FenomAI lilcheaty commited on
Commit
1564656
·
0 Parent(s):

Duplicate from lilcheaty/MiniMax-H3-NVFP4

Browse files

Co-authored-by: Safzan Pirani <lilcheaty@users.noreply.huggingface.co>

.gitattributes ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ *.7z filter=lfs diff=lfs merge=lfs -text
2
+ *.arrow filter=lfs diff=lfs merge=lfs -text
3
+ *.bin filter=lfs diff=lfs merge=lfs -text
4
+ *.bz2 filter=lfs diff=lfs merge=lfs -text
5
+ *.ckpt filter=lfs diff=lfs merge=lfs -text
6
+ *.ftz filter=lfs diff=lfs merge=lfs -text
7
+ *.gz filter=lfs diff=lfs merge=lfs -text
8
+ *.h5 filter=lfs diff=lfs merge=lfs -text
9
+ *.joblib filter=lfs diff=lfs merge=lfs -text
10
+ *.lfs.* filter=lfs diff=lfs merge=lfs -text
11
+ *.mlmodel filter=lfs diff=lfs merge=lfs -text
12
+ *.model filter=lfs diff=lfs merge=lfs -text
13
+ *.msgpack filter=lfs diff=lfs merge=lfs -text
14
+ *.npy filter=lfs diff=lfs merge=lfs -text
15
+ *.npz filter=lfs diff=lfs merge=lfs -text
16
+ *.onnx filter=lfs diff=lfs merge=lfs -text
17
+ *.ot filter=lfs diff=lfs merge=lfs -text
18
+ *.parquet filter=lfs diff=lfs merge=lfs -text
19
+ *.pb filter=lfs diff=lfs merge=lfs -text
20
+ *.pickle filter=lfs diff=lfs merge=lfs -text
21
+ *.pkl filter=lfs diff=lfs merge=lfs -text
22
+ *.pt filter=lfs diff=lfs merge=lfs -text
23
+ *.pth filter=lfs diff=lfs merge=lfs -text
24
+ *.rar filter=lfs diff=lfs merge=lfs -text
25
+ *.safetensors filter=lfs diff=lfs merge=lfs -text
26
+ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
27
+ *.tar.* filter=lfs diff=lfs merge=lfs -text
28
+ *.tar filter=lfs diff=lfs merge=lfs -text
29
+ *.tflite filter=lfs diff=lfs merge=lfs -text
30
+ *.tgz filter=lfs diff=lfs merge=lfs -text
31
+ *.wasm filter=lfs diff=lfs merge=lfs -text
32
+ *.xz filter=lfs diff=lfs merge=lfs -text
33
+ *.zip filter=lfs diff=lfs merge=lfs -text
34
+ *.zst filter=lfs diff=lfs merge=lfs -text
35
+ *tfevents* filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,198 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ license_name: minimax-h3-community-license-agreement
4
+ license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
5
+ tags:
6
+ - comfyui
7
+ - nvfp4
8
+ - quantized
9
+ - video
10
+ - text-to-video
11
+ base_model: MiniMaxAI/MiniMax-H3
12
+ base_model_relation: quantized
13
+ ---
14
+
15
+ # MiniMax H3 — NVFP4
16
+
17
+ NVFP4 quantizations of the **MiniMax-H3** ref2va diffusion transformer for ComfyUI.
18
+
19
+ - Original model: https://huggingface.co/MiniMaxAI/MiniMax-H3
20
+ - ComfyUI repackage these derive from: https://huggingface.co/Comfy-Org/MiniMax-H3
21
+
22
+ **NVFP4 requires an NVIDIA Blackwell GPU** (RTX 50-series, RTX PRO 6000, B200). On Ada,
23
+ Hopper or older the NVFP4 path is emulated — use Comfy-Org's `int8_convrot` files instead.
24
+
25
+ ## Which file do I want?
26
+
27
+ | file | size | s/it | VRAM (DiT) | notes |
28
+ |---|---|---|---|---|
29
+ | **`minimax_h3_ref2va_pruned_nvfp4.safetensors`** | **12.5 GB** | **1.90** | 11.9 GB | **recommended** |
30
+ | `minimax_h3_ref2va_nvfp4_mixed.safetensors` | 24.4 GB | 1.92 | ~20 GB | from unpruned bf16 |
31
+ | `minimax_h3_ref2va_nvfp4_full.safetensors` | 18.7 GB | 1.91 | ~16 GB | experimental |
32
+
33
+ Take **`pruned_nvfp4`** unless you have a specific reason not to: half the size of the
34
+ alternatives at identical speed, and it leaves the modulation path at full precision.
35
+
36
+ ## Why the pruned base is the right one to quantize
37
+
38
+ Comfy-Org's `pruned` checkpoint is not lossily pruned — it is a **structural refactor of
39
+ AdaLN**, and understanding it explains the whole table above.
40
+
41
+ In the bf16 model, AdaLN modulation dominates the parameter count:
42
+
43
+ | group | bf16 | pruned |
44
+ |---|---|---|
45
+ | `adaln_proj` | **13.04B (39.4%)** | **0.04B (0.2%)** |
46
+ | `mlp` | 12.02B | 11.56B |
47
+ | `attn` | 8.02B | 7.71B |
48
+ | token_refiner / norms / embedders | 0.05B | 0.80B |
49
+ | **total** | **33.12B** | **20.11B** |
50
+
51
+ The bf16 model projects a 5376-dim conditioning vector into modulation parameters per block.
52
+ The pruned model replaces this with an 8-dim timestep table (`adaln_t_table`, shape
53
+ `[1025, 8]`) feeding `adaln_proj.linear` of shape `[96768, 8]`. Because modulation depends
54
+ only on the timestep, that 5376-wide projection was almost entirely redundant — 13.04B
55
+ parameters collapse to 0.04B, a ~326x reduction.
56
+
57
+ This matters for quantization because AdaLN is the part you least want to quantize: it emits
58
+ the scale and shift applied to every residual stream, so error there is multiplicative and
59
+ compounds across all 50 blocks and every sampling step. In the bf16 model you face a bad
60
+ choice — protect AdaLN and produce a ~36 GB file (larger than the 34 GB int8 it should beat),
61
+ or quantize 39% of the model and hope. **In the pruned model the problem disappears**: AdaLN
62
+ is already tiny, so you keep it at full precision for free and quantize only attn+mlp, which
63
+ are error-tolerant.
64
+
65
+ Comfy-Org's `pruned_int8_convrot` quantizes exactly those 200 attn/mlp layers to
66
+ int8_convrot and leaves everything else alone. `pruned_nvfp4` takes that same set to NVFP4.
67
+
68
+ ## Measured
69
+
70
+ RTX PRO 6000 Blackwell (96 GB), ComfyUI 0.30.0, ref2va, 864x480, 39 frames, 20 steps,
71
+ `res_multistep` / `beta`, three matched seeds:
72
+
73
+ | model | size | staged VRAM | s/it |
74
+ |---|---|---|---|
75
+ | `pruned_int8_convrot` (Comfy-Org) | 21.0 GB | 19,995 MB | 2.17 |
76
+ | **`pruned_nvfp4` (this repo)** | **12.5 GB** | **11,944 MB** | **1.90** |
77
+
78
+ **-40% file size, -8.0 GB VRAM, -12.4% sampling time.**
79
+
80
+ At ~12 GB for the DiT, a 32 GB card (RTX 5090) becomes viable if the text encoder is
81
+ offloaded to CPU after encoding — it runs once per prompt, not once per sampling step.
82
+
83
+ ## How `pruned_nvfp4` was built
84
+
85
+ The obvious tool does **not** work. The StarNodes model converter passes non-floating-point
86
+ tensors through untouched:
87
+
88
+ ```python
89
+ if not tensor.dtype.is_floating_point:
90
+ return tensor # already-int8 weights are copied verbatim
91
+ ```
92
+
93
+ so running it on an int8 checkpoint silently produces a byte-identical file. A real
94
+ dequantize -> requantize is required. `comfy-kitchen` exposes both halves:
95
+
96
+ ```python
97
+ params = ckt.TensorWiseINT8Layout.Params(
98
+ scale=weight_scale, orig_dtype=torch.bfloat16, orig_shape=qdata.shape,
99
+ is_weight=True, convrot=True, convrot_groupsize=256)
100
+ deq = ckt.TensorWiseINT8Layout.dequantize(qdata, params) # -> bf16
101
+ nq, nparams = ckt.TensorCoreNVFP4Layout.quantize(deq.contiguous()) # -> NVFP4
102
+ tensors = ckt.TensorCoreNVFP4Layout.state_dict_tensors(nq, nparams)
103
+ ```
104
+
105
+ Per-layer config lives in a `comfy_quant` uint8 tensor holding JSON, e.g.
106
+ `{"format": "int8_tensorwise", "convrot": true, "convrot_groupsize": 256}`, rewritten to
107
+ `{"format": "nvfp4"}` on output. Note `quantize()` requires bf16/fp16 — float32 raises
108
+ `Unsupported dtype code`.
109
+
110
+ Full script: `pruned_to_nvfp4.py` in this repo. 200 layers, ~6 seconds on one GPU.
111
+
112
+ ## Prompting: H3 wants a structured IR, not prose
113
+
114
+ **Read this before blaming the weights for bad output.** H3 was trained on the structured
115
+ output of **H3-Context-IR**, a preprocessing model that rewrites a plain request into
116
+ labelled sections; MiniMax's model card calls it "critical to the quality of the final
117
+ output". ComfyUI passes your raw string straight to the DiT, so you must write that
118
+ structure yourself.
119
+
120
+ Official guides:
121
+ [base](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md) ·
122
+ [ref](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md)
123
+
124
+ ```text
125
+ For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
126
+
127
+ integrated_multimodal_description: [Shot 1] Live-action, cinematic, the young woman shown in
128
+ <Picture 1> remains beside the rain-covered train window, preserving her appearance and the
129
+ carriage layout. The camera trucks right with small amplitude at slow speed as she lifts her
130
+ gaze toward the passing city lights. The quiet, breathy young woman (S1) says:
131
+ <d>[English] I get off at the next station.</d> She folds the letter along its existing crease.
132
+
133
+ overall_soundscape: The train wheels produce a steady metallic rhythm beneath a low
134
+ ventilation hum. Rain ticks against the window while paper rustles softly in her hands.
135
+
136
+ non_diegetic_music: Sustained cello notes at a slow tempo with widely spaced piano tones.
137
+ ```
138
+
139
+ **Dialogue must be explicit or you get gibberish.** Speech is generated jointly with video,
140
+ so saying *that* someone speaks without giving the words yields correct prosody and mouth
141
+ shapes with no lexical content. Speaker identity, action and delivery go *outside* `<d>`;
142
+ only the language tag and verbatim words go *inside*. Use stable IDs `(S1)`, `(S2)`, and
143
+ `(S1,S2)` for simultaneous speech.
144
+
145
+ Other essentials: `[Shot 1]` carries no timestamp, later shots use `[Shot N] At MM:SS.mmm`;
146
+ aim for 350-500 words of description; write camera motion as type + amplitude + speed;
147
+ reference tags must appear in the order the inputs were connected. ref2va accepts up to 9
148
+ reference images, and 3-4 varied shots hold identity far better than one.
149
+
150
+ ## Companion files (mirrored, not ours)
151
+
152
+ | file | precision | origin |
153
+ |---|---|---|
154
+ | `text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors` | NVFP4-AWQ | **Comfy-Org, unmodified** |
155
+ | `vae/minimax_h3_video_vae_fp16.safetensors` | FP16 | **Comfy-Org, unmodified** |
156
+ | `vae/minimax_h3_audio_vae_fp32.safetensors` | FP32 | **Comfy-Org, unmodified** |
157
+
158
+ **The NVFP4 text encoder is Comfy-Org's work, not ours.** Only the
159
+ `minimax_h3_ref2va_*nvfp4*.safetensors` diffusion models here are new.
160
+
161
+ VAEs are deliberately **not** quantized: they are small, run once per generation rather than
162
+ per step, and decode straight to pixels and audio samples where error is immediately
163
+ visible. The text encoder being NVFP4 buys VRAM, not speed — it also runs once per prompt.
164
+
165
+ ## Usage
166
+
167
+ ```
168
+ 📂 ComfyUI/models/
169
+ ├── 📂 diffusion_models/
170
+ │ └── minimax_h3_ref2va_pruned_nvfp4.safetensors
171
+ ├── 📂 text_encoders/
172
+ │ └── qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
173
+ └── 📂 vae/
174
+ ├── minimax_h3_video_vae_fp16.safetensors
175
+ └── minimax_h3_audio_vae_fp32.safetensors
176
+ ```
177
+
178
+ Recommended stack total: **33.7 GB**. Use the official
179
+ [R2V template](https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_minimax_h3_r2v.json),
180
+ swapping the diffusion model. Requires ComfyUI >= 0.30.0 (native H3 support in
181
+ `comfy/ldm/minimax/`). `CLIPLoader` type must be `minimax`; sampler `res_multistep`;
182
+ frame `length` must satisfy 17n+5.
183
+
184
+ ## Honest limitations
185
+
186
+ - Quality was compared against `pruned_int8_convrot` at three matched seeds with no visible
187
+ degradation, but this is **not a rigorous evaluation** — no FVD, no human study, no
188
+ long-duration or 2K testing.
189
+ - `pruned_nvfp4` is **doubly quantized** (bf16 -> int8_convrot by Comfy-Org -> NVFP4 here).
190
+ Error from both passes compounds. It held up in testing, but that is a real caveat.
191
+ - Only **ref2va** is converted. `fl2va` is not included.
192
+ - Benchmarks are single-GPU, one card, one resolution.
193
+
194
+ Failure cases are welcome in the discussions tab — concrete artifacts beat aggregate scores.
195
+
196
+ ## License
197
+
198
+ Inherits the MiniMax-H3 Community License Agreement from the original model.
make_profile.py ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Generate StarNodes converter profiles for MiniMax-H3 from the bf16 safetensors header.
2
+
3
+ Two variants:
4
+ minimax_h3_nvfp4_mixed.json - AdaLN kept at FP8, attn/mlp at NVFP4 (recommended)
5
+ minimax_h3_nvfp4_full.json - everything incl. AdaLN at NVFP4 (aggressive)
6
+
7
+ Kept at BF16 in both: all biases, all norms, rope freqs, patch/time/condition
8
+ embedders and the final output heads (~0.04B params total, 0.1% of the model).
9
+ """
10
+ import json, struct, sys, os, datetime
11
+
12
+ SRC = "/workspace/ComfyUI/models/diffusion_models/minimax_h3_ref2va_bf16.safetensors"
13
+ OUT_DIR = "/workspace/ComfyUI/custom_nodes/comfyui-starnodes-modelconverter/profiles"
14
+
15
+ with open(SRC, "rb") as f:
16
+ n = struct.unpack("<Q", f.read(8))[0]
17
+ hdr = json.loads(f.read(n))
18
+ keys = sorted(k for k in hdr if k != "__metadata__")
19
+
20
+
21
+ def classify(key, adaln_fmt):
22
+ # non-weight tensors and everything tiny stays bf16
23
+ if not key.endswith(".weight"):
24
+ return "BF16"
25
+ if "norm" in key or key.endswith("inv_freq"):
26
+ return "BF16"
27
+ if any(t in key for t in ("patch_proj", "time_embedder", "condition_proj", "final_layer")):
28
+ return "BF16"
29
+ if "adaln" in key:
30
+ return adaln_fmt
31
+ if ".attn." in key or ".mlp." in key:
32
+ return "NVFP4"
33
+ return "BF16"
34
+
35
+
36
+ def build(adaln_fmt, name):
37
+ layers, counts = {}, {}
38
+ for k in keys:
39
+ fmt = classify(k, adaln_fmt)
40
+ layers[k] = fmt
41
+ counts[fmt] = counts.get(fmt, 0) + 1
42
+ if fmt != "BF16":
43
+ base = k[: -len(".weight")]
44
+ layers[f"{base}.weight_scale"] = "FP32_SCALE"
45
+ layers[f"{base}.comfy_quant"] = "METADATA"
46
+ prof = {
47
+ "__metadata__": {
48
+ "original_model_name": name,
49
+ "original_model_path": SRC,
50
+ "timestamp": datetime.datetime.now().isoformat(),
51
+ "total_layers": len(layers),
52
+ "created_by": "hand-authored for MiniMax-H3 (33.12B: adaln 39.4%, mlp 36.3%, attn 24.2%)",
53
+ },
54
+ "layers": layers,
55
+ }
56
+ path = os.path.join(OUT_DIR, f"{name}.json")
57
+ with open(path, "w") as f:
58
+ json.dump(prof, f, indent=1)
59
+ print(f"{name}: {counts} -> {path}")
60
+
61
+
62
+ build("FP8_E4M3FN + SCALE", "minimax_h3_nvfp4_mixed")
63
+ build("NVFP4", "minimax_h3_nvfp4_full")
minimax_h3_ref2va_nvfp4_full.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5f50168a8eb86583d24531169d40baf34879d58cb9877217390725a282b50e9e
3
+ size 18745492256
minimax_h3_ref2va_nvfp4_mixed.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7a810cecf28317e17c7009e4e33c2d547a298db12ac14bb9f2a79bac7e65a9b2
3
+ size 24435437368
minimax_h3_ref2va_pruned_nvfp4.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8eea02f43902e69904990c4405968d01a13c6656aa392d37ab80331e79b5df2f
3
+ size 12528636800
profiles/minimax_h3_nvfp4_full.json ADDED
@@ -0,0 +1,1062 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "__metadata__": {
3
+ "original_model_name": "minimax_h3_nvfp4_full",
4
+ "original_model_path": "/workspace/ComfyUI/models/diffusion_models/minimax_h3_ref2va_bf16.safetensors",
5
+ "timestamp": "2026-08-03T06:06:50.561944",
6
+ "total_layers": 1051,
7
+ "created_by": "hand-authored for MiniMax-H3 (33.12B: adaln 39.4%, mlp 36.3%, attn 24.2%)"
8
+ },
9
+ "layers": {
10
+ "audio_patch_proj.bias": "BF16",
11
+ "audio_patch_proj.weight": "BF16",
12
+ "blocks.0.adaln_proj.linear.bias": "BF16",
13
+ "blocks.0.adaln_proj.linear.weight": "NVFP4",
14
+ "blocks.0.adaln_proj.linear.weight_scale": "FP32_SCALE",
15
+ "blocks.0.adaln_proj.linear.comfy_quant": "METADATA",
16
+ "blocks.0.attn.k_norm.weight": "BF16",
17
+ "blocks.0.attn.out_proj.weight": "NVFP4",
18
+ "blocks.0.attn.out_proj.weight_scale": "FP32_SCALE",
19
+ "blocks.0.attn.out_proj.comfy_quant": "METADATA",
20
+ "blocks.0.attn.q_norm.weight": "BF16",
21
+ "blocks.0.attn.qkv_proj.weight": "NVFP4",
22
+ "blocks.0.attn.qkv_proj.weight_scale": "FP32_SCALE",
23
+ "blocks.0.attn.qkv_proj.comfy_quant": "METADATA",
24
+ "blocks.0.mlp.fc1.weight": "NVFP4",
25
+ "blocks.0.mlp.fc1.weight_scale": "FP32_SCALE",
26
+ "blocks.0.mlp.fc1.comfy_quant": "METADATA",
27
