From aada0082d129743c3fc34f52a869506e7ec9d5d3 Mon Sep 17 00:00:00 2001 From: hsliuustc0106 Date: Wed, 26 Aug 2026 21:15:31 +0800 Subject: [PATCH] docs: clarify FLUX.2-klein HWR validation scope Signed-off-by: hsliuustc0106 --- .../feature/offloader/distributed_layerwise_offload.md | 9 ++++++--- .../diffusion/offloader/distributed_layerwise_offload.md | 7 ++++--- 2 files changed, 10 insertions(+), 6 deletions(-) diff --git a/docs/design/feature/offloader/distributed_layerwise_offload.md b/docs/design/feature/offloader/distributed_layerwise_offload.md index 4500b17c7..c47b9187b 100644 --- a/docs/design/feature/offloader/distributed_layerwise_offload.md +++ b/docs/design/feature/offloader/distributed_layerwise_offload.md @@ -223,9 +223,12 @@ The current producer/consumer boundary is the model-declared final-layout BF16 contract (MiniMax H3 and `black-forest-labs/FLUX.2-klein-4B`). The FLUX.2-klein contract covers both transformer block stacks, constructor-stable packed QKV mapping state, BF16 parameters, and any persistent loader-owned `beta`/`eps` -buffers. It supports ordinary-loader final layouts for TP1 and TP2 rank -identities plus SP layout identities; online quantization, HSDP, LoRA/adapted -weights, and non-default load formats remain ineligible. HWR mode +buffers. The shared HWR machinery supports ordinary-loader final layouts for +TP1 and TP2 rank identities plus SP layout identities. FLUX.2-klein evidence +in this change is TP1-only; its TP2/SP layouts rely on the existing +per-coordinate identity mechanics and remain unmeasured. Online quantization, +HSDP, LoRA/adapted weights, and non-default load formats remain ineligible. +HWR mode `disabled`, DLO-disabled, and DLO AllGather configurations stop before source identity or store construction and retain the existing checkpoint-mmap or ordinary-loader path. diff --git a/docs/user_guide/diffusion/offloader/distributed_layerwise_offload.md b/docs/user_guide/diffusion/offloader/distributed_layerwise_offload.md index cdc1c4a2c..45ff8df2d 100644 --- a/docs/user_guide/diffusion/offloader/distributed_layerwise_offload.md +++ b/docs/user_guide/diffusion/offloader/distributed_layerwise_offload.md @@ -113,9 +113,10 @@ HWR is an opt-in startup optimization for models that declare the final-layout BF16 restore contract. Use it only with no-AllGather DLO: The validated BF16 model contracts currently cover MiniMax H3 and -`black-forest-labs/FLUX.2-klein-4B`. FLUX.2-klein-9B, FLUX.2-dev, online FP8, -HSDP, LoRA/adapted weights, and non-default load formats are not enabled by -this contract. +`black-forest-labs/FLUX.2-klein-4B`. FLUX.2-klein-9B shares the same model +class but has not been validated against this contract. FLUX.2-dev, online +FP8, HSDP, LoRA/adapted weights, and non-default load formats remain outside +the validated scope. ```bash vllm serve /path/to/model --omni \