This repository was archived by the owner on Oct 6, 2026. It is now read-only.

Repository navigation

Reduce NVFP4-CSF scale reconstruction overhead - #459

Merged

lukealonso merged 23 commits into

perf/nvfp4-csf-operands

Oct 6, 2026

Merged

Reduce NVFP4-CSF scale reconstruction overhead#459
lukealonso merged 23 commits into
masterfrom
perf/nvfp4-csf-operands

Commits on Sep 29, 2026
- Serve DS4.1 X4T with native W4A16 and fused scale expansion
  voipmonitor
  committed
  07d7448
Commits on Sep 30, 2026
Commits on Oct 1, 2026
- Decode losslessly compressed NVFP4 expert scales into shared native scratch
  
  voipmonitor
  and
  codex
  committed
  c2797ba
- Name compressed FP4 scale interfaces MXFP4-CSF and NVFP4-CSF
  
  Martin Vit
  and
  codex
  committed
  5c13c2e
- Preserve FP32 NVFP4 calibration under BF16 model loading
  voipmonitor
  committed
  8cb43b9
- Support canonical CSF checkpoints and native NVFP4 scale expansion
  voipmonitor
  committed
  1b32d92
- Separate CSF tensor preparation from model checkpoint discovery
  
  voipmonitor
  and
  codex
  committed
  157cfd2
- Prepare CSF scales through fused-MoE weight post-processing
  
  voipmonitor
  and
  codex
  committed
  096464c
- Preserve native MXFP8 expert execution for compressed MXFP4 scales
  voipmonitor
  committed
  7dc9174
- Reconstruct NVFP4-CSF scale operands in dynamic MoE
  voipmonitor
  committed
  2fa9fef
- Document CSF serving results and remaining decode overhead
  voipmonitor
  committed
  bb5f4c4
Commits on Oct 2, 2026
Commits on Oct 3, 2026
- docs: qualify CSF operand reconstruction without load-time scale compression
  
  voipmonitor
  and
  claude
  committed
  993eac6
Commits on Oct 4, 2026
- release: b12x-450-moe-routing lists its models as in the Karmic beta (#461)
  
  voipmonitor
  and
  claude
  committed
  567b3b8
Commits on Oct 6, 2026