Recipe

nvidia-nemotron-labs-3-puzzle-75b-a9b-bf16-ud-iq4-xl-intel-arc-pro-b70-32gb-llama-cpp-tp2

nvidia-nemotron-labs-3-puzzle-75b-a9b-bf16-ud-iq4-xl-intel-arc-pro-b70-32gb-llama-cpp-tp2

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

Record

Status
candidate
Source
localmaxxing
Engine
llama.cpp
Engine version
fork YanissAmz/llama.cpp branch puzzle-port (base b566325) + local SSM_SCAN d_state=96 Vulkan patch (required for this arch)
Accelerators
2
Tensor parallel
2
Context tokens
2,048
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16
Repository
nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16
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/cmrwextc6009mnz012vxmsj3d

No tokenized launch fields. This is measured or documented compatibility, not a Docker launch.

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
12,0481,046.239.1observednvidia-nemotron-labs-3-puzzle-75b-a9b-bf16-ud-iq4-xl-intel-arc-pro-b70-32gb-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
intel-arc-pro-b70-32gb
id
nvidia-nemotron-labs-3-puzzle-75b-a9b-bf16-ud-iq4-xl-intel-arc-pro-b70-32gb-llama-cpp-tp2
model instance id
nvidia-nvidia-nemotron-labs-3-puzzle-75b-a9b-bf16--ud-iq4-xl
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
nvidia-nemotron-labs-3-puzzle-75b-a9b-bf16-ud-iq4-xl-intel-arc-pro-b70-32gb-llama-cpp-tp2-sweep
status
candidate

capabilities

engine

name
llama.cpp
version
fork YanissAmz/llama.cpp branch puzzle-port (base b566325) + local SSM_SCAN d_state=96 Vulkan patch (required for this arch)

serving

max context tokens
2,048
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-not-evidenced
serving.max concurrency · state unknown

metadata

localmaxxing · backend
vulkan
localmaxxing · batch size
1
localmaxxing · hardware label
Intel Arc Pro B70
localmaxxing · notes
First public run of this model? nemotron_h_moe (hybrid Mamba2+attn+MoE, 75B/9B-active) cannot load on mainline llama.cpp; requires the puzzle-port fork PLUS a local 4-file patch adding SSM_SCAN d_state=96 Vulkan pipelines (without it: ~3.5 t/s CPU fallback; with it: 39 t/s, ~11x). Decode: llama-bench tg64 r=5, 7 clean reps 38.90-39.33. Prefill: pp2048, tuned -ub 1024 -ts 0.9/1 (default gives 947). -sm layer only (tensor-split gated off upstream for hybrid archs; Vulkan lacks row-split). Sched pipeline parallelism active (copies=4). GPUs are SR-IOV VFs: power draw not measurable (sysman PF-only). Sustained-serving caveat: long-lived multi-GPU decode sags ~4x over ~6 requests without a periodic chunky GPU burst (clock-governor behavior); single-GPU immune. Mesa 26.1.4 (kisak), kernel 7.0.0-28.
localmaxxing · run id
cmrwextc6009mnz012vxmsj3d

provenance

captured at
2026-08-30T09:10:02Z

sources

captured atkindurl
2026-08-30T09:10:02Znormalized-recipewww.localmaxxing.com/en/runs/cmrwextc6009mnz012vxmsj3d