Repository navigation

[2/2] Export vLLM fake-quant checkpoints through Megatron-Bridge - #2889

Merged

kevalmorabia97 merged 11 commits into

mainfrom

kmorabia/mbridge-vllm-fakequant-export

Oct 24, 2026

Merged

[2/2] Export vLLM fake-quant checkpoints through Megatron-Bridge#2889
kevalmorabia97 merged 11 commits into
mainfrom
kmorabia/mbridge-vllm-fakequant-export

Commits on Oct 23, 2026
- Export vLLM fake-quant checkpoints through Megatron-Bridge
  
  kevalmorabia97
  and
  claude
  committed
  b190d28
- Drop every weight-quantizer buffer when folding weights
  
  kevalmorabia97
  and
  claude
  committed
  c5c3433
- Cover per-expert Qwen3-MoE in the fake-quant export parallelism test
  
  kevalmorabia97
  and
  claude
  committed
  4c5fd63
- Keep the SVDQuant low-rank term when folding weights
  
  kevalmorabia97
  and
  claude
  committed
  916395d
- Mark the export as saved in one place for both export paths
  
  kevalmorabia97
  and
  claude
  committed
  45cca84
- Fold sequential weight quantizers (e.g. W4A8) in mtq.fold_weight
  
  kevalmorabia97
  and
  claude
  committed
  f882803
- Fold the fake-quant export tests into the existing Megatron example tests
  
  kevalmorabia97
  and
  claude
  committed
  5c1fc9e
- Compare against the versioned recipe of the Megatron-LM fake-quant export
  
  kevalmorabia97
  and
  claude
  committed
  08ce4b2
- Check the quantizers once before the Megatron-LM fake-quant export
  
  kevalmorabia97
  and
  claude
  committed
  f13e5cc
- Test the Megatron examples and GPU tests against a newer Megatron-Bridge
  
  kevalmorabia97
  and
  claude
  committed
  65ca9f2
Commits on Oct 24, 2026
- Keep a quantized output_layer's checkpoint state on newer Megatron-Core
  
  kevalmorabia97
  and
  claude
  committed
  449cffb