๊ฐ€์ž… ํ›„ ์ดˆ๋Œ€ ๋งํฌ๋ฅผ ๊ณต์œ ํ•˜๋ฉด ๋™์˜์ƒ ์žฌ์ƒ ๋ฐ ์ดˆ๋Œ€ ๋ณด์ƒ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

Tech2Wild
@Tech2Wild
๐ŸŽฎ Tech, gaming, AI, and everything in between. ๐Ÿค– Building with it, not just talking about it. ๐Ÿ”ฅ From the mind of @ToNYD2WiLD
๊ฐ€์ž… March 2026
222 ํŒ”๋กœ์ž‰ ์ค‘    4.8K ํŒฌ
๐Ÿง  GLM-5.3 743B (Int4-Int8Mix) on 4ร— DGX Spark just got TWO new lanes ๐Ÿš€ ๐ŸŸฆ Lane 2: DCP4 + NVFP4 KV + DFlash2 ๐Ÿ“ฆ 689,772-token KV pool @ 262K ctx: 2.35ร— our old best ๐Ÿ“ 512K window boots: 594K pool, needle 3/3 at 490K tokens โœ… ๐ŸŸง Lane 1: DCP4 + fp8 KV + DFlash2 ๐Ÿ“ฆ 462,308-token pool @ 262K โšก Same decode on both: 54 tok/s counting ยท 40 code ยท 36 json ยท 17 prose (single stream) ๐Ÿ“ˆ peak 63 tok/s aggregate (C5, mixed real prompts) ๐ŸŽฏ 254K needle 3/3 on both lanes ๐Ÿงช NVFP4 + DCP merge checked token-by-token vs DCP1: inside run-to-run noise โš ๏ธ Trade-off: prefill ~400 tok/s ๐Ÿ™ DCP4 + DFlash2 is @ajclark's work, now in the repo with full credit. Thanks @omanuke, @wuwenthink, @knapcio, @brah_ddah & incoai ๐Ÿ”—
๋” ๋ณด๊ธฐ