+ "blocks.0.mlp.fc2.weight": "NVFP4",
28
+ "blocks.0.mlp.fc2.weight_scale": "FP32_SCALE",
29
+ "blocks.0.mlp.fc2.comfy_quant": "METADATA",
30
+ "blocks.0.norm1.weight": "BF16",
31
+ "blocks.0.norm2.weight": "BF16",
32
+ "blocks.1.adaln_proj.linear.bias": "BF16",
33
+ "blocks.1.adaln_proj.linear.weight": "NVFP4",
34
+ "blocks.1.adaln_proj.linear.weight_scale": "FP32_SCALE",
35
+ "blocks.1.adaln_proj.linear.comfy_quant": "METADATA",
36
+ "blocks.1.attn.k_norm.weight": "BF16",
37
+ "blocks.1.attn.out_proj.weight": "NVFP4",
38
+ "blocks.1.attn.out_proj.weight_scale": "FP32_SCALE",
39
+ "blocks.1.attn.out_proj.comfy_quant": "METADATA",
40
+ "blocks.1.attn.q_norm.weight": "BF16",
41
+ "blocks.1.attn.qkv_proj.weight": "NVFP4",
42
+ "blocks.1.attn.qkv_proj.weight_scale": "FP32_SCALE",
43
+ "blocks.1.attn.qkv_proj.comfy_quant": "METADATA",
44
+ "blocks.1.mlp.fc1.weight": "NVFP4",
45
+ "blocks.1.mlp.fc1.weight_scale": "FP32_SCALE",
46
+ "blocks.1.mlp.fc1.comfy_quant": "METADATA",
47
+ "blocks.1.mlp.fc2.weight": "NVFP4",
48
+ "blocks.1.mlp.fc2.weight_scale": "FP32_SCALE",
49
+ "blocks.1.mlp.fc2.comfy_quant": "METADATA",
50
+ "blocks.1.norm1.weight": "BF16",
51
+ "blocks.1.norm2.weight": "BF16",
52
+ "blocks.10.adaln_proj.linear.bias": "BF16",
53
+ "blocks.10.adaln_proj.linear.weight": "NVFP4",
54
+ "blocks.10.adaln_proj.linear.weight_scale": "FP32_SCALE",
55
+ "blocks.10.adaln_proj.linear.comfy_quant": "METADATA",
56
+ "blocks.10.attn.k_norm.weight": "BF16",
57
+ "blocks.10.attn.out_proj.weight": "NVFP4",
58
+ "blocks.10.attn.out_proj.weight_scale": "FP32_SCALE",
59
+ "blocks.10.attn.out_proj.comfy_quant": "METADATA",
60
+ "blocks.10.attn.q_norm.weight": "BF16",
61
+ "blocks.10.attn.qkv_proj.weight": "NVFP4",
62
+ "blocks.10.attn.qkv_proj.weight_scale": "FP32_SCALE",
63
+ "blocks.10.attn.qkv_proj.comfy_quant": "METADATA",
64
+ "blocks.10.mlp.fc1.weight": "NVFP4",
65
+ "blocks.10.mlp.fc1.weight_scale": "FP32_SCALE",
66
+ "blocks.10.mlp.fc1.comfy_quant": "METADATA",
67
+ "blocks.10.mlp.fc2.weight": "NVFP4",
68
+ "blocks.10.mlp.fc2.weight_scale": "FP32_SCALE",
69
+ "blocks.10.mlp.fc2.comfy_quant": "METADATA",
70
+ "blocks.10.norm1.weight": "BF16",
71
+ "blocks.10.norm2.weight": "BF16",
72
+ "blocks.11.adaln_proj.linear.bias": "BF16",
73
+ "blocks.11.adaln_proj.linear.weight": "NVFP4",
74
+ "blocks.11.adaln_proj.linear.weight_scale": "FP32_SCALE",
75
+ "blocks.11.adaln_proj.linear.comfy_quant": "METADATA",
76
+ "blocks.11.attn.k_norm.weight": "BF16",
77
+ "blocks.11.attn.out_proj.weight": "NVFP4",
78
+ "blocks.11.attn.out_proj.weight_scale": "FP32_SCALE",
79
+ "blocks.11.attn.out_proj.comfy_quant": "METADATA",
80
+ "blocks.11.attn.q_norm.weight": "BF16",
81
+ "blocks.11.attn.qkv_proj.weight": "NVFP4",
82
+ "blocks.11.attn.qkv_proj.weight_scale": "FP32_SCALE",
83
+ "blocks.11.attn.qkv_proj.comfy_quant": "METADATA",
84
+ "blocks.11.mlp.fc1.weight": "NVFP4",
85
+ "blocks.11.mlp.fc1.weight_scale": "FP32_SCALE",
86
+ "blocks.11.mlp.fc1.comfy_quant": "METADATA",
87
+ "blocks.11.mlp.fc2.weight": "NVFP4",
88
+ "blocks.11.mlp.fc2.weight_scale": "FP32_SCALE",
89
+ "blocks.11.mlp.fc2.comfy_quant": "METADATA",
90
+ "blocks.11.norm1.weight": "BF16",
91
+ "blocks.11.norm2.weight": "BF16",
92
+ "blocks.12.adaln_proj.linear.bias": "BF16",
93
+ "blocks.12.adaln_proj.linear.weight": "NVFP4",
94
+ "blocks.12.adaln_proj.linear.weight_scale": "FP32_SCALE",
95
+ "blocks.12.adaln_proj.linear.comfy_quant": "METADATA",
96
+ "blocks.12.attn.k_norm.weight": "BF16",
97
+ "blocks.12.attn.out_proj.weight": "NVFP4",
98
+ "blocks.12.attn.out_proj.weight_scale": "FP32_SCALE",
99
+ "blocks.12.attn.out_proj.comfy_quant": "METADATA",
100
+ "blocks.12.attn.q_norm.weight": "BF16",
101
+ "blocks.12.attn.qkv_proj.weight": "NVFP4",
102
+ "blocks.12.attn.qkv_proj.weight_scale": "FP32_SCALE",
103
+ "blocks.12.attn.qkv_proj.comfy_quant": "METADATA",
104
+ "blocks.12.mlp.fc1.weight": "NVFP4",
105
+ "blocks.12.mlp.fc1.weight_scale": "FP32_SCALE",
106
+ "blocks.12.mlp.fc1.comfy_quant": "METADATA",
107
+ "blocks.12.mlp.fc2.weight": "NVFP4",
108
+ "blocks.12.mlp.fc2.weight_scale": "FP32_SCALE",
109
+ "blocks.12.mlp.fc2.comfy_quant": "METADATA",
110
+ "blocks.12.norm1.weight": "BF16",
111
+ "blocks.12.norm2.weight": "BF16",
112
+ "blocks.13.adaln_proj.linear.bias": "BF16",
113
+ "blocks.13.adaln_proj.linear.weight": "NVFP4",
114
+ "blocks.13.adaln_proj.linear.weight_scale": "FP32_SCALE",
115
+ "blocks.13.adaln_proj.linear.comfy_quant": "METADATA",
116
+ "blocks.13.attn.k_norm.weight": "BF16",
117
+ "blocks.13.attn.out_proj.weight": "NVFP4",
118
+ "blocks.13.attn.out_proj.weight_scale": "FP32_SCALE",
119
+ "blocks.13.attn.out_proj.comfy_quant": "METADATA",
120
+ "blocks.13.attn.q_norm.weight": "BF16",
121
+ "blocks.13.attn.qkv_proj.weight": "NVFP4",
122
+ "blocks.13.attn.qkv_proj.weight_scale": "FP32_SCALE",
123
+ "blocks.13.attn.qkv_proj.comfy_quant": "METADATA",
124
+ "blocks.13.mlp.fc1.weight": "NVFP4",
125
+ "blocks.13.mlp.fc1.weight_scale": "FP32_SCALE",
126
+ "blocks.13.mlp.fc1.comfy_quant": "METADATA",
127
+ "blocks.13.mlp.fc2.weight": "NVFP4",
128
+ "blocks.13.mlp.fc2.weight_scale": "FP32_SCALE",
129
+ "blocks.13.mlp.fc2.comfy_quant": "METADATA",
130
+ "blocks.13.norm1.weight": "BF16",
131
+ "blocks.13.norm2.weight": "BF16",
132
+ "blocks.14.adaln_proj.linear.bias": "BF16",
133
+ "blocks.14.adaln_proj.linear.weight": "NVFP4",
134
+ "blocks.14.adaln_proj.linear.weight_scale": "FP32_SCALE",
135
+ "blocks.14.adaln_proj.linear.comfy_quant": "METADATA",
136
+ "blocks.14.attn.k_norm.weight": "BF16",
137
+ "blocks.14.attn.out_proj.weight": "NVFP4",
138
+ "blocks.14.attn.out_proj.weight_scale": "FP32_SCALE",
139
+ "blocks.14.attn.out_proj.comfy_quant": "METADATA",
140
+ "blocks.14.attn.q_norm.weight": "BF16",
141
+ "blocks.14.attn.qkv_proj.weight": "NVFP4",
142
+ "blocks.14.attn.qkv_proj.weight_scale": "FP32_SCALE",
143
+ "blocks.14.attn.qkv_proj.comfy_quant": "METADATA",
144
+ "blocks.14.mlp.fc1.weight": "NVFP4",
145
+ "blocks.14.mlp.fc1.weight_scale": "FP32_SCALE",
146
+ "blocks.14.mlp.fc1.comfy_quant": "METADATA",
147
+ "blocks.14.mlp.fc2.weight": "NVFP4",
148
+ "blocks.14.mlp.fc2.weight_scale": "FP32_SCALE",
149
+ "blocks.14.mlp.fc2.comfy_quant": "METADATA",
150
+ "blocks.14.norm1.weight": "BF16",
151
+ "blocks.14.norm2.weight": "BF16",
152
+ "blocks.15.adaln_proj.linear.bias": "BF16",
153
+ "blocks.15.adaln_proj.linear.weight": "NVFP4",
154
+ "blocks.15.adaln_proj.linear.weight_scale": "FP32_SCALE",
155
+ "blocks.15.adaln_proj.linear.comfy_quant": "METADATA",
156
+ "blocks.15.attn.k_norm.weight": "BF16",
157
+ "blocks.15.attn.out_proj.weight": "NVFP4",
158
+ "blocks.15.attn.out_proj.weight_scale": "FP32_SCALE",
159
+ "blocks.15.attn.out_proj.comfy_quant": "METADATA",
160
+ "blocks.15.attn.q_norm.weight": "BF16",
161
+ "blocks.15.attn.qkv_proj.weight": "NVFP4",
162
+ "blocks.15.attn.qkv_proj.weight_scale": "FP32_SCALE",
163
+ "blocks.15.attn.qkv_proj.comfy_quant": "METADATA",
164
+ "blocks.15.mlp.fc1.weight": "NVFP4",
165
+ "blocks.15.mlp.fc1.weight_scale": "FP32_SCALE",
166
+ "blocks.15.mlp.fc1.comfy_quant": "METADATA",
167
+ "blocks.15.mlp.fc2.weight": "NVFP4",
168
+ "blocks.15.mlp.fc2.weight_scale": "FP32_SCALE",
169
+ "blocks.15.mlp.fc2.comfy_quant": "METADATA",
170
+ "blocks.15.norm1.weight": "BF16",
171
+ "blocks.15.norm2.weight": "BF16",
172
+ "blocks.16.adaln_proj.linear.bias": "BF16",
173
+ "blocks.16.adaln_proj.linear.weight": "NVFP4",
174
+ "blocks.16.adaln_proj.linear.weight_scale": "FP32_SCALE",
175
+ "blocks.16.adaln_proj.linear.comfy_quant": "METADATA",
176
+ "blocks.16.attn.k_norm.weight": "BF16",
177
+ "blocks.16.attn.out_proj.weight": "NVFP4",
178
+ "blocks.16.attn.out_proj.weight_scale": "FP32_SCALE",
179
+ "blocks.16.attn.out_proj.comfy_quant": "METADATA",
180
+ "blocks.16.attn.q_norm.weight": "BF16",
181
+ "blocks.16.attn.qkv_proj.weight": "NVFP4",
182
+ "blocks.16.attn.qkv_proj.weight_scale": "FP32_SCALE",
183
+ "blocks.16.attn.qkv_proj.comfy_quant": "METADATA",
184
+ "blocks.16.mlp.fc1.weight": "NVFP4",
185
+ "blocks.16.mlp.fc1.weight_scale": "FP32_SCALE",
186
+ "blocks.16.mlp.fc1.comfy_quant": "METADATA",
187
+ "blocks.16.mlp.fc2.weight": "NVFP4",
188
+ "blocks.16.mlp.fc2.weight_scale": "FP32_SCALE",
189
+ "blocks.16.mlp.fc2.comfy_quant": "METADATA",
190
+ "blocks.16.norm1.weight": "BF16",
191
+ "blocks.16.norm2.weight": "BF16",
192
+ "blocks.17.adaln_proj.linear.bias": "BF16",
193
+ "blocks.17.adaln_proj.linear.weight": "NVFP4",
194
+ "blocks.17.adaln_proj.linear.weight_scale": "FP32_SCALE",
195
+ "blocks.17.adaln_proj.linear.comfy_quant": "METADATA",
196
+ "blocks.17.attn.k_norm.weight": "BF16",
197
+ "blocks.17.attn.out_proj.weight": "NVFP4",
198
+ "blocks.17.attn.out_proj.weight_scale": "FP32_SCALE",
199
+ "blocks.17.attn.out_proj.comfy_quant": "METADATA",
200
+ "blocks.17.attn.q_norm.weight": "BF16",
201
+ "blocks.17.attn.qkv_proj.weight": "NVFP4",
202
+ "blocks.17.attn.qkv_proj.weight_scale": "FP32_SCALE",
203
+ "blocks.17.attn.qkv_proj.comfy_quant": "METADATA",
204
+ "blocks.17.mlp.fc1.weight": "NVFP4",
205
+ "blocks.17.mlp.fc1.weight_scale": "FP32_SCALE",
206
+ "blocks.17.mlp.fc1.comfy_quant": "METADATA",
207
+ "blocks.17.mlp.fc2.weight": "NVFP4",
208
+ "blocks.17.mlp.fc2.weight_scale": "FP32_SCALE",
209
+ "blocks.17.mlp.fc2.comfy_quant": "METADATA",
210
+ "blocks.17.norm1.weight": "BF16",
211
+ "blocks.17.norm2.weight": "BF16",
212
+ "blocks.18.adaln_proj.linear.bias": "BF16",
213
+ "blocks.18.adaln_proj.linear.weight": "NVFP4",
214
+ "blocks.18.adaln_proj.linear.weight_scale": "FP32_SCALE",
215
+ "blocks.18.adaln_proj.linear.comfy_quant": "METADATA",
216
+ "blocks.18.attn.k_norm.weight": "BF16",
217
+ "blocks.18.attn.out_proj.weight": "NVFP4",
218
+ "blocks.18.attn.out_proj.weight_scale": "FP32_SCALE",
219
+ "blocks.18.attn.out_proj.comfy_quant": "METADATA",
220
+ "blocks.18.attn.q_norm.weight": "BF16",
221
+ "blocks.18.attn.qkv_proj.weight": "NVFP4",
222
+ "blocks.18.attn.qkv_proj.weight_scale": "FP32_SCALE",
223
+ "blocks.18.attn.qkv_proj.comfy_quant": "METADATA",
224
+ "blocks.18.mlp.fc1.weight": "NVFP4",
225
+ "blocks.18.mlp.fc1.weight_scale": "FP32_SCALE",
226
+ "blocks.18.mlp.fc1.comfy_quant": "METADATA",
227
+ "blocks.18.mlp.fc2.weight": "NVFP4",
228
+ "blocks.18.mlp.fc2.weight_scale": "FP32_SCALE",
229
+ "blocks.18.mlp.fc2.comfy_quant": "METADATA",
230
+ "blocks.18.norm1.weight": "BF16",
231
+ "blocks.18.norm2.weight": "BF16",
232
+ "blocks.19.adaln_proj.linear.bias": "BF16",
233
+ "blocks.19.adaln_proj.linear.weight": "NVFP4",
234
+ "blocks.19.adaln_proj.linear.weight_scale": "FP32_SCALE",
235
+ "blocks.19.adaln_proj.linear.comfy_quant": "METADATA",
236
+ "blocks.19.attn.k_norm.weight": "BF16",
237
+ "blocks.19.attn.out_proj.weight": "NVFP4",
238
+ "blocks.19.attn.out_proj.weight_scale": "FP32_SCALE",
239
+ "blocks.19.attn.out_proj.comfy_quant": "METADATA",
240
+ "blocks.19.attn.q_norm.weight": "BF16",
241
+ "blocks.19.attn.qkv_proj.weight": "NVFP4",
242
+ "blocks.19.attn.qkv_proj.weight_scale": "FP32_SCALE",
243
+ "blocks.19.attn.qkv_proj.comfy_quant": "METADATA",
244
+ "blocks.19.mlp.fc1.weight": "NVFP4",
245
+ "blocks.19.mlp.fc1.weight_scale": "FP32_SCALE",
246
+ "blocks.19.mlp.fc1.comfy_quant": "METADATA",
247
+ "blocks.19.mlp.fc2.weight": "NVFP4",
248
+ "blocks.19.mlp.fc2.weight_scale": "FP32_SCALE",
249
+ "blocks.19.mlp.fc2.comfy_quant": "METADATA",
250
+ "blocks.19.norm1.weight": "BF16",
251
+ "blocks.19.norm2.weight": "BF16",
252
+ "blocks.2.adaln_proj.linear.bias": "BF16",
253
+ "blocks.2.adaln_proj.linear.weight": "NVFP4",
254
+ "blocks.2.adaln_proj.linear.weight_scale": "FP32_SCALE",
255
+ "blocks.2.adaln_proj.linear.comfy_quant": "METADATA",
256
+ "blocks.2.attn.k_norm.weight": "BF16",
257
+ "blocks.2.attn.out_proj.weight": "NVFP4",
258
+ "blocks.2.attn.out_proj.weight_scale": "FP32_SCALE",
259
+ "blocks.2.attn.out_proj.comfy_quant": "METADATA",
260
+ "blocks.2.attn.q_norm.weight": "BF16",
261
+ "blocks.2.attn.qkv_proj.weight": "NVFP4",
262
+ "blocks.2.attn.qkv_proj.weight_scale": "FP32_SCALE",
263
+ "blocks.2.attn.qkv_proj.comfy_quant": "METADATA",
264
+ "blocks.2.mlp.fc1.weight": "NVFP4",
265
+ "blocks.2.mlp.fc1.weight_scale": "FP32_SCALE",
266
+ "blocks.2.mlp.fc1.comfy_quant": "METADATA",
267
+ "blocks.2.mlp.fc2.weight": "NVFP4",
268
+ "blocks.2.mlp.fc2.weight_scale": "FP32_SCALE",
269
+ "blocks.2.mlp.fc2.comfy_quant": "METADATA",
270
+ "blocks.2.norm1.weight": "BF16",
271
+ "blocks.2.norm2.weight": "BF16",
272
+ "blocks.20.adaln_proj.linear.bias": "BF16",
273
+ "blocks.20.adaln_proj.linear.weight": "NVFP4",
274
+ "blocks.20.adaln_proj.linear.weight_scale": "FP32_SCALE",
275
+ "blocks.20.adaln_proj.linear.comfy_quant": "METADATA",
276
+ "blocks.20.attn.k_norm.weight": "BF16",
277
+ "blocks.20.attn.out_proj.weight": "NVFP4",
278
+ "blocks.20.attn.out_proj.weight_scale": "FP32_SCALE",
279
+ "blocks.20.attn.out_proj.comfy_quant": "METADATA",
280
+ "blocks.20.attn.q_norm.weight": "BF16",
281
+ "blocks.20.attn.qkv_proj.weight": "NVFP4",
282
+ "blocks.20.attn.qkv_proj.weight_scale": "FP32_SCALE",
283
+ "blocks.20.attn.qkv_proj.comfy_quant": "METADATA",
284
+ "blocks.20.mlp.fc1.weight": "NVFP4",
285
+ "blocks.20.mlp.fc1.weight_scale": "FP32_SCALE",
286
+ "blocks.20.mlp.fc1.comfy_quant": "METADATA",
287
+ "blocks.20.mlp.fc2.weight": "NVFP4",
288
+ "blocks.20.mlp.fc2.weight_scale": "FP32_SCALE",
289
+ "blocks.20.mlp.fc2.comfy_quant": "METADATA",
290
+ "blocks.20.norm1.weight": "BF16",
291
+ "blocks.20.norm2.weight": "BF16",
292
+ "blocks.21.adaln_proj.linear.bias": "BF16",
293
+ "blocks.21.adaln_proj.linear.weight": "NVFP4",
294
+ "blocks.21.adaln_proj.linear.weight_scale": "FP32_SCALE",
295
+ "blocks.21.adaln_proj.linear.comfy_quant": "METADATA",
296
+ "blocks.21.attn.k_norm.weight": "BF16",
297
+ "blocks.21.attn.out_proj.weight": "NVFP4",
298
+ "blocks.21.attn.out_proj.weight_scale": "FP32_SCALE",
299
+ "blocks.21.attn.out_proj.comfy_quant": "METADATA",
300
+ "blocks.21.attn.q_norm.weight": "BF16",
301
+ "blocks.21.attn.qkv_proj.weight": "NVFP4",
302
+ "blocks.21.attn.qkv_proj.weight_scale": "FP32_SCALE",
303
+ "blocks.21.attn.qkv_proj.comfy_quant": "METADATA",
304
+ "blocks.21.mlp.fc1.weight": "NVFP4",
305
+ "blocks.21.mlp.fc1.weight_scale": "FP32_SCALE",
306
+ "blocks.21.mlp.fc1.comfy_quant": "METADATA",
307
+ "blocks.21.mlp.fc2.weight": "NVFP4",
308
+ "blocks.21.mlp.fc2.weight_scale": "FP32_SCALE",
309
+ "blocks.21.mlp.fc2.comfy_quant": "METADATA",
310
+ "blocks.21.norm1.weight": "BF16",
311
+ "blocks.21.norm2.weight": "BF16",
312
+ "blocks.22.adaln_proj.linear.bias": "BF16",
313
+ "blocks.22.adaln_proj.linear.weight": "NVFP4",
314
+ "blocks.22.adaln_proj.linear.weight_scale": "FP32_SCALE",
315
+ "blocks.22.adaln_proj.linear.comfy_quant": "METADATA",
316
+ "blocks.22.attn.k_norm.weight": "BF16",
317
+ "blocks.22.attn.out_proj.weight": "NVFP4",
318
+ "blocks.22.attn.out_proj.weight_scale": "FP32_SCALE",
319
+ "blocks.22.attn.out_proj.comfy_quant": "METADATA",
320
+ "blocks.22.attn.q_norm.weight": "BF16",
321
+ "blocks.22.attn.qkv_proj.weight": "NVFP4",
322
+ "blocks.22.attn.qkv_proj.weight_scale": "FP32_SCALE",
323
+ "blocks.22.attn.qkv_proj.comfy_quant": "METADATA",
324
+ "blocks.22.mlp.fc1.weight": "NVFP4",
325
+ "blocks.22.mlp.fc1.weight_scale": "FP32_SCALE",
326
+ "blocks.22.mlp.fc1.comfy_quant": "METADATA",
327
+ "blocks.22.mlp.fc2.weight": "NVFP4",
328
+ "blocks.22.mlp.fc2.weight_scale": "FP32_SCALE",
329
+ "blocks.22.mlp.fc2.comfy_quant": "METADATA",
330
+ "blocks.22.norm1.weight": "BF16",
331
+ "blocks.22.norm2.weight": "BF16",
332
+ "blocks.23.adaln_proj.linear.bias": "BF16",
333
+ "blocks.23.adaln_proj.linear.weight": "NVFP4",
334
+ "blocks.23.adaln_proj.linear.weight_scale": "FP32_SCALE",
335
+ "blocks.23.adaln_proj.linear.comfy_quant": "METADATA",
336
+ "blocks.23.attn.k_norm.weight": "BF16",
337
+ "blocks.23.attn.out_proj.weight": "NVFP4",
338
+ "blocks.23.attn.out_proj.weight_scale": "FP32_SCALE",
339
+ "blocks.23.attn.out_proj.comfy_quant": "METADATA",
340
+ "blocks.23.attn.q_norm.weight": "BF16",
341
+ "blocks.23.attn.qkv_proj.weight": "NVFP4",
342
+ "blocks.23.attn.qkv_proj.weight_scale": "FP32_SCALE",
