Remove f8 flux, instead configure at load, improved quality & corrected configs
Browse files- float8_quantize.py +18 -18
- modules/conditioner.py +10 -10
- modules/flux_model.py +232 -44
- modules/flux_model_f8.py +0 -491
- util.py +6 -9
float8_quantize.py
CHANGED
@@ -424,28 +424,28 @@ def quantize_flow_transformer_and_dispatch_float8(
|
|
424 |
continue
|
425 |
m_extra.to(device)
|
426 |
m_extra.eval()
|
427 |
-
if (
|
428 |
-
|
429 |
-
and not isinstance(m_extra, (F8Linear, CublasLinear))
|
430 |
-
and quantize_flow_embedder_layers
|
431 |
):
|
432 |
-
|
433 |
-
|
434 |
-
|
435 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
436 |
m_extra,
|
437 |
float8_dtype=float8_dtype,
|
438 |
input_float8_dtype=input_float8_dtype,
|
439 |
-
|
440 |
-
|
441 |
-
del m_extra
|
442 |
-
elif module != "final_layer" and not quantize_flow_embedder_layers:
|
443 |
-
recursive_swap_linears(
|
444 |
-
m_extra,
|
445 |
-
float8_dtype=float8_dtype,
|
446 |
-
input_float8_dtype=input_float8_dtype,
|
447 |
-
quantize_modulation=quantize_modulation,
|
448 |
-
)
|
449 |
torch.cuda.empty_cache()
|
450 |
if swap_linears_with_cublaslinear and flow_dtype == torch.float16:
|
451 |
swap_to_cublaslinear(flow_model)
|
|
|
424 |
continue
|
425 |
m_extra.to(device)
|
426 |
m_extra.eval()
|
427 |
+
if isinstance(m_extra, nn.Linear) and not isinstance(
|
428 |
+
m_extra, (F8Linear, CublasLinear)
|
|
|
|
|
429 |
):
|
430 |
+
if quantize_flow_embedder_layers:
|
431 |
+
setattr(
|
432 |
+
flow_model,
|
433 |
+
module,
|
434 |
+
F8Linear.from_linear(
|
435 |
+
m_extra,
|
436 |
+
float8_dtype=float8_dtype,
|
437 |
+
input_float8_dtype=input_float8_dtype,
|
438 |
+
),
|
439 |
+
)
|
440 |
+
del m_extra
|
441 |
+
elif module != "final_layer":
|
442 |
+
if quantize_flow_embedder_layers:
|
443 |
+
recursive_swap_linears(
|
444 |
m_extra,
|
445 |
float8_dtype=float8_dtype,
|
446 |
input_float8_dtype=input_float8_dtype,
|
447 |
+
quantize_modulation=quantize_modulation,
|
448 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
449 |
torch.cuda.empty_cache()
|
450 |
if swap_linears_with_cublaslinear and flow_dtype == torch.float16:
|
451 |
swap_to_cublaslinear(flow_model)
|
modules/conditioner.py
CHANGED
@@ -56,6 +56,16 @@ class HFEmbedder(nn.Module):
|
|
56 |
self.max_length = max_length
|
57 |
self.output_key = "pooler_output" if self.is_clip else "last_hidden_state"
|
58 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
if self.is_clip:
|
60 |
self.tokenizer: CLIPTokenizer = CLIPTokenizer.from_pretrained(
|
61 |
version, max_length=max_length
|
@@ -64,11 +74,6 @@ class HFEmbedder(nn.Module):
|
|
64 |
self.hf_module: CLIPTextModel = CLIPTextModel.from_pretrained(
|
65 |
version,
|
66 |
**hf_kwargs,
|
67 |
-
quantization_config=(
|
68 |
-
auto_quantization_config(quantization_dtype)
|
69 |
-
if quantization_dtype
|
70 |
-
else None
|
71 |
-
),
|
72 |
)
|
73 |
|
74 |
else:
|
@@ -78,11 +83,6 @@ class HFEmbedder(nn.Module):
|
|
78 |
self.hf_module: T5EncoderModel = T5EncoderModel.from_pretrained(
|
79 |
version,
|
80 |
**hf_kwargs,
|
81 |
-
quantization_config=(
|
82 |
-
auto_quantization_config(quantization_dtype)
|
83 |
-
if quantization_dtype
|
84 |
-
else None
|
85 |
-
),
|
86 |
)
|
87 |
|
88 |
def offload(self):
|
|
|
56 |
self.max_length = max_length
|
57 |
self.output_key = "pooler_output" if self.is_clip else "last_hidden_state"
|
58 |
|
59 |
+
auto_quant_config = (
|
60 |
+
auto_quantization_config(quantization_dtype) if quantization_dtype else None
|
61 |
+
)
|
62 |
+
|
63 |
+
# BNB will move to cuda:0 by default if not specified
|
64 |
+
if isinstance(auto_quant_config, BitsAndBytesConfig):
|
65 |
+
hf_kwargs["device_map"] = {"": self.device.index}
|
66 |
+
if auto_quant_config is not None:
|
67 |
+
hf_kwargs["quantization_config"] = auto_quant_config
|
68 |
+
|
69 |
if self.is_clip:
|
70 |
self.tokenizer: CLIPTokenizer = CLIPTokenizer.from_pretrained(
|
71 |
version, max_length=max_length
|
|
|
74 |
self.hf_module: CLIPTextModel = CLIPTextModel.from_pretrained(
|
75 |
version,
|
76 |
**hf_kwargs,
|
|
|
|
|
|
|
|
|
|
|
77 |
)
|
78 |
|
79 |
else:
|
|
|
83 |
self.hf_module: T5EncoderModel = T5EncoderModel.from_pretrained(
|
84 |
version,
|
85 |
**hf_kwargs,
|
|
|
|
|
|
|
|
|
|
|
86 |
)
|
87 |
|
88 |
def offload(self):
|
modules/flux_model.py
CHANGED
@@ -1,7 +1,11 @@
|
|
1 |
