docs/qwen21.md

@@ -21,6 +21,14 @@ Custom node: `custom_nodes\ComfyUI-GGUF` (`UnetLoaderGGUF`). Do not install a se
 powershell -ExecutionPolicy Bypass -File scripts\setup-qwen21.ps1
 ```
 
+DiT download uses the exact host invocation:
+
+```powershell
+hf download hf://abenzerps/Qwen-Image-2.1-Uncensored-GGUF/qwen-image-2.1-Q8_0.gguf
+```
+
+That lands in the Hugging Face hub cache; the setup script copies it into Shared `diffusion_models`. The abenzerps GGUF ships with `kv_count=0` (no `general.architecture`) **and** Q8_0-quantized 1D RMSNorm weights (logical 128 → packed 136), which breaks Comfy’s fused `rms_rope`. `scripts/tag-qwen21-gguf.py` rewrites the file with `general.architecture=qwen_image` and promotes small/1D tensors to F32 so city96 `UnetLoaderGGUF` can load it.
+
 Restart Comfy **only when idle** (`COMFY_CONTROL_URL/status` → `gpu.busy=false`). Confirm `object_info` lists `UnetLoaderGGUF` and the three filenames.
 
 ## App
@@ -28,7 +36,7 @@ Restart Comfy **only when idle** (`COMFY_CONTROL_URL/status` → `gpu.busy=false
 - Engine key: `qwen21` · UI label: **Qwen 2.1**
 - Generate (T2I) only in this build. Edit / Compose / Iterate / Video disable with: “Qwen 2.1 is T2I in this build”.
 - Sampler defaults: euler / simple / cfg **1** / steps **25** · `ModelSamplingAuraFlow` shift **3.1**
-- Default canvas **1024×1024** (aspect 16:9 / 9:16 / 1:1 honored, multiples of 32). Drop to 768 if VRAM errors.
+- Default canvas **1024×1024** (aspect 16:9 / 9:16 / 1:1 → long-edge square via `TextEncodeQwenImage21` resolution; multiples of 32). Drop to 768 if VRAM errors.
 - No hero / locks / Klein LoRA stack on this engine.
 
 Graph: `server/assets/studio2_qwen21_t2i.json`.

scripts/setup-qwen21.ps1

@@ -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

scripts/tag-qwen21-gguf.py

@@ -0,0 +1,132 @@
+#!/usr/bin/env python3
+"""Make abenzerps Qwen Image 2.1 DiT GGUF loadable in city96 ComfyUI-GGUF.
+
+abenzerps/Qwen-Image-2.1-Uncensored-GGUF ships Q8_0 with:
+  - kv_count=0 (no general.architecture)
+  - 1D RMSNorm weights quantized to Q8_0 (logical 128 -> packed 136),
+    which breaks Comfy's fused rms_rope path
+
+This rewrite:
+  1. Adds general.architecture=qwen_image
+  2. Dequantizes small / 1D tensors to F32 (city96 convert keeps them hiprec)
+"""
+from __future__ import annotations
+
+import argparse
+import sys
+from pathlib import Path
+
+# Match city96 ComfyUI-GGUF/tools/convert.py QUANTIZATION_THRESHOLD
+QUANTIZATION_THRESHOLD = 1024
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("src", type=Path)
+    ap.add_argument("dst", type=Path, nargs="?", default=None)
+    ap.add_argument("--arch", default="qwen_image")
+    ap.add_argument("--name", default="qwen-image-2.1")
+    ap.add_argument("--inplace", action="store_true", help="Replace src after a successful tag")
+    args = ap.parse_args()
+
+    import gguf
+    import numpy as np
+
+    src = args.src.resolve()
+    if not src.is_file():
+        print(f"missing source: {src}", file=sys.stderr)
+        return 1
+
+    dst = (args.dst.resolve() if args.dst else src.with_name(src.stem + ".tagged.gguf"))
+
+    reader = gguf.GGUFReader(str(src))
+
+    def get_field(name: str):
+        field = reader.fields.get(name)
+        if field is None:
+            return None
+        try:
+            return field.contents()
+        except Exception:
+            return None
+
+    if dst == src:
+        dst = src.with_name(src.stem + ".retag.gguf")
+
+    f32 = gguf.GGMLQuantizationType.F32
+    compat = {gguf.GGMLQuantizationType.F32, gguf.GGMLQuantizationType.F16, gguf.GGMLQuantizationType.BF16}
+
+    print(f"rewriting {src} -> {dst} (arch={args.arch}, tensors={len(reader.tensors)})")
+    writer = gguf.GGUFWriter(str(dst), arch=args.arch, use_temp_file=True)
+    writer.add_name(args.name)
+    try:
+        writer.add_type("model")
+    except Exception:
+        pass
+
+    promoted = 0
+    for tensor in reader.tensors:
+        logical = tuple(int(x) for x in tensor.shape)
+        # GGUF stores dims reversed vs torch; logical numel is what matters
+        numel = 1
+        for d in logical:
+            numel *= d
+        qtype = tensor.tensor_type
+        data = tensor.data
+        if hasattr(data, "copy"):
+            data = data.copy()
+
+        needs_f32 = qtype not in compat and (len(logical) <= 1 or numel <= QUANTIZATION_THRESHOLD)
+        if needs_f32:
+            # Dequantize packed blocks -> float32 with logical shape (GGUF order)
+            dequant = gguf.quants.dequantize(np.asarray(data), qtype).astype(np.float32, copy=False)
+            expected = numel
+            if dequant.size != expected:
+                print(
