Steve Goldsby

The things that got you where you are today may not get you where you want to be tomorrow. Franks and beans.

[Steve Goldsby @GoldsbySteve]

If you’re on dual Spark with 0731:

• Want stable 64k agent path, simple ops → baseline is fine (~23 tok/s, serializes under load) • Want real multi-user / multi-agent throughput + 1M ctx → Tech2Wild stack is the win

Huge credit to @Tech2Wild / tonyd2wild (+ Patch 4 shared-expert fix). Without that loader fix, 0731 DSpark acceptance collapses and you get “correct but half-speed.”

Same metal. Different config. Night and day under concurrency.

Optional shorter single-post version
A/B: DeepSeek-V4-Flash-0731 on 2× DGX Spark TP=2

My baseline (seqs=1, no spec, fp8 KV, 64k): ~23 tok/s single-stream ~22 tok/s aggregate even at C=10 (queues)

@Tech2Wild DSpark k=5 + NVFP4 KV + seqs=6 + 1M: ~47 tok/s single-stream (~2×) ~72 tok/s aggregate short/mid C=10 (~3.2–3.4×) TTFT p95 short C=10: 51s → 12s

seqs=1 silently serializes concurrency. DSpark buys single-stream. NVFP4 buys context headroom.

100/100 ok, 0 preempt. Measure with non-stream usage under speculation.

[Steve Goldsby @GoldsbySteve]

Ran Deepseek-v4-flash on my own configuration vs. @Tech2Wild and @MiaAI_lab configs. Embarrassing. They're doing God's work.

For all the haters: it's not about TPS. It's about keeping your IP local so @sama and @DarioAmodei can't monetize it.

Pick your harness. Keep it local. I use @NousResearch Hermes. Do your thing.

[Steve Goldsby @GoldsbySteve]

seqs=1 is the silent killer. Concurrent requests queue. Aggregate stays pinned ~22 tok/s no matter how hard you hit it. Raising seqs is what unlocks concurrency.

DSpark k=5 is the single-stream story. ~2× decode on structured output. Acceptance-driven; content mix matters.

NVFP4 KV is a context lever, not a free speed lever. Bigger pool (511k → 1.36M) + 1M ceiling. Long/max stay prefill-bound either way.

Measure carefully under speculation. Streaming ITL under-reports DSpark (chunks ≈ steps). Trust non-stream usage.completion_tokens / wall and aggregate tok/s.

[Steve Goldsby @GoldsbySteve]

Honest single-stream decode (non-stream, usage/wall): • baseline ~23 tok/s • Tech2Wild ~47 tok/s → ~2.0×

Aggregate throughput (short ~1k in / 128 out): • C=10 baseline 22.5 tok/s (flat — pure queue) • C=10 Tech2Wild 71.5 tok/s → 3.2× • mid C=10: 21.4 → 72.3 tok/s → 3.4×

TTFT p95 short C=10: • 51.5s → 11.6s

100/100 requests ok. 0 preemptions.

[Steve Goldsby @GoldsbySteve]

My baseline • image: aidendle sparkrun-vllm-ds4-gb10 • speculation: off • KV: fp8_ds_mla • max_model_len: 65k • max_num_seqs: 1 (queues, no real batching) • util 0.85 · batched_tokens 8192 • live KV pool: 511k tok

Tech2Wild • image: dspark-nvfp4-stage-c (+ Patch 4) • speculation: dspark k=5 • KV: nvfp4_ds_mla • max_model_len: 1M • max_num_seqs: 6 • util 0.78 · capture 36 · async sched • live KV pool: 1.36M tok

[Steve Goldsby @GoldsbySteve]

My baseline • image: aidendle sparkrun-vllm-ds4-gb10 • speculation: off • KV: fp8_ds_mla • max_model_len: 65k • max_num_seqs: 1 (queues, no real batching) • util 0.85 · batched_tokens 8192 • live KV pool: 511k tok

Tech2Wild • image: dspark-nvfp4-stage-c (+ Patch 4) • speculation: dspark k=5 • KV: nvfp4_ds_mla • max_model_len: 1M • max_num_seqs: 6 • util 0.78 · capture 36 · async sched • live KV pool: 1.36M tok

[Steve Goldsby @GoldsbySteve]

@Tech2Wild @MiaAI_lab @sama @DarioAmodei Ran an A/B on DeepSeek-V4-Flash-0731 across 2× DGX Spark (GB10), TP=2.

Same model weights. Same nodes. Same prompt set. Only the serve config changed.

Baseline (what I had) vs @Tech2Wild / tonyd2wild DSpark+NVFP4 recipe.

Spoiler: concurrency is where it stops being subtle.

[Steve Goldsby @GoldsbySteve]

@TheAhmadOsman Rage bait. I’m running deep seek flash on 2 sparks. Quiet. c8. Fast. It’s the harness not the model.

[Steve Goldsby @GoldsbySteve]

Claim your https://t.co/RQ09lmQvK3 name. Move quick! You can use my referral code if you want (or not) -- XTCLNYDAUR

[Steve Goldsby @GoldsbySteve]

@davepl1968 Outstanding work

[Steve Goldsby @GoldsbySteve]

@mr_r0b0t Does this count?