from collections import namedtuple
|
2 |
import os
|
|
|
3 |
import torch
|
4 |
|
|
|
|
|
|
|
5 |
DISABLE_COMPILE = os.getenv("DISABLE_COMPILE", "0") == "1"
|
6 |
torch.backends.cuda.matmul.allow_tf32 = True
|
7 |
torch.backends.cudnn.allow_tf32 = True
|
@@ -111,11 +115,39 @@ def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 10
|
|
111 |
|
112 |
|
113 |
class MLPEmbedder(nn.Module):
|
114 |
-
def __init__(
|
|
|
|
|
|
|
|
|
115 |
super().__init__()
|
116 |
-
self.in_layer =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
117 |
self.silu = nn.SiLU()
|
118 |
-
self.out_layer =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
119 |
|
120 |
def forward(self, x: Tensor) -> Tensor:
|
121 |
return self.out_layer(self.silu(self.in_layer(x)))
|
@@ -143,14 +175,38 @@ class QKNorm(torch.nn.Module):
|
|
143 |
|
144 |
|
145 |
class SelfAttention(nn.Module):
|
146 |
-
def __init__(
|
|
|
|
|
|
|
|
|
|
|
|
|
147 |
super().__init__()
|
|
|
|
|
148 |
self.num_heads = num_heads
|
149 |
head_dim = dim // num_heads
|
150 |
|
151 |
-
self.qkv =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
self.norm = QKNorm(head_dim)
|
153 |
-
self.proj =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
154 |
self.K = 3
|
155 |
self.H = self.num_heads
|
156 |
self.KH = self.K * self.H
|
@@ -173,11 +229,21 @@ ModulationOut = namedtuple("ModulationOut", ["shift", "scale", "gate"])
|
|
173 |
|
174 |
|
175 |
class Modulation(nn.Module):
|
176 |
-
def __init__(self, dim: int, double: bool):
|
177 |
super().__init__()
|
|
|
|
|
178 |
self.is_double = double
|
179 |
self.multiplier = 6 if double else 3
|
180 |
-
self.lin =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
181 |
self.act = nn.SiLU()
|
182 |
|
183 |
def forward(self, vec: Tensor) -> tuple[ModulationOut, ModulationOut | None]:
|
@@ -197,37 +263,83 @@ class DoubleStreamBlock(nn.Module):
|
|
197 |
mlp_ratio: float,
|
198 |
qkv_bias: bool = False,
|
199 |
dtype: torch.dtype = torch.float16,
|
|
|
|
|
200 |
):
|
201 |
super().__init__()
|
|
|
|
|
202 |
self.dtype = dtype
|
203 |
|
204 |
mlp_hidden_dim = int(hidden_size * mlp_ratio)
|
205 |
self.num_heads = num_heads
|
206 |
self.hidden_size = hidden_size
|
207 |
-
self.img_mod = Modulation(
|
|
|
|
|
208 |
self.img_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
209 |
self.img_attn = SelfAttention(
|
210 |
-
dim=hidden_size,
|
|
|
|
|
|
|
211 |
)
|
212 |
|
213 |
self.img_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
214 |
self.img_mlp = nn.Sequential(
|
215 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
216 |
nn.GELU(approximate="tanh"),
|
217 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
218 |
)
|
219 |
|
220 |
-
self.txt_mod = Modulation(
|
|
|
|
|
221 |
self.txt_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
222 |
self.txt_attn = SelfAttention(
|
223 |
-
dim=hidden_size,
|
|
|
|
|
|
|
224 |
)
|
225 |
|
226 |
self.txt_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
227 |
self.txt_mlp = nn.Sequential(
|
228 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
229 |
nn.GELU(approximate="tanh"),
|
230 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
231 |
)
|
232 |
self.K = 3
|
233 |
self.H = self.num_heads
|
@@ -296,8 +408,12 @@ class SingleStreamBlock(nn.Module):
|
|
296 |
mlp_ratio: float = 4.0,
|
297 |
qk_scale: float | None = None,
|
298 |
dtype: torch.dtype = torch.float16,
|
|
|
|
|
299 |
):
|
300 |
super().__init__()
|
|
|
|
|
301 |
self.dtype = dtype
|
302 |
self.hidden_dim = hidden_size
|
303 |
self.num_heads = num_heads
|
@@ -306,9 +422,25 @@ class SingleStreamBlock(nn.Module):
|
|
306 |
|
307 |
self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
|
308 |
# qkv and mlp_in
|
309 |
-
self.linear1 =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
310 |
# proj and mlp_out
|
311 |
-
self.linear2 =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
312 |
|
313 |
self.norm = QKNorm(head_dim)
|
314 |
|
@@ -316,7 +448,11 @@ class SingleStreamBlock(nn.Module):
|
|
316 |
self.pre_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
317 |
|
318 |
self.mlp_act = nn.GELU(approximate="tanh")
|
319 |
-
self.modulation = Modulation(
|
|
|
|
|
|
|
|
|
320 |
|
321 |
self.K = 3
|
322 |
self.H = self.num_heads
|
@@ -364,50 +500,96 @@ class Flux(nn.Module):
|
|
364 |
Transformer model for flow matching on sequences.