343
+ "blocks.23.attn.qkv_proj.comfy_quant": "METADATA",
344
+ "blocks.23.mlp.fc1.weight": "NVFP4",
345
+ "blocks.23.mlp.fc1.weight_scale": "FP32_SCALE",
346
+ "blocks.23.mlp.fc1.comfy_quant": "METADATA",
347
+ "blocks.23.mlp.fc2.weight": "NVFP4",
348
+ "blocks.23.mlp.fc2.weight_scale": "FP32_SCALE",
349
+ "blocks.23.mlp.fc2.comfy_quant": "METADATA",
350
+ "blocks.23.norm1.weight": "BF16",
351
+ "blocks.23.norm2.weight": "BF16",
352
+ "blocks.24.adaln_proj.linear.bias": "BF16",
353
+ "blocks.24.adaln_proj.linear.weight": "NVFP4",
354
+ "blocks.24.adaln_proj.linear.weight_scale": "FP32_SCALE",
355
+ "blocks.24.adaln_proj.linear.comfy_quant": "METADATA",
356
+ "blocks.24.attn.k_norm.weight": "BF16",
357
+ "blocks.24.attn.out_proj.weight": "NVFP4",
358
+ "blocks.24.attn.out_proj.weight_scale": "FP32_SCALE",
359
+ "blocks.24.attn.out_proj.comfy_quant": "METADATA",
360
+ "blocks.24.attn.q_norm.weight": "BF16",
361
+ "blocks.24.attn.qkv_proj.weight": "NVFP4",
362
+ "blocks.24.attn.qkv_proj.weight_scale": "FP32_SCALE",
363
+ "blocks.24.attn.qkv_proj.comfy_quant": "METADATA",
364
+ "blocks.24.mlp.fc1.weight": "NVFP4",
365
+ "blocks.24.mlp.fc1.weight_scale": "FP32_SCALE",
366
+ "blocks.24.mlp.fc1.comfy_quant": "METADATA",
367
+ "blocks.24.mlp.fc2.weight": "NVFP4",
368
+ "blocks.24.mlp.fc2.weight_scale": "FP32_SCALE",
369
+ "blocks.24.mlp.fc2.comfy_quant": "METADATA",
370
+ "blocks.24.norm1.weight": "BF16",
371
+ "blocks.24.norm2.weight": "BF16",
372
+ "blocks.25.adaln_proj.linear.bias": "BF16",
373
+ "blocks.25.adaln_proj.linear.weight": "NVFP4",
374
+ "blocks.25.adaln_proj.linear.weight_scale": "FP32_SCALE",
375
+ "blocks.25.adaln_proj.linear.comfy_quant": "METADATA",
376
+ "blocks.25.attn.k_norm.weight": "BF16",
377
+ "blocks.25.attn.out_proj.weight": "NVFP4",
378
+ "blocks.25.attn.out_proj.weight_scale": "FP32_SCALE",
379
+ "blocks.25.attn.out_proj.comfy_quant": "METADATA",
380
+ "blocks.25.attn.q_norm.weight": "BF16",
381
+ "blocks.25.attn.qkv_proj.weight": "NVFP4",
382
+ "blocks.25.attn.qkv_proj.weight_scale": "FP32_SCALE",
383
+ "blocks.25.attn.qkv_proj.comfy_quant": "METADATA",
384
+ "blocks.25.mlp.fc1.weight": "NVFP4",
385
+ "blocks.25.mlp.fc1.weight_scale": "FP32_SCALE",
386
+ "blocks.25.mlp.fc1.comfy_quant": "METADATA",
387
+ "blocks.25.mlp.fc2.weight": "NVFP4",
388
+ "blocks.25.mlp.fc2.weight_scale": "FP32_SCALE",
389
+ "blocks.25.mlp.fc2.comfy_quant": "METADATA",
390
+ "blocks.25.norm1.weight": "BF16",
391
+ "blocks.25.norm2.weight": "BF16",
392
+ "blocks.26.adaln_proj.linear.bias": "BF16",
393
+ "blocks.26.adaln_proj.linear.weight": "NVFP4",
394
+ "blocks.26.adaln_proj.linear.weight_scale": "FP32_SCALE",
395
+ "blocks.26.adaln_proj.linear.comfy_quant": "METADATA",
396
+ "blocks.26.attn.k_norm.weight": "BF16",
397
+ "blocks.26.attn.out_proj.weight": "NVFP4",
398
+ "blocks.26.attn.out_proj.weight_scale": "FP32_SCALE",
399
+ "blocks.26.attn.out_proj.comfy_quant": "METADATA",
400
+ "blocks.26.attn.q_norm.weight": "BF16",
401
+ "blocks.26.attn.qkv_proj.weight": "NVFP4",
402
+ "blocks.26.attn.qkv_proj.weight_scale": "FP32_SCALE",
403
+ "blocks.26.attn.qkv_proj.comfy_quant": "METADATA",
404
+ "blocks.26.mlp.fc1.weight": "NVFP4",
405
+ "blocks.26.mlp.fc1.weight_scale": "FP32_SCALE",
406
+ "blocks.26.mlp.fc1.comfy_quant": "METADATA",
407
+ "blocks.26.mlp.fc2.weight": "NVFP4",
408
+ "blocks.26.mlp.fc2.weight_scale": "FP32_SCALE",
409
+ "blocks.26.mlp.fc2.comfy_quant": "METADATA",
410
+ "blocks.26.norm1.weight": "BF16",
411
+ "blocks.26.norm2.weight": "BF16",
412
+ "blocks.27.adaln_proj.linear.bias": "BF16",
413
+ "blocks.27.adaln_proj.linear.weight": "NVFP4",
414
+ "blocks.27.adaln_proj.linear.weight_scale": "FP32_SCALE",
415
+ "blocks.27.adaln_proj.linear.comfy_quant": "METADATA",
416
+ "blocks.27.attn.k_norm.weight": "BF16",
417
+ "blocks.27.attn.out_proj.weight": "NVFP4",
418
+ "blocks.27.attn.out_proj.weight_scale": "FP32_SCALE",
419
+ "blocks.27.attn.out_proj.comfy_quant": "METADATA",
420
+ "blocks.27.attn.q_norm.weight": "BF16",
421
+ "blocks.27.attn.qkv_proj.weight": "NVFP4",
422
+ "blocks.27.attn.qkv_proj.weight_scale": "FP32_SCALE",
423
+ "blocks.27.attn.qkv_proj.comfy_quant": "METADATA",
424
+ "blocks.27.mlp.fc1.weight": "NVFP4",
425
+ "blocks.27.mlp.fc1.weight_scale": "FP32_SCALE",
426
+ "blocks.27.mlp.fc1.comfy_quant": "METADATA",
427
+ "blocks.27.mlp.fc2.weight": "NVFP4",
428
+ "blocks.27.mlp.fc2.weight_scale": "FP32_SCALE",
429
+ "blocks.27.mlp.fc2.comfy_quant": "METADATA",
430
+ "blocks.27.norm1.weight": "BF16",
431
+ "blocks.27.norm2.weight": "BF16",
432
+ "blocks.28.adaln_proj.linear.bias": "BF16",
433
+ "blocks.28.adaln_proj.linear.weight": "NVFP4",
434
+ "blocks.28.adaln_proj.linear.weight_scale": "FP32_SCALE",
435
+ "blocks.28.adaln_proj.linear.comfy_quant": "METADATA",
436
+ "blocks.28.attn.k_norm.weight": "BF16",
437
+ "blocks.28.attn.out_proj.weight": "NVFP4",
438
+ "blocks.28.attn.out_proj.weight_scale": "FP32_SCALE",
439
+ "blocks.28.attn.out_proj.comfy_quant": "METADATA",
440
+ "blocks.28.attn.q_norm.weight": "BF16",
441
+ "blocks.28.attn.qkv_proj.weight": "NVFP4",
442
+ "blocks.28.attn.qkv_proj.weight_scale": "FP32_SCALE",
443
+ "blocks.28.attn.qkv_proj.comfy_quant": "METADATA",
444
+ "blocks.28.mlp.fc1.weight": "NVFP4",
445
+ "blocks.28.mlp.fc1.weight_scale": "FP32_SCALE",
446
+ "blocks.28.mlp.fc1.comfy_quant": "METADATA",
447
+ "blocks.28.mlp.fc2.weight": "NVFP4",
448
+ "blocks.28.mlp.fc2.weight_scale": "FP32_SCALE",
449
+ "blocks.28.mlp.fc2.comfy_quant": "METADATA",
450
+ "blocks.28.norm1.weight": "BF16",
451
+ "blocks.28.norm2.weight": "BF16",
452
+ "blocks.29.adaln_proj.linear.bias": "BF16",
453
+ "blocks.29.adaln_proj.linear.weight": "NVFP4",
454
+ "blocks.29.adaln_proj.linear.weight_scale": "FP32_SCALE",
455
+ "blocks.29.adaln_proj.linear.comfy_quant": "METADATA",
456
+ "blocks.29.attn.k_norm.weight": "BF16",
457
+ "blocks.29.attn.out_proj.weight": "NVFP4",
458
+ "blocks.29.attn.out_proj.weight_scale": "FP32_SCALE",
459
+ "blocks.29.attn.out_proj.comfy_quant": "METADATA",
460
+ "blocks.29.attn.q_norm.weight": "BF16",
461
+ "blocks.29.attn.qkv_proj.weight": "NVFP4",
462
+ "blocks.29.attn.qkv_proj.weight_scale": "FP32_SCALE",
463
+ "blocks.29.attn.qkv_proj.comfy_quant": "METADATA",
464
+ "blocks.29.mlp.fc1.weight": "NVFP4",
465
+ "blocks.29.mlp.fc1.weight_scale": "FP32_SCALE",
466
+ "blocks.29.mlp.fc1.comfy_quant": "METADATA",
467
+ "blocks.29.mlp.fc2.weight": "NVFP4",
468
+ "blocks.29.mlp.fc2.weight_scale": "FP32_SCALE",
469
+ "blocks.29.mlp.fc2.comfy_quant": "METADATA",
470
+ "blocks.29.norm1.weight": "BF16",
471
+ "blocks.29.norm2.weight": "BF16",
472
+ "blocks.3.adaln_proj.linear.bias": "BF16",
473
+ "blocks.3.adaln_proj.linear.weight": "NVFP4",
474
+ "blocks.3.adaln_proj.linear.weight_scale": "FP32_SCALE",
475
+ "blocks.3.adaln_proj.linear.comfy_quant": "METADATA",
476
+ "blocks.3.attn.k_norm.weight": "BF16",
477
+ "blocks.3.attn.out_proj.weight": "NVFP4",
478
+ "blocks.3.attn.out_proj.weight_scale": "FP32_SCALE",
479
+ "blocks.3.attn.out_proj.comfy_quant": "METADATA",
480
+ "blocks.3.attn.q_norm.weight": "BF16",
481
+ "blocks.3.attn.qkv_proj.weight": "NVFP4",
482
+ "blocks.3.attn.qkv_proj.weight_scale": "FP32_SCALE",
483
+ "blocks.3.attn.qkv_proj.comfy_quant": "METADATA",
484
+ "blocks.3.mlp.fc1.weight": "NVFP4",
485
+ "blocks.3.mlp.fc1.weight_scale": "FP32_SCALE",
486
+ "blocks.3.mlp.fc1.comfy_quant": "METADATA",
487
+ "blocks.3.mlp.fc2.weight": "NVFP4",
488
+ "blocks.3.mlp.fc2.weight_scale": "FP32_SCALE",
489
+ "blocks.3.mlp.fc2.comfy_quant": "METADATA",
490
+ "blocks.3.norm1.weight": "BF16",
491
+ "blocks.3.norm2.weight": "BF16",
492
+ "blocks.30.adaln_proj.linear.bias": "BF16",
493
+ "blocks.30.adaln_proj.linear.weight": "NVFP4",
494
+ "blocks.30.adaln_proj.linear.weight_scale": "FP32_SCALE",
495
+ "blocks.30.adaln_proj.linear.comfy_quant": "METADATA",
496
+ "blocks.30.attn.k_norm.weight": "BF16",
497
+ "blocks.30.attn.out_proj.weight": "NVFP4",
498
+ "blocks.30.attn.out_proj.weight_scale": "FP32_SCALE",
499
+ "blocks.30.attn.out_proj.comfy_quant": "METADATA",
500
+ "blocks.30.attn.q_norm.weight": "BF16",
501
+ "blocks.30.attn.qkv_proj.weight": "NVFP4",
502
+ "blocks.30.attn.qkv_proj.weight_scale": "FP32_SCALE",
503
+ "blocks.30.attn.qkv_proj.comfy_quant": "METADATA",
504
+ "blocks.30.mlp.fc1.weight": "NVFP4",
505
+ "blocks.30.mlp.fc1.weight_scale": "FP32_SCALE",
506
+ "blocks.30.mlp.fc1.comfy_quant": "METADATA",
507
+ "blocks.30.mlp.fc2.weight": "NVFP4",
508
+ "blocks.30.mlp.fc2.weight_scale": "FP32_SCALE",
509
+ "blocks.30.mlp.fc2.comfy_quant": "METADATA",
510
+ "blocks.30.norm1.weight": "BF16",
511
+ "blocks.30.norm2.weight": "BF16",
512
+ "blocks.31.adaln_proj.linear.bias": "BF16",
513
+ "blocks.31.adaln_proj.linear.weight": "NVFP4",
514
+ "blocks.31.adaln_proj.linear.weight_scale": "FP32_SCALE",
515
+ "blocks.31.adaln_proj.linear.comfy_quant": "METADATA",
516
+ "blocks.31.attn.k_norm.weight": "BF16",
517
+ "blocks.31.attn.out_proj.weight": "NVFP4",
518
+ "blocks.31.attn.out_proj.weight_scale": "FP32_SCALE",
519
+ "blocks.31.attn.out_proj.comfy_quant": "METADATA",
520
+ "blocks.31.attn.q_norm.weight": "BF16",
521
+ "blocks.31.attn.qkv_proj.weight": "NVFP4",
522
+ "blocks.31.attn.qkv_proj.weight_scale": "FP32_SCALE",
523
+ "blocks.31.attn.qkv_proj.comfy_quant": "METADATA",
524
+ "blocks.31.mlp.fc1.weight": "NVFP4",
525
+ "blocks.31.mlp.fc1.weight_scale": "FP32_SCALE",
526
+ "blocks.31.mlp.fc1.comfy_quant": "METADATA",
527
+ "blocks.31.mlp.fc2.weight": "NVFP4",
528
+ "blocks.31.mlp.fc2.weight_scale": "FP32_SCALE",
529
+ "blocks.31.mlp.fc2.comfy_quant": "METADATA",
530
+ "blocks.31.norm1.weight": "BF16",
531
+ "blocks.31.norm2.weight": "BF16",
532
+ "blocks.32.adaln_proj.linear.bias": "BF16",
533
+ "blocks.32.adaln_proj.linear.weight": "NVFP4",
534
+ "blocks.32.adaln_proj.linear.weight_scale": "FP32_SCALE",
535
+ "blocks.32.adaln_proj.linear.comfy_quant": "METADATA",
536
+ "blocks.32.attn.k_norm.weight": "BF16",
537
+ "blocks.32.attn.out_proj.weight": "NVFP4",
538
+ "blocks.32.attn.out_proj.weight_scale": "FP32_SCALE",
539
+ "blocks.32.attn.out_proj.comfy_quant": "METADATA",
540
+ "blocks.32.attn.q_norm.weight": "BF16",
541
+ "blocks.32.attn.qkv_proj.weight": "NVFP4",
542
+ "blocks.32.attn.qkv_proj.weight_scale": "FP32_SCALE",
543
+ "blocks.32.attn.qkv_proj.comfy_quant": "METADATA",
544
+ "blocks.32.mlp.fc1.weight": "NVFP4",
545
+ "blocks.32.mlp.fc1.weight_scale": "FP32_SCALE",
546
+ "blocks.32.mlp.fc1.comfy_quant": "METADATA",
547
+ "blocks.32.mlp.fc2.weight": "NVFP4",
548
+ "blocks.32.mlp.fc2.weight_scale": "FP32_SCALE",
549
+ "blocks.32.mlp.fc2.comfy_quant": "METADATA",
550
+ "blocks.32.norm1.weight": "BF16",
551
+ "blocks.32.norm2.weight": "BF16",
552
+ "blocks.33.adaln_proj.linear.bias": "BF16",
553
+ "blocks.33.adaln_proj.linear.weight": "NVFP4",
554
+ "blocks.33.adaln_proj.linear.weight_scale": "FP32_SCALE",
555
+ "blocks.33.adaln_proj.linear.comfy_quant": "METADATA",
556
+ "blocks.33.attn.k_norm.weight": "BF16",
557
+ "blocks.33.attn.out_proj.weight": "NVFP4",
558
+ "blocks.33.attn.out_proj.weight_scale": "FP32_SCALE",
559
+ "blocks.33.attn.out_proj.comfy_quant": "METADATA",
560
+ "blocks.33.attn.q_norm.weight": "BF16",
561
+ "blocks.33.attn.qkv_proj.weight": "NVFP4",
562
+ "blocks.33.attn.qkv_proj.weight_scale": "FP32_SCALE",
563
+ "blocks.33.attn.qkv_proj.comfy_quant": "METADATA",
564
+ "blocks.33.mlp.fc1.weight": "NVFP4",
565
+ "blocks.33.mlp.fc1.weight_scale": "FP32_SCALE",
566
+ "blocks.33.mlp.fc1.comfy_quant": "METADATA",
567
+ "blocks.33.mlp.fc2.weight": "NVFP4",
568
+ "blocks.33.mlp.fc2.weight_scale": "FP32_SCALE",
569
+ "blocks.33.mlp.fc2.comfy_quant": "METADATA",
570
+ "blocks.33.norm1.weight": "BF16",
571
+ "blocks.33.norm2.weight": "BF16",
572
+ "blocks.34.adaln_proj.linear.bias": "BF16",
573
+ "blocks.34.adaln_proj.linear.weight": "NVFP4",
574
+ "blocks.34.adaln_proj.linear.weight_scale": "FP32_SCALE",
575
+ "blocks.34.adaln_proj.linear.comfy_quant": "METADATA",
576
+ "blocks.34.attn.k_norm.weight": "BF16",
577
+ "blocks.34.attn.out_proj.weight": "NVFP4",
578
+ "blocks.34.attn.out_proj.weight_scale": "FP32_SCALE",
579
+ "blocks.34.attn.out_proj.comfy_quant": "METADATA",
580
+ "blocks.34.attn.q_norm.weight": "BF16",
581
+ "blocks.34.attn.qkv_proj.weight": "NVFP4",
582
+ "blocks.34.attn.qkv_proj.weight_scale": "FP32_SCALE",
583
+ "blocks.34.attn.qkv_proj.comfy_quant": "METADATA",
584
+ "blocks.34.mlp.fc1.weight": "NVFP4",
585
+ "blocks.34.mlp.fc1.weight_scale": "FP32_SCALE",
586
+ "blocks.34.mlp.fc1.comfy_quant": "METADATA",
587
+ "blocks.34.mlp.fc2.weight": "NVFP4",
588
+ "blocks.34.mlp.fc2.weight_scale": "FP32_SCALE",
589
+ "blocks.34.mlp.fc2.comfy_quant": "METADATA",
590
+ "blocks.34.norm1.weight": "BF16",
591
+ "blocks.34.norm2.weight": "BF16",
592
+ "blocks.35.adaln_proj.linear.bias": "BF16",
593
+ "blocks.35.adaln_proj.linear.weight": "NVFP4",
594
+ "blocks.35.adaln_proj.linear.weight_scale": "FP32_SCALE",
595
+ "blocks.35.adaln_proj.linear.comfy_quant": "METADATA",
596
+ "blocks.35.attn.k_norm.weight": "BF16",
597
+ "blocks.35.attn.out_proj.weight": "NVFP4",
598
+ "blocks.35.attn.out_proj.weight_scale": "FP32_SCALE",
599
+ "blocks.35.attn.out_proj.comfy_quant": "METADATA",
600
+ "blocks.35.attn.q_norm.weight": "BF16",
601
+ "blocks.35.attn.qkv_proj.weight": "NVFP4",
602
+ "blocks.35.attn.qkv_proj.weight_scale": "FP32_SCALE",
603
+ "blocks.35.attn.qkv_proj.comfy_quant": "METADATA",
604
+ "blocks.35.mlp.fc1.weight": "NVFP4",
605
+ "blocks.35.mlp.fc1.weight_scale": "FP32_SCALE",
606
+ "blocks.35.mlp.fc1.comfy_quant": "METADATA",
607
+ "blocks.35.mlp.fc2.weight": "NVFP4",
608
+ "blocks.35.mlp.fc2.weight_scale": "FP32_SCALE",
609
+ "blocks.35.mlp.fc2.comfy_quant": "METADATA",
610
+ "blocks.35.norm1.weight": "BF16",
611
+ "blocks.35.norm2.weight": "BF16",
612
+ "blocks.36.adaln_proj.linear.bias": "BF16",
613
+ "blocks.36.adaln_proj.linear.weight": "NVFP4",
614
+ "blocks.36.adaln_proj.linear.weight_scale": "FP32_SCALE",
615
+ "blocks.36.adaln_proj.linear.comfy_quant": "METADATA",
616
+ "blocks.36.attn.k_norm.weight": "BF16",
617
+ "blocks.36.attn.out_proj.weight": "NVFP4",
618
+ "blocks.36.attn.out_proj.weight_scale": "FP32_SCALE",
619
+ "blocks.36.attn.out_proj.comfy_quant": "METADATA",
620
+ "blocks.36.attn.q_norm.weight": "BF16",
621
+ "blocks.36.attn.qkv_proj.weight": "NVFP4",
622
+ "blocks.36.attn.qkv_proj.weight_scale": "FP32_SCALE",
623
+ "blocks.36.attn.qkv_proj.comfy_quant": "METADATA",
624
+ "blocks.36.mlp.fc1.weight": "NVFP4",
625
+ "blocks.36.mlp.fc1.weight_scale": "FP32_SCALE",
626
+ "blocks.36.mlp.fc1.comfy_quant": "METADATA",
627
+ "blocks.36.mlp.fc2.weight": "NVFP4",
628
+ "blocks.36.mlp.fc2.weight_scale": "FP32_SCALE",
629
+ "blocks.36.mlp.fc2.comfy_quant": "METADATA",
630
+ "blocks.36.norm1.weight": "BF16",
631
+ "blocks.36.norm2.weight": "BF16",
632
+ "blocks.37.adaln_proj.linear.bias": "BF16",
633
+ "blocks.37.adaln_proj.linear.weight": "NVFP4",
634
+ "blocks.37.adaln_proj.linear.weight_scale": "FP32_SCALE",
635
+ "blocks.37.adaln_proj.linear.comfy_quant": "METADATA",
636
+ "blocks.37.attn.k_norm.weight": "BF16",
637
+ "blocks.37.attn.out_proj.weight": "NVFP4",
638
+ "blocks.37.attn.out_proj.weight_scale": "FP32_SCALE",
639
+ "blocks.37.attn.out_proj.comfy_quant": "METADATA",
640
+ "blocks.37.attn.q_norm.weight": "BF16",
641
+ "blocks.37.attn.qkv_proj.weight": "NVFP4",
642
+ "blocks.37.attn.qkv_proj.weight_scale": "FP32_SCALE",
643
+ "blocks.37.attn.qkv_proj.comfy_quant": "METADATA",
644
+ "blocks.37.mlp.fc1.weight": "NVFP4",
645
+ "blocks.37.mlp.fc1.weight_scale": "FP32_SCALE",
646
+ "blocks.37.mlp.fc1.comfy_quant": "METADATA",
647
+ "blocks.37.mlp.fc2.weight": "NVFP4",
648
+ "blocks.37.mlp.fc2.weight_scale": "FP32_SCALE",
649
+ "blocks.37.mlp.fc2.comfy_quant": "METADATA",
650
+ "blocks.37.norm1.weight": "BF16",
651
+ "blocks.37.norm2.weight": "BF16",
652
+ "blocks.38.adaln_proj.linear.bias": "BF16",
653
+ "blocks.38.adaln_proj.linear.weight": "NVFP4",
654
+ "blocks.38.adaln_proj.linear.weight_scale": "FP32_SCALE",
655
+ "blocks.38.adaln_proj.linear.comfy_quant": "METADATA",
656
