Paul Cozzolino

🏡 Dad & Partner 💜 💻Founder (Fast Fix 123, Zezo Media AudienceMob, Sunday Market) 📍NY🗽Transplant in 🇨🇦 🎤 Ex-Emo 🔬No Gods No Magic

[Paul Cozzolino @codoublez]

The world awaits your smile. =D

[Paul Cozzolino @codoublez]

@thdxr +1 Bright Eyes

[Paul Cozzolino @codoublez]

Yeah good call — Dockerfile alone wasn't enough. Just pushed a single-file
bootstrap script that does the whole zero-to-serving on dual Spark TP=2:
curl -fsSLO https://t.co/tMEgvkQa1d
chmod +x bootstrap_dsv4_spark.sh
./bootstrap_dsv4_spark.sh --head-host spark-a --worker-host spark-b
Idempotent — handles SSH check, model download, QSFP /30 setup, image build
(eugr/spark-vllm-docker scaffold + our DSV4 Dockerfile + patch), scp-distribute
to the worker, launch on both nodes, waits for /health=200. ~30-50 min first
run, ~7 min on re-runs with --skip-build. Verified the URL works (HTTP 200,
syntax clean).
Script: https://t.co/Wn5HGJ7syc
Quickstart: https://t.co/iv03vUmV6L
Note for anyone tracking the upstream PR: vllm-project/vllm#40991 was closed
today and replaced by #41834 ("[New Model][Nvidia] Add SM12x support for
DeepSeek V4 Flash with essential fixes"). New PR targets branch
codex/ds4-sm120-min-enable. Our build is on jasl/vllm@ds4-sm120-experimental
which still works for SM12x DSV4 today.

[Paul Cozzolino @codoublez]

Got DeepSeek V4-Flash running on dual DGX Spark TP=2.

W4A16 GPTQ + FP8 attention, 14-15 tok/s, full harness PASS.

Key fix: workspace pre-reservation patch — vLLM locks the attention scratch at warmup-decode size, which crashes on real prefill.

🤗 https://t.co/eJegt9QcK7

[Paul Cozzolino @codoublez]

@JASONMCNAB @MatthewBerman @grok you need 2 of them..

https://t.co/L2cRrfLkqq

[Paul Cozzolino @codoublez]

@ArtiIntelligent Now 1M context benchmarked..quickstart guide - https://t.co/L2cRrfLkqq

[Paul Cozzolino @codoublez]

@ArtiIntelligent you can run up to 512k with seq=1

[Paul Cozzolino @codoublez]

@garyfung Got DeepSeek V4-Flash running on dual DGX Spark TP=2.

W4A16 GPTQ + FP8 attention, 14-15 tok/s, full harness PASS.

Key fix: workspace pre-reservation patch — vLLM locks the attention scratch at warmup-decode size, which crashes on real prefill.

🤗 https://t.co/eJegt9QcK7

[Paul Cozzolino @codoublez]

@TheAhmadOsman Got DeepSeek V4-Flash running on dual DGX Spark TP=2.

W4A16 GPTQ + FP8 attention, 14-15 tok/s, full harness PASS.

Key fix: workspace pre-reservation patch — vLLM locks the attention scratch at warmup-decode size, which crashes on real prefill.

🤗 https://t.co/eJegt9QcK7

[Paul Cozzolino @codoublez]

@TheAhmadOsman https://t.co/lDLRf0JjKW

[Paul Cozzolino @codoublez]

@TheAhmadOsman hope this works

[Paul Cozzolino @codoublez]

@kshvbgde @diabrowser @PerplexityComet whichever one lets me use it

[Paul Cozzolino @codoublez]

@aaditsh @paraschopra y not just use cline?

[Reid Jackson @reidjjackson]

How it started / How it’s going

[Paul Cozzolino @codoublez]

@Powercommitment @Sumanth_077 @firecrawl sudo crawl bulbapedia for all japaneset sets and cards by number and create me app to analyze my cards from an image

[Paul Cozzolino @codoublez]

@CarmineOptions @lordOfAFew @cursor_ai We're having trouble connecting to anthropic. This might be temporary - please try again in a moment.

[Paul Cozzolino @codoublez]

@lacybuilds We're having trouble connecting to anthropic. This might be temporary - please try again in a moment. is all i see

[Paul Cozzolino @codoublez]

Happy Halloween from the Cozzomadrigals 👻 https://t.co/8pTdylIvOE

[Paul Cozzolino @codoublez]

@155pod this is v sad

["Junior" Salazar for FL State House District 70 @Junior4FL]

We're not free until we're all free.

We're not The Sunshine State until that light shines on all.

We don't stop till that happens.

Let's get to work, Florida.

#BansOffOurBodies #RoeVWade #DeSantisDestroysFlorida

[Paul Cozzolino @codoublez]

Some friends just never go away.
I love you all. Please get me home. MONKEY!!!! @ Las Vegas Nevada https://t.co/xPxuY2MILp