Model techniques map
Techniquesmodel architecturecontext capacity

implementation detail · filed under model architecture

N-gram embedding host-memory offload and prefetch

N-gram embedding tables are stored in host memory and asynchronously prefetched or offloaded to reduce GPU-memory use while overlapping transfers with model computation.

Also called asynchronously offloaded to host memory, asynchronous prefetching, Host-memory offload for N-gram embeddings.

sources
4
model
1
lab adopt it
1
strongest
core

How sources treat it

One count per evidence span, weakest treatment to strongest.

optional 3used 2core 1

Documented in

Evidence

6 spans quoted from the sources, strongest treatment first.

Capacity is further added outside the backbone by a single n-gram embedding layer whose tables are prefetched from host memory

coreunclearin Qwen3.8-Flash-NextQwen

The N-gram Embedding can reside entirely in host memory to save GPU memory

usedunclearin Qwen3.8-Flash-NextSGLang

lookup positions are computed in advance and asynchronously prefetched, so it never permanently occupies GPU memory.

usedunclearin Qwen3.8-Flash-NextSGLang

can be asynchronously offloaded to host memory.

optionalinference servingin Qwen3.8-Flash-NextQwen

the embedding table can be offloaded to host memory

optionalinference servingin Qwen3.8-Flash-NextQwen

overlapped with model computation through asynchronous prefetching

optionalinference servingin Qwen3.8-Flash-NextQwen

Filed alongside

Other methods under model architecture :: context capacity.