+ "blocks.38.attn.k_norm.weight": "BF16",
657
+ "blocks.38.attn.out_proj.weight": "NVFP4",
658
+ "blocks.38.attn.out_proj.weight_scale": "FP32_SCALE",
659
+ "blocks.38.attn.out_proj.comfy_quant": "METADATA",
660
+ "blocks.38.attn.q_norm.weight": "BF16",
661
+ "blocks.38.attn.qkv_proj.weight": "NVFP4",
662
+ "blocks.38.attn.qkv_proj.weight_scale": "FP32_SCALE",
663
+ "blocks.38.attn.qkv_proj.comfy_quant": "METADATA",
664
+ "blocks.38.mlp.fc1.weight": "NVFP4",
665
+ "blocks.38.mlp.fc1.weight_scale": "FP32_SCALE",
666
+ "blocks.38.mlp.fc1.comfy_quant": "METADATA",
667
+ "blocks.38.mlp.fc2.weight": "NVFP4",
668
+ "blocks.38.mlp.fc2.weight_scale": "FP32_SCALE",
669
+ "blocks.38.mlp.fc2.comfy_quant": "METADATA",
670
+ "blocks.38.norm1.weight": "BF16",
671
+ "blocks.38.norm2.weight": "BF16",
672
+ "blocks.39.adaln_proj.linear.bias": "BF16",
673
+ "blocks.39.adaln_proj.linear.weight": "NVFP4",
674
+ "blocks.39.adaln_proj.linear.weight_scale": "FP32_SCALE",
675
+ "blocks.39.adaln_proj.linear.comfy_quant": "METADATA",
676
+ "blocks.39.attn.k_norm.weight": "BF16",
677
+ "blocks.39.attn.out_proj.weight": "NVFP4",
678
+ "blocks.39.attn.out_proj.weight_scale": "FP32_SCALE",
679
+ "blocks.39.attn.out_proj.comfy_quant": "METADATA",
680
+ "blocks.39.attn.q_norm.weight": "BF16",
681
+ "blocks.39.attn.qkv_proj.weight": "NVFP4",
682
+ "blocks.39.attn.qkv_proj.weight_scale": "FP32_SCALE",
683
+ "blocks.39.attn.qkv_proj.comfy_quant": "METADATA",
684
+ "blocks.39.mlp.fc1.weight": "NVFP4",
685
+ "blocks.39.mlp.fc1.weight_scale": "FP32_SCALE",
686
+ "blocks.39.mlp.fc1.comfy_quant": "METADATA",
687
+ "blocks.39.mlp.fc2.weight": "NVFP4",
688
+ "blocks.39.mlp.fc2.weight_scale": "FP32_SCALE",
689
+ "blocks.39.mlp.fc2.comfy_quant": "METADATA",
690
+ "blocks.39.norm1.weight": "BF16",
691
+ "blocks.39.norm2.weight": "BF16",
692
+ "blocks.4.adaln_proj.linear.bias": "BF16",
693
+ "blocks.4.adaln_proj.linear.weight": "NVFP4",
694
+ "blocks.4.adaln_proj.linear.weight_scale": "FP32_SCALE",
695
+ "blocks.4.adaln_proj.linear.comfy_quant": "METADATA",
696
+ "blocks.4.attn.k_norm.weight": "BF16",
697
+ "blocks.4.attn.out_proj.weight": "NVFP4",
698
+ "blocks.4.attn.out_proj.weight_scale": "FP32_SCALE",
699
+ "blocks.4.attn.out_proj.comfy_quant": "METADATA",
700
+ "blocks.4.attn.q_norm.weight": "BF16",
701
+ "blocks.4.attn.qkv_proj.weight": "NVFP4",
702
+ "blocks.4.attn.qkv_proj.weight_scale": "FP32_SCALE",
703
+ "blocks.4.attn.qkv_proj.comfy_quant": "METADATA",
704
+ "blocks.4.mlp.fc1.weight": "NVFP4",
705
+ "blocks.4.mlp.fc1.weight_scale": "FP32_SCALE",
706
+ "blocks.4.mlp.fc1.comfy_quant": "METADATA",
707
+ "blocks.4.mlp.fc2.weight": "NVFP4",
708
+ "blocks.4.mlp.fc2.weight_scale": "FP32_SCALE",
709
+ "blocks.4.mlp.fc2.comfy_quant": "METADATA",
710
+ "blocks.4.norm1.weight": "BF16",
711
+ "blocks.4.norm2.weight": "BF16",
712
+ "blocks.40.adaln_proj.linear.bias": "BF16",
713
+ "blocks.40.adaln_proj.linear.weight": "NVFP4",
714
+ "blocks.40.adaln_proj.linear.weight_scale": "FP32_SCALE",
715
+ "blocks.40.adaln_proj.linear.comfy_quant": "METADATA",
716
+ "blocks.40.attn.k_norm.weight": "BF16",
717
+ "blocks.40.attn.out_proj.weight": "NVFP4",
718
+ "blocks.40.attn.out_proj.weight_scale": "FP32_SCALE",
719
+ "blocks.40.attn.out_proj.comfy_quant": "METADATA",
720
+ "blocks.40.attn.q_norm.weight": "BF16",
721
+ "blocks.40.attn.qkv_proj.weight": "NVFP4",
722
+ "blocks.40.attn.qkv_proj.weight_scale": "FP32_SCALE",
723
+ "blocks.40.attn.qkv_proj.comfy_quant": "METADATA",
724
+ "blocks.40.mlp.fc1.weight": "NVFP4",
725
+ "blocks.40.mlp.fc1.weight_scale": "FP32_SCALE",
726
+ "blocks.40.mlp.fc1.comfy_quant": "METADATA",
727
+ "blocks.40.mlp.fc2.weight": "NVFP4",
728
+ "blocks.40.mlp.fc2.weight_scale": "FP32_SCALE",
729
+ "blocks.40.mlp.fc2.comfy_quant": "METADATA",
730
+ "blocks.40.norm1.weight": "BF16",
731
+ "blocks.40.norm2.weight": "BF16",
732
+ "blocks.41.adaln_proj.linear.bias": "BF16",
733
+ "blocks.41.adaln_proj.linear.weight": "NVFP4",
734
+ "blocks.41.adaln_proj.linear.weight_scale": "FP32_SCALE",
735
+ "blocks.41.adaln_proj.linear.comfy_quant": "METADATA",
736
+ "blocks.41.attn.k_norm.weight": "BF16",
737
+ "blocks.41.attn.out_proj.weight": "NVFP4",
738
+ "blocks.41.attn.out_proj.weight_scale": "FP32_SCALE",
739
+ "blocks.41.attn.out_proj.comfy_quant": "METADATA",
740
+ "blocks.41.attn.q_norm.weight": "BF16",
741
+ "blocks.41.attn.qkv_proj.weight": "NVFP4",
742
+ "blocks.41.attn.qkv_proj.weight_scale": "FP32_SCALE",
743
+ "blocks.41.attn.qkv_proj.comfy_quant": "METADATA",
744
+ "blocks.41.mlp.fc1.weight": "NVFP4",
745
+ "blocks.41.mlp.fc1.weight_scale": "FP32_SCALE",
746
+ "blocks.41.mlp.fc1.comfy_quant": "METADATA",
747
+ "blocks.41.mlp.fc2.weight": "NVFP4",
748
+ "blocks.41.mlp.fc2.weight_scale": "FP32_SCALE",
749
+ "blocks.41.mlp.fc2.comfy_quant": "METADATA",
750
+ "blocks.41.norm1.weight": "BF16",
751
+ "blocks.41.norm2.weight": "BF16",
752
+ "blocks.42.adaln_proj.linear.bias": "BF16",
753
+ "blocks.42.adaln_proj.linear.weight": "NVFP4",
754
+ "blocks.42.adaln_proj.linear.weight_scale": "FP32_SCALE",
755
+ "blocks.42.adaln_proj.linear.comfy_quant": "METADATA",
756
+ "blocks.42.attn.k_norm.weight": "BF16",
757
+ "blocks.42.attn.out_proj.weight": "NVFP4",
758
+ "blocks.42.attn.out_proj.weight_scale": "FP32_SCALE",
759
+ "blocks.42.attn.out_proj.comfy_quant": "METADATA",
760
+ "blocks.42.attn.q_norm.weight": "BF16",
761
+ "blocks.42.attn.qkv_proj.weight": "NVFP4",
762
+ "blocks.42.attn.qkv_proj.weight_scale": "FP32_SCALE",
763
+ "blocks.42.attn.qkv_proj.comfy_quant": "METADATA",
764
+ "blocks.42.mlp.fc1.weight": "NVFP4",
765
+ "blocks.42.mlp.fc1.weight_scale": "FP32_SCALE",
766
+ "blocks.42.mlp.fc1.comfy_quant": "METADATA",
767
+ "blocks.42.mlp.fc2.weight": "NVFP4",
768
+ "blocks.42.mlp.fc2.weight_scale": "FP32_SCALE",
769
+ "blocks.42.mlp.fc2.comfy_quant": "METADATA",
770
+ "blocks.42.norm1.weight": "BF16",
771
+ "blocks.42.norm2.weight": "BF16",
772
+ "blocks.43.adaln_proj.linear.bias": "BF16",
773
+ "blocks.43.adaln_proj.linear.weight": "NVFP4",
774
+ "blocks.43.adaln_proj.linear.weight_scale": "FP32_SCALE",
775
+ "blocks.43.adaln_proj.linear.comfy_quant": "METADATA",
776
+ "blocks.43.attn.k_norm.weight": "BF16",
777
+ "blocks.43.attn.out_proj.weight": "NVFP4",
778
+ "blocks.43.attn.out_proj.weight_scale": "FP32_SCALE",
779
+ "blocks.43.attn.out_proj.comfy_quant": "METADATA",
780
+ "blocks.43.attn.q_norm.weight": "BF16",
781
+ "blocks.43.attn.qkv_proj.weight": "NVFP4",
782
+ "blocks.43.attn.qkv_proj.weight_scale": "FP32_SCALE",
783
+ "blocks.43.attn.qkv_proj.comfy_quant": "METADATA",
784
+ "blocks.43.mlp.fc1.weight": "NVFP4",
785
+ "blocks.43.mlp.fc1.weight_scale": "FP32_SCALE",
786
+ "blocks.43.mlp.fc1.comfy_quant": "METADATA",
787
+ "blocks.43.mlp.fc2.weight": "NVFP4",
788
+ "blocks.43.mlp.fc2.weight_scale": "FP32_SCALE",
789
+ "blocks.43.mlp.fc2.comfy_quant": "METADATA",
790
+ "blocks.43.norm1.weight": "BF16",
791
+ "blocks.43.norm2.weight": "BF16",
792
+ "blocks.44.adaln_proj.linear.bias": "BF16",
793
+ "blocks.44.adaln_proj.linear.weight": "NVFP4",
794
+ "blocks.44.adaln_proj.linear.weight_scale": "FP32_SCALE",
795
+ "blocks.44.adaln_proj.linear.comfy_quant": "METADATA",
796
+ "blocks.44.attn.k_norm.weight": "BF16",
797
+ "blocks.44.attn.out_proj.weight": "NVFP4",
798
+ "blocks.44.attn.out_proj.weight_scale": "FP32_SCALE",
799
+ "blocks.44.attn.out_proj.comfy_quant": "METADATA",
800
+ "blocks.44.attn.q_norm.weight": "BF16",
801
+ "blocks.44.attn.qkv_proj.weight": "NVFP4",
802
+ "blocks.44.attn.qkv_proj.weight_scale": "FP32_SCALE",
803
+ "blocks.44.attn.qkv_proj.comfy_quant": "METADATA",
804
+ "blocks.44.mlp.fc1.weight": "NVFP4",
805
+ "blocks.44.mlp.fc1.weight_scale": "FP32_SCALE",
806
+ "blocks.44.mlp.fc1.comfy_quant": "METADATA",
807
+ "blocks.44.mlp.fc2.weight": "NVFP4",
808
+ "blocks.44.mlp.fc2.weight_scale": "FP32_SCALE",
809
+ "blocks.44.mlp.fc2.comfy_quant": "METADATA",
810
+ "blocks.44.norm1.weight": "BF16",
811
+ "blocks.44.norm2.weight": "BF16",
812
+ "blocks.45.adaln_proj.linear.bias": "BF16",
813
+ "blocks.45.adaln_proj.linear.weight": "NVFP4",
814
+ "blocks.45.adaln_proj.linear.weight_scale": "FP32_SCALE",
815
+ "blocks.45.adaln_proj.linear.comfy_quant": "METADATA",
816
+ "blocks.45.attn.k_norm.weight": "BF16",
817
+ "blocks.45.attn.out_proj.weight": "NVFP4",
818
+ "blocks.45.attn.out_proj.weight_scale": "FP32_SCALE",
819
+ "blocks.45.attn.out_proj.comfy_quant": "METADATA",
820
+ "blocks.45.attn.q_norm.weight": "BF16",
821
+ "blocks.45.attn.qkv_proj.weight": "NVFP4",
822
+ "blocks.45.attn.qkv_proj.weight_scale": "FP32_SCALE",
823
+ "blocks.45.attn.qkv_proj.comfy_quant": "METADATA",
824
+ "blocks.45.mlp.fc1.weight": "NVFP4",
825
+ "blocks.45.mlp.fc1.weight_scale": "FP32_SCALE",
826
+ "blocks.45.mlp.fc1.comfy_quant": "METADATA",
827
+ "blocks.45.mlp.fc2.weight": "NVFP4",
828
+ "blocks.45.mlp.fc2.weight_scale": "FP32_SCALE",
829
+ "blocks.45.mlp.fc2.comfy_quant": "METADATA",
830
+ "blocks.45.norm1.weight": "BF16",
831
+ "blocks.45.norm2.weight": "BF16",
832
+ "blocks.46.adaln_proj.linear.bias": "BF16",
833
+ "blocks.46.adaln_proj.linear.weight": "NVFP4",
834
+ "blocks.46.adaln_proj.linear.weight_scale": "FP32_SCALE",
835
+ "blocks.46.adaln_proj.linear.comfy_quant": "METADATA",
836
+ "blocks.46.attn.k_norm.weight": "BF16",
837
+ "blocks.46.attn.out_proj.weight": "NVFP4",
838
+ "blocks.46.attn.out_proj.weight_scale": "FP32_SCALE",
839
+ "blocks.46.attn.out_proj.comfy_quant": "METADATA",
840
+ "blocks.46.attn.q_norm.weight": "BF16",
841
+ "blocks.46.attn.qkv_proj.weight": "NVFP4",
842
+ "blocks.46.attn.qkv_proj.weight_scale": "FP32_SCALE",
843
+ "blocks.46.attn.qkv_proj.comfy_quant": "METADATA",
844
+ "blocks.46.mlp.fc1.weight": "NVFP4",
845
+ "blocks.46.mlp.fc1.weight_scale": "FP32_SCALE",
846
+ "blocks.46.mlp.fc1.comfy_quant": "METADATA",
847
+ "blocks.46.mlp.fc2.weight": "NVFP4",
848
+ "blocks.46.mlp.fc2.weight_scale": "FP32_SCALE",
849
+ "blocks.46.mlp.fc2.comfy_quant": "METADATA",
850
+ "blocks.46.norm1.weight": "BF16",
851
+ "blocks.46.norm2.weight": "BF16",
852
+ "blocks.47.adaln_proj.linear.bias": "BF16",
853
+ "blocks.47.adaln_proj.linear.weight": "NVFP4",
854
+ "blocks.47.adaln_proj.linear.weight_scale": "FP32_SCALE",
855
+ "blocks.47.adaln_proj.linear.comfy_quant": "METADATA",
856
+ "blocks.47.attn.k_norm.weight": "BF16",
857
+ "blocks.47.attn.out_proj.weight": "NVFP4",
858
+ "blocks.47.attn.out_proj.weight_scale": "FP32_SCALE",
859
+ "blocks.47.attn.out_proj.comfy_quant": "METADATA",
860
+ "blocks.47.attn.q_norm.weight": "BF16",
861
+ "blocks.47.attn.qkv_proj.weight": "NVFP4",
862
+ "blocks.47.attn.qkv_proj.weight_scale": "FP32_SCALE",
863
+ "blocks.47.attn.qkv_proj.comfy_quant": "METADATA",
864
+ "blocks.47.mlp.fc1.weight": "NVFP4",
865
+ "blocks.47.mlp.fc1.weight_scale": "FP32_SCALE",
866
+ "blocks.47.mlp.fc1.comfy_quant": "METADATA",
867
+ "blocks.47.mlp.fc2.weight": "NVFP4",
868
+ "blocks.47.mlp.fc2.weight_scale": "FP32_SCALE",
869
+ "blocks.47.mlp.fc2.comfy_quant": "METADATA",
870
+ "blocks.47.norm1.weight": "BF16",
871
+ "blocks.47.norm2.weight": "BF16",
872
+ "blocks.48.adaln_proj.linear.bias": "BF16",
873
+ "blocks.48.adaln_proj.linear.weight": "NVFP4",
874
+ "blocks.48.adaln_proj.linear.weight_scale": "FP32_SCALE",
875
+ "blocks.48.adaln_proj.linear.comfy_quant": "METADATA",
876
+ "blocks.48.attn.k_norm.weight": "BF16",
877
+ "blocks.48.attn.out_proj.weight": "NVFP4",
878
+ "blocks.48.attn.out_proj.weight_scale": "FP32_SCALE",
879
+ "blocks.48.attn.out_proj.comfy_quant": "METADATA",
880
+ "blocks.48.attn.q_norm.weight": "BF16",
881
+ "blocks.48.attn.qkv_proj.weight": "NVFP4",
882
+ "blocks.48.attn.qkv_proj.weight_scale": "FP32_SCALE",
883
+ "blocks.48.attn.qkv_proj.comfy_quant": "METADATA",
884
+ "blocks.48.mlp.fc1.weight": "NVFP4",
885
+ "blocks.48.mlp.fc1.weight_scale": "FP32_SCALE",
886
+ "blocks.48.mlp.fc1.comfy_quant": "METADATA",
887
+ "blocks.48.mlp.fc2.weight": "NVFP4",
888
+ "blocks.48.mlp.fc2.weight_scale": "FP32_SCALE",
889
+ "blocks.48.mlp.fc2.comfy_quant": "METADATA",
890
+ "blocks.48.norm1.weight": "BF16",
891
+ "blocks.48.norm2.weight": "BF16",
892
+ "blocks.49.adaln_proj.linear.bias": "BF16",
893
+ "blocks.49.adaln_proj.linear.weight": "NVFP4",
894
+ "blocks.49.adaln_proj.linear.weight_scale": "FP32_SCALE",
895
+ "blocks.49.adaln_proj.linear.comfy_quant": "METADATA",
896
+ "blocks.49.attn.k_norm.weight": "BF16",
897
+ "blocks.49.attn.out_proj.weight": "NVFP4",
898
+ "blocks.49.attn.out_proj.weight_scale": "FP32_SCALE",
899
+ "blocks.49.attn.out_proj.comfy_quant": "METADATA",
900
+ "blocks.49.attn.q_norm.weight": "BF16",
901
+ "blocks.49.attn.qkv_proj.weight": "NVFP4",
902
+ "blocks.49.attn.qkv_proj.weight_scale": "FP32_SCALE",
903
+ "blocks.49.attn.qkv_proj.comfy_quant": "METADATA",
904
+ "blocks.49.mlp.fc1.weight": "NVFP4",
905
+ "blocks.49.mlp.fc1.weight_scale": "FP32_SCALE",
906
+ "blocks.49.mlp.fc1.comfy_quant": "METADATA",
907
+ "blocks.49.mlp.fc2.weight": "NVFP4",
908
+ "blocks.49.mlp.fc2.weight_scale": "FP32_SCALE",
909
+ "blocks.49.mlp.fc2.comfy_quant": "METADATA",
910
+ "blocks.49.norm1.weight": "BF16",
911
+ "blocks.49.norm2.weight": "BF16",
912
+ "blocks.5.adaln_proj.linear.bias": "BF16",
913
+ "blocks.5.adaln_proj.linear.weight": "NVFP4",
914
+ "blocks.5.adaln_proj.linear.weight_scale": "FP32_SCALE",
915
+ "blocks.5.adaln_proj.linear.comfy_quant": "METADATA",
916
+ "blocks.5.attn.k_norm.weight": "BF16",
917
+ "blocks.5.attn.out_proj.weight": "NVFP4",
918
+ "blocks.5.attn.out_proj.weight_scale": "FP32_SCALE",
919
+ "blocks.5.attn.out_proj.comfy_quant": "METADATA",
920
+ "blocks.5.attn.q_norm.weight": "BF16",
921
+ "blocks.5.attn.qkv_proj.weight": "NVFP4",
922
+ "blocks.5.attn.qkv_proj.weight_scale": "FP32_SCALE",
923
+ "blocks.5.attn.qkv_proj.comfy_quant": "METADATA",
924
+ "blocks.5.mlp.fc1.weight": "NVFP4",
925
+ "blocks.5.mlp.fc1.weight_scale": "FP32_SCALE",
926
+ "blocks.5.mlp.fc1.comfy_quant": "METADATA",
927
+ "blocks.5.mlp.fc2.weight": "NVFP4",
928
+ "blocks.5.mlp.fc2.weight_scale": "FP32_SCALE",
929
+ "blocks.5.mlp.fc2.comfy_quant": "METADATA",
930
+ "blocks.5.norm1.weight": "BF16",
931
+ "blocks.5.norm2.weight": "BF16",
932
+ "blocks.6.adaln_proj.linear.bias": "BF16",
933
+ "blocks.6.adaln_proj.linear.weight": "NVFP4",
934
+ "blocks.6.adaln_proj.linear.weight_scale": "FP32_SCALE",
935
+ "blocks.6.adaln_proj.linear.comfy_quant": "METADATA",
936
+ "blocks.6.attn.k_norm.weight": "BF16",
937
+ "blocks.6.attn.out_proj.weight": "NVFP4",
938
+ "blocks.6.attn.out_proj.weight_scale": "FP32_SCALE",
939
+ "blocks.6.attn.out_proj.comfy_quant": "METADATA",
940
+ "blocks.6.attn.q_norm.weight": "BF16",
941
+ "blocks.6.attn.qkv_proj.weight": "NVFP4",
942
+ "blocks.6.attn.qkv_proj.weight_scale": "FP32_SCALE",
943
+ "blocks.6.attn.qkv_proj.comfy_quant": "METADATA",
944
+ "blocks.6.mlp.fc1.weight": "NVFP4",
945
+ "blocks.6.mlp.fc1.weight_scale": "FP32_SCALE",
946
+ "blocks.6.mlp.fc1.comfy_quant": "METADATA",
947
+ "blocks.6.mlp.fc2.weight": "NVFP4",
948
+ "blocks.6.mlp.fc2.weight_scale": "FP32_SCALE",
949
+ "blocks.6.mlp.fc2.comfy_quant": "METADATA",
950
+ "blocks.6.norm1.weight": "BF16",
951
+ "blocks.6.norm2.weight": "BF16",
952
+ "blocks.7.adaln_proj.linear.bias": "BF16",
953
+ "blocks.7.adaln_proj.linear.weight": "NVFP4",
954
+ "blocks.7.adaln_proj.linear.weight_scale": "FP32_SCALE",
955
+ "blocks.7.adaln_proj.linear.comfy_quant": "METADATA",
956
+ "blocks.7.attn.k_norm.weight": "BF16",
957
+ "blocks.7.attn.out_proj.weight": "NVFP4",
958
+ "blocks.7.attn.out_proj.weight_scale": "FP32_SCALE",
959
+ "blocks.7.attn.out_proj.comfy_quant": "METADATA",
960
+ "blocks.7.attn.q_norm.weight": "BF16",
961
+ "blocks.7.attn.qkv_proj.weight": "NVFP4",
962
+ "blocks.7.attn.qkv_proj.weight_scale": "FP32_SCALE",
963
+ "blocks.7.attn.qkv_proj.comfy_quant": "METADATA",
964
+ "blocks.7.mlp.fc1.weight": "NVFP4",
965
+ "blocks.7.mlp.fc1.weight_scale": "FP32_SCALE",
966
+ "blocks.7.mlp.fc1.comfy_quant": "METADATA",
967
+ "blocks.7.mlp.fc2.weight": "NVFP4",
968
+ "blocks.7.mlp.fc2.weight_scale": "FP32_SCALE",
969
+ "blocks.7.mlp.fc2.comfy_quant": "METADATA",
970
+ "blocks.7.norm1.weight": "BF16",
971
