Merged
Conversation
Member
|
@bot /style |
Contributor
|
Style bot fixed some files and pushed the changes. |
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
yiyixuxu
approved these changes
Mar 12, 2026
Collaborator
yiyixuxu
left a comment
yiyixuxu
left a comment
There was a problem hiding this comment.
thanks a ton!
I left some very small feedbacks!
|
Awesome! |
Pfannkuchensack
added a commit
to Pfannkuchensack/InvokeAI
that referenced
this pull request
Mar 17, 2026
…onditioning) Prepare support for the FLUX.2-klein-9b-kv model which uses KV-cached attention for up to 2.5x faster multi-reference editing. On step 0, reference image tokens are included in the full forward pass and their K/V projections are cached per-layer. On subsequent steps, cached K/V are reused without recomputing reference tokens. Based on diffusers PR huggingface/diffusers#13262 (not yet released). Will become functional once diffusers releases Flux2KleinKVPipeline support. Changes: - Add Klein9BKV variant to Flux2VariantType taxonomy - Detect Flux2KleinKVPipeline in model config identification - Add starter model entry for black-forest-labs/FLUX.2-klein-9b-kv - Add denoise_kv() backend function with KV-cache denoising loop - Add Flux2KVDenoiseInvocation node with full inpainting support - Update Qwen3 encoder variant validation to include Klein9BKV
GaoxiangLuo
added a commit
to GaoxiangLuo/sglang
that referenced
this pull request
Mar 18, 2026
Add Flux2KleinKVCachePipeline for KV-cached reference image conditioning. On step 0, reference tokens go through the full forward pass and their post-RoPE K/V are cached per-layer. Steps 1+ reuse cached K/V, skipping redundant ref token computation. Based on diffusers PR huggingface/diffusers#13262, adapted to SGLang (USPAttention, SP, flashinfer RoPE). Key changes: - Flux2KVLayerCache / Flux2KVCache for per-layer K/V storage - Causal attention in extract mode (ref tokens self-attend only) - Modulation blending with fixed timestep for ref tokens - Flux2KleinKVCacheDenoisingStage for step-dependent cache switching - Expose --pipeline-class-name CLI flag (existing ServerArgs field) - Integration test case for flux_2_klein_9b_kv_ti2i Target: black-forest-labs/FLUX.2-klein-9b-kv (rope_theta=2000).
GaoxiangLuo
added a commit
to GaoxiangLuo/sglang
that referenced
this pull request
Mar 18, 2026
Add Flux2KleinKVCachePipeline for KV-cached reference image conditioning. On step 0, reference tokens go through the full forward pass and their post-RoPE K/V are cached per-layer. Steps 1+ reuse cached K/V, skipping redundant ref token computation. Based on diffusers PR huggingface/diffusers#13262, adapted to SGLang (USPAttention, SP, flashinfer RoPE). Key changes: - Flux2KVLayerCache / Flux2KVCache for per-layer K/V storage - Causal attention in extract mode (ref tokens self-attend only) - Modulation blending with fixed timestep for ref tokens - Flux2KleinKVCacheDenoisingStage for step-dependent cache switching - Expose --pipeline-class-name CLI flag (existing ServerArgs field) - Integration test case for flux_2_klein_9b_kv_ti2i Target: black-forest-labs/FLUX.2-klein-9b-kv (rope_theta=2000).
5 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Adds
Flux2KleinKVPipeline— a KV-cached reference image conditioning pipeline for Flux2 Klein 9B.On the first denoising step, reference image tokens are included in the full transformer forward pass and their post-RoPE attention K/V projections are cached per-layer. On subsequent steps, only the noise latents are forwarded and the cached reference K/V are injected during attention, avoiding redundant recomputation of reference tokens.
Transformer changes (
transformer_flux2.py):Flux2KVLayerCache/Flux2KVCache— per-layer and global KV cache containers_flux2_kv_causal_attention— causal attention dispatch (ref tokens self-attend only; txt+img attend to all)_blend_mod_params,_blend_double_block_mods,_blend_single_block_mods) so reference tokens receive fixed-timestep modulationFlux2KVAttnProcessor/Flux2KVParallelSelfAttnProcessor— KV-aware attention processors for double/single stream blocksFlux2Transformer2DModel.forwardextended withkv_cache,kv_cache_mode,num_ref_tokens,ref_fixed_timestepparametersPipeline (
pipeline_flux2_klein_kv.py):Before submitting
documentation guidelines, and
here are tips on formatting docstrings.