Pinned post
Latest posts 5
An honour to introduce and question the security masterminds on stage at @EFDevcon.
Two days of shared knowledge under a huge Ethereum logo.
Thanks for securing our ecosystem 🫡 @0xMattmatt @DevPramoth @Elliot0x @kaanuzdogan @griffgreen @Montyly @0xlouistsai @rsquare @ngp2311 @Fricoben @mijovic988 @SagivMooly @_rossstephenson
Two days of shared knowledge under a huge Ethereum logo.
Thanks for securing our ecosystem 🫡 @0xMattmatt @DevPramoth @Elliot0x @kaanuzdogan @griffgreen @Montyly @0xlouistsai @rsquare @ngp2311 @Fricoben @mijovic988 @SagivMooly @_rossstephenson
This was so much out of my comfort zone, but so much fun!
Thank you @_rossstephenson for your guidance and energy that made the Devcon 🔥
Thank you @_rossstephenson for your guidance and energy that made the Devcon 🔥
I am so proud to have been able to lead anincredible group of Emcees for Devcon VII this year in India! ♥️
- 30 Emcees!
- From > 20 different countries!
- Allocated across 4 days of Devcon!
- Emceeing for 7 different stages!
- For 100s of speaker sessions!
The BIGGEST and most DIVERSE group of Emcees for any single crypto conference in the world!
Coming together to become the GLUE that holds Devcon together 🔥
From technical talks, to intense lightning stages, to hands on workshops where anything could happen!
Theese Emcees conducted themselves with grace and professionalism! Thank you all for trusting me as your Lead and becoming the EMCEES OF DEVCON!
@abena_eth @aledecentral1 @alexanderlee314 @buzea200 @binji_x @0xChiranjeev @chrishobcroft @d_gusakov @DmytroMatviiv @mijovic988 @eshpvt @dutraguss @huvoliveira @JamesNewell @jiminleex @JorkLeonhardt @jgattermayer @kirkthebaird @louviq @MartinBreiten @MissPurrple_ @OfirEliasi @ottokekalainen @pranaytej @robdotrego @zemse_ @solangegueiros @tech_mingler @yq_acc @zaryab_eth @EFDevconView quoted post on X
We also maxed out a single RX 7900 XTX: Strix Halo + one card:
DeepSeek-V4.1-Flash-1017 (EXL3 2.25 bpw) at 1M context, ~130K prompt:
• Decode: 54.8 tok/s on text, 40.2 on code
• Prefill: 942 tok/s on text, 604.5 on code
Even a 1,040,000-token prompt reads at 718 tok/s, then decodes at 51.1 tok/s.
Recipe: github.com/pepuscz/strix-…
github.com/mw00/project-m…
DeepSeek-V4.1-Flash-1017 (EXL3 2.25 bpw) at 1M context, ~130K prompt:
• Decode: 54.8 tok/s on text, 40.2 on code
• Prefill: 942 tok/s on text, 604.5 on code
Even a 1,040,000-token prompt reads at 718 tok/s, then decodes at 51.1 tok/s.
Recipe: github.com/pepuscz/strix-…
github.com/mw00/project-m…
EXL3 quant excels at efficiency but so far was available only for NVIDIA. So we wrote AMD-optimised EXL3 kernels for Project Maya:View quoted post on X
DeepSeek-V4.1-Flash-1017 (EXL3 2.25 bpw) at 1M context, ~130K prompt:
• Decode: 75.3 tok/s on text, 50.3 on code
• Prefill: 1,654 tok/s on text, 1,434 on code
Even a 1,040,000-token prompt reads at 980 tok/s, then decodes at 67 tok/s.
Decode on Maya's own benchmark is 2.7x faster (22.3 → 59.6 tok/s).
Measured on hybrid setup with Strix Halo (128GB Ryzen AI Max+ 395) + 2× RX 7900 XTX over OCuLink, model on an NVMe over USB4.
For context, the same model on a larger 2.9 bpw EXL3 quant decodes at 80.7 tok/s on two DGX Sparks (per @MiaAI_lab) and 60.1 tok/s on one RTX PRO 6000 Blackwell in Project Maya.
One-command Ansible install: github.com/pepuscz/strix-…
Thanks @PeasantSmith for merging and @turboderp_ for EXL3.
github.com/mw00/project-m…
EXL3 quant excels at efficiency but so far was available only for NVIDIA. So we wrote AMD-optimised EXL3 kernels for Project Maya:
DeepSeek-V4.1-Flash-1017 (EXL3 2.25 bpw) at 1M context, ~130K prompt:
• Decode: 75.3 tok/s on text, 50.3 on code
• Prefill: 1,654 tok/s on text, 1,434 on code
Even a 1,040,000-token prompt reads at 980 tok/s, then decodes at 67 tok/s.
Decode on Maya's own benchmark is 2.7x faster (22.3 → 59.6 tok/s).
Measured on hybrid setup with Strix Halo (128GB Ryzen AI Max+ 395) + 2× RX 7900 XTX over OCuLink, model on an NVMe over USB4.
For context, the same model on a larger 2.9 bpw EXL3 quant decodes at 80.7 tok/s on two DGX Sparks (per @MiaAI_lab) and 60.1 tok/s on one RTX PRO 6000 Blackwell in Project Maya.
One-command Ansible install: github.com/pepuscz/strix-…
Thanks @PeasantSmith for merging and @turboderp_ for EXL3.
github.com/mw00/project-m…
DeepSeek-V4.1-Flash-1017 (EXL3 2.25 bpw) at 1M context, ~130K prompt:
• Decode: 75.3 tok/s on text, 50.3 on code
• Prefill: 1,654 tok/s on text, 1,434 on code
Even a 1,040,000-token prompt reads at 980 tok/s, then decodes at 67 tok/s.
Decode on Maya's own benchmark is 2.7x faster (22.3 → 59.6 tok/s).
Measured on hybrid setup with Strix Halo (128GB Ryzen AI Max+ 395) + 2× RX 7900 XTX over OCuLink, model on an NVMe over USB4.
For context, the same model on a larger 2.9 bpw EXL3 quant decodes at 80.7 tok/s on two DGX Sparks (per @MiaAI_lab) and 60.1 tok/s on one RTX PRO 6000 Blackwell in Project Maya.
One-command Ansible install: github.com/pepuscz/strix-…
Thanks @PeasantSmith for merging and @turboderp_ for EXL3.
github.com/mw00/project-m…
github.com GitHub - pepuscz/strix-halo-llm: Winning Qwen and DeepSeek setups for AMD Strix Halo: pinned... Winning Qwen and DeepSeek setups for AMD Strix Halo: pinned Ansible recipes and matched benchmarks on a 128 GiB Ryzen AI Max+ 395 / Radeon 8060S. - pepuscz/strix-halo-llm