|
365 |
"""
|
366 |
|
367 |
-
def __init__(self,
|
368 |
super().__init__()
|
369 |
|
370 |
self.dtype = dtype
|
371 |
-
self.params = params
|
372 |
-
self.in_channels = params.in_channels
|
373 |
self.out_channels = self.in_channels
|
374 |
-
|
|
|
|
|
|
|
|
|
|
|
375 |
raise ValueError(
|
376 |
-
f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}"
|
377 |
)
|
378 |
-
pe_dim = params.hidden_size // params.num_heads
|
379 |
-
if sum(params.axes_dim) != pe_dim:
|
380 |
raise ValueError(
|
381 |
-
f"Got {params.axes_dim} but expected positional dim {pe_dim}"
|
382 |
)
|
383 |
-
self.hidden_size = params.hidden_size
|
384 |
-
self.num_heads = params.num_heads
|
385 |
self.pe_embedder = EmbedND(
|
386 |
dim=pe_dim,
|
387 |
-
theta=params.theta,
|
388 |
-
axes_dim=params.axes_dim,
|
389 |
dtype=self.dtype,
|
390 |
)
|
391 |
-
self.img_in =
|
392 |
-
|
393 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
394 |
self.guidance_in = (
|
395 |
-
MLPEmbedder(
|
396 |
-
|
|
|
|
|
|
|
|
|
|
|
397 |
else nn.Identity()
|
398 |
)
|
399 |
-
self.txt_in =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
400 |
|
401 |
self.double_blocks = nn.ModuleList(
|
402 |
[
|
403 |
DoubleStreamBlock(
|
404 |
self.hidden_size,
|
405 |
self.num_heads,
|
406 |
-
mlp_ratio=params.mlp_ratio,
|
407 |
-
qkv_bias=params.qkv_bias,
|
408 |
dtype=self.dtype,
|
|
|
|
|
409 |
)
|
410 |
-
for _ in range(params.depth)
|
411 |
]
|
412 |
)
|
413 |
|
@@ -416,10 +598,12 @@ class Flux(nn.Module):
|
|
416 |
SingleStreamBlock(
|
417 |
self.hidden_size,
|
418 |
self.num_heads,
|
419 |
-
mlp_ratio=params.mlp_ratio,
|
420 |
dtype=self.dtype,
|
|
|
|
|
421 |
)
|
422 |
-
for _ in range(params.depth_single_blocks)
|
423 |
]
|
424 |
)
|
425 |
|
@@ -472,13 +656,17 @@ class Flux(nn.Module):
|
|
472 |
return img
|
473 |
|
474 |
@classmethod
|
475 |
-
def from_pretrained(
|
|
|
|
|
476 |
from util import load_config_from_path
|
477 |
from safetensors.torch import load_file
|
478 |
|
479 |
config = load_config_from_path(path)
|
480 |
with torch.device("meta"):
|
481 |
-
klass = cls(
|
|
|
|
|
482 |
|
483 |
ckpt = load_file(config.ckpt_path, device="cpu")
|
484 |
klass.load_state_dict(ckpt, assign=True)
|
|
|
1 |
from collections import namedtuple
|
2 |
import os
|
3 |
+
from typing import TYPE_CHECKING
|
4 |
import torch
|
5 |
|
6 |
+
if TYPE_CHECKING:
|
7 |
+
from util import ModelSpec
|
8 |
+
|
9 |
DISABLE_COMPILE = os.getenv("DISABLE_COMPILE", "0") == "1"
|
10 |
torch.backends.cuda.matmul.allow_tf32 = True
|
11 |
torch.backends.cudnn.allow_tf32 = True
|
|
|
115 |
|
116 |
|
117 |
class MLPEmbedder(nn.Module):
|
118 |
+
def __init__(
|
119 |
+
self, in_dim: int, hidden_dim: int, prequantized: bool = False, quantized=False
|
120 |
+
):
|
121 |
+
from float8_quantize import F8Linear
|
122 |
+
|
123 |
super().__init__()
|
124 |
+
self.in_layer = (
|
125 |
+
nn.Linear(in_dim, hidden_dim, bias=True)
|
126 |
+
if not prequantized
|
127 |
+
else (
|
128 |
+
F8Linear(
|
129 |
+
in_features=in_dim,
|
130 |
+
out_features=hidden_dim,
|
131 |
+
bias=True,
|
132 |
+
)
|
133 |
+
if quantized
|
134 |
+
else nn.Linear(in_dim, hidden_dim, bias=True)
|
135 |
+
)
|
136 |
+
)
|
137 |
self.silu = nn.SiLU()
|
138 |
+
self.out_layer = (
|
139 |
+
nn.Linear(hidden_dim, hidden_dim, bias=True)
|
140 |
+
if not prequantized
|
141 |
+
else (
|
142 |
+
F8Linear(
|
143 |
+
in_features=hidden_dim,
|
144 |
+
out_features=hidden_dim,
|
145 |
+
bias=True,
|
146 |
+
)
|
147 |
+
if quantized
|
148 |
+
else nn.Linear(hidden_dim, hidden_dim, bias=True)
|
149 |
+
)
|
150 |
+
)
|
151 |
|
152 |
def forward(self, x: Tensor) -> Tensor:
|
153 |
return self.out_layer(self.silu(self.in_layer(x)))
|
|
|
175 |
|
176 |
|
177 |
class SelfAttention(nn.Module):
|
178 |
+
def __init__(
|
179 |
+
self,
|
180 |
+
dim: int,
|
181 |
+