[Steve Goldsby @GoldsbySteve]

If you're going to setup your Nvidia DGX Spark Cluster, use Nvidia Cluster Assisant! I did mine by hand, took a few hours and screwed up my netplan a few times in the process (which makes it a paint for a headless cluster). The cluster assistant does it in 5 minutes and it's all pointy clicky draggy droppy vs. terminal/CLI. https://t.co/aUMSwotsQC

[Steve Goldsby @GoldsbySteve]

It took me a while to migrate to Hermes desktop from the CLI, but with all the recent updates, it’s my daily driver now. The keyboard shortcuts, the ability to switch profiles really quickly, the ability to drag and drop files into it instead of saving to a shared folder somewhere and the “hey Hermes” feature is amazing. I have the CLI on every node and laptop and desktop in the house, all pointing to one gateway. Really simplifies things a whole lot. You have all outdone yourself.

[Steve Goldsby @GoldsbySteve]

Today I reached critical mass. I've been building my Hermes infrastructure and config @Teknium for a couple of months. Focused on local infrastructure and models for IP protection. 2 Mac Studios w/512G + exo, 3 DGX sparks in a cluster.

It's all true. Once you work with Hermes long enough, it gets better, faster, smarter. I'm definitely past 10x but not at 100x. Remaining local aux models downloading now, online by the weekend, and variable token costs drops to zero.

I literally did what used to take me 3-5 days tonight in 3 hours.

Recommendation: use openrouter, nous, etc. to develop your solution with frontier models, have Hermes tell you how to localize, and cut the cord.

Buckle up. It's about to get real.

[Steve Goldsby @GoldsbySteve]

@elder_plinius You're doing God's work. Carry on

[Steve Goldsby @GoldsbySteve]

@elonmusk Well well well

[Steve Goldsby @GoldsbySteve]

Try Grok. https://t.co/6EeTyg676V

Yes, ivermectin shows antiviral activity against certain negative-sense RNA viruses in laboratory (in vitro) studies, but this is not a broad or clinically proven effect, and it is not recommended or approved for treating any viral infections in humans or animals.
What the evidence shows
Negative-sense (negative-strand) RNA viruses have genomes that must first be transcribed into positive-sense mRNA by a viral RNA-dependent RNA polymerase. Examples include influenza viruses, paramyxoviruses (e.g., measles, mumps, Hendra, Newcastle disease, respiratory syncytial virus/RSV), rhabdoviruses (e.g., rabies), and filoviruses (e.g., Ebola).
Ivermectin (an FDA-approved antiparasitic) has demonstrated inhibitory effects in cell culture against several of these:

Influenza A (including avian strains, family Orthomyxoviridae): Ivermectin inhibits nuclear import of the viral ribonucleoprotein (vRNP) complex via blockade of importin α/β1 (IMPα/β1). This disrupts a critical step in the viral life cycle (Götz et al., 2016).
Hendra virus (family Paramyxoviridae, genus Henipavirus): It blocks nuclear import of the viral V protein, reducing infection in mammalian cells (Atkinson et al., 2018). Reviews list it among viruses inhibited in vitro.
Newcastle disease virus (NDV, family Paramyxoviridae): In vitro antiviral activity reported in cell-based assays.
Bovine respiratory syncytial virus (BRSV) and related paramyxoviruses (e.g., bovine parainfluenza virus 3): Dose-dependent inhibition of replication and progeny virus production in vitro.

Mechanism (host-directed, not virus-specific): Many RNA viruses hijack the host's nuclear import machinery (IMPα/β1) to traffic viral proteins into the nucleus for replication, immune evasion, or assembly. Ivermectin disrupts this interaction. This explains activity against both some positive-sense (e.g., dengue, Zika, SARS-CoV-2 in vitro) and negative-sense viruses that rely on it (especially those with nuclear phases like influenza). Cytoplasmic negative-sense viruses (e.g., rabies, Ebola) show little or no reported activity.
Important limitations

All strong data is in vitro (cell culture). Effective concentrations are often high (e.g., ~2–5 μM range in SARS-CoV-2 and influenza studies).
No robust clinical evidence supports use against any negative-sense RNA virus in humans. Pharmacokinetic studies show standard antiparasitic doses (e.g., 200 μg/kg) produce plasma levels far below those needed for antiviral effects in lab tests. Higher doses risk toxicity (neurological effects via GABA/glutamate channels).
In vivo/animal data is sparse and mixed; no large controlled trials exist for influenza, RSV, Hendra, or similar viruses.
Claims promoting ivermectin for flu, RSV, or other respiratory viruses (e.g., by some advocacy groups) lack supporting clinical trials and contradict regulatory assessments (FDA, WHO, etc.).
For comparison: Even against SARS-CoV-2 (positive-sense), large randomized trials found no meaningful clinical benefit at safe doses despite strong in vitro data.

Bottom line
Ivermectin can inhibit replication of specific negative-sense RNA viruses (notably influenza A and certain paramyxoviruses like Hendra/NDV) in the lab by interfering with host nuclear transport. It does not work universally against all negative-sense RNA viruses, and there is no established clinical role for it in treating or preventing these infections.
Standard medical care, vaccines (where available), and approved antivirals (e.g., oseltamivir for influenza, supportive care for RSV) remain the evidence-based approaches. Self-medication or off-label use carries risks and is not advised. Consult a physician for any viral illness. Research into host-directed antivirals continues, but ivermectin has not translated from bench to bedside for these viruses.

[Steve Goldsby @GoldsbySteve]

@Jack_W_Lindsey Harness. Not model

[Steve Goldsby @GoldsbySteve]

Hey @elonmusk Maybe I'm missing something. My X feed is fleek. I'd like to be able to share it with others (without polluting their feed) so they can see what I see. I'm hard on AI. They're hard on slop. They're hard on gardening. etc. Should I / can I subscribe to their way of thinking?

[Steve Goldsby @GoldsbySteve]

@Cloud1a7 At $20-40/rd that's super expensive.