOkay, that is enough thinking. Let me get to work.

Environment

VariableValue
CUDA_VISIBLE_DEVICES0,1,2,3
DSV4_AGENT_CKPT_TAIL1
DSV4_CONSTANT_SHAPE1
DSV4_DECODE_FUSED_IDX1
DSV4_DECODE_RADIX_TOPK1
DSV4_FA_UNION1
DSV4_GLU_FUSE1
DSV4_IDX_SKIP1
DSV4_MMVQ_SMALLK1
DSV4_MOE_FUSE1
DSV4_MOE_RESIDENT1
DSV4_MOE_TILE1
DSV4_PREFILL_GRAPHS1
DSV4_PREFILL_RADIX_TOPK1
DSV4_SPARSE_FA1
DSV4_STABLE_TOPO1
GGML_CUDA_P2P1
GGML_GALLOC_STICKY1

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
1131,0721,745.441.118,774observeddeepseek-v4-flash-0731-ud-iq2-m-rtx-3090-24gb-llama-cpp-tp4-sweep

Remaining fields

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

hardware count
4
hardware id
rtx-3090-24gb
id
deepseek-v4-flash-0731-ud-iq2-m-rtx-3090-24gb-llama-cpp-tp4
model instance id
unsloth-deepseek-v4-flash-0731-gguf--ud-iq2-m
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
deepseek-v4-flash-0731-ud-iq2-m-rtx-3090-24gb-llama-cpp-tp4-sweep
status
candidate

capabilities

engine

name
llama.cpp
version
69 (fork ds4-longctx @ e6718011c)

serving

max concurrency
1
max context tokens
131,072
tensor parallel
4
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
cuda
localmaxxing · batch size
1
localmaxxing · hardware label
RTX 3090
localmaxxing · notes
DeepSeek-V4-Flash-0731 (304B MoE, Unsloth UD-IQ2_M, ~90GB GGUF) on 4x RTX 3090 24GB (96GB total, ~90.3GiB used at 130K ctx). Engine: llama.cpp fork with DeepSeek-V4-Flash optimizations - github.com/alesha-pro/llama.cpp branch ds4-longctx @ e6718011c. Launcher: scripts/ds4-prod-serve.sh (full env flag set in commandSnippet). Custom work vs upstream: prefill CUDA graphs, sparse FA + indexer skip, fused MoE (GLU fuse + tiling + resident experts), radix top-k, register-based MLA head-compression sinkhorn kernel, 2-pass fp8 KV quantize kernel, sticky graph allocator, startup warm pass (moves allocator climb out of first request). GPU tuning: PL 270W/card + locked SM clocks (scripts/ds4-gpu-clocks.sh). Full story: DS4HANDOFF.md in the repo. Measured 2026-08-06, server fully warm. This run: fresh 32768-token prompt -> prefill 1745.4 t/s (TTFT 18.8s, ~265W/card), then decode 41.07 t/s at 32K depth (~210W/card). Same config decodes 44.6 t/s at short context.
localmaxxing · observed command
CUDA_VISIBLE_DEVICES=0,1,2,3 GGML_CUDA_P2P=1 DSV4_CONSTANT_SHAPE=1 DSV4_SPARSE_FA=1 DSV4_FA_UNION=1 DSV4_IDX_SKIP=1 DSV4_MOE_TILE=1 DSV4_MOE_RESIDENT=1 DSV4_GLU_FUSE=1 DSV4_MOE_FUSE=1 DSV4_DECODE_FUSED_IDX=1 DSV4_DECODE_RADIX_TOPK=1 DSV4_PREFILL_RADIX_TOPK=1 DSV4_MMVQ_SMALLK=1 DSV4_STABLE_TOPO=1 GGML_GALLOC_STICKY=1 DSV4_PREFILL_GRAPHS=1 DSV4_AGENT_CKPT_TAIL=1 llama-server -m DeepSeek-V4-Flash-0731-UD-IQ2_M-00001-of-00003.gguf -ngl 999 -ts 1,1,0.95,1.05 --split-mode layer --flash-attn on --no-repack --ctx-size 131072 --batch-size 8192 --ubatch-size 384 --ctx-checkpoints 1 -t 8 --parallel 1 --jinja --reasoning-budget 1024 --reasoning-budget-message 'Okay, that is enough thinking. Let me get to work.'
localmaxxing · run id
cmsi0l2sc01lepp019uelu1ga
localmaxxing · tokenized · arguments
llama-server, -m, DeepSeek-V4-Flash-0731-UD-IQ2_M-00001-of-00003.gguf, -ngl, 999, -ts, 1,1,0.95,1.05, --split-mode, layer, --flash-attn, on, --no-repack, --ctx-size, 131072, --batch-size, 8192, --ubatch-size, 384, --ctx-checkpoints, 1, -t, 8, --parallel, 1, --jinja, --reasoning-budget, 1024, --reasoning-budget-message, Okay, that is enough thinking. Let me get to work.
localmaxxing · tokenized · environment · CUDA VISIBLE DEVICES
0,1,2,3
localmaxxing · tokenized · environment · DSV4 AGENT CKPT TAIL
1
localmaxxing · tokenized · environment · DSV4 CONSTANT SHAPE
1
localmaxxing · tokenized · environment · DSV4 DECODE FUSED IDX
1
localmaxxing · tokenized · environment · DSV4 DECODE RADIX TOPK
1
localmaxxing · tokenized · environment · DSV4 FA UNION
1
localmaxxing · tokenized · environment · DSV4 GLU FUSE
1
localmaxxing · tokenized · environment · DSV4 IDX SKIP
1
localmaxxing · tokenized · environment · DSV4 MMVQ SMALLK
1
localmaxxing · tokenized · environment · DSV4 MOE FUSE
1
localmaxxing · tokenized · environment · DSV4 MOE RESIDENT
1
localmaxxing · tokenized · environment · DSV4 MOE TILE
1
localmaxxing · tokenized · environment · DSV4 PREFILL GRAPHS
1
localmaxxing · tokenized · environment · DSV4 PREFILL RADIX TOPK
1
localmaxxing · tokenized · environment · DSV4 SPARSE FA
1
localmaxxing · tokenized · environment · DSV4 STABLE TOPO
1
localmaxxing · tokenized · environment · GGML CUDA P2P
1
localmaxxing · tokenized · environment · GGML GALLOC STICKY
1
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/cmsi0l2sc01lepp019uelu1ga ↗