xM_My (@NullBurn) profile picture

Local LLM

DGX Spark 2x
Joined June 2013
发现 Muse App 居然藏了 slash 命令,之前一直以为没有 /clear — 下一条消息起重置上下文 /compact — 主动压缩上下文,不用等自动触发 /status — 查看当前会话状态 /usage on|off — 开关 token 用量显示 /subagents — 查看子代理 在输入框打 / 就能弹出来
1
1
228
Alma这么强?排第一!吊打codex 、pi、claude code!
You’re unable to view this Post because this account owner limits who can view their Posts. Learn more
85
2020: Obsidian, Evernote, Joplin, Logseq, Roam, Notion, OneNote 2026: Rex, Tern, Herdr, Paseo, T3 Code, Orca, Claude Code, Codex, pi, hermes Six years. New tools, same guy.
1
1
390
GPU clock cap alone isn't enough — cap the CPU too. On 2x DGX Spark (GB10) running DeepSeek-V4.1-Flash-1017, I capped the Cortex-X925 performance cores from 3900 to 2808 MHz and left the A725 efficiency cores alone. What changed: · 10-core CPU load: SoC peak 94°C → 74°C · prefill 32k: 2082 → 1984 tok/s (-5%) · decode x4: 95.3 → 94.6 tok/s (within noise) The GPU cap is nvidia-smi -lgc 0,2000. Why it works: prefill/decode here are GPU/bandwidth bound, so a 28% drop in pure CPU throughput costs almost nothing in inference. The CPU cap is a fuse for when software pegs the CPU, not a performance knob.
9
10
2
86
7,964
代价与边界: · 纯 CPU 吞吐 -28%(3900→2808 的线性结果) · 推理 prefill -5%、decode 噪声内——瓶颈在 GPU/带宽,不在 CPU · 空闲温度基本不变,热量主要不来自 CPU 核 所以:GPU cap 是功耗上限,CPU cap 是“软件把 CPU 跑飞”时的保险丝。
10
最大的坑:默认 auto_select=1,固件自己选频。 此时 scaling_max_freq 只是请求值,硬件照样 3.9 GHz,而 scaling_cur_freq 却显示 2808000,骗你以为生效。 必须先 echo 0 > auto_select 再写 max。验证要在负载下看 cpuinfo_avg_freq。
6
做法: GPU:nvidia-smi -lgc 0,2000 CPU 只压 X925: echo 0 > $p/auto_select echo 2808000 > $p/scaling_max_freq 遍历 policy*,只处理 cpuinfo_max_freq > 2808000 的核。包成 systemd oneshot 即可开机生效、一键还原。
78
Correction: I found my earlier 100+ tokens/s result was inflated by an API proxy buffering tokens and delivering them in bursts, which distorted SparkDash’s client-side decode measurement.
Running GLM-5.3-Flash on 2× DGX Spark with NVIDIA Flare ✨ - 1M context window, 2.32M-token shared KV pool - Prose decode, single stream: 117.2 tok/s - Prose decode, 2 concurrent: 167.4 tok/s aggregate - Prose decode, 4 concurrent: 215.8 tok/s aggregate - 32K prefill: 2,327.9 tok/s - Available memory after loading: ~13 GiB on head /worker each Checkpoint: GLM-5.3-Flash-NVFP4-MXFP8-CSF-QAD github.com/NVIDIA/flare-loca…
11
2
2
17
2,924
Running GLM-5.3-Flash on 2× DGX Spark with NVIDIA Flare ✨ - 1M context window, 2.32M-token shared KV pool - Prose decode, single stream: 117.2 tok/s - Prose decode, 2 concurrent: 167.4 tok/s aggregate - Prose decode, 4 concurrent: 215.8 tok/s aggregate - 32K prefill: 2,327.9 tok/s - Available memory after loading: ~13 GiB on head /worker each Checkpoint: GLM-5.3-Flash-NVFP4-MXFP8-CSF-QAD github.com/NVIDIA/flare-loca…
NVIDIA just released Flare, a local inference server built for DGX Spark and RTX Blackwell GPUs - Built for long chats, agents and tool calls - Serves from Rust, no Python - Plans memory up front - FP8 KV cache by default Interesting timing... right around when DeepSeek open-sourced their inference engine. Link below.
66
28
18
267
43,587
prefill 2300tps @ 32k
4
1
1
11
2,736
卧槽? 4500tokne/s??? qwen3.8 flash next on 4*R9700
Did I go a bit overboard?
2
301
推荐有DGX Spark的人都去安装!
sparkDash ⚡️ v2.0 is now LIVE!! This is a HUGE release and a COMPLETE revamp! Here's what's new: User Interface & UX - Completely redesigned UI: modern, clean, minimal, yet powerful - Live fleet overview with VRAM, GPU, temps, power and tok/s on every card Run models directly from sparkDash - Add your own start/stop scripts and run your local models straight from the dashboard - Start a model from its Spark's card and see when it's still loading Benchmarks - A dedicated page for each benchmark: Decode, Prefill, Quality and Tool Eval Bench - Built-in quality benchmarks to check how good a model really is: GSM8K, MMLU and instruction following - Tool Eval Bench built in, with ready-made runs - Compare two runs side by side - More accurate Prefill tok/s - Shareable result cards Monitoring - GPU history that survives restarts - VRAM breakdown showing what the model holds, what the system uses, and what's free - Three new pages: Token Totals, Fleet Energy and Activity Total tokens used - Generated, prompt and cached tokens across your whole fleet - Break it down by type, model or Spark - Ranges: 24h, 7d, 30d and all time - Cache hit rate and change vs the previous period - Auto insights: busiest day, busiest model, and which Spark did the most work - Sortable tables and CSV export Fleet Energy - kWh per node, by hour or by day - Average power and peak draw - Efficiency in Wh per 1,000 tokens - Coverage, showing how much of the period was measured - Estimated cost, using the electricity price and currency you set in Settings Activity - Full event history for your fleet: Sparks going online or offline, throttling, and benchmark and model start/stop events - Filter by Spark and event type, with paging - The latest events also appear on the Overview And lots and lots of other cool features! More screenshots below. Get it here: github.com/MiaAI-Lab/sparkDa…
209
This is how and when I will die tinyurl.com/howwillyoudie2 .
1
#ngv #yayoikusama #melbourne @ National Gallery of Victoria instagram.com/p/Bgk-XLeBqZt/ – at Melbourne, Victoria
#cairns @ QT Port Douglas instagram.com/p/BWMlD9tAijz/ – at Port Douglas - Craiglie, Queensland