[diffusion] feat: support quant-videogen prq kv-cache quantization (memory-saving) for causal-dit (#32581)

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: Mick <mickjagger19@icloud.com>
This commit is contained in:
Haocheng Xi
2026-08-08 12:56:54 +08:00
committed by GitHub
co-authored by Claude Opus 4.8 Mick
parent 24c84dfa68
commit f64328c7f6
16 changed files with 1174 additions and 54 deletions
@@ -32,6 +32,16 @@ sglang serve \
--enable-torch-compile false
```
### Optional KV-Cache Compression
Long-running sessions can enable lossy int4 PRQ compression for completed
causal KV-cache chunks by installing `quant-videogen` as described in the
quantization guide and adding `--kv-cache-quant int4` to the server command.
The current and most recent completed chunks remain in BF16. See
[Causal KV-Cache Quantization](/docs/sglang-diffusion/quantization#causal-kv-cache-quantization)
for the algorithm, tuning options, measured memory-latency tradeoff, and
support limits.
## 3. Realtime WebUI
The lightweight local WebUI is useful for validating latency, frame transport, and camera control behavior.