Repository navigation

Run fused-MoE expert GPTQ updates on all of this process's GPUs - #3002

Merged

cjluo-nv merged 7 commits into

mainfrom

chenjiel/gptq-parallel-expert-updates

Oct 31, 2026

Merged

Run fused-MoE expert GPTQ updates on all of this process's GPUs#3002
cjluo-nv merged 7 commits into
mainfrom
chenjiel/gptq-parallel-expert-updates

Commits on Oct 30, 2026
- Run fused-MoE expert GPTQ updates on all of this process's GPUs
  
  cjluo-nv
  and
  claude
  committed
  2a0039f
- Disable autograd in the GPTQ expert-update worker threads
  
  cjluo-nv
  and
  claude
  committed
  46262c8
- Keep GPTQ expert updates' modes, memory and weight lifetime in check
  
  cjluo-nv
  and
  claude
  committed
  88de90b
- Move GPTQ expert updates only where the room check covers the encoder
  
  cjluo-nv
  and
  claude
  committed
  df3f78f
- Fail the GPTQ expert-update GPU tests when the GGML CUDA extension is missing
  
  cjluo-nv
  and
  claude
  committed
  958a02d
- Count GPTQ's block and group buffers in the expert-update room check
  
  cjluo-nv
  and
  claude
  committed
  3d90e69
Commits on Oct 31, 2026
- Share one per-GPU thread runner between imatrix encodes and GPTQ updates
  
  cjluo-nv
  and
  claude
  committed
  e3123f9