+ "blocks.7.norm2.weight": "BF16",
972
+ "blocks.8.adaln_proj.linear.bias": "BF16",
973
+ "blocks.8.adaln_proj.linear.weight": "NVFP4",
974
+ "blocks.8.adaln_proj.linear.weight_scale": "FP32_SCALE",
975
+ "blocks.8.adaln_proj.linear.comfy_quant": "METADATA",
976
+ "blocks.8.attn.k_norm.weight": "BF16",
977
+ "blocks.8.attn.out_proj.weight": "NVFP4",
978
+ "blocks.8.attn.out_proj.weight_scale": "FP32_SCALE",
979
+ "blocks.8.attn.out_proj.comfy_quant": "METADATA",
980
+ "blocks.8.attn.q_norm.weight": "BF16",
981
+ "blocks.8.attn.qkv_proj.weight": "NVFP4",
982
+ "blocks.8.attn.qkv_proj.weight_scale": "FP32_SCALE",
983
+ "blocks.8.attn.qkv_proj.comfy_quant": "METADATA",
984
+ "blocks.8.mlp.fc1.weight": "NVFP4",
985
+ "blocks.8.mlp.fc1.weight_scale": "FP32_SCALE",
986
+ "blocks.8.mlp.fc1.comfy_quant": "METADATA",
987
+ "blocks.8.mlp.fc2.weight": "NVFP4",
988
+ "blocks.8.mlp.fc2.weight_scale": "FP32_SCALE",
989
+ "blocks.8.mlp.fc2.comfy_quant": "METADATA",
990
+ "blocks.8.norm1.weight": "BF16",
991
+ "blocks.8.norm2.weight": "BF16",
992
+ "blocks.9.adaln_proj.linear.bias": "BF16",
993
+ "blocks.9.adaln_proj.linear.weight": "NVFP4",
994
+ "blocks.9.adaln_proj.linear.weight_scale": "FP32_SCALE",
995
+ "blocks.9.adaln_proj.linear.comfy_quant": "METADATA",
996
+ "blocks.9.attn.k_norm.weight": "BF16",
997
+ "blocks.9.attn.out_proj.weight": "NVFP4",
998
+ "blocks.9.attn.out_proj.weight_scale": "FP32_SCALE",
999
+ "blocks.9.attn.out_proj.comfy_quant": "METADATA",
1000
+ "blocks.9.attn.q_norm.weight": "BF16",
1001
+ "blocks.9.attn.qkv_proj.weight": "NVFP4",
1002
+ "blocks.9.attn.qkv_proj.weight_scale": "FP32_SCALE",
1003
+ "blocks.9.attn.qkv_proj.comfy_quant": "METADATA",
1004
+ "blocks.9.mlp.fc1.weight": "NVFP4",
1005
+ "blocks.9.mlp.fc1.weight_scale": "FP32_SCALE",
1006
+ "blocks.9.mlp.fc1.comfy_quant": "METADATA",
1007
+ "blocks.9.mlp.fc2.weight": "NVFP4",
1008
+ "blocks.9.mlp.fc2.weight_scale": "FP32_SCALE",
1009
+ "blocks.9.mlp.fc2.comfy_quant": "METADATA",
1010
+ "blocks.9.norm1.weight": "BF16",
1011
+ "blocks.9.norm2.weight": "BF16",
1012
+ "condition_proj.bias": "BF16",
1013
+ "condition_proj.weight": "BF16",
1014
+ "final_layer.adaln_proj.linear.bias": "BF16",
1015
+ "final_layer.adaln_proj.linear.weight": "BF16",
1016
+ "final_layer.audio_out.bias": "BF16",
1017
+ "final_layer.audio_out.weight": "BF16",
1018
+ "final_layer.norm.weight": "BF16",
1019
+ "final_layer.video_out.bias": "BF16",
1020
+ "final_layer.video_out.weight": "BF16",
1021
+ "rope.inv_freq": "BF16",
1022
+ "time_embedder.proj_in.bias": "BF16",
1023
+ "time_embedder.proj_in.weight": "BF16",
1024
+ "time_embedder.proj_out.bias": "BF16",
1025
+ "time_embedder.proj_out.weight": "BF16",
1026
+ "token_refiner.blocks.0.attn.k_norm.weight": "BF16",
1027
+ "token_refiner.blocks.0.attn.out_proj.weight": "NVFP4",
1028
+ "token_refiner.blocks.0.attn.out_proj.weight_scale": "FP32_SCALE",
1029
+ "token_refiner.blocks.0.attn.out_proj.comfy_quant": "METADATA",
1030
+ "token_refiner.blocks.0.attn.q_norm.weight": "BF16",
1031
+ "token_refiner.blocks.0.attn.qkv_proj.weight": "NVFP4",
1032
+ "token_refiner.blocks.0.attn.qkv_proj.weight_scale": "FP32_SCALE",
1033
+ "token_refiner.blocks.0.attn.qkv_proj.comfy_quant": "METADATA",
1034
+ "token_refiner.blocks.0.mlp.fc1.weight": "NVFP4",
1035
+ "token_refiner.blocks.0.mlp.fc1.weight_scale": "FP32_SCALE",
1036
+ "token_refiner.blocks.0.mlp.fc1.comfy_quant": "METADATA",
1037
+ "token_refiner.blocks.0.mlp.fc2.weight": "NVFP4",
1038
+ "token_refiner.blocks.0.mlp.fc2.weight_scale": "FP32_SCALE",
1039
+ "token_refiner.blocks.0.mlp.fc2.comfy_quant": "METADATA",
1040
+ "token_refiner.blocks.0.norm1.weight": "BF16",
1041
+ "token_refiner.blocks.0.norm2.weight": "BF16",
1042
+ "token_refiner.blocks.1.attn.k_norm.weight": "BF16",
1043
+ "token_refiner.blocks.1.attn.out_proj.weight": "NVFP4",
1044
+ "token_refiner.blocks.1.attn.out_proj.weight_scale": "FP32_SCALE",
1045
+ "token_refiner.blocks.1.attn.out_proj.comfy_quant": "METADATA",
1046
+ "token_refiner.blocks.1.attn.q_norm.weight": "BF16",
1047
+ "token_refiner.blocks.1.attn.qkv_proj.weight": "NVFP4",
1048
+ "token_refiner.blocks.1.attn.qkv_proj.weight_scale": "FP32_SCALE",
1049
+ "token_refiner.blocks.1.attn.qkv_proj.comfy_quant": "METADATA",
1050
+ "token_refiner.blocks.1.mlp.fc1.weight": "NVFP4",
1051
+ "token_refiner.blocks.1.mlp.fc1.weight_scale": "FP32_SCALE",
1052
+ "token_refiner.blocks.1.mlp.fc1.comfy_quant": "METADATA",
1053
+ "token_refiner.blocks.1.mlp.fc2.weight": "NVFP4",
1054
+ "token_refiner.blocks.1.mlp.fc2.weight_scale": "FP32_SCALE",
1055
+ "token_refiner.blocks.1.mlp.fc2.comfy_quant": "METADATA",
1056
+ "token_refiner.blocks.1.norm1.weight": "BF16",
1057
+ "token_refiner.blocks.1.norm2.weight": "BF16",
1058
+ "token_refiner.final_norm.weight": "BF16",
1059
+ "video_patch_proj.bias": "BF16",
1060
+ "video_patch_proj.weight": "BF16"
1061
+ }
1062
+ }
profiles/minimax_h3_nvfp4_mixed.json ADDED
@@ -0,0 +1,1062 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "__metadata__": {
3
+ "original_model_name": "minimax_h3_nvfp4_mixed",
4
+ "original_model_path": "/workspace/ComfyUI/models/diffusion_models/minimax_h3_ref2va_bf16.safetensors",
5
+ "timestamp": "2026-08-03T06:06:50.554208",
6
+ "total_layers": 1051,
7
+ "created_by": "hand-authored for MiniMax-H3 (33.12B: adaln 39.4%, mlp 36.3%, attn 24.2%)"
8
+ },
9
+ "layers": {
10
+ "audio_patch_proj.bias": "BF16",
11
+ "audio_patch_proj.weight": "BF16",
12
+ "blocks.0.adaln_proj.linear.bias": "BF16",
13
+ "blocks.0.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
14
+ "blocks.0.adaln_proj.linear.weight_scale": "FP32_SCALE",
15
+ "blocks.0.adaln_proj.linear.comfy_quant": "METADATA",
16
+ "blocks.0.attn.k_norm.weight": "BF16",
17
+ "blocks.0.attn.out_proj.weight": "NVFP4",
18
+ "blocks.0.attn.out_proj.weight_scale": "FP32_SCALE",
19
+ "blocks.0.attn.out_proj.comfy_quant": "METADATA",
20
+ "blocks.0.attn.q_norm.weight": "BF16",
21
+ "blocks.0.attn.qkv_proj.weight": "NVFP4",
22
+ "blocks.0.attn.qkv_proj.weight_scale": "FP32_SCALE",
23
+ "blocks.0.attn.qkv_proj.comfy_quant": "METADATA",
24
+ "blocks.0.mlp.fc1.weight": "NVFP4",
25
+ "blocks.0.mlp.fc1.weight_scale": "FP32_SCALE",
26
+ "blocks.0.mlp.fc1.comfy_quant": "METADATA",
27
+ "blocks.0.mlp.fc2.weight": "NVFP4",
28
+ "blocks.0.mlp.fc2.weight_scale": "FP32_SCALE",
29
+ "blocks.0.mlp.fc2.comfy_quant": "METADATA",
30
+ "blocks.0.norm1.weight": "BF16",
31
+ "blocks.0.norm2.weight": "BF16",
32
+ "blocks.1.adaln_proj.linear.bias": "BF16",
33
+ "blocks.1.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
34
+ "blocks.1.adaln_proj.linear.weight_scale": "FP32_SCALE",
35
+ "blocks.1.adaln_proj.linear.comfy_quant": "METADATA",
36
+ "blocks.1.attn.k_norm.weight": "BF16",
37
+ "blocks.1.attn.out_proj.weight": "NVFP4",
38
+ "blocks.1.attn.out_proj.weight_scale": "FP32_SCALE",
39
+ "blocks.1.attn.out_proj.comfy_quant": "METADATA",
40
+ "blocks.1.attn.q_norm.weight": "BF16",
41
+ "blocks.1.attn.qkv_proj.weight": "NVFP4",
42
+ "blocks.1.attn.qkv_proj.weight_scale": "FP32_SCALE",
43
+ "blocks.1.attn.qkv_proj.comfy_quant": "METADATA",
44
+ "blocks.1.mlp.fc1.weight": "NVFP4",
45
+ "blocks.1.mlp.fc1.weight_scale": "FP32_SCALE",
46
+ "blocks.1.mlp.fc1.comfy_quant": "METADATA",
47
+ "blocks.1.mlp.fc2.weight": "NVFP4",
48
+ "blocks.1.mlp.fc2.weight_scale": "FP32_SCALE",
49
+ "blocks.1.mlp.fc2.comfy_quant": "METADATA",
50
+ "blocks.1.norm1.weight": "BF16",
51
+ "blocks.1.norm2.weight": "BF16",
52
+ "blocks.10.adaln_proj.linear.bias": "BF16",
53
+ "blocks.10.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
54
+ "blocks.10.adaln_proj.linear.weight_scale": "FP32_SCALE",
55
+ "blocks.10.adaln_proj.linear.comfy_quant": "METADATA",
56
+ "blocks.10.attn.k_norm.weight": "BF16",
57
+ "blocks.10.attn.out_proj.weight": "NVFP4",
58
+ "blocks.10.attn.out_proj.weight_scale": "FP32_SCALE",
59
+ "blocks.10.attn.out_proj.comfy_quant": "METADATA",
60
+ "blocks.10.attn.q_norm.weight": "BF16",
61
+ "blocks.10.attn.qkv_proj.weight": "NVFP4",
62
+ "blocks.10.attn.qkv_proj.weight_scale": "FP32_SCALE",
63
+ "blocks.10.attn.qkv_proj.comfy_quant": "METADATA",
64
+ "blocks.10.mlp.fc1.weight": "NVFP4",
65
+ "blocks.10.mlp.fc1.weight_scale": "FP32_SCALE",
66
+ "blocks.10.mlp.fc1.comfy_quant": "METADATA",
67
+ "blocks.10.mlp.fc2.weight": "NVFP4",
68
+ "blocks.10.mlp.fc2.weight_scale": "FP32_SCALE",
69
+ "blocks.10.mlp.fc2.comfy_quant": "METADATA",
70
+ "blocks.10.norm1.weight": "BF16",
71
+ "blocks.10.norm2.weight": "BF16",
72
+ "blocks.11.adaln_proj.linear.bias": "BF16",
73
+ "blocks.11.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
74
+ "blocks.11.adaln_proj.linear.weight_scale": "FP32_SCALE",
75
+ "blocks.11.adaln_proj.linear.comfy_quant": "METADATA",
76
+ "blocks.11.attn.k_norm.weight": "BF16",
77
+ "blocks.11.attn.out_proj.weight": "NVFP4",
78
+ "blocks.11.attn.out_proj.weight_scale": "FP32_SCALE",
79
+ "blocks.11.attn.out_proj.comfy_quant": "METADATA",
80
+ "blocks.11.attn.q_norm.weight": "BF16",
81
+ "blocks.11.attn.qkv_proj.weight": "NVFP4",
82
+ "blocks.11.attn.qkv_proj.weight_scale": "FP32_SCALE",
83
+ "blocks.11.attn.qkv_proj.comfy_quant": "METADATA",
84
+ "blocks.11.mlp.fc1.weight": "NVFP4",
85
+ "blocks.11.mlp.fc1.weight_scale": "FP32_SCALE",
86
+ "blocks.11.mlp.fc1.comfy_quant": "METADATA",
87
+ "blocks.11.mlp.fc2.weight": "NVFP4",
88
+ "blocks.11.mlp.fc2.weight_scale": "FP32_SCALE",
89
+ "blocks.11.mlp.fc2.comfy_quant": "METADATA",
90
+ "blocks.11.norm1.weight": "BF16",
91
+ "blocks.11.norm2.weight": "BF16",
92
+ "blocks.12.adaln_proj.linear.bias": "BF16",
93
+ "blocks.12.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
94
+ "blocks.12.adaln_proj.linear.weight_scale": "FP32_SCALE",
95
+ "blocks.12.adaln_proj.linear.comfy_quant": "METADATA",
96
+ "blocks.12.attn.k_norm.weight": "BF16",
97
+ "blocks.12.attn.out_proj.weight": "NVFP4",
98
+ "blocks.12.attn.out_proj.weight_scale": "FP32_SCALE",
99
+ "blocks.12.attn.out_proj.comfy_quant": "METADATA",
100
+ "blocks.12.attn.q_norm.weight": "BF16",
101
+ "blocks.12.attn.qkv_proj.weight": "NVFP4",
102
+ "blocks.12.attn.qkv_proj.weight_scale": "FP32_SCALE",
103
+ "blocks.12.attn.qkv_proj.comfy_quant": "METADATA",
104
+ "blocks.12.mlp.fc1.weight": "NVFP4",
105
+ "blocks.12.mlp.fc1.weight_scale": "FP32_SCALE",
106
+ "blocks.12.mlp.fc1.comfy_quant": "METADATA",
107
+ "blocks.12.mlp.fc2.weight": "NVFP4",
108
+ "blocks.12.mlp.fc2.weight_scale": "FP32_SCALE",
109
+ "blocks.12.mlp.fc2.comfy_quant": "METADATA",
110
+ "blocks.12.norm1.weight": "BF16",
111
+ "blocks.12.norm2.weight": "BF16",
112
+ "blocks.13.adaln_proj.linear.bias": "BF16",
113
+ "blocks.13.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
114
+ "blocks.13.adaln_proj.linear.weight_scale": "FP32_SCALE",
115
+ "blocks.13.adaln_proj.linear.comfy_quant": "METADATA",
116
+ "blocks.13.attn.k_norm.weight": "BF16",
117
+ "blocks.13.attn.out_proj.weight": "NVFP4",
118
+ "blocks.13.attn.out_proj.weight_scale": "FP32_SCALE",
119
+ "blocks.13.attn.out_proj.comfy_quant": "METADATA",
120
+ "blocks.13.attn.q_norm.weight": "BF16",
121
+ "blocks.13.attn.qkv_proj.weight": "NVFP4",
122
+ "blocks.13.attn.qkv_proj.weight_scale": "FP32_SCALE",
123
+ "blocks.13.attn.qkv_proj.comfy_quant": "METADATA",
124
+ "blocks.13.mlp.fc1.weight": "NVFP4",
125
+ "blocks.13.mlp.fc1.weight_scale": "FP32_SCALE",
126
+ "blocks.13.mlp.fc1.comfy_quant": "METADATA",
127
+ "blocks.13.mlp.fc2.weight": "NVFP4",
128
+ "blocks.13.mlp.fc2.weight_scale": "FP32_SCALE",
129
+ "blocks.13.mlp.fc2.comfy_quant": "METADATA",
130
+ "blocks.13.norm1.weight": "BF16",
131
+ "blocks.13.norm2.weight": "BF16",
132
+ "blocks.14.adaln_proj.linear.bias": "BF16",
133
+ "blocks.14.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
134
+ "blocks.14.adaln_proj.linear.weight_scale": "FP32_SCALE",
135
+ "blocks.14.adaln_proj.linear.comfy_quant": "METADATA",
136
+ "blocks.14.attn.k_norm.weight": "BF16",
137
+ "blocks.14.attn.out_proj.weight": "NVFP4",
138
+ "blocks.14.attn.out_proj.weight_scale": "FP32_SCALE",
139
+ "blocks.14.attn.out_proj.comfy_quant": "METADATA",
140
+ "blocks.14.attn.q_norm.weight": "BF16",
141
+ "blocks.14.attn.qkv_proj.weight": "NVFP4",
142
+ "blocks.14.attn.qkv_proj.weight_scale": "FP32_SCALE",
143
+ "blocks.14.attn.qkv_proj.comfy_quant": "METADATA",
144
+ "blocks.14.mlp.fc1.weight": "NVFP4",
145
+ "blocks.14.mlp.fc1.weight_scale": "FP32_SCALE",
146
+ "blocks.14.mlp.fc1.comfy_quant": "METADATA",
147
+ "blocks.14.mlp.fc2.weight": "NVFP4",
148
+ "blocks.14.mlp.fc2.weight_scale": "FP32_SCALE",
149
+ "blocks.14.mlp.fc2.comfy_quant": "METADATA",
150
+ "blocks.14.norm1.weight": "BF16",
151
+ "blocks.14.norm2.weight": "BF16",
152
+ "blocks.15.adaln_proj.linear.bias": "BF16",
153
+ "blocks.15.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
154
+ "blocks.15.adaln_proj.linear.weight_scale": "FP32_SCALE",
155
+ "blocks.15.adaln_proj.linear.comfy_quant": "METADATA",
156
+ "blocks.15.attn.k_norm.weight": "BF16",
157
+ "blocks.15.attn.out_proj.weight": "NVFP4",
158
+ "blocks.15.attn.out_proj.weight_scale": "FP32_SCALE",
159
+ "blocks.15.attn.out_proj.comfy_quant": "METADATA",
160
+ "blocks.15.attn.q_norm.weight": "BF16",
161
+ "blocks.15.attn.qkv_proj.weight": "NVFP4",
162
+ "blocks.15.attn.qkv_proj.weight_scale": "FP32_SCALE",
163
+ "blocks.15.attn.qkv_proj.comfy_quant": "METADATA",
164
+ "blocks.15.mlp.fc1.weight": "NVFP4",
165
+ "blocks.15.mlp.fc1.weight_scale": "FP32_SCALE",
166
+ "blocks.15.mlp.fc1.comfy_quant": "METADATA",
167
+ "blocks.15.mlp.fc2.weight": "NVFP4",
168
+ "blocks.15.mlp.fc2.weight_scale": "FP32_SCALE",
169
+ "blocks.15.mlp.fc2.comfy_quant": "METADATA",
170
+ "blocks.15.norm1.weight": "BF16",
171
+ "blocks.15.norm2.weight": "BF16",
172
+ "blocks.16.adaln_proj.linear.bias": "BF16",
173
+ "blocks.16.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
174
+ "blocks.16.adaln_proj.linear.weight_scale": "FP32_SCALE",
175
+ "blocks.16.adaln_proj.linear.comfy_quant": "METADATA",
176
+ "blocks.16.attn.k_norm.weight": "BF16",
177
+ "blocks.16.attn.out_proj.weight": "NVFP4",
178
+ "blocks.16.attn.out_proj.weight_scale": "FP32_SCALE",
179
+ "blocks.16.attn.out_proj.comfy_quant": "METADATA",
180
+ "blocks.16.attn.q_norm.weight": "BF16",
181
+ "blocks.16.attn.qkv_proj.weight": "NVFP4",
182
+ "blocks.16.attn.qkv_proj.weight_scale": "FP32_SCALE",
183
+ "blocks.16.attn.qkv_proj.comfy_quant": "METADATA",
184
+ "blocks.16.mlp.fc1.weight": "NVFP4",
185
+ "blocks.16.mlp.fc1.weight_scale": "FP32_SCALE",
186
+ "blocks.16.mlp.fc1.comfy_quant": "METADATA",
187
+ "blocks.16.mlp.fc2.weight": "NVFP4",
188
+ "blocks.16.mlp.fc2.weight_scale": "FP32_SCALE",
189
+ "blocks.16.mlp.fc2.comfy_quant": "METADATA",
190
+ "blocks.16.norm1.weight": "BF16",
191
+ "blocks.16.norm2.weight": "BF16",
192
+ "blocks.17.adaln_proj.linear.bias": "BF16",
193
+ "blocks.17.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
194
+ "blocks.17.adaln_proj.linear.weight_scale": "FP32_SCALE",
195
+ "blocks.17.adaln_proj.linear.comfy_quant": "METADATA",
196
+ "blocks.17.attn.k_norm.weight": "BF16",
197
+ "blocks.17.attn.out_proj.weight": "NVFP4",
198
+ "blocks.17.attn.out_proj.weight_scale": "FP32_SCALE",
199
+ "blocks.17.attn.out_proj.comfy_quant": "METADATA",
200
+ "blocks.17.attn.q_norm.weight": "BF16",
201
+ "blocks.17.attn.qkv_proj.weight": "NVFP4",
202
+ "blocks.17.attn.qkv_proj.weight_scale": "FP32_SCALE",
203
+ "blocks.17.attn.qkv_proj.comfy_quant": "METADATA",
204
+ "blocks.17.mlp.fc1.weight": "NVFP4",
205
+ "blocks.17.mlp.fc1.weight_scale": "FP32_SCALE",
206
+ "blocks.17.mlp.fc1.comfy_quant": "METADATA",
207
+ "blocks.17.mlp.fc2.weight": "NVFP4",
208
+ "blocks.17.mlp.fc2.weight_scale": "FP32_SCALE",
209
+ "blocks.17.mlp.fc2.comfy_quant": "METADATA",
210
+ "blocks.17.norm1.weight": "BF16",
211
+ "blocks.17.norm2.weight": "BF16",
212
+ "blocks.18.adaln_proj.linear.bias": "BF16",
213
+ "blocks.18.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
214
+ "blocks.18.adaln_proj.linear.weight_scale": "FP32_SCALE",
215
+ "blocks.18.adaln_proj.linear.comfy_quant": "METADATA",
216
+ "blocks.18.attn.k_norm.weight": "BF16",
217