+                    f"dequant size mismatch {tensor.name}: got {dequant.size} want {expected} "
+                    f"shape={logical} qtype={qtype}",
+                    file=sys.stderr,
+                )
+                return 3
+            data = dequant.reshape(logical)
+            qtype = f32
+            promoted += 1
+
+        writer.add_tensor(tensor.name, data, raw_dtype=qtype)
+
+    writer.write_header_to_file()
+    writer.write_kv_data_to_file()
+    writer.write_tensors_to_file(progress=True)
+    writer.close()
+
+    check = gguf.GGUFReader(str(dst))
+    field = check.fields.get("general.architecture")
+    got = field.contents() if field is not None else None
+    print(
+        f"verified architecture={got!r} tensors={len(check.tensors)} "
+        f"promoted_f32={promoted} size={dst.stat().st_size}"
+    )
+    if got != args.arch or len(check.tensors) != len(reader.tensors):
+        print("tag failed", file=sys.stderr)
+        return 2
+
+    # Spot-check a known norm weight is F32 @ 128
+    sample = next((t for t in check.tensors if t.name.endswith("attn.norm_q.weight")), None)
+    if sample is not None:
+        print(
+            f"sample {sample.name}: shape={tuple(int(x) for x in sample.shape)} "
+            f"type={sample.tensor_type.name} data_shape={tuple(sample.data.shape)}"
+        )
+
+    if args.inplace:
+        bak = src.with_suffix(src.suffix + ".untagged.bak")
+        if bak.exists():
+            bak.unlink()
+        src.replace(bak)
+        dst.replace(src)
+        print(f"inplace: {src} (backup {bak})")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

server/assets/studio2_qwen21_t2i.json

@@ -6,6 +6,14 @@
     "class_type": "UnetLoaderGGUF",
     "_meta": { "title": "Load Qwen 2.1 GGUF" }
   },
+  "3": {
+    "inputs": {
+      "shift": 3.1,
+      "model": ["4", 0]
+    },
+    "class_type": "ModelSamplingAuraFlow",
+    "_meta": { "title": "Qwen 2.1 shift" }
+  },
   "5": {
     "inputs": {
       "clip_name": "qwen3vl_8b_int8_convrot.safetensors",
@@ -22,38 +30,15 @@
     "class_type": "VAELoader",
     "_meta": { "title": "Load Qwen 2.1 VAE" }
   },
-  "66": {
-    "inputs": {
-      "shift": 3.1,
-      "model": ["4", 0]
-    },
-    "class_type": "ModelSamplingAuraFlow",
-    "_meta": { "title": "ModelSamplingAuraFlow" }
-  },
   "9": {
     "inputs": {
-      "text": "",
-      "clip": ["5", 0]
+      "clip": ["5", 0],
+      "prompt": "",
+      "negative_prompt": "",
+      "resolution": 1024
     },
-    "class_type": "CLIPTextEncode",
-    "_meta": { "title": "Positive Prompt" }
-  },
-  "10": {
-    "inputs": {
-      "text": "",
-      "clip": ["5", 0]
-    },
-    "class_type": "CLIPTextEncode",
-    "_meta": { "title": "Negative Prompt" }
-  },
-  "14": {
-    "inputs": {
-      "width": 1024,
-      "height": 1024,
-      "batch_size": 1
-    },
-    "class_type": "EmptySD3LatentImage",
-    "_meta": { "title": "Empty Latent" }
+    "class_type": "TextEncodeQwenImage21",
+    "_meta": { "title": "Text Encode Qwen Image 2.1" }
   },
   "15": {
     "inputs": {
@@ -63,10 +48,10 @@
       "sampler_name": "euler",
       "scheduler": "simple",
       "denoise": 1,
-      "model": ["66", 0],
+      "model": ["3", 0],
       "positive": ["9", 0],
-      "negative": ["10", 0],
-      "latent_image": ["14", 0]
+      "negative": ["9", 1],
+      "latent_image": ["9", 2]
     },
     "class_type": "KSampler",
     "_meta": { "title": "KSampler" }

server/utils/studio2/runner.ts

@@ -135,17 +135,17 @@ async function prepareGraph(r: any) {
         s.width = size.width;
         s.height = size.height;
         graph = structuredClone(qwen21Template);
-        graph['9'].inputs.text = q.compiledPrompt;
-        graph['10'].inputs.text = stylePrompt(q.imageStyles, true);
-        graph['14'].inputs.width = size.width;
-        graph['14'].inputs.height = size.height;
+        graph['9'].inputs.prompt = q.compiledPrompt;
+        graph['9'].inputs.negative_prompt = stylePrompt(q.imageStyles, true);
+        // TextEncodeQwenImage21 builds the 64-ch empty latent from resolution (square T2I).
+        graph['9'].inputs.resolution = Math.max(size.width, size.height);
         graph['15'].inputs.seed = s.seed;
         graph['15'].inputs.steps = s.steps || 25;
         graph['15'].inputs.cfg = s.cfg ?? 1;
         graph['15'].inputs.sampler_name = 'euler';
         graph['15'].inputs.scheduler = 'simple';
         graph['21'].inputs.filename_prefix = prefix + '/image';
-        r.sampleLatent = 'empty latent';
+        r.sampleLatent = 'qwen21 textencode latent';
         r.sampleDenoise = null;
         r.heroReferenceAttached = false;
         r.graphId = 'studio2_qwen21_t2i.json';