num_heads: int = 8,
|
182 |
+
qkv_bias: bool = False,
|
183 |
+
prequantized: bool = False,
|
184 |
+
):
|
185 |
super().__init__()
|
186 |
+
from float8_quantize import F8Linear
|
187 |
+
|
188 |
self.num_heads = num_heads
|
189 |
head_dim = dim // num_heads
|
190 |
|
191 |
+
self.qkv = (
|
192 |
+
nn.Linear(dim, dim * 3, bias=qkv_bias)
|
193 |
+
if not prequantized
|
194 |
+
else F8Linear(
|
195 |
+
in_features=dim,
|
196 |
+
out_features=dim * 3,
|
197 |
+
bias=qkv_bias,
|
198 |
+
)
|
199 |
+
)
|
200 |
self.norm = QKNorm(head_dim)
|
201 |
+
self.proj = (
|
202 |
+
nn.Linear(dim, dim)
|
203 |
+
if not prequantized
|
204 |
+
else F8Linear(
|
205 |
+
in_features=dim,
|
206 |
+
out_features=dim,
|
207 |
+
bias=True,
|
208 |
+
)
|
209 |
+
)
|
210 |
self.K = 3
|
211 |
self.H = self.num_heads
|
212 |
self.KH = self.K * self.H
|
|
|
229 |
|
230 |
|
231 |
class Modulation(nn.Module):
|
232 |
+
def __init__(self, dim: int, double: bool, quantized_modulation: bool = False):
|
233 |
super().__init__()
|
234 |
+
from float8_quantize import F8Linear
|
235 |
+
|
236 |
self.is_double = double
|
237 |
self.multiplier = 6 if double else 3
|
238 |
+
self.lin = (
|
239 |
+
nn.Linear(dim, self.multiplier * dim, bias=True)
|
240 |
+
if not quantized_modulation
|
241 |
+
else F8Linear(
|
242 |
+
in_features=dim,
|
243 |
+
out_features=self.multiplier * dim,
|
244 |
+
bias=True,
|
245 |
+
)
|
246 |
+
)
|
247 |
self.act = nn.SiLU()
|
248 |
|
249 |
def forward(self, vec: Tensor) -> tuple[ModulationOut, ModulationOut | None]:
|
|
|
263 |
mlp_ratio: float,
|
264 |
qkv_bias: bool = False,
|
265 |
dtype: torch.dtype = torch.float16,
|
266 |
+
quantized_modulation: bool = False,
|
267 |
+
prequantized: bool = False,
|
268 |
):
|
269 |
super().__init__()
|
270 |
+
from float8_quantize import F8Linear
|
271 |
+
|
272 |
self.dtype = dtype
|
273 |
|
274 |
mlp_hidden_dim = int(hidden_size * mlp_ratio)
|
275 |
self.num_heads = num_heads
|
276 |
self.hidden_size = hidden_size
|
277 |
+
self.img_mod = Modulation(
|
278 |
+
hidden_size, double=True, quantized_modulation=quantized_modulation
|
279 |
+
)
|
280 |
self.img_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
281 |
self.img_attn = SelfAttention(
|
282 |
+
dim=hidden_size,
|
283 |
+
num_heads=num_heads,
|
284 |
+
qkv_bias=qkv_bias,
|
285 |
+
prequantized=prequantized,
|
286 |
)
|
287 |
|
288 |
self.img_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
289 |
self.img_mlp = nn.Sequential(
|
290 |
+
(
|
291 |
+
nn.Linear(hidden_size, mlp_hidden_dim, bias=True)
|
292 |
+
if not prequantized
|
293 |
+
else F8Linear(
|
294 |
+
in_features=hidden_size,
|
295 |
+
out_features=mlp_hidden_dim,
|
296 |
+
bias=True,
|
297 |
+
)
|
298 |
+
),
|
299 |
nn.GELU(approximate="tanh"),
|
300 |
+
(
|
301 |
+
nn.Linear(mlp_hidden_dim, hidden_size, bias=True)
|
302 |
+
if not prequantized
|
303 |
+
else F8Linear(
|
304 |
+
in_features=mlp_hidden_dim,
|
305 |
+
out_features=hidden_size,
|
306 |
+
bias=True,
|
307 |
+
)
|
308 |
+
),
|
309 |
)
|
310 |
|
311 |
+
self.txt_mod = Modulation(
|
312 |
+
hidden_size, double=True, quantized_modulation=quantized_modulation
|
313 |
+
)
|
314 |
self.txt_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
315 |
self.txt_attn = SelfAttention(
|
316 |
+
dim=hidden_size,
|
317 |
+
num_heads=num_heads,
|
318 |
+
qkv_bias=qkv_bias,
|
319 |
+
prequantized=prequantized,
|
320 |
)
|
321 |
|
322 |
self.txt_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
323 |
self.txt_mlp = nn.Sequential(
|
324 |
+
(
|
325 |
+
nn.Linear(hidden_size, mlp_hidden_dim, bias=True)
|
326 |
+
if not prequantized
|
327 |
+
else F8Linear(
|
328 |
+
in_features=hidden_size,
|
329 |
+
out_features=mlp_hidden_dim,
|
330 |
+
bias=True,
|
331 |
+
)
|
332 |
+
),
|
333 |
nn.GELU(approximate="tanh"),
|
334 |
+
(
|
335 |
+