+ "blocks.18.attn.out_proj.weight": "NVFP4",
218
+ "blocks.18.attn.out_proj.weight_scale": "FP32_SCALE",
219
+ "blocks.18.attn.out_proj.comfy_quant": "METADATA",
220
+ "blocks.18.attn.q_norm.weight": "BF16",
221
+ "blocks.18.attn.qkv_proj.weight": "NVFP4",
222
+ "blocks.18.attn.qkv_proj.weight_scale": "FP32_SCALE",
223
+ "blocks.18.attn.qkv_proj.comfy_quant": "METADATA",
224
+ "blocks.18.mlp.fc1.weight": "NVFP4",
225
+ "blocks.18.mlp.fc1.weight_scale": "FP32_SCALE",
226
+ "blocks.18.mlp.fc1.comfy_quant": "METADATA",
227
+ "blocks.18.mlp.fc2.weight": "NVFP4",
228
+ "blocks.18.mlp.fc2.weight_scale": "FP32_SCALE",
229
+ "blocks.18.mlp.fc2.comfy_quant": "METADATA",
230
+ "blocks.18.norm1.weight": "BF16",
231
+ "blocks.18.norm2.weight": "BF16",
232
+ "blocks.19.adaln_proj.linear.bias": "BF16",
233
+ "blocks.19.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
234
+ "blocks.19.adaln_proj.linear.weight_scale": "FP32_SCALE",
235
+ "blocks.19.adaln_proj.linear.comfy_quant": "METADATA",
236
+ "blocks.19.attn.k_norm.weight": "BF16",
237
+ "blocks.19.attn.out_proj.weight": "NVFP4",
238
+ "blocks.19.attn.out_proj.weight_scale": "FP32_SCALE",
239
+ "blocks.19.attn.out_proj.comfy_quant": "METADATA",
240
+ "blocks.19.attn.q_norm.weight": "BF16",
241
+ "blocks.19.attn.qkv_proj.weight": "NVFP4",
242
+ "blocks.19.attn.qkv_proj.weight_scale": "FP32_SCALE",
243
+ "blocks.19.attn.qkv_proj.comfy_quant": "METADATA",
244
+ "blocks.19.mlp.fc1.weight": "NVFP4",
245
+ "blocks.19.mlp.fc1.weight_scale": "FP32_SCALE",
246
+ "blocks.19.mlp.fc1.comfy_quant": "METADATA",
247
+ "blocks.19.mlp.fc2.weight": "NVFP4",
248
+ "blocks.19.mlp.fc2.weight_scale": "FP32_SCALE",
249
+ "blocks.19.mlp.fc2.comfy_quant": "METADATA",
250
+ "blocks.19.norm1.weight": "BF16",
251
+ "blocks.19.norm2.weight": "BF16",
252
+ "blocks.2.adaln_proj.linear.bias": "BF16",
253
+ "blocks.2.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
254
+ "blocks.2.adaln_proj.linear.weight_scale": "FP32_SCALE",
255
+ "blocks.2.adaln_proj.linear.comfy_quant": "METADATA",
256
+ "blocks.2.attn.k_norm.weight": "BF16",
257
+ "blocks.2.attn.out_proj.weight": "NVFP4",
258
+ "blocks.2.attn.out_proj.weight_scale": "FP32_SCALE",
259
+ "blocks.2.attn.out_proj.comfy_quant": "METADATA",
260
+ "blocks.2.attn.q_norm.weight": "BF16",
261
+ "blocks.2.attn.qkv_proj.weight": "NVFP4",
262
+ "blocks.2.attn.qkv_proj.weight_scale": "FP32_SCALE",
263
+ "blocks.2.attn.qkv_proj.comfy_quant": "METADATA",
264
+ "blocks.2.mlp.fc1.weight": "NVFP4",
265
+ "blocks.2.mlp.fc1.weight_scale": "FP32_SCALE",
266
+ "blocks.2.mlp.fc1.comfy_quant": "METADATA",
267
+ "blocks.2.mlp.fc2.weight": "NVFP4",
268
+ "blocks.2.mlp.fc2.weight_scale": "FP32_SCALE",
269
+ "blocks.2.mlp.fc2.comfy_quant": "METADATA",
270
+ "blocks.2.norm1.weight": "BF16",
271
+ "blocks.2.norm2.weight": "BF16",
272
+ "blocks.20.adaln_proj.linear.bias": "BF16",
273
+ "blocks.20.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
274
+ "blocks.20.adaln_proj.linear.weight_scale": "FP32_SCALE",
275
+ "blocks.20.adaln_proj.linear.comfy_quant": "METADATA",
276
+ "blocks.20.attn.k_norm.weight": "BF16",
277
+ "blocks.20.attn.out_proj.weight": "NVFP4",
278
+ "blocks.20.attn.out_proj.weight_scale": "FP32_SCALE",
279
+ "blocks.20.attn.out_proj.comfy_quant": "METADATA",
280
+ "blocks.20.attn.q_norm.weight": "BF16",
281
+ "blocks.20.attn.qkv_proj.weight": "NVFP4",
282
+ "blocks.20.attn.qkv_proj.weight_scale": "FP32_SCALE",
283
+ "blocks.20.attn.qkv_proj.comfy_quant": "METADATA",
284
+ "blocks.20.mlp.fc1.weight": "NVFP4",
285
+ "blocks.20.mlp.fc1.weight_scale": "FP32_SCALE",
286
+ "blocks.20.mlp.fc1.comfy_quant": "METADATA",
287
+ "blocks.20.mlp.fc2.weight": "NVFP4",
288
+ "blocks.20.mlp.fc2.weight_scale": "FP32_SCALE",
289
+ "blocks.20.mlp.fc2.comfy_quant": "METADATA",
290
+ "blocks.20.norm1.weight": "BF16",
291
+ "blocks.20.norm2.weight": "BF16",
292
+ "blocks.21.adaln_proj.linear.bias": "BF16",
293
+ "blocks.21.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
294
+ "blocks.21.adaln_proj.linear.weight_scale": "FP32_SCALE",
295
+ "blocks.21.adaln_proj.linear.comfy_quant": "METADATA",
296
+ "blocks.21.attn.k_norm.weight": "BF16",
297
+ "blocks.21.attn.out_proj.weight": "NVFP4",
298
+ "blocks.21.attn.out_proj.weight_scale": "FP32_SCALE",
299
+ "blocks.21.attn.out_proj.comfy_quant": "METADATA",
300
+ "blocks.21.attn.q_norm.weight": "BF16",
301
+ "blocks.21.attn.qkv_proj.weight": "NVFP4",
302
+ "blocks.21.attn.qkv_proj.weight_scale": "FP32_SCALE",
303
+ "blocks.21.attn.qkv_proj.comfy_quant": "METADATA",
304
+ "blocks.21.mlp.fc1.weight": "NVFP4",
305
+ "blocks.21.mlp.fc1.weight_scale": "FP32_SCALE",
306
+ "blocks.21.mlp.fc1.comfy_quant": "METADATA",
307
+ "blocks.21.mlp.fc2.weight": "NVFP4",
308
+ "blocks.21.mlp.fc2.weight_scale": "FP32_SCALE",
309
+ "blocks.21.mlp.fc2.comfy_quant": "METADATA",
310
+ "blocks.21.norm1.weight": "BF16",
311
+ "blocks.21.norm2.weight": "BF16",
312
+ "blocks.22.adaln_proj.linear.bias": "BF16",
313
+ "blocks.22.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
314
+ "blocks.22.adaln_proj.linear.weight_scale": "FP32_SCALE",
315
+ "blocks.22.adaln_proj.linear.comfy_quant": "METADATA",
316
+ "blocks.22.attn.k_norm.weight": "BF16",
317
+ "blocks.22.attn.out_proj.weight": "NVFP4",
318
+ "blocks.22.attn.out_proj.weight_scale": "FP32_SCALE",
319
+ "blocks.22.attn.out_proj.comfy_quant": "METADATA",
320
+ "blocks.22.attn.q_norm.weight": "BF16",
321
+ "blocks.22.attn.qkv_proj.weight": "NVFP4",
322
+ "blocks.22.attn.qkv_proj.weight_scale": "FP32_SCALE",
323
+ "blocks.22.attn.qkv_proj.comfy_quant": "METADATA",
324
+ "blocks.22.mlp.fc1.weight": "NVFP4",
325
+ "blocks.22.mlp.fc1.weight_scale": "FP32_SCALE",
326
+ "blocks.22.mlp.fc1.comfy_quant": "METADATA",
327
+ "blocks.22.mlp.fc2.weight": "NVFP4",
328
+ "blocks.22.mlp.fc2.weight_scale": "FP32_SCALE",
329
+ "blocks.22.mlp.fc2.comfy_quant": "METADATA",
330
+ "blocks.22.norm1.weight": "BF16",
331
+ "blocks.22.norm2.weight": "BF16",
332
+ "blocks.23.adaln_proj.linear.bias": "BF16",
333
+ "blocks.23.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
334
+ "blocks.23.adaln_proj.linear.weight_scale": "FP32_SCALE",
335
+ "blocks.23.adaln_proj.linear.comfy_quant": "METADATA",
336
+ "blocks.23.attn.k_norm.weight": "BF16",
337
+ "blocks.23.attn.out_proj.weight": "NVFP4",
338
+ "blocks.23.attn.out_proj.weight_scale": "FP32_SCALE",
339
+ "blocks.23.attn.out_proj.comfy_quant": "METADATA",
340
+ "blocks.23.attn.q_norm.weight": "BF16",
341
+ "blocks.23.attn.qkv_proj.weight": "NVFP4",
342
+ "blocks.23.attn.qkv_proj.weight_scale": "FP32_SCALE",
343
+ "blocks.23.attn.qkv_proj.comfy_quant": "METADATA",
344
+ "blocks.23.mlp.fc1.weight": "NVFP4",
345
+ "blocks.23.mlp.fc1.weight_scale": "FP32_SCALE",
346
+ "blocks.23.mlp.fc1.comfy_quant": "METADATA",
347
+ "blocks.23.mlp.fc2.weight": "NVFP4",
348
+ "blocks.23.mlp.fc2.weight_scale": "FP32_SCALE",
349
+ "blocks.23.mlp.fc2.comfy_quant": "METADATA",
350
+ "blocks.23.norm1.weight": "BF16",
351
+ "blocks.23.norm2.weight": "BF16",
352
+ "blocks.24.adaln_proj.linear.bias": "BF16",
353
+ "blocks.24.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
354
+ "blocks.24.adaln_proj.linear.weight_scale": "FP32_SCALE",
355
+ "blocks.24.adaln_proj.linear.comfy_quant": "METADATA",
356
+ "blocks.24.attn.k_norm.weight": "BF16",
357
+ "blocks.24.attn.out_proj.weight": "NVFP4",
358
+ "blocks.24.attn.out_proj.weight_scale": "FP32_SCALE",
359
+ "blocks.24.attn.out_proj.comfy_quant": "METADATA",
360
+ "blocks.24.attn.q_norm.weight": "BF16",
361
+ "blocks.24.attn.qkv_proj.weight": "NVFP4",
362
+ "blocks.24.attn.qkv_proj.weight_scale": "FP32_SCALE",
363
+ "blocks.24.attn.qkv_proj.comfy_quant": "METADATA",
364
+ "blocks.24.mlp.fc1.weight": "NVFP4",
365
+ "blocks.24.mlp.fc1.weight_scale": "FP32_SCALE",
366
+ "blocks.24.mlp.fc1.comfy_quant": "METADATA",
367
+ "blocks.24.mlp.fc2.weight": "NVFP4",
368
+ "blocks.24.mlp.fc2.weight_scale": "FP32_SCALE",
369
+ "blocks.24.mlp.fc2.comfy_quant": "METADATA",
370
+ "blocks.24.norm1.weight": "BF16",
371
+ "blocks.24.norm2.weight": "BF16",
372
+ "blocks.25.adaln_proj.linear.bias": "BF16",
373
+ "blocks.25.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
374
+ "blocks.25.adaln_proj.linear.weight_scale": "FP32_SCALE",
375
+ "blocks.25.adaln_proj.linear.comfy_quant": "METADATA",
376
+ "blocks.25.attn.k_norm.weight": "BF16",
377
+ "blocks.25.attn.out_proj.weight": "NVFP4",
378
+ "blocks.25.attn.out_proj.weight_scale": "FP32_SCALE",
379
+ "blocks.25.attn.out_proj.comfy_quant": "METADATA",
380
+ "blocks.25.attn.q_norm.weight": "BF16",
381
+ "blocks.25.attn.qkv_proj.weight": "NVFP4",
382
+ "blocks.25.attn.qkv_proj.weight_scale": "FP32_SCALE",
383
+ "blocks.25.attn.qkv_proj.comfy_quant": "METADATA",
384
+ "blocks.25.mlp.fc1.weight": "NVFP4",
385
+ "blocks.25.mlp.fc1.weight_scale": "FP32_SCALE",
386
+ "blocks.25.mlp.fc1.comfy_quant": "METADATA",
387
+ "blocks.25.mlp.fc2.weight": "NVFP4",
388
+ "blocks.25.mlp.fc2.weight_scale": "FP32_SCALE",
389
+ "blocks.25.mlp.fc2.comfy_quant": "METADATA",
390
+ "blocks.25.norm1.weight": "BF16",
391
+ "blocks.25.norm2.weight": "BF16",
392
+ "blocks.26.adaln_proj.linear.bias": "BF16",
393
+ "blocks.26.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
394
+ "blocks.26.adaln_proj.linear.weight_scale": "FP32_SCALE",
395
+ "blocks.26.adaln_proj.linear.comfy_quant": "METADATA",
396
+ "blocks.26.attn.k_norm.weight": "BF16",
397
+ "blocks.26.attn.out_proj.weight": "NVFP4",
398
+ "blocks.26.attn.out_proj.weight_scale": "FP32_SCALE",
399
+ "blocks.26.attn.out_proj.comfy_quant": "METADATA",
400
+ "blocks.26.attn.q_norm.weight": "BF16",
401
+ "blocks.26.attn.qkv_proj.weight": "NVFP4",
402
+ "blocks.26.attn.qkv_proj.weight_scale": "FP32_SCALE",
403
+ "blocks.26.attn.qkv_proj.comfy_quant": "METADATA",
404
+ "blocks.26.mlp.fc1.weight": "NVFP4",
405
+ "blocks.26.mlp.fc1.weight_scale": "FP32_SCALE",
406
+ "blocks.26.mlp.fc1.comfy_quant": "METADATA",
407
+ "blocks.26.mlp.fc2.weight": "NVFP4",
408
+ "blocks.26.mlp.fc2.weight_scale": "FP32_SCALE",
409
+ "blocks.26.mlp.fc2.comfy_quant": "METADATA",
410
+ "blocks.26.norm1.weight": "BF16",
411
+ "blocks.26.norm2.weight": "BF16",
412
+ "blocks.27.adaln_proj.linear.bias": "BF16",
413
+ "blocks.27.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
414
+ "blocks.27.adaln_proj.linear.weight_scale": "FP32_SCALE",
415
+ "blocks.27.adaln_proj.linear.comfy_quant": "METADATA",
416
+ "blocks.27.attn.k_norm.weight": "BF16",
417
+ "blocks.27.attn.out_proj.weight": "NVFP4",
418
+ "blocks.27.attn.out_proj.weight_scale": "FP32_SCALE",
419
+ "blocks.27.attn.out_proj.comfy_quant": "METADATA",
420
+ "blocks.27.attn.q_norm.weight": "BF16",
421
+ "blocks.27.attn.qkv_proj.weight": "NVFP4",
422
+ "blocks.27.attn.qkv_proj.weight_scale": "FP32_SCALE",
423
+ "blocks.27.attn.qkv_proj.comfy_quant": "METADATA",
424
+ "blocks.27.mlp.fc1.weight": "NVFP4",
425
+ "blocks.27.mlp.fc1.weight_scale": "FP32_SCALE",
426
+ "blocks.27.mlp.fc1.comfy_quant": "METADATA",
427
+ "blocks.27.mlp.fc2.weight": "NVFP4",
428
+ "blocks.27.mlp.fc2.weight_scale": "FP32_SCALE",
429
+ "blocks.27.mlp.fc2.comfy_quant": "METADATA",
430
+ "blocks.27.norm1.weight": "BF16",
431
+ "blocks.27.norm2.weight": "BF16",
432
+ "blocks.28.adaln_proj.linear.bias": "BF16",
433
+ "blocks.28.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
434
+ "blocks.28.adaln_proj.linear.weight_scale": "FP32_SCALE",
435
+ "blocks.28.adaln_proj.linear.comfy_quant": "METADATA",
436
+ "blocks.28.attn.k_norm.weight": "BF16",
437
+ "blocks.28.attn.out_proj.weight": "NVFP4",
438
+ "blocks.28.attn.out_proj.weight_scale": "FP32_SCALE",
439
+ "blocks.28.attn.out_proj.comfy_quant": "METADATA",
440
+ "blocks.28.attn.q_norm.weight": "BF16",
441
+ "blocks.28.attn.qkv_proj.weight": "NVFP4",
442
+ "blocks.28.attn.qkv_proj.weight_scale": "FP32_SCALE",
443
+ "blocks.28.attn.qkv_proj.comfy_quant": "METADATA",
444
+ "blocks.28.mlp.fc1.weight": "NVFP4",
445
+ "blocks.28.mlp.fc1.weight_scale": "FP32_SCALE",
446
+ "blocks.28.mlp.fc1.comfy_quant": "METADATA",
447
+ "blocks.28.mlp.fc2.weight": "NVFP4",
448
+ "blocks.28.mlp.fc2.weight_scale": "FP32_SCALE",
449
+ "blocks.28.mlp.fc2.comfy_quant": "METADATA",
450
+ "blocks.28.norm1.weight": "BF16",
451
+ "blocks.28.norm2.weight": "BF16",
452
+ "blocks.29.adaln_proj.linear.bias": "BF16",
453
+ "blocks.29.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
454
+ "blocks.29.adaln_proj.linear.weight_scale": "FP32_SCALE",
455
+ "blocks.29.adaln_proj.linear.comfy_quant": "METADATA",
456
+ "blocks.29.attn.k_norm.weight": "BF16",
457
+ "blocks.29.attn.out_proj.weight": "NVFP4",
458
+ "blocks.29.attn.out_proj.weight_scale": "FP32_SCALE",
459
+ "blocks.29.attn.out_proj.comfy_quant": "METADATA",
460
+ "blocks.29.attn.q_norm.weight": "BF16",
461
+ "blocks.29.attn.qkv_proj.weight": "NVFP4",
462
+ "blocks.29.attn.qkv_proj.weight_scale": "FP32_SCALE",
463
+ "blocks.29.attn.qkv_proj.comfy_quant": "METADATA",
464
+ "blocks.29.mlp.fc1.weight": "NVFP4",
465
+ "blocks.29.mlp.fc1.weight_scale": "FP32_SCALE",
466
+ "blocks.29.mlp.fc1.comfy_quant": "METADATA",
467
+ "blocks.29.mlp.fc2.weight": "NVFP4",
468
+ "blocks.29.mlp.fc2.weight_scale": "FP32_SCALE",
469
+ "blocks.29.mlp.fc2.comfy_quant": "METADATA",
470
+ "blocks.29.norm1.weight": "BF16",
471
+ "blocks.29.norm2.weight": "BF16",
472
+ "blocks.3.adaln_proj.linear.bias": "BF16",
473
+ "blocks.3.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
474
+ "blocks.3.adaln_proj.linear.weight_scale": "FP32_SCALE",
475
+ "blocks.3.adaln_proj.linear.comfy_quant": "METADATA",
476
+ "blocks.3.attn.k_norm.weight": "BF16",
477
+ "blocks.3.attn.out_proj.weight": "NVFP4",
478
+ "blocks.3.attn.out_proj.weight_scale": "FP32_SCALE",
479
+ "blocks.3.attn.out_proj.comfy_quant": "METADATA",
480
+ "blocks.3.attn.q_norm.weight": "BF16",
481
+ "blocks.3.attn.qkv_proj.weight": "NVFP4",
482
+ "blocks.3.attn.qkv_proj.weight_scale": "FP32_SCALE",
483
+ "blocks.3.attn.qkv_proj.comfy_quant": "METADATA",
484
+ "blocks.3.mlp.fc1.weight": "NVFP4",
485
+ "blocks.3.mlp.fc1.weight_scale": "FP32_SCALE",
486
+ "blocks.3.mlp.fc1.comfy_quant": "METADATA",
487
+ "blocks.3.mlp.fc2.weight": "NVFP4",
488
+ "blocks.3.mlp.fc2.weight_scale": "FP32_SCALE",
489
+ "blocks.3.mlp.fc2.comfy_quant": "METADATA",
490
+ "blocks.3.norm1.weight": "BF16",
491
+ "blocks.3.norm2.weight": "BF16",
492
+ "blocks.30.adaln_proj.linear.bias": "BF16",
493
+ "blocks.30.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
494
+ "blocks.30.adaln_proj.linear.weight_scale": "FP32_SCALE",
495
+ "blocks.30.adaln_proj.linear.comfy_quant": "METADATA",
496
+ "blocks.30.attn.k_norm.weight": "BF16",
497
+ "blocks.30.attn.out_proj.weight": "NVFP4",
498
+ "blocks.30.attn.out_proj.weight_scale": "FP32_SCALE",
499
+ "blocks.30.attn.out_proj.comfy_quant": "METADATA",
500
+ "blocks.30.attn.q_norm.weight": "BF16",
501
+ "blocks.30.attn.qkv_proj.weight": "NVFP4",
502
+ "blocks.30.attn.qkv_proj.weight_scale": "FP32_SCALE",
503
+ "blocks.30.attn.qkv_proj.comfy_quant": "METADATA",
504
+ "blocks.30.mlp.fc1.weight": "NVFP4",
505
+ "blocks.30.mlp.fc1.weight_scale": "FP32_SCALE",
506
+ "blocks.30.mlp.fc1.comfy_quant": "METADATA",
507
+ "blocks.30.mlp.fc2.weight": "NVFP4",
508
+ "blocks.30.mlp.fc2.weight_scale": "FP32_SCALE",
509
+ "blocks.30.mlp.fc2.comfy_quant": "METADATA",
510
+ "blocks.30.norm1.weight": "BF16",
511
+ "blocks.30.norm2.weight": "BF16",
512
+ "blocks.31.adaln_proj.linear.bias": "BF16",
513
+ "blocks.31.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
514
+ "blocks.31.adaln_proj.linear.weight_scale": "FP32_SCALE",
515
+ "blocks.31.adaln_proj.linear.comfy_quant": "METADATA",
516
+ "blocks.31.attn.k_norm.weight": "BF16",
517
+ "blocks.31.attn.out_proj.weight": "NVFP4",
518
+ "blocks.31.attn.out_proj.weight_scale": "FP32_SCALE",
519
+ "blocks.31.attn.out_proj.comfy_quant": "METADATA",
520
+ "blocks.31.attn.q_norm.weight": "BF16",
521
+ "blocks.31.attn.qkv_proj.weight": "NVFP4",
522
+ "blocks.31.attn.qkv_proj.weight_scale": "FP32_SCALE",
523
+ "blocks.31.attn.qkv_proj.comfy_quant": "METADATA",
524
+ "blocks.31.mlp.fc1.weight": "NVFP4",
525
+ "blocks.31.mlp.fc1.weight_scale": "FP32_SCALE",
