mold/config.example.toml at main · utensils/mold · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
# mold configuration example
# Default location: ~/.mold/config.toml
#
# Model compatibility notes:
#   ✅ GGUF quantized transformers (Q4, Q8) — fully supported
#   ✅ BF16 / FP16 safetensors — supported, but full-size FLUX (23GB) requires
#      >24GB VRAM due to activation memory. Q8 GGUF (~12GB) is recommended.
#   ❌ FP8 safetensors (F8_E4M3) — requires CUDA FP8 kernels not yet compiled
#      into this candle build. Use a GGUF-quantized version instead.
#   ❌ Scaled FP8 (fp8_scaled) — ComfyUI-only custom format, not supported.

default_model = "flux2-klein"
server_port   = 7680
default_width  = 768
default_height = 768
default_steps  = 4

# ── FLUX.1 Schnell Q8 ─────────────────────────────────────────────────────────
# Fast 4-step model. Great for quick iteration and previews.
[models.flux-schnell]
transformer    = "/path/to/AI/models/unet/flux1-schnell-Q8_0.gguf"
vae            = "/path/to/AI/models/vae/ae.safetensors"
t5_encoder     = "/path/to/AI/models/text_encoders/t5xxl_fp16.safetensors"
clip_encoder   = "/path/to/AI/models/text_encoders/clip_l.safetensors"
t5_tokenizer   = "/path/to/AI/models/tokenizers/t5-v1_1-xxl.tokenizer.json"
clip_tokenizer = "/path/to/AI/models/tokenizers/clip-vit-large-patch14.tokenizer.json"
is_schnell     = true
default_steps  = 4
default_guidance = 0.0
default_width  = 768
default_height = 768
family         = "flux"
description    = "FLUX.1 Schnell Q8 — fast 4-step, general purpose"

# ── FLUX.1 Dev Q8 ─────────────────────────────────────────────────────────────
# High quality, 25-step dev model. Best detail and prompt adherence.
[models.flux-dev]
transformer    = "/path/to/AI/models/unet/flux1-dev-Q8_0.gguf"
vae            = "/path/to/AI/models/vae/ae.safetensors"
t5_encoder     = "/path/to/AI/models/text_encoders/t5xxl_fp16.safetensors"
clip_encoder   = "/path/to/AI/models/text_encoders/clip_l.safetensors"
t5_tokenizer   = "/path/to/AI/models/tokenizers/t5-v1_1-xxl.tokenizer.json"
clip_tokenizer = "/path/to/AI/models/tokenizers/clip-vit-large-patch14.tokenizer.json"
is_schnell     = false
default_steps  = 25
default_guidance = 3.5
default_width  = 768
default_height = 768
family         = "flux"
description    = "FLUX.1 Dev Q8 — full quality, best detail"

# ── FLUX.1 Dev Q4 ─────────────────────────────────────────────────────────────
# Smaller/faster dev model. Good balance of speed and quality.
[models.flux-dev-q4]
transformer    = "/path/to/AI/models/unet/flux1-dev-Q4_1.gguf"
vae            = "/path/to/AI/models/vae/ae.safetensors"
t5_encoder     = "/path/to/AI/models/text_encoders/t5xxl_fp16.safetensors"
clip_encoder   = "/path/to/AI/models/text_encoders/clip_l.safetensors"
t5_tokenizer   = "/path/to/AI/models/tokenizers/t5-v1_1-xxl.tokenizer.json"
clip_tokenizer = "/path/to/AI/models/tokenizers/clip-vit-large-patch14.tokenizer.json"
is_schnell     = false
default_steps  = 20
default_guidance = 3.5
default_width  = 768
default_height = 768
family         = "flux"
description    = "FLUX.1 Dev Q4 — smaller/faster, good balance"

# ── Custom finetune example (GGUF) ────────────────────────────────────────────
# Any FLUX-based finetune in GGUF format can be added here.
# FP8 safetensors finetuned models (e.g. UltraReal v2 fp8, jibMix fp8,
# iniverseMix fp8) are NOT currently supported — convert to GGUF Q8 first.
#
# [models.my-finetune]
# transformer    = "/path/to/AI/models/unet/my-finetune-Q8_0.gguf"
# vae            = "/path/to/AI/models/vae/ae.safetensors"
# t5_encoder     = "/path/to/AI/models/text_encoders/t5xxl_fp16.safetensors"
# clip_encoder   = "/path/to/AI/models/text_encoders/clip_l.safetensors"
# t5_tokenizer   = "/path/to/AI/models/tokenizers/t5-v1_1-xxl.tokenizer.json"
# clip_tokenizer = "/path/to/AI/models/tokenizers/clip-vit-large-patch14.tokenizer.json"
# is_schnell     = false
# default_steps  = 25
# default_guidance = 3.5
# default_width  = 768
# default_height = 768
# family         = "flux"
# description    = "My custom FLUX finetune"