nn.Linear(mlp_hidden_dim, hidden_size, bias=True)
|
336 |
+
if not prequantized
|
337 |
+
else F8Linear(
|
338 |
+
in_features=mlp_hidden_dim,
|
339 |
+
out_features=hidden_size,
|
340 |
+
bias=True,
|
341 |
+
)
|
342 |
+
),
|
343 |
)
|
344 |
self.K = 3
|
345 |
self.H = self.num_heads
|
|
|
408 |
mlp_ratio: float = 4.0,
|
409 |
qk_scale: float | None = None,
|
410 |
dtype: torch.dtype = torch.float16,
|
411 |
+
quantized_modulation: bool = False,
|
412 |
+
prequantized: bool = False,
|
413 |
):
|
414 |
super().__init__()
|
415 |
+
from float8_quantize import F8Linear
|
416 |
+
|
417 |
self.dtype = dtype
|
418 |
self.hidden_dim = hidden_size
|
419 |
self.num_heads = num_heads
|
|
|
422 |
|
423 |
self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
|
424 |
# qkv and mlp_in
|
425 |
+
self.linear1 = (
|
426 |
+
nn.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim)
|
427 |
+
if not prequantized
|
428 |
+
else F8Linear(
|
429 |
+
in_features=hidden_size,
|
430 |
+
out_features=hidden_size * 3 + self.mlp_hidden_dim,
|
431 |
+
bias=True,
|
432 |
+
)
|
433 |
+
)
|
434 |
# proj and mlp_out
|
435 |
+
self.linear2 = (
|
436 |
+
nn.Linear(hidden_size + self.mlp_hidden_dim, hidden_size)
|
437 |
+
if not prequantized
|
438 |
+
else F8Linear(
|
439 |
+
in_features=hidden_size + self.mlp_hidden_dim,
|
440 |
+
out_features=hidden_size,
|
441 |
+
bias=True,
|
442 |
+
)
|
443 |
+
)
|
444 |
|
445 |
self.norm = QKNorm(head_dim)
|
446 |
|
|
|
448 |
self.pre_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
449 |
|
450 |
self.mlp_act = nn.GELU(approximate="tanh")
|
451 |
+
self.modulation = Modulation(
|
452 |
+
hidden_size,
|
453 |
+
double=False,
|
454 |
+
quantized_modulation=quantized_modulation and prequantized,
|
455 |
+
)
|
456 |
|
457 |
self.K = 3
|
458 |
self.H = self.num_heads
|
|
|
500 |
Transformer model for flow matching on sequences.
|
501 |
"""
|
502 |
|
503 |
+
def __init__(self, config: "ModelSpec", dtype: torch.dtype = torch.float16):
|
504 |
super().__init__()
|
505 |
|
506 |
self.dtype = dtype
|
507 |
+
self.params = config.params
|
508 |
+
self.in_channels = config.params.in_channels
|
509 |
self.out_channels = self.in_channels
|
510 |
+
prequantized_flow = config.prequantized_flow
|
511 |
+
quantized_embedders = config.quantize_flow_embedder_layers and prequantized_flow
|
512 |
+
quantized_modulation = config.quantize_modulation and prequantized_flow
|
513 |
+
from float8_quantize import F8Linear
|
514 |
+
|
515 |
+
if config.params.hidden_size % config.params.num_heads != 0:
|
516 |
raise ValueError(
|
517 |
+
f"Hidden size {config.params.hidden_size} must be divisible by num_heads {config.params.num_heads}"
|
518 |
)
|
519 |
+
pe_dim = config.params.hidden_size // config.params.num_heads
|
520 |
+
if sum(config.params.axes_dim) != pe_dim:
|
521 |
raise ValueError(
|
522 |
+
f"Got {config.params.axes_dim} but expected positional dim {pe_dim}"
|
523 |
)
|
524 |
+
self.hidden_size = config.params.hidden_size
|
525 |
+
self.num_heads = config.params.num_heads
|
526 |
self.pe_embedder = EmbedND(
|
527 |
dim=pe_dim,
|
528 |
+
theta=config.params.theta,
|
529 |
+
axes_dim=config.params.axes_dim,
|
530 |
dtype=self.dtype,
|
531 |
)
|
532 |
+
self.img_in = (
|
533 |
+
nn.Linear(self.in_channels, self.hidden_size, bias=True)
|
534 |
+
if not prequantized_flow
|
535 |
+
else (
|
536 |
+
F8Linear(
|
537 |
+
in_features=self.in_channels,
|
538 |
+
out_features=self.hidden_size,
|
539 |
+
bias=True,
|
540 |
+
)
|
541 |
+
if quantized_embedders
|
542 |
+
else nn.Linear(self.in_channels, self.hidden_size, bias=True)
|
543 |
+
)
|
544 |
+
)
|
545 |
+
self.time_in = MLPEmbedder(
|
546 |
+
in_dim=256,
|
547 |
+
hidden_dim=self.hidden_size,
|
548 |
+
prequantized=prequantized_flow,
|
549 |
+
quantized=quantized_embedders,
|
550 |
+
)
|
551 |
+
