New Quantized Inference Optimizations for AMD

huggingface · x · 2026-07-14

Hugging Face shared an update on new models and quantization schemes for AMD Strix Halo users.

Key points include:

• The 298B parameter model Hy3 uses a new 2-bit FPX codebook, aiming to map more efficiently to AMD hardware's INT8 channels.
• Compared to IQ2M, the new quantization scheme is 2.55% smaller in size.
• End-to-end latency without MTP drops by roughly 2.4% on short coding prompts, and by about 6.2% on a 19,654-token coding prompt.
• 14.05GB of reusable state can be offloaded to SSD instead of taking up scarce UMA cache memory.
• Unweighted control scores are 81 for HermesAgent-20 and 88 for the full Tool-Eval, which the author claims outperforms most Qwen models.
• Run speed is about 17–25 tok/s, with prefill at roughly 200 tok/s; unlocking these optimizations requires the latest FTX inference-engine (llama fork) update.

Original post →

More from Infra

• Anthropic S-1: training infra edge cut per-task cost 40% and doubled revenue growth — BenBajarin · 2026-10-26
• Rewriting scaling law study: optimal synthetic data share rises from 10% to 32% with compute — heghbalz · 2026-10-26
• US states halt new data centers as 72% of Americans oppose local builds — YvesMulkers · 2026-10-26
• $5M Seed Round Can't Afford One DeepSeek Training Run: GPU-less AI Startup's Three Paths — FinanceYF5 · 2026-10-26
• HBM demand shifts: low-stack inference plans shelved, Rubin to go 12-high only — BenBajarin · 2026-10-26
• CircuitHub unveils The Grid, a factory-scale robot making PCBs in the US twice as fast as Shenzhen — garrytan · 2026-10-26

Infra channel →