526
+ "blocks.31.mlp.fc1.comfy_quant": "METADATA",
527
+ "blocks.31.mlp.fc2.weight": "NVFP4",
528
+ "blocks.31.mlp.fc2.weight_scale": "FP32_SCALE",
529
+ "blocks.31.mlp.fc2.comfy_quant": "METADATA",
530
+ "blocks.31.norm1.weight": "BF16",
531
+ "blocks.31.norm2.weight": "BF16",
532
+ "blocks.32.adaln_proj.linear.bias": "BF16",
533
+ "blocks.32.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
534
+ "blocks.32.adaln_proj.linear.weight_scale": "FP32_SCALE",
535
+ "blocks.32.adaln_proj.linear.comfy_quant": "METADATA",
536
+ "blocks.32.attn.k_norm.weight": "BF16",
537
+ "blocks.32.attn.out_proj.weight": "NVFP4",
538
+ "blocks.32.attn.out_proj.weight_scale": "FP32_SCALE",
539
+ "blocks.32.attn.out_proj.comfy_quant": "METADATA",
540
+ "blocks.32.attn.q_norm.weight": "BF16",
541
+ "blocks.32.attn.qkv_proj.weight": "NVFP4",
542
+ "blocks.32.attn.qkv_proj.weight_scale": "FP32_SCALE",
543
+ "blocks.32.attn.qkv_proj.comfy_quant": "METADATA",
544
+ "blocks.32.mlp.fc1.weight": "NVFP4",
545
+ "blocks.32.mlp.fc1.weight_scale": "FP32_SCALE",
546
+ "blocks.32.mlp.fc1.comfy_quant": "METADATA",
547
+ "blocks.32.mlp.fc2.weight": "NVFP4",
548
+ "blocks.32.mlp.fc2.weight_scale": "FP32_SCALE",
549
+ "blocks.32.mlp.fc2.comfy_quant": "METADATA",
550
+ "blocks.32.norm1.weight": "BF16",
551
+ "blocks.32.norm2.weight": "BF16",
552
+ "blocks.33.adaln_proj.linear.bias": "BF16",
553
+ "blocks.33.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
554
+ "blocks.33.adaln_proj.linear.weight_scale": "FP32_SCALE",
555
+ "blocks.33.adaln_proj.linear.comfy_quant": "METADATA",
556
+ "blocks.33.attn.k_norm.weight": "BF16",
557
+ "blocks.33.attn.out_proj.weight": "NVFP4",
558
+ "blocks.33.attn.out_proj.weight_scale": "FP32_SCALE",
559
+ "blocks.33.attn.out_proj.comfy_quant": "METADATA",
560
+ "blocks.33.attn.q_norm.weight": "BF16",
561
+ "blocks.33.attn.qkv_proj.weight": "NVFP4",
562
+ "blocks.33.attn.qkv_proj.weight_scale": "FP32_SCALE",
563
+ "blocks.33.attn.qkv_proj.comfy_quant": "METADATA",
564
+ "blocks.33.mlp.fc1.weight": "NVFP4",
565
+ "blocks.33.mlp.fc1.weight_scale": "FP32_SCALE",
566
+ "blocks.33.mlp.fc1.comfy_quant": "METADATA",
567
+ "blocks.33.mlp.fc2.weight": "NVFP4",
568
+ "blocks.33.mlp.fc2.weight_scale": "FP32_SCALE",
569
+ "blocks.33.mlp.fc2.comfy_quant": "METADATA",
570
+ "blocks.33.norm1.weight": "BF16",
571
+ "blocks.33.norm2.weight": "BF16",
572
+ "blocks.34.adaln_proj.linear.bias": "BF16",
573
+ "blocks.34.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
574
+ "blocks.34.adaln_proj.linear.weight_scale": "FP32_SCALE",
575
+ "blocks.34.adaln_proj.linear.comfy_quant": "METADATA",
576
+ "blocks.34.attn.k_norm.weight": "BF16",
577
+ "blocks.34.attn.out_proj.weight": "NVFP4",
578
+ "blocks.34.attn.out_proj.weight_scale": "FP32_SCALE",
579
+ "blocks.34.attn.out_proj.comfy_quant": "METADATA",
580
+ "blocks.34.attn.q_norm.weight": "BF16",
581
+ "blocks.34.attn.qkv_proj.weight": "NVFP4",
582
+ "blocks.34.attn.qkv_proj.weight_scale": "FP32_SCALE",
583
+ "blocks.34.attn.qkv_proj.comfy_quant": "METADATA",
584
+ "blocks.34.mlp.fc1.weight": "NVFP4",
585
+ "blocks.34.mlp.fc1.weight_scale": "FP32_SCALE",
586
+ "blocks.34.mlp.fc1.comfy_quant": "METADATA",
587
+ "blocks.34.mlp.fc2.weight": "NVFP4",
588
+ "blocks.34.mlp.fc2.weight_scale": "FP32_SCALE",
589
+ "blocks.34.mlp.fc2.comfy_quant": "METADATA",
590
+ "blocks.34.norm1.weight": "BF16",
591
+ "blocks.34.norm2.weight": "BF16",
592
+ "blocks.35.adaln_proj.linear.bias": "BF16",
593
+ "blocks.35.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
594
+ "blocks.35.adaln_proj.linear.weight_scale": "FP32_SCALE",
595
+ "blocks.35.adaln_proj.linear.comfy_quant": "METADATA",
596
+ "blocks.35.attn.k_norm.weight": "BF16",
597
+ "blocks.35.attn.out_proj.weight": "NVFP4",
598
+ "blocks.35.attn.out_proj.weight_scale": "FP32_SCALE",
599
+ "blocks.35.attn.out_proj.comfy_quant": "METADATA",
600
+ "blocks.35.attn.q_norm.weight": "BF16",
601
+ "blocks.35.attn.qkv_proj.weight": "NVFP4",
602
+ "blocks.35.attn.qkv_proj.weight_scale": "FP32_SCALE",
603
+ "blocks.35.attn.qkv_proj.comfy_quant": "METADATA",
604
+ "blocks.35.mlp.fc1.weight": "NVFP4",
605
+ "blocks.35.mlp.fc1.weight_scale": "FP32_SCALE",
606
+ "blocks.35.mlp.fc1.comfy_quant": "METADATA",
607
+ "blocks.35.mlp.fc2.weight": "NVFP4",
608
+ "blocks.35.mlp.fc2.weight_scale": "FP32_SCALE",
609
+ "blocks.35.mlp.fc2.comfy_quant": "METADATA",
610
+ "blocks.35.norm1.weight": "BF16",
611
+ "blocks.35.norm2.weight": "BF16",
612
+ "blocks.36.adaln_proj.linear.bias": "BF16",
613
+ "blocks.36.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
614
+ "blocks.36.adaln_proj.linear.weight_scale": "FP32_SCALE",
615
+ "blocks.36.adaln_proj.linear.comfy_quant": "METADATA",
616
+ "blocks.36.attn.k_norm.weight": "BF16",
617
+ "blocks.36.attn.out_proj.weight": "NVFP4",
618
+ "blocks.36.attn.out_proj.weight_scale": "FP32_SCALE",
619
+ "blocks.36.attn.out_proj.comfy_quant": "METADATA",
620
+ "blocks.36.attn.q_norm.weight": "BF16",
621
+ "blocks.36.attn.qkv_proj.weight": "NVFP4",
622
+ "blocks.36.attn.qkv_proj.weight_scale": "FP32_SCALE",
623
+ "blocks.36.attn.qkv_proj.comfy_quant": "METADATA",
624
+ "blocks.36.mlp.fc1.weight": "NVFP4",
625
+ "blocks.36.mlp.fc1.weight_scale": "FP32_SCALE",
626
+ "blocks.36.mlp.fc1.comfy_quant": "METADATA",
627
+ "blocks.36.mlp.fc2.weight": "NVFP4",
628
+ "blocks.36.mlp.fc2.weight_scale": "FP32_SCALE",
629
+ "blocks.36.mlp.fc2.comfy_quant": "METADATA",
630
+ "blocks.36.norm1.weight": "BF16",
631
+ "blocks.36.norm2.weight": "BF16",
632
+ "blocks.37.adaln_proj.linear.bias": "BF16",
633
+ "blocks.37.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
634
+ "blocks.37.adaln_proj.linear.weight_scale": "FP32_SCALE",
635
+ "blocks.37.adaln_proj.linear.comfy_quant": "METADATA",
636
+ "blocks.37.attn.k_norm.weight": "BF16",
637
+ "blocks.37.attn.out_proj.weight": "NVFP4",
638
+ "blocks.37.attn.out_proj.weight_scale": "FP32_SCALE",
639
+ "blocks.37.attn.out_proj.comfy_quant": "METADATA",
640
+ "blocks.37.attn.q_norm.weight": "BF16",
641
+ "blocks.37.attn.qkv_proj.weight": "NVFP4",
642
+ "blocks.37.attn.qkv_proj.weight_scale": "FP32_SCALE",
643
+ "blocks.37.attn.qkv_proj.comfy_quant": "METADATA",
644
+ "blocks.37.mlp.fc1.weight": "NVFP4",
645
+ "blocks.37.mlp.fc1.weight_scale": "FP32_SCALE",
646
+ "blocks.37.mlp.fc1.comfy_quant": "METADATA",
647
+ "blocks.37.mlp.fc2.weight": "NVFP4",
648
+ "blocks.37.mlp.fc2.weight_scale": "FP32_SCALE",
649
+ "blocks.37.mlp.fc2.comfy_quant": "METADATA",
650
+ "blocks.37.norm1.weight": "BF16",
651
+ "blocks.37.norm2.weight": "BF16",
652
+ "blocks.38.adaln_proj.linear.bias": "BF16",
653
+ "blocks.38.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
654
+ "blocks.38.adaln_proj.linear.weight_scale": "FP32_SCALE",
655
+ "blocks.38.adaln_proj.linear.comfy_quant": "METADATA",
656
+ "blocks.38.attn.k_norm.weight": "BF16",
657
+ "blocks.38.attn.out_proj.weight": "NVFP4",
658
+ "blocks.38.attn.out_proj.weight_scale": "FP32_SCALE",
659
+ "blocks.38.attn.out_proj.comfy_quant": "METADATA",
660
+ "blocks.38.attn.q_norm.weight": "BF16",
661
+ "blocks.38.attn.qkv_proj.weight": "NVFP4",
662
+ "blocks.38.attn.qkv_proj.weight_scale": "FP32_SCALE",
663
+ "blocks.38.attn.qkv_proj.comfy_quant": "METADATA",
664
+ "blocks.38.mlp.fc1.weight": "NVFP4",
665
+ "blocks.38.mlp.fc1.weight_scale": "FP32_SCALE",
666
+ "blocks.38.mlp.fc1.comfy_quant": "METADATA",
667
+ "blocks.38.mlp.fc2.weight": "NVFP4",
668
+ "blocks.38.mlp.fc2.weight_scale": "FP32_SCALE",
669
+ "blocks.38.mlp.fc2.comfy_quant": "METADATA",
670
+ "blocks.38.norm1.weight": "BF16",
671
+ "blocks.38.norm2.weight": "BF16",
672
+ "blocks.39.adaln_proj.linear.bias": "BF16",
673
+ "blocks.39.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
674
+ "blocks.39.adaln_proj.linear.weight_scale": "FP32_SCALE",
675
+ "blocks.39.adaln_proj.linear.comfy_quant": "METADATA",
676
+ "blocks.39.attn.k_norm.weight": "BF16",
677
+ "blocks.39.attn.out_proj.weight": "NVFP4",
678
+ "blocks.39.attn.out_proj.weight_scale": "FP32_SCALE",
679
+ "blocks.39.attn.out_proj.comfy_quant": "METADATA",
680
+ "blocks.39.attn.q_norm.weight": "BF16",
681
+ "blocks.39.attn.qkv_proj.weight": "NVFP4",
682
+ "blocks.39.attn.qkv_proj.weight_scale": "FP32_SCALE",
683
+ "blocks.39.attn.qkv_proj.comfy_quant": "METADATA",
684
+ "blocks.39.mlp.fc1.weight": "NVFP4",
685
+ "blocks.39.mlp.fc1.weight_scale": "FP32_SCALE",
686
+ "blocks.39.mlp.fc1.comfy_quant": "METADATA",
687
+ "blocks.39.mlp.fc2.weight": "NVFP4",
688
+ "blocks.39.mlp.fc2.weight_scale": "FP32_SCALE",
689
+ "blocks.39.mlp.fc2.comfy_quant": "METADATA",
690
+ "blocks.39.norm1.weight": "BF16",
691
+ "blocks.39.norm2.weight": "BF16",
692
+ "blocks.4.adaln_proj.linear.bias": "BF16",
693
+ "blocks.4.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
694
+ "blocks.4.adaln_proj.linear.weight_scale": "FP32_SCALE",
695
+ "blocks.4.adaln_proj.linear.comfy_quant": "METADATA",
696
+ "blocks.4.attn.k_norm.weight": "BF16",
697
+ "blocks.4.attn.out_proj.weight": "NVFP4",
698
+ "blocks.4.attn.out_proj.weight_scale": "FP32_SCALE",
699
+ "blocks.4.attn.out_proj.comfy_quant": "METADATA",
700
+ "blocks.4.attn.q_norm.weight": "BF16",
701
+ "blocks.4.attn.qkv_proj.weight": "NVFP4",
702
+ "blocks.4.attn.qkv_proj.weight_scale": "FP32_SCALE",
703
+ "blocks.4.attn.qkv_proj.comfy_quant": "METADATA",
704
+ "blocks.4.mlp.fc1.weight": "NVFP4",
705
+ "blocks.4.mlp.fc1.weight_scale": "FP32_SCALE",
706
+ "blocks.4.mlp.fc1.comfy_quant": "METADATA",
707
+ "blocks.4.mlp.fc2.weight": "NVFP4",
708
+ "blocks.4.mlp.fc2.weight_scale": "FP32_SCALE",
709
+ "blocks.4.mlp.fc2.comfy_quant": "METADATA",
710
+ "blocks.4.norm1.weight": "BF16",
711
+ "blocks.4.norm2.weight": "BF16",
712
+ "blocks.40.adaln_proj.linear.bias": "BF16",
713
+ "blocks.40.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
714
+ "blocks.40.adaln_proj.linear.weight_scale": "FP32_SCALE",
715
+ "blocks.40.adaln_proj.linear.comfy_quant": "METADATA",
716
+ "blocks.40.attn.k_norm.weight": "BF16",
717
+ "blocks.40.attn.out_proj.weight": "NVFP4",
718
+ "blocks.40.attn.out_proj.weight_scale": "FP32_SCALE",
719
+ "blocks.40.attn.out_proj.comfy_quant": "METADATA",
720
+ "blocks.40.attn.q_norm.weight": "BF16",
721
+ "blocks.40.attn.qkv_proj.weight": "NVFP4",
722
+ "blocks.40.attn.qkv_proj.weight_scale": "FP32_SCALE",
723
+ "blocks.40.attn.qkv_proj.comfy_quant": "METADATA",
724
+ "blocks.40.mlp.fc1.weight": "NVFP4",
725
+ "blocks.40.mlp.fc1.weight_scale": "FP32_SCALE",
726
+ "blocks.40.mlp.fc1.comfy_quant": "METADATA",
727
+ "blocks.40.mlp.fc2.weight": "NVFP4",
728
+ "blocks.40.mlp.fc2.weight_scale": "FP32_SCALE",
729
+ "blocks.40.mlp.fc2.comfy_quant": "METADATA",
730
+ "blocks.40.norm1.weight": "BF16",
731
+ "blocks.40.norm2.weight": "BF16",
732
+ "blocks.41.adaln_proj.linear.bias": "BF16",
733
+ "blocks.41.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
734
+ "blocks.41.adaln_proj.linear.weight_scale": "FP32_SCALE",
735
+ "blocks.41.adaln_proj.linear.comfy_quant": "METADATA",
736
+ "blocks.41.attn.k_norm.weight": "BF16",
737
+ "blocks.41.attn.out_proj.weight": "NVFP4",
738
+ "blocks.41.attn.out_proj.weight_scale": "FP32_SCALE",
739
+ "blocks.41.attn.out_proj.comfy_quant": "METADATA",
740
+ "blocks.41.attn.q_norm.weight": "BF16",
741
+ "blocks.41.attn.qkv_proj.weight": "NVFP4",
742
+ "blocks.41.attn.qkv_proj.weight_scale": "FP32_SCALE",
743
+ "blocks.41.attn.qkv_proj.comfy_quant": "METADATA",
744
+ "blocks.41.mlp.fc1.weight": "NVFP4",
745
+ "blocks.41.mlp.fc1.weight_scale": "FP32_SCALE",
746
+ "blocks.41.mlp.fc1.comfy_quant": "METADATA",
747
+ "blocks.41.mlp.fc2.weight": "NVFP4",
748
+ "blocks.41.mlp.fc2.weight_scale": "FP32_SCALE",
749
+ "blocks.41.mlp.fc2.comfy_quant": "METADATA",
750
+ "blocks.41.norm1.weight": "BF16",
751
+ "blocks.41.norm2.weight": "BF16",
752
+ "blocks.42.adaln_proj.linear.bias": "BF16",
753
+ "blocks.42.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
754
+ "blocks.42.adaln_proj.linear.weight_scale": "FP32_SCALE",
755
+ "blocks.42.adaln_proj.linear.comfy_quant": "METADATA",
756
+ "blocks.42.attn.k_norm.weight": "BF16",
757
+ "blocks.42.attn.out_proj.weight": "NVFP4",
758
+ "blocks.42.attn.out_proj.weight_scale": "FP32_SCALE",
759
+ "blocks.42.attn.out_proj.comfy_quant": "METADATA",
760
+ "blocks.42.attn.q_norm.weight": "BF16",
761
+ "blocks.42.attn.qkv_proj.weight": "NVFP4",
762
+ "blocks.42.attn.qkv_proj.weight_scale": "FP32_SCALE",
763
+ "blocks.42.attn.qkv_proj.comfy_quant": "METADATA",
764
+ "blocks.42.mlp.fc1.weight": "NVFP4",
765
+ "blocks.42.mlp.fc1.weight_scale": "FP32_SCALE",
766
+ "blocks.42.mlp.fc1.comfy_quant": "METADATA",
767
+ "blocks.42.mlp.fc2.weight": "NVFP4",
768
+ "blocks.42.mlp.fc2.weight_scale": "FP32_SCALE",
769
+ "blocks.42.mlp.fc2.comfy_quant": "METADATA",
770
+ "blocks.42.norm1.weight": "BF16",
771
+ "blocks.42.norm2.weight": "BF16",
772
+ "blocks.43.adaln_proj.linear.bias": "BF16",
773
+ "blocks.43.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
774
+ "blocks.43.adaln_proj.linear.weight_scale": "FP32_SCALE",
775
+ "blocks.43.adaln_proj.linear.comfy_quant": "METADATA",
776
+ "blocks.43.attn.k_norm.weight": "BF16",
777
+ "blocks.43.attn.out_proj.weight": "NVFP4",
778
+ "blocks.43.attn.out_proj.weight_scale": "FP32_SCALE",
779
+ "blocks.43.attn.out_proj.comfy_quant": "METADATA",
780
+ "blocks.43.attn.q_norm.weight": "BF16",
781
+ "blocks.43.attn.qkv_proj.weight": "NVFP4",
782
+ "blocks.43.attn.qkv_proj.weight_scale": "FP32_SCALE",
783
+ "blocks.43.attn.qkv_proj.comfy_quant": "METADATA",
784
+ "blocks.43.mlp.fc1.weight": "NVFP4",
785
+ "blocks.43.mlp.fc1.weight_scale": "FP32_SCALE",
786
+ "blocks.43.mlp.fc1.comfy_quant": "METADATA",
787
+ "blocks.43.mlp.fc2.weight": "NVFP4",
788
+ "blocks.43.mlp.fc2.weight_scale": "FP32_SCALE",
789
+ "blocks.43.mlp.fc2.comfy_quant": "METADATA",
790
+ "blocks.43.norm1.weight": "BF16",
791
+ "blocks.43.norm2.weight": "BF16",
792
+ "blocks.44.adaln_proj.linear.bias": "BF16",
793
+ "blocks.44.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
794
+ "blocks.44.adaln_proj.linear.weight_scale": "FP32_SCALE",
795
+ "blocks.44.adaln_proj.linear.comfy_quant": "METADATA",
796
+ "blocks.44.attn.k_norm.weight": "BF16",
797
+ "blocks.44.attn.out_proj.weight": "NVFP4",
798
+ "blocks.44.attn.out_proj.weight_scale": "FP32_SCALE",
799
+ "blocks.44.attn.out_proj.comfy_quant": "METADATA",
800
+ "blocks.44.attn.q_norm.weight": "BF16",
801
+ "blocks.44.attn.qkv_proj.weight": "NVFP4",
802
+ "blocks.44.attn.qkv_proj.weight_scale": "FP32_SCALE",
803
+ "blocks.44.attn.qkv_proj.comfy_quant": "METADATA",
804
+ "blocks.44.mlp.fc1.weight": "NVFP4",
805
+ "blocks.44.mlp.fc1.weight_scale": "FP32_SCALE",
806
+ "blocks.44.mlp.fc1.comfy_quant": "METADATA",
807
+ "blocks.44.mlp.fc2.weight": "NVFP4",
808
+ "blocks.44.mlp.fc2.weight_scale": "FP32_SCALE",
809
+ "blocks.44.mlp.fc2.comfy_quant": "METADATA",
810
+ "blocks.44.norm1.weight": "BF16",
811
+ "blocks.44.norm2.weight": "BF16",
812
+ "blocks.45.adaln_proj.linear.bias": "BF16",
813
+ "blocks.45.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
814
+ "blocks.45.adaln_proj.linear.weight_scale": "FP32_SCALE",
815
+ "blocks.45.adaln_proj.linear.comfy_quant": "METADATA",
816
+ "blocks.45.attn.k_norm.weight": "BF16",
817
+ "blocks.45.attn.out_proj.weight": "NVFP4",
818
+ "blocks.45.attn.out_proj.weight_scale": "FP32_SCALE",
819
+ "blocks.45.attn.out_proj.comfy_quant": "METADATA",
820
+ "blocks.45.attn.q_norm.weight": "BF16",
821
+ "blocks.45.attn.qkv_proj.weight": "NVFP4",
822
+ "blocks.45.attn.qkv_proj.weight_scale": "FP32_SCALE",
823
+ "blocks.45.attn.qkv_proj.comfy_quant": "METADATA",
824
+ "blocks.45.mlp.fc1.weight": "NVFP4",
825
+ "blocks.45.mlp.fc1.weight_scale": "FP32_SCALE",
826
+ "blocks.45.mlp.fc1.comfy_quant": "METADATA",
827
+ "blocks.45.mlp.fc2.weight": "NVFP4",
828
+ "blocks.45.mlp.fc2.weight_scale": "FP32_SCALE",
829
+ "blocks.45.mlp.fc2.comfy_quant": "METADATA",
830
+ "blocks.45.norm1.weight": "BF16",
831
+ "blocks.45.norm2.weight": "BF16",
832
+ "blocks.46.adaln_proj.linear.bias": "BF16",
833
+ "blocks.46.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