self.vector_in = MLPEmbedder(
|
552 |
+
config.params.vec_in_dim,
|
553 |
+
self.hidden_size,
|
554 |
+
prequantized=prequantized_flow,
|
555 |
+
quantized=quantized_embedders,
|
556 |
+
)
|
557 |
self.guidance_in = (
|
558 |
+
MLPEmbedder(
|
559 |
+
in_dim=256,
|
560 |
+
hidden_dim=self.hidden_size,
|
561 |
+
prequantized=prequantized_flow,
|
562 |
+
quantized=quantized_embedders,
|
563 |
+
)
|
564 |
+
if config.params.guidance_embed
|
565 |
else nn.Identity()
|
566 |
)
|
567 |
+
self.txt_in = (
|
568 |
+
nn.Linear(config.params.context_in_dim, self.hidden_size)
|
569 |
+
if not quantized_embedders
|
570 |
+
else (
|
571 |
+
F8Linear(
|
572 |
+
in_features=config.params.context_in_dim,
|
573 |
+
out_features=self.hidden_size,
|
574 |
+
bias=True,
|
575 |
+
)
|
576 |
+
if quantized_embedders
|
577 |
+
else nn.Linear(config.params.context_in_dim, self.hidden_size)
|
578 |
+
)
|
579 |
+
)
|
580 |
|
581 |
self.double_blocks = nn.ModuleList(
|
582 |
[
|
583 |
DoubleStreamBlock(
|
584 |
self.hidden_size,
|
585 |
self.num_heads,
|
586 |
+
mlp_ratio=config.params.mlp_ratio,
|
587 |
+
qkv_bias=config.params.qkv_bias,
|
588 |
dtype=self.dtype,
|
589 |
+
quantized_modulation=quantized_modulation,
|
590 |
+
prequantized=prequantized_flow,
|
591 |
)
|
592 |
+
for _ in range(config.params.depth)
|
593 |
]
|
594 |
)
|
595 |
|
|
|
598 |
SingleStreamBlock(
|
599 |
self.hidden_size,
|
600 |
self.num_heads,
|
601 |
+
mlp_ratio=config.params.mlp_ratio,
|
602 |
dtype=self.dtype,
|
603 |
+
quantized_modulation=quantized_modulation,
|
604 |
+
prequantized=prequantized_flow,
|
605 |
)
|
606 |
+
for _ in range(config.params.depth_single_blocks)
|
607 |
]
|
608 |
)
|
609 |
|
|
|
656 |
return img
|
657 |
|
658 |
@classmethod
|
659 |
+
def from_pretrained(
|
660 |
+
cls: "Flux", path: str, dtype: torch.dtype = torch.float16
|
661 |
+
) -> "Flux":
|
662 |
from util import load_config_from_path
|
663 |
from safetensors.torch import load_file
|
664 |
|
665 |
config = load_config_from_path(path)
|
666 |
with torch.device("meta"):
|
667 |
+
klass = cls(config=config, dtype=dtype)
|
668 |
+
if not config.prequantized_flow:
|
669 |
+
klass.type(dtype)
|
670 |
|
671 |
ckpt = load_file(config.ckpt_path, device="cpu")
|
672 |
klass.load_state_dict(ckpt, assign=True)
|
modules/flux_model_f8.py
DELETED
@@ -1,491 +0,0 @@
|
|
1 |
-
from collections import namedtuple
|
2 |
-
import os
|
3 |
-
import torch
|
4 |
-
|
5 |
-
DISABLE_COMPILE = os.getenv("DISABLE_COMPILE", "0") == "1"
|
6 |
-
torch.backends.cuda.matmul.allow_tf32 = True
|
7 |
-
torch.backends.cudnn.allow_tf32 = True
|
8 |
-
torch.backends.cudnn.benchmark = True
|
9 |
-
torch.backends.cudnn.benchmark_limit = 20
|
10 |
-
torch.set_float32_matmul_precision("high")
|
11 |
-
import math
|
12 |
-
|
13 |
-
from torch import Tensor, nn
|
14 |
-
from pydantic import BaseModel
|
15 |
-
from torch.nn import functional as F
|
16 |
-
from float8_quantize import F8Linear
|
17 |
-
|
18 |
-
try:
|
19 |
-
from cublas_ops import CublasLinear
|
20 |
-
except ImportError:
|
21 |
-
CublasLinear = nn.Linear
|
22 |
-
|
23 |
-
|
24 |
-
class FluxParams(BaseModel):
|
25 |
-
in_channels: int
|
26 |
-
vec_in_dim: int
|
27 |
-
context_in_dim: int
|
28 |
-
hidden_size: int
|
29 |
-
mlp_ratio: float
|
30 |
-
num_heads: int
|
31 |
-
depth: int
|
32 |
-
depth_single_blocks: int
|
33 |
-
axes_dim: list[int]
|
34 |
-
theta: int
|
35 |
-
qkv_bias: bool
|
36 |
-
guidance_embed: bool
|
37 |
-
|
38 |
-
|
39 |
-
# attention is always same shape each time it's called per H*W, so compile with fullgraph
|
40 |
-
# @torch.compile(mode="reduce-overhead", fullgraph=True, disable=DISABLE_COMPILE)
|
41 |
-
def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor) -> Tensor:
|
42 |
-
q, k = apply_rope(q, k, pe)
|
43 |
-
x = F.scaled_dot_product_attention(q, k, v).transpose(1, 2)
