fix: Hunyuan3D 2.1 batch size crashes in attention and forward pass by Kivylius · Pull Request #13699 · Comfy-Org/ComfyUI

Kivylius · 2026-05-04T12:06:25Z

CrossAttention.forward: hardcoded 1 in kv.view() replaced with actual batch size b
Attention.forward: hardcoded 1 in qkv_combined.view() replaced with actual batch size B
HunYuanDiTPlain.forward: context.chunk(2) and output.chunk(2) now guarded with shape[0] >= 2 check to avoid crash when running without negative conditioning

Fixes #10142

coderabbitai · 2026-05-04T12:08:58Z

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: c0a20d2b-5b5d-411b-9901-6e1a74d033fa

📥 Commits

Reviewing files that changed from the base of the PR and between cc98e29 and 43b0dab.

📒 Files selected for processing (1)

comfy/ldm/hunyuan3dv2_1/hunyuandit.py

✅ Files skipped from review due to trivial changes (1)

comfy/ldm/hunyuan3dv2_1/hunyuandit.py

📝 Walkthrough

Walkthrough

This pull request corrects batch-dimension handling in attention layers and adds guards for classifier-free guidance operations. CrossAttention.forward and Attention.forward now reshape concatenated kv/qkv tensors using the runtime batch size instead of a hard-coded 1. In HunYuanDiTPlain.forward, context chunking and final output swapping for classifier-free guidance are performed only when the batch size is at least 2; otherwise context and output are left unchanged.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately and concisely describes the main fix: correcting batch size handling in attention and forward pass of Hunyuan3D 2.1.
Description check	✅ Passed	The description clearly outlines the specific changes made across three functions and references the fixed issue, directly corresponding to the changeset.
Linked Issues check	✅ Passed	The PR addresses the core requirements from issue `#10142` by replacing hardcoded batch dimensions and guarding chunk operations to prevent crashes when context lacks expected shape.
Out of Scope Changes check	✅ Passed	All changes are directly scoped to fixing batch size crashes in attention and forward pass functions; no unrelated modifications are present.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Review rate limit: 6/8 reviews remaining, refill in 9 minutes and 59 seconds.}

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

- CrossAttention.forward: hardcoded `1` in kv.view() replaced with actual batch size `b` - Attention.forward: hardcoded `1` in qkv_combined.view() replaced with actual batch size `B` - HunYuanDiTPlain.forward: context.chunk(2) and output.chunk(2) now guarded with shape[0] >= 2 check to avoid crash when running without negative conditioning Fixes Comfy-Org#10142

Alanaktion · 2026-05-15T18:25:59Z

Can confirm that this fixes issues I was having with running the Hunyuan3D 2.1 template. Went from errors or invalid results to correct behavior, using MPS on a Mac Studio.

alexisrolland · 2026-05-16T01:23:06Z

Thanks @Kivylius do you have a way to reproduce the issue before fixing it? It would be helpful for me to test.

Kivylius · 2026-05-18T13:56:59Z

@alexisrolland very mutch like @Alanaktion

select Hunyuan3D 2.1 template
download models
generate and it just give error(s)

More details in #10142

Im the Macbook Pro M2

alexisrolland · 2026-05-20T10:56:43Z

@alexisrolland very mutch like @Alanaktion

select Hunyuan3D 2.1 template

download models

generate and it just give error(s)

More details in #10142

Im the Macbook Pro M2

That's what I tried before posting my previous comment, the template worked fine for me ;)

Kivylius · 2026-05-20T11:18:09Z

That's what I tried before posting my previous comment, the template worked fine for me ;)

Im not sure about you configuration, but for me, it was just fresh install, no plugins, no extesnions, not even changed any varables at all in the template. I'm not sure what different from your to mine, but i recomend to try it on virtual machene from scatch if possible. I'm on MacOS 26.4.1 and Python 3.13.5 personaly but not sure if that makes mutch of difference. In that thread it seems like several other people are having the same issue, but im not sure of there setups or versions, plugins ect, but all seem to be the exact same error.

One interesting observation;, sometimes its get further then other attemps, to me that screams of invalid outputs that not handled properly, or more so that that specific erroneous output is more prominent is specific setups over others else this would have already been fixed.

If there any other logs that are missing that not in that thread already, let me know and ill try my best to provide it here.

alexisrolland · 2026-05-20T11:57:47Z

I tested with batch size 1, 2, 3, and different resolutions 1024, 2048. I could not reproduce the issue either before or after the fix. Since multiple people have reported this fixes it for them, I am merging it.