834
+ "blocks.46.adaln_proj.linear.weight_scale": "FP32_SCALE",
835
+ "blocks.46.adaln_proj.linear.comfy_quant": "METADATA",
836
+ "blocks.46.attn.k_norm.weight": "BF16",
837
+ "blocks.46.attn.out_proj.weight": "NVFP4",
838
+ "blocks.46.attn.out_proj.weight_scale": "FP32_SCALE",
839
+ "blocks.46.attn.out_proj.comfy_quant": "METADATA",
840
+ "blocks.46.attn.q_norm.weight": "BF16",
841
+ "blocks.46.attn.qkv_proj.weight": "NVFP4",
842
+ "blocks.46.attn.qkv_proj.weight_scale": "FP32_SCALE",
843
+ "blocks.46.attn.qkv_proj.comfy_quant": "METADATA",
844
+ "blocks.46.mlp.fc1.weight": "NVFP4",
845
+ "blocks.46.mlp.fc1.weight_scale": "FP32_SCALE",
846
+ "blocks.46.mlp.fc1.comfy_quant": "METADATA",
847
+ "blocks.46.mlp.fc2.weight": "NVFP4",
848
+ "blocks.46.mlp.fc2.weight_scale": "FP32_SCALE",
849
+ "blocks.46.mlp.fc2.comfy_quant": "METADATA",
850
+ "blocks.46.norm1.weight": "BF16",
851
+ "blocks.46.norm2.weight": "BF16",
852
+ "blocks.47.adaln_proj.linear.bias": "BF16",
853
+ "blocks.47.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
854
+ "blocks.47.adaln_proj.linear.weight_scale": "FP32_SCALE",
855
+ "blocks.47.adaln_proj.linear.comfy_quant": "METADATA",
856
+ "blocks.47.attn.k_norm.weight": "BF16",
857
+ "blocks.47.attn.out_proj.weight": "NVFP4",
858
+ "blocks.47.attn.out_proj.weight_scale": "FP32_SCALE",
859
+ "blocks.47.attn.out_proj.comfy_quant": "METADATA",
860
+ "blocks.47.attn.q_norm.weight": "BF16",
861
+ "blocks.47.attn.qkv_proj.weight": "NVFP4",
862
+ "blocks.47.attn.qkv_proj.weight_scale": "FP32_SCALE",
863
+ "blocks.47.attn.qkv_proj.comfy_quant": "METADATA",
864
+ "blocks.47.mlp.fc1.weight": "NVFP4",
865
+ "blocks.47.mlp.fc1.weight_scale": "FP32_SCALE",
866
+ "blocks.47.mlp.fc1.comfy_quant": "METADATA",
867
+ "blocks.47.mlp.fc2.weight": "NVFP4",
868
+ "blocks.47.mlp.fc2.weight_scale": "FP32_SCALE",
869
+ "blocks.47.mlp.fc2.comfy_quant": "METADATA",
870
+ "blocks.47.norm1.weight": "BF16",
871
+ "blocks.47.norm2.weight": "BF16",
872
+ "blocks.48.adaln_proj.linear.bias": "BF16",
873
+ "blocks.48.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
874
+ "blocks.48.adaln_proj.linear.weight_scale": "FP32_SCALE",
875
+ "blocks.48.adaln_proj.linear.comfy_quant": "METADATA",
876
+ "blocks.48.attn.k_norm.weight": "BF16",
877
+ "blocks.48.attn.out_proj.weight": "NVFP4",
878
+ "blocks.48.attn.out_proj.weight_scale": "FP32_SCALE",
879
+ "blocks.48.attn.out_proj.comfy_quant": "METADATA",
880
+ "blocks.48.attn.q_norm.weight": "BF16",
881
+ "blocks.48.attn.qkv_proj.weight": "NVFP4",
882
+ "blocks.48.attn.qkv_proj.weight_scale": "FP32_SCALE",
883
+ "blocks.48.attn.qkv_proj.comfy_quant": "METADATA",
884
+ "blocks.48.mlp.fc1.weight": "NVFP4",
885
+ "blocks.48.mlp.fc1.weight_scale": "FP32_SCALE",
886
+ "blocks.48.mlp.fc1.comfy_quant": "METADATA",
887
+ "blocks.48.mlp.fc2.weight": "NVFP4",
888
+ "blocks.48.mlp.fc2.weight_scale": "FP32_SCALE",
889
+ "blocks.48.mlp.fc2.comfy_quant": "METADATA",
890
+ "blocks.48.norm1.weight": "BF16",
891
+ "blocks.48.norm2.weight": "BF16",
892
+ "blocks.49.adaln_proj.linear.bias": "BF16",
893
+ "blocks.49.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
894
+ "blocks.49.adaln_proj.linear.weight_scale": "FP32_SCALE",
895
+ "blocks.49.adaln_proj.linear.comfy_quant": "METADATA",
896
+ "blocks.49.attn.k_norm.weight": "BF16",
897
+ "blocks.49.attn.out_proj.weight": "NVFP4",
898
+ "blocks.49.attn.out_proj.weight_scale": "FP32_SCALE",
899
+ "blocks.49.attn.out_proj.comfy_quant": "METADATA",
900
+ "blocks.49.attn.q_norm.weight": "BF16",
901
+ "blocks.49.attn.qkv_proj.weight": "NVFP4",
902
+ "blocks.49.attn.qkv_proj.weight_scale": "FP32_SCALE",
903
+ "blocks.49.attn.qkv_proj.comfy_quant": "METADATA",
904
+ "blocks.49.mlp.fc1.weight": "NVFP4",
905
+ "blocks.49.mlp.fc1.weight_scale": "FP32_SCALE",
906
+ "blocks.49.mlp.fc1.comfy_quant": "METADATA",
907
+ "blocks.49.mlp.fc2.weight": "NVFP4",
908
+ "blocks.49.mlp.fc2.weight_scale": "FP32_SCALE",
909
+ "blocks.49.mlp.fc2.comfy_quant": "METADATA",
910
+ "blocks.49.norm1.weight": "BF16",
911
+ "blocks.49.norm2.weight": "BF16",
912
+ "blocks.5.adaln_proj.linear.bias": "BF16",
913
+ "blocks.5.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
914
+ "blocks.5.adaln_proj.linear.weight_scale": "FP32_SCALE",
915
+ "blocks.5.adaln_proj.linear.comfy_quant": "METADATA",
916
+ "blocks.5.attn.k_norm.weight": "BF16",
917
+ "blocks.5.attn.out_proj.weight": "NVFP4",
918
+ "blocks.5.attn.out_proj.weight_scale": "FP32_SCALE",
919
+ "blocks.5.attn.out_proj.comfy_quant": "METADATA",
920
+ "blocks.5.attn.q_norm.weight": "BF16",
921
+ "blocks.5.attn.qkv_proj.weight": "NVFP4",
922
+ "blocks.5.attn.qkv_proj.weight_scale": "FP32_SCALE",
923
+ "blocks.5.attn.qkv_proj.comfy_quant": "METADATA",
924
+ "blocks.5.mlp.fc1.weight": "NVFP4",
925
+ "blocks.5.mlp.fc1.weight_scale": "FP32_SCALE",
926
+ "blocks.5.mlp.fc1.comfy_quant": "METADATA",
927
+ "blocks.5.mlp.fc2.weight": "NVFP4",
928
+ "blocks.5.mlp.fc2.weight_scale": "FP32_SCALE",
929
+ "blocks.5.mlp.fc2.comfy_quant": "METADATA",
930
+ "blocks.5.norm1.weight": "BF16",
931
+ "blocks.5.norm2.weight": "BF16",
932
+ "blocks.6.adaln_proj.linear.bias": "BF16",
933
+ "blocks.6.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
934
+ "blocks.6.adaln_proj.linear.weight_scale": "FP32_SCALE",
935
+ "blocks.6.adaln_proj.linear.comfy_quant": "METADATA",
936
+ "blocks.6.attn.k_norm.weight": "BF16",
937
+ "blocks.6.attn.out_proj.weight": "NVFP4",
938
+ "blocks.6.attn.out_proj.weight_scale": "FP32_SCALE",
939
+ "blocks.6.attn.out_proj.comfy_quant": "METADATA",
940
+ "blocks.6.attn.q_norm.weight": "BF16",
941
+ "blocks.6.attn.qkv_proj.weight": "NVFP4",
942
+ "blocks.6.attn.qkv_proj.weight_scale": "FP32_SCALE",
943
+ "blocks.6.attn.qkv_proj.comfy_quant": "METADATA",
944
+ "blocks.6.mlp.fc1.weight": "NVFP4",
945
+ "blocks.6.mlp.fc1.weight_scale": "FP32_SCALE",
946
+ "blocks.6.mlp.fc1.comfy_quant": "METADATA",
947
+ "blocks.6.mlp.fc2.weight": "NVFP4",
948
+ "blocks.6.mlp.fc2.weight_scale": "FP32_SCALE",
949
+ "blocks.6.mlp.fc2.comfy_quant": "METADATA",
950
+ "blocks.6.norm1.weight": "BF16",
951
+ "blocks.6.norm2.weight": "BF16",
952
+ "blocks.7.adaln_proj.linear.bias": "BF16",
953
+ "blocks.7.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
954
+ "blocks.7.adaln_proj.linear.weight_scale": "FP32_SCALE",
955
+ "blocks.7.adaln_proj.linear.comfy_quant": "METADATA",
956
+ "blocks.7.attn.k_norm.weight": "BF16",
957
+ "blocks.7.attn.out_proj.weight": "NVFP4",
958
+ "blocks.7.attn.out_proj.weight_scale": "FP32_SCALE",
959
+ "blocks.7.attn.out_proj.comfy_quant": "METADATA",
960
+ "blocks.7.attn.q_norm.weight": "BF16",
961
+ "blocks.7.attn.qkv_proj.weight": "NVFP4",
962
+ "blocks.7.attn.qkv_proj.weight_scale": "FP32_SCALE",
963
+ "blocks.7.attn.qkv_proj.comfy_quant": "METADATA",
964
+ "blocks.7.mlp.fc1.weight": "NVFP4",
965
+ "blocks.7.mlp.fc1.weight_scale": "FP32_SCALE",
966
+ "blocks.7.mlp.fc1.comfy_quant": "METADATA",
967
+ "blocks.7.mlp.fc2.weight": "NVFP4",
968
+ "blocks.7.mlp.fc2.weight_scale": "FP32_SCALE",
969
+ "blocks.7.mlp.fc2.comfy_quant": "METADATA",
970
+ "blocks.7.norm1.weight": "BF16",
971
+ "blocks.7.norm2.weight": "BF16",
972
+ "blocks.8.adaln_proj.linear.bias": "BF16",
973
+ "blocks.8.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
974
+ "blocks.8.adaln_proj.linear.weight_scale": "FP32_SCALE",
975
+ "blocks.8.adaln_proj.linear.comfy_quant": "METADATA",
976
+ "blocks.8.attn.k_norm.weight": "BF16",
977
+ "blocks.8.attn.out_proj.weight": "NVFP4",
978
+ "blocks.8.attn.out_proj.weight_scale": "FP32_SCALE",
979
+ "blocks.8.attn.out_proj.comfy_quant": "METADATA",
980
+ "blocks.8.attn.q_norm.weight": "BF16",
981
+ "blocks.8.attn.qkv_proj.weight": "NVFP4",
982
+ "blocks.8.attn.qkv_proj.weight_scale": "FP32_SCALE",
983
+ "blocks.8.attn.qkv_proj.comfy_quant": "METADATA",
984
+ "blocks.8.mlp.fc1.weight": "NVFP4",
985
+ "blocks.8.mlp.fc1.weight_scale": "FP32_SCALE",
986
+ "blocks.8.mlp.fc1.comfy_quant": "METADATA",
987
+ "blocks.8.mlp.fc2.weight": "NVFP4",
988
+ "blocks.8.mlp.fc2.weight_scale": "FP32_SCALE",
989
+ "blocks.8.mlp.fc2.comfy_quant": "METADATA",
990
+ "blocks.8.norm1.weight": "BF16",
991
+ "blocks.8.norm2.weight": "BF16",
992
+ "blocks.9.adaln_proj.linear.bias": "BF16",
993
+ "blocks.9.adaln_proj.linear.weight": "FP8_E4M3FN + SCALE",
994
+ "blocks.9.adaln_proj.linear.weight_scale": "FP32_SCALE",
995
+ "blocks.9.adaln_proj.linear.comfy_quant": "METADATA",
996
+ "blocks.9.attn.k_norm.weight": "BF16",
997
+ "blocks.9.attn.out_proj.weight": "NVFP4",
998
+ "blocks.9.attn.out_proj.weight_scale": "FP32_SCALE",
999
+ "blocks.9.attn.out_proj.comfy_quant": "METADATA",
1000
+ "blocks.9.attn.q_norm.weight": "BF16",
1001
+ "blocks.9.attn.qkv_proj.weight": "NVFP4",
1002
+ "blocks.9.attn.qkv_proj.weight_scale": "FP32_SCALE",
1003
+ "blocks.9.attn.qkv_proj.comfy_quant": "METADATA",
1004
+ "blocks.9.mlp.fc1.weight": "NVFP4",
1005
+ "blocks.9.mlp.fc1.weight_scale": "FP32_SCALE",
1006
+ "blocks.9.mlp.fc1.comfy_quant": "METADATA",
1007
+ "blocks.9.mlp.fc2.weight": "NVFP4",
1008
+ "blocks.9.mlp.fc2.weight_scale": "FP32_SCALE",
1009
+ "blocks.9.mlp.fc2.comfy_quant": "METADATA",
1010
+ "blocks.9.norm1.weight": "BF16",
1011
+ "blocks.9.norm2.weight": "BF16",
1012
+ "condition_proj.bias": "BF16",
1013
+ "condition_proj.weight": "BF16",
1014
+ "final_layer.adaln_proj.linear.bias": "BF16",
1015
+ "final_layer.adaln_proj.linear.weight": "BF16",
1016
+ "final_layer.audio_out.bias": "BF16",
1017
+ "final_layer.audio_out.weight": "BF16",
1018
+ "final_layer.norm.weight": "BF16",
1019
+ "final_layer.video_out.bias": "BF16",
1020
+ "final_layer.video_out.weight": "BF16",
1021
+ "rope.inv_freq": "BF16",
1022
+ "time_embedder.proj_in.bias": "BF16",
1023
+ "time_embedder.proj_in.weight": "BF16",
1024
+ "time_embedder.proj_out.bias": "BF16",
1025
+ "time_embedder.proj_out.weight": "BF16",
1026
+ "token_refiner.blocks.0.attn.k_norm.weight": "BF16",
1027
+ "token_refiner.blocks.0.attn.out_proj.weight": "NVFP4",
1028
+ "token_refiner.blocks.0.attn.out_proj.weight_scale": "FP32_SCALE",
1029
+ "token_refiner.blocks.0.attn.out_proj.comfy_quant": "METADATA",
1030
+ "token_refiner.blocks.0.attn.q_norm.weight": "BF16",
1031
+ "token_refiner.blocks.0.attn.qkv_proj.weight": "NVFP4",
1032
+ "token_refiner.blocks.0.attn.qkv_proj.weight_scale": "FP32_SCALE",
1033
+ "token_refiner.blocks.0.attn.qkv_proj.comfy_quant": "METADATA",
1034
+ "token_refiner.blocks.0.mlp.fc1.weight": "NVFP4",
1035
+ "token_refiner.blocks.0.mlp.fc1.weight_scale": "FP32_SCALE",
1036
+ "token_refiner.blocks.0.mlp.fc1.comfy_quant": "METADATA",
1037
+ "token_refiner.blocks.0.mlp.fc2.weight": "NVFP4",
1038
+ "token_refiner.blocks.0.mlp.fc2.weight_scale": "FP32_SCALE",
1039
+ "token_refiner.blocks.0.mlp.fc2.comfy_quant": "METADATA",
1040
+ "token_refiner.blocks.0.norm1.weight": "BF16",
1041
+ "token_refiner.blocks.0.norm2.weight": "BF16",
1042
+ "token_refiner.blocks.1.attn.k_norm.weight": "BF16",
1043
+ "token_refiner.blocks.1.attn.out_proj.weight": "NVFP4",
1044
+ "token_refiner.blocks.1.attn.out_proj.weight_scale": "FP32_SCALE",
1045
+ "token_refiner.blocks.1.attn.out_proj.comfy_quant": "METADATA",
1046
+ "token_refiner.blocks.1.attn.q_norm.weight": "BF16",
1047
+ "token_refiner.blocks.1.attn.qkv_proj.weight": "NVFP4",
1048
+ "token_refiner.blocks.1.attn.qkv_proj.weight_scale": "FP32_SCALE",
1049
+ "token_refiner.blocks.1.attn.qkv_proj.comfy_quant": "METADATA",
1050
+ "token_refiner.blocks.1.mlp.fc1.weight": "NVFP4",
1051
+ "token_refiner.blocks.1.mlp.fc1.weight_scale": "FP32_SCALE",
1052
+ "token_refiner.blocks.1.mlp.fc1.comfy_quant": "METADATA",
1053
+ "token_refiner.blocks.1.mlp.fc2.weight": "NVFP4",
1054
+ "token_refiner.blocks.1.mlp.fc2.weight_scale": "FP32_SCALE",
1055
+ "token_refiner.blocks.1.mlp.fc2.comfy_quant": "METADATA",
1056
+ "token_refiner.blocks.1.norm1.weight": "BF16",
1057
+ "token_refiner.blocks.1.norm2.weight": "BF16",
1058
+ "token_refiner.final_norm.weight": "BF16",
1059
+ "video_patch_proj.bias": "BF16",
1060
+ "video_patch_proj.weight": "BF16"
1061
+ }
1062
+ }
pruned_to_nvfp4.py ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Requantize MiniMax-H3 pruned_int8_convrot -> NVFP4.
2
+
3
+ The StarNodes converter cannot do this: it passes non-floating-point tensors through
4
+ untouched, so an already-int8 model comes out byte-identical. This walks the state dict
5
+ and does a real dequantize -> requantize on the 200 attn/mlp layers, leaving AdaLN, the
6
+ norms, embedders and token_refiner exactly as they are.
7
+
8
+ Per-layer input: .weight (I8), .weight_scale (F32), .comfy_quant (U8 json bytes)
9
+ Per-layer output: .weight + NVFP4 layout suffixes, .comfy_quant = {"format": "nvfp4"}
10
+ """
11
+ import json, os, sys, time, collections
12
+ import torch
13
+ from safetensors import safe_open
14
+ from safetensors.torch import save_file
15
+ from comfy_kitchen import tensor as ckt
16
+
17
+ SRC = "/ComfyUI/models/diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors"
18
+ DST = "/ComfyUI/models/diffusion_models/minimax_h3_ref2va_pruned_nvfp4.safetensors"
19
+ DEV = "cuda"
20
+ FP8_DTYPES = (torch.float8_e4m3fn, torch.float8_e5m2)
21
+
22
+ f = safe_open(SRC, "pt")
23
+ keys = list(f.keys())
24
+
25
+ # layers that carry a comfy_quant blob are the quantized ones
26
+ quant_bases = sorted({k[: -len(".comfy_quant")] for k in keys if k.endswith(".comfy_quant")})
27
+ print(f"quantized layers found: {len(quant_bases)}", flush=True)
28
+
29
+ out, stats = {}, collections.Counter()
30
+ consumed = set()
31
+ t0 = time.time()
32
+
33
+ for i, base in enumerate(quant_bases):
34
+ wk, sk, ck_ = f"{base}.weight", f"{base}.weight_scale", f"{base}.comfy_quant"
35
+ cfg = json.loads(bytes(f.get_tensor(ck_).tolist()).decode())
36
+ if cfg.get("format") != "int8_tensorwise":
37
+ print(f" SKIP {base}: unexpected format {cfg}", flush=True)
38
+ stats["skipped_unknown_format"] += 1
39
+ continue
40
+
41
+ qdata = f.get_tensor(wk).to(DEV)
42
+ scale = f.get_tensor(sk).to(DEV)
43
+
44
+ params = ckt.TensorWiseINT8Layout.Params(
45
+ scale=scale,
46
+ orig_dtype=torch.bfloat16,
47
+ orig_shape=tuple(qdata.shape),
48
+ is_weight=True,
49
+ convrot=cfg.get("convrot", False),
50
+ convrot_groupsize=cfg.get("convrot_groupsize", 256),
51
+ )
52
+ deq = ckt.TensorWiseINT8Layout.dequantize(qdata, params) # -> bf16
53
+
54
+ nq, nparams = ckt.TensorCoreNVFP4Layout.quantize(deq.float().contiguous())
55
+ tensors = ckt.TensorCoreNVFP4Layout.state_dict_tensors(nq, nparams)
56
+
57
+ for suffix, t in tensors.items():
58
+ name = f"{base}.weight{suffix}"
59
+ if t.dtype == torch.float8_e8m0fnu:
60
+ out[name] = t.view(torch.uint8).cpu()
61
+ elif t.dtype in FP8_DTYPES:
62
+ out[name] = t.view(torch.uint8).cpu().view(t.dtype)
63
+ else:
64
+ out[name] = t.cpu()
65
+
66
+ blob = json.dumps({"format": "nvfp4"}).encode()
67
+ out[ck_] = torch.tensor(list(blob), dtype=torch.uint8)
68
+
69
+ consumed.update({wk, sk, ck_})
70
+ stats["nvfp4"] += 1
71
+ del qdata, scale, deq, nq
72
+ if (i + 1) % 25 == 0:
73
+ torch.cuda.empty_cache()
74
+ print(f" {i+1}/{len(quant_bases)} {time.time()-t0:.0f}s", flush=True)
75
+
76
+ # copy everything untouched
77
+ for k in keys:
78
+ if k in consumed or k in out:
79
+ continue
80
+ out[k] = f.get_tensor(k)
81
+ stats["copied"] += 1
82
+
83
+ print(f"stats: {dict(stats)}", flush=True)
84
+ print(f"writing {DST} ...", flush=True)
85
+ save_file(out, DST)
86
+ gb = os.path.getsize(DST) / 1e9
87
+ print(f"DONE {gb:.2f} GB in {time.time()-t0:.0f}s", flush=True)
88
+
89
+ c = collections.Counter(str(v.dtype) for v in out.values())
90
+ print("dtype census:", dict(c), flush=True)
text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:35a88d51044231fe332301d7a62aa81e3f2cba62febeb446e2c1e3e0ef76f2c6
3
+ size 15687142551
vae/minimax_h3_audio_vae_fp32.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8e505d95dd1561d47abd43d4238fd40d9bb1ae9e147ed0a4cba778d76ae4db48
3
+ size 605254808
vae/minimax_h3_video_vae_fp16.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7c1f131492e7eddacaac9069a61b81bdd39de5cc96561e677c5eab1cdce5e522
3
+ size 5207808496