|
44 |
-
x = x.reshape(*x.shape[:-2], -1)
|
45 |
-
return x
|
46 |
-
|
47 |
-
|
48 |
-
# @torch.compile(mode="reduce-overhead", disable=DISABLE_COMPILE)
|
49 |
-
def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
|
50 |
-
scale = torch.arange(0, dim, 2, dtype=torch.float32, device=pos.device) / dim
|
51 |
-
omega = 1.0 / (theta**scale)
|
52 |
-
out = torch.einsum("...n,d->...nd", pos, omega)
|
53 |
-
out = torch.stack(
|
54 |
-
[torch.cos(out), -torch.sin(out), torch.sin(out), torch.cos(out)], dim=-1
|
55 |
-
)
|
56 |
-
out = out.reshape(*out.shape[:-1], 2, 2)
|
57 |
-
return out
|
58 |
-
|
59 |
-
|
60 |
-
def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor) -> tuple[Tensor, Tensor]:
|
61 |
-
xq_ = xq.reshape(*xq.shape[:-1], -1, 1, 2)
|
62 |
-
xk_ = xk.reshape(*xk.shape[:-1], -1, 1, 2)
|
63 |
-
xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
|
64 |
-
xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
|
65 |
-
return xq_out.reshape(*xq.shape), xk_out.reshape(*xk.shape)
|
66 |
-
|
67 |
-
|
68 |
-
class EmbedND(nn.Module):
|
69 |
-
def __init__(
|
70 |
-
self,
|
71 |
-
dim: int,
|
72 |
-
theta: int,
|
73 |
-
axes_dim: list[int],
|
74 |
-
dtype: torch.dtype = torch.bfloat16,
|
75 |
-
):
|
76 |
-
super().__init__()
|
77 |
-
self.dim = dim
|
78 |
-
self.theta = theta
|
79 |
-
self.axes_dim = axes_dim
|
80 |
-
self.dtype = dtype
|
81 |
-
|
82 |
-
def forward(self, ids: Tensor) -> Tensor:
|
83 |
-
n_axes = ids.shape[-1]
|
84 |
-
emb = torch.cat(
|
85 |
-
[
|
86 |
-
rope(ids[..., i], self.axes_dim[i], self.theta).type(self.dtype)
|
87 |
-
for i in range(n_axes)
|
88 |
-
],
|
89 |
-
dim=-3,
|
90 |
-
)
|
91 |
-
|
92 |
-
return emb.unsqueeze(1)
|
93 |
-
|
94 |
-
|
95 |
-
def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 1000.0):
|
96 |
-
"""
|
97 |
-
Create sinusoidal timestep embeddings.
|
98 |
-
:param t: a 1-D Tensor of N indices, one per batch element.
|
99 |
-
These may be fractional.
|
100 |
-
:param dim: the dimension of the output.
|
101 |
-
:param max_period: controls the minimum frequency of the embeddings.
|
102 |
-
:return: an (N, D) Tensor of positional embeddings.
|
103 |
-
"""
|
104 |
-
t = time_factor * t
|
105 |
-
half = dim // 2
|
106 |
-
freqs = torch.exp(
|
107 |
-
-math.log(max_period)
|
108 |
-
* torch.arange(start=0, end=half, dtype=torch.float32, device=t.device)
|
109 |
-
/ half
|
110 |
-
)
|
111 |
-
|
112 |
-
args = t[:, None].float() * freqs[None]
|
113 |
-
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
114 |
-
if dim % 2:
|
115 |
-
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
|
116 |
-
return embedding
|
117 |
-
|
118 |
-
|
119 |
-
class MLPEmbedder(nn.Module):
|
120 |
-
def __init__(self, in_dim: int, hidden_dim: int):
|
121 |
-
super().__init__()
|
122 |
-
self.in_layer = F8Linear(in_dim, hidden_dim, bias=True)
|
123 |
-
self.silu = nn.SiLU()
|
124 |
-
self.out_layer = F8Linear(hidden_dim, hidden_dim, bias=True)
|
125 |
-
|
126 |
-
def forward(self, x: Tensor) -> Tensor:
|
127 |
-
return self.out_layer(self.silu(self.in_layer(x)))
|
128 |
-
|
129 |
-
|
130 |
-
class RMSNorm(torch.nn.Module):
|
131 |
-
def __init__(self, dim: int):
|
132 |
-
super().__init__()
|
133 |
-
self.scale = nn.Parameter(torch.ones(dim))
|
134 |
-
|
135 |
-
def forward(self, x: Tensor):
|
136 |
-
return F.rms_norm(x, self.scale.shape, self.scale, eps=1e-6)
|
137 |
-
|
138 |
-
|
139 |
-
class QKNorm(torch.nn.Module):
|
140 |
-
def __init__(self, dim: int):
|
141 |
-
super().__init__()
|
142 |
-
self.query_norm = RMSNorm(dim)
|
143 |
-
self.key_norm = RMSNorm(dim)
|
144 |
-
|
145 |
-
def forward(self, q: Tensor, k: Tensor, v: Tensor) -> tuple[Tensor, Tensor]:
|
146 |
-
q = self.query_norm(q)
|
147 |
-
k = self.key_norm(k)
|
148 |
-
return q, k
|
149 |
-
|
150 |
-
|
151 |
-
