All pull requests

Pull requests

• Fix vLLM fake-quant serving of FP8 checkpoints, MLA KV presets, CUDA graph.

  #2633·sychen52 opened Oct 1, 2026Contributor·

  ·
  4

• Add Cosmos3 Edge DFlash training recipe

  #2632·ConstBob opened Oct 1, 2026·

  ·
  7

• [5/6] Preserve source config in HF fakequant export

  #2630·kinjalpatel27 opened Oct 1, 2026Contributor·

  ·
  2

• [4/6] Reload vLLM fakequant checkpoints automatically

  #2629·kinjalpatel27 opened Oct 1, 2026Contributor·

  ·
  2

• [3/6] Fold grouped expert weights during fakequant export

  #2628·kinjalpatel27 opened Oct 1, 2026Contributor·

  ·
  2

• [2/6] Export Megatron fakequant quantizer state and YAML

  #2627·kinjalpatel27 opened Oct 1, 2026Contributor·

  ·
  2

• [1/6] Add vLLM fakequant CLI wrapper and app entry point

  #2626·kinjalpatel27 opened Oct 1, 2026Contributor·

  ·
  7

• [6/6] Prevent EP ranks from overwriting Megatron export shards

  #2625·kinjalpatel27 opened Oct 1, 2026Contributor·

  ·
  2

• [2/3] Quantize the DSA KV cache and export GLM-5 / GLM-5.2 from Megatron-Bridge

  #2624·kevalmorabia97 opened Oct 1, 2026Collaborator·

  ·
  16

• [1/3] Export expert-parallel Megatron-Bridge MoE checkpoints and fix exporter / QAD gaps

  #2623·kevalmorabia97 opened Oct 1, 2026Collaborator·

  ·
  19

• Clarify Day 0 sub-agent serving ownership

  #2622·chadvoegele opened Oct 1, 2026Contributor·

  ·
  2

• Add vLLM NVFP4 MLA KV-cache fake quant

  #2620·sychen52 opened Oct 1, 2026Contributor·

  ·
  2

• Fix q/k/v quantizer wiring in register_attention_for_kv_quant

  #2618·javierdejesusda opened Oct 1, 2026Contributor·

  ·
  1

  6

• fix(speculative): keep the DFlash draft's RoPE frequencies in fp32 on a fresh convert

  #2617·h-guo18 opened Oct 1, 2026Contributor·

  ·
  4

• fix(onnx): support per-tensor scales in qdq_to_dq

  #2613·Yi-111-a opened Oct 1, 2026·

  ·
  1

  3

• docs: fix TensorRT-LLM and autotune links

  #2612·pratikgx opened Sep 30, 2026·

  ·
  2

• [5/5] llm_sparsity: load the model from local disk, downloading it first if needed

  #2611·shengliangxu opened Sep 30, 2026Collaborator·5/5·

  ·
  3

• [4/5] megatron_bridge: load the HF model from local disk; Megatron export copies its non-model files

  #2610·shengliangxu opened Sep 30, 2026Collaborator·4/5·

  ·
  4

• [3/5] hf_ptq: load the model from local disk, downloading it first if needed

  #2609·shengliangxu opened Sep 30, 2026Collaborator·3/5·

  ·
  3

• [2/5] HF exporters write off-index safetensors and non-model files

  #2608·shengliangxu opened Sep 30, 2026Collaborator·2/5·

  ·
  4

• [1/5] Add ensure_local_checkpoint to load models from local disk, and copy_non_model_files

  #2607·shengliangxu opened Sep 30, 2026Collaborator·1/5·

  ·
  7

• docs(qwen3.6): add the decode-time fix for SciCode non-termination

  #2606·kevalmorabia97 opened Sep 30, 2026Collaborator·

  ·
  40

• docs(ptq): explicitly select PyTorch base image for HF PTQ

  #2605·chadvoegele opened Sep 30, 2026Contributor·

  ·
  5

• Pack each IQ weight once and decode the IQ formats on CUDA

  #2604·cjluo-nv opened Sep 30, 2026Collaborator·

  ·
  5

• Keep an unset ONNX dimension in the reported shape

  #2603·SashaMIT opened Sep 30, 2026·

  ·
  2