REAL-Q: E2E LLM Quantization via Dynamic Gradient Descent

Name: REAL-Q: E2E LLM Quantization via Dynamic Gradient Descent
Uploaded: 2026-09-07T01:22:15.855675+00:00
Duration: 418 s
Description: The discussion of 'REAL-Q' addresses a flaw in LLM quantization known as 'information misalignment,' where static methods fail to account for cumulative error drift. The authors conclude that by using dynamic, block-wise gradient descent and a global surrogate loss function, they can achieve reliable, real-time model compression.

The discussion of 'REAL-Q' addresses a flaw in LLM quantization known as 'information misalignment,' where static methods fail to account for cumulative error drift. The authors conclude that by using dynamic, block-wise gradient descent and a global surrogate loss function, they can achieve reliable, real-time model compression.

Open the full episode