The previous gate (len(cond_or_uncond) == 2 and set == {0, 1}) was intended to skip the cond/uncond swap when only one half was present under MultiGPU CFG Split, but it was too restrictive: it also skipped batch_size > 1 + CFG (cond_or_uncond like [0, 0, 1, 1] or [0,0,0,0, 1,1,1,1]), where chunk(2) still splits the batch cleanly into a cond half and an uncond half and the swap is still required. Switch to context.shape[0] >= 2, matching the parallel fix landed on master in #13699. The swap is a permutation-invariant no-op when the two halves don't form a CFG pair (since the output swap_cfg_halves block immediately undoes the permutation), so the only thing the gate actually needs to do is guard against chunk(2) on a batch of one. Amp-Thread-ID: https://ampcode.com/threads/T-019e4a00-fe3d-76bd-a2f2-a8c8c4040082 Co-authored-by: Amp <amp@ampcode.com>

CrossAttention.kv.view and Attention.qkv_combined.view both hardcoded batch=1 in the reshape, crashing or silently mis-shaping whenever the actual batch dimension was greater than 1. These were fixed on master in #13699 as part of the same patch that gated the chunk(2) swap, but worksplit-multigpu only picked up the chunk(2) gate. Bring the two view() fixes over so we have parity with master. Amp-Thread-ID: https://ampcode.com/threads/T-019e4a00-fe3d-76bd-a2f2-a8c8c4040082 Co-authored-by: Amp <amp@ampcode.com>

Brings in 18 commits from master so worksplit-multigpu does not regress fixes that landed on main since the last sync: - #13699 Hunyuan 3D 2.1 batch-size fixes (overlap with our own backport; conflict resolved in favor of the shape>=2 gate that binds swap_cfg_halves once and reuses it for the output swap-back) - #14031 ModelPatcherDynamic lora reshape / backup restore fix - #13802 Multi-threaded model load (memory_management / pinned_memory / model_management / aimdo plumbing) - #12679 lanczos single-channel tensor fix - #14010 Stable Audio 3 support - assorted partner-node, openapi, workflow-template, and tooling updates Amp-Thread-ID: https://ampcode.com/threads/T-019e4a00-fe3d-76bd-a2f2-a8c8c4040082 Co-authored-by: Amp <amp@ampcode.com>

…omfy-Org#13699)

Kivylius requested review from Kosinkadink, alexisrolland, comfyanonymous, guill, kijai and rattus128 as code owners May 4, 2026 12:06

Kivylius force-pushed the fix/hunyuan3dv2-batch-size-crashes branch from cc98e29 to 43b0dab Compare May 4, 2026 12:11

Merge branch 'master' into fix/hunyuan3dv2-batch-size-crashes

20103e1

Merge branch 'master' into fix/hunyuan3dv2-batch-size-crashes

0443cef

alexisrolland approved these changes May 20, 2026

View reviewed changes

alexisrolland merged commit 78b5dec into Comfy-Org:master May 20, 2026
14 checks passed

This was referenced May 22, 2026

set_attr_param re-wraps Parameter subclasses, breaking bnb Params4bit (NF4/FP4) at partially_unload #14046

Open

Multi-GPU non-CUDA: unconditional torch.cuda.set_device() in worksplit-multigpu hot paths #14069

Open

simonri pushed a commit to simonri/ComfyUI-flash-attention-3 that referenced this pull request May 26, 2026

fix: Hunyuan3D 2.1 batch size crashes in attention and forward pass (C��

ff0aaca

…omfy-Org#13699)

coderabbitai Bot mentioned this pull request May 26, 2026

VAE.decode chunked loop shape mismatch with TAEHV on multi-frame Wan latents #14114

Open

This was referenced Jun 10, 2026

FluxKVCache: batch size mismatch crash when cond batching changes between steps (torch.cat dim-2 error) #14389

Closed

[XPU] GGUF Q6_K dequantization segfault on Intel GPU during model loading #14515

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix: Hunyuan3D 2.1 batch size crashes in attention and forward pass#13699

fix: Hunyuan3D 2.1 batch size crashes in attention and forward pass#13699
alexisrolland merged 3 commits into
Comfy-Org:masterfrom
Kivylius:fix/hunyuan3dv2-batch-size-crashes

Kivylius commented May 4, 2026

coderabbitai Bot commented May 4, 2026 •

edited

Loading

Walkthrough

❌ Failed checks (1 warning)

Alanaktion commented May 15, 2026

alexisrolland commented May 16, 2026

Kivylius commented May 18, 2026 •

edited

Loading

alexisrolland commented May 20, 2026

Kivylius commented May 20, 2026 •

edited

Loading

alexisrolland commented May 20, 2026 •

edited

Loading

Uh oh!

Labels

3 participants

Uh oh!

Conversation

Kivylius commented May 4, 2026

coderabbitai Bot commented May 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Walkthrough

❌ Failed checks (1 warning)

Alanaktion commented May 15, 2026

alexisrolland commented May 16, 2026

Kivylius commented May 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

alexisrolland commented May 20, 2026

Kivylius commented May 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

alexisrolland commented May 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Labels

3 participants

coderabbitai Bot commented May 4, 2026 •

edited

Loading

Kivylius commented May 18, 2026 •

edited

Loading

Kivylius commented May 20, 2026 •

edited

Loading

alexisrolland commented May 20, 2026 •

edited

Loading