Fix Qwen 2.1 GGUF load: promote Q8 norms to F32 and wire TextEncode latent.

abenzerps Q8_0 ships 1D RMSNorms as packed Q8 (136 vs 128), which breaks
Comfy rms_rope; tagger now dequantizes small tensors and the graph uses
TextEncodeQwenImage21's 64-ch latent plus AuraFlow shift.

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Towsty
2026-09-20 16:58:00 -05:00
co-authored by Cursor
parent 68731916ca
commit 608f7d8d47
5 changed files with 200 additions and 39 deletions
+37 -1
View File
@@ -146,7 +146,43 @@ if (-not $SkipDownload) {
Write-Ok "HF CLI: $cli"
# DiT GGUF only (no Q6/Q5/Q4, no full bf16 DiT)
Invoke-HfFile -Cli $cli -Repo 'abenzerps/Qwen-Image-2.1-Uncensored-GGUF' -File 'qwen-image-2.1-Q8_0.gguf' -LocalDir $Diff
# Exact invocation required on this host:
# hf download hf://abenzerps/Qwen-Image-2.1-Uncensored-GGUF/qwen-image-2.1-Q8_0.gguf
# That form writes into the Hugging Face hub cache (not --local-dir).
Write-Step "hf download hf://abenzerps/Qwen-Image-2.1-Uncensored-GGUF/qwen-image-2.1-Q8_0.gguf"
$ditName = 'qwen-image-2.1-Q8_0.gguf'
$ditDest = Join-Path $Diff $ditName
if ((Test-Path $ditDest) -and -not $Force -and (Get-Item $ditDest).Length -gt 1MB) {
Write-Ok "Already present: $ditDest ($((Get-Item $ditDest).Length) bytes)"
} else {
& $cli download "hf://abenzerps/Qwen-Image-2.1-Uncensored-GGUF/$ditName"
if ($LASTEXITCODE -ne 0) { throw "hf download failed for DiT GGUF (exit $LASTEXITCODE)" }
$cacheRoot = Join-Path $env:USERPROFILE '.cache\huggingface\hub\models--abenzerps--Qwen-Image-2.1-Uncensored-GGUF'
$found = Get-ChildItem $cacheRoot -Recurse -File -Filter $ditName -ErrorAction SilentlyContinue |
Sort-Object LastWriteTime -Descending |
Select-Object -First 1
if (-not $found) { throw "hf download finished but $ditName was not found under $cacheRoot" }
Ensure-Dir $Diff
Copy-Item -Force $found.FullName $ditDest
Write-Ok "Copied $($found.FullName) -> $ditDest ($((Get-Item $ditDest).Length) bytes)"
}
# abenzerps ships kv_count=0 + Q8_0 1D RMSNorms (packed 136 vs logical 128).
# Tag architecture=qwen_image and promote small/1D tensors to F32 for ComfyUI-GGUF.
Write-Step "Tag DiT GGUF (architecture + F32 norms)"
$tagger = Join-Path $PSScriptRoot 'tag-qwen21-gguf.py'
$pyCandidates = @(
(Join-Path $root '.venv\Scripts\python.exe'),
(Join-Path (Split-Path $root) 'standalone-env\python.exe'),
(Get-Command python -ErrorAction SilentlyContinue | Select-Object -ExpandProperty Source)
) | Where-Object { $_ -and (Test-Path $_) }
$py = $pyCandidates | Select-Object -First 1
if (-not $py) { throw "No Python found to run $tagger" }
$fixed = Join-Path $Diff 'qwen-image-2.1-Q8_0.fixed.gguf'
& $py $tagger $ditDest $fixed
if ($LASTEXITCODE -ne 0) { throw "tag-qwen21-gguf.py failed (exit $LASTEXITCODE)" }
Copy-Item -Force $fixed $ditDest
Write-Ok "Installed tagged GGUF at $ditDest"
# TE + VAE companions from Comfy-Org (INT8 TE for 16 GB VRAM)
Invoke-HfFile -Cli $cli -Repo 'Comfy-Org/Qwen-Image-2.1' -File 'text_encoders/qwen3vl_8b_int8_convrot.safetensors' -LocalDir $TE