class SelfAttention(nn.Module):
|
152 |
-
def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False):
|
153 |
-
super().__init__()
|
154 |
-
self.num_heads = num_heads
|
155 |
-
head_dim = dim // num_heads
|
156 |
-
|
157 |
-
self.qkv = F8Linear(dim, dim * 3, bias=qkv_bias)
|
158 |
-
self.norm = QKNorm(head_dim)
|
159 |
-
self.proj = F8Linear(dim, dim)
|
160 |
-
self.K = 3
|
161 |
-
self.H = self.num_heads
|
162 |
-
self.KH = self.K * self.H
|
163 |
-
|
164 |
-
def rearrange_for_norm(self, x: Tensor) -> tuple[Tensor, Tensor, Tensor]:
|
165 |
-
B, L, D = x.shape
|
166 |
-
q, k, v = x.reshape(B, L, self.K, self.H, D // self.KH).permute(2, 0, 3, 1, 4)
|
167 |
-
return q, k, v
|
168 |
-
|
169 |
-
def forward(self, x: Tensor, pe: Tensor) -> Tensor:
|
170 |
-
qkv = self.qkv(x)
|
171 |
-
q, k, v = self.rearrange_for_norm(qkv)
|
172 |
-
q, k = self.norm(q, k, v)
|
173 |
-
x = attention(q, k, v, pe=pe)
|
174 |
-
x = self.proj(x)
|
175 |
-
return x
|
176 |
-
|
177 |
-
|
178 |
-
ModulationOut = namedtuple("ModulationOut", ["shift", "scale", "gate"])
|
179 |
-
|
180 |
-
|
181 |
-
class Modulation(nn.Module):
|
182 |
-
def __init__(self, dim: int, double: bool):
|
183 |
-
super().__init__()
|
184 |
-
self.is_double = double
|
185 |
-
self.multiplier = 6 if double else 3
|
186 |
-
self.lin = F8Linear(dim, self.multiplier * dim, bias=True)
|
187 |
-
self.act = nn.SiLU()
|
188 |
-
|
189 |
-
def forward(self, vec: Tensor) -> tuple[ModulationOut, ModulationOut | None]:
|
190 |
-
out = self.lin(self.act(vec))[:, None, :].chunk(self.multiplier, dim=-1)
|
191 |
-
|
192 |
-
return (
|
193 |
-
ModulationOut(*out[:3]),
|
194 |
-
ModulationOut(*out[3:]) if self.is_double else None,
|
195 |
-
)
|
196 |
-
|
197 |
-
|
198 |
-
class DoubleStreamBlock(nn.Module):
|
199 |
-
def __init__(
|
200 |
-
self,
|
201 |
-
hidden_size: int,
|
202 |
-
num_heads: int,
|
203 |
-
mlp_ratio: float,
|
204 |
-
qkv_bias: bool = False,
|
205 |
-
dtype: torch.dtype = torch.float16,
|
206 |
-
):
|
207 |
-
super().__init__()
|
208 |
-
self.dtype = dtype
|
209 |
-
|
210 |
-
mlp_hidden_dim = int(hidden_size * mlp_ratio)
|
211 |
-
self.num_heads = num_heads
|
212 |
-
self.hidden_size = hidden_size
|
213 |
-
self.img_mod = Modulation(hidden_size, double=True)
|
214 |
-
self.img_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
215 |
-
self.img_attn = SelfAttention(
|
216 |
-
dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias
|
217 |
-
)
|
218 |
-
|
219 |
-
self.img_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
220 |
-
self.img_mlp = nn.Sequential(
|
221 |
-
F8Linear(hidden_size, mlp_hidden_dim, bias=True),
|
222 |
-
nn.GELU(approximate="tanh"),
|
223 |
-
F8Linear(mlp_hidden_dim, hidden_size, bias=True),
|
224 |
-
)
|
225 |
-
|
226 |
-
self.txt_mod = Modulation(hidden_size, double=True)
|
227 |
-
self.txt_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
228 |
-
self.txt_attn = SelfAttention(
|
229 |
-
dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias
|
230 |
-
)
|
231 |
-
|
232 |
-
self.txt_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
233 |
-
self.txt_mlp = nn.Sequential(
|
234 |
-
F8Linear(hidden_size, mlp_hidden_dim, bias=True),
|
235 |
-
nn.GELU(approximate="tanh"),
|
236 |
-
F8Linear(mlp_hidden_dim, hidden_size, bias=True),
|
237 |
-
)
|
238 |
-
self.K = 3
|
239 |
-
self.H = self.num_heads
|
240 |
-
self.KH = self.K * self.H
|
241 |
-
|
242 |
-
def rearrange_for_norm(self, x: Tensor) -> tuple[Tensor, Tensor, Tensor]:
|
243 |
-
B, L, D = x.shape
|
244 |
-
q, k, v = x.reshape(B, L, self.K, self.H, D // self.KH).permute(2, 0, 3, 1, 4)
|
245 |
-
return q, k, v
|
246 |
-
|
247 |
-
def forward(
|
248 |
-
self,
|
249 |
-
img: Tensor,
|
250 |
-
txt: Tensor,
|
251 |
-
vec: Tensor,
|
252 |
-
pe: Tensor,
|
253 |
-
) -> tuple[Tensor, Tensor]:
|
254 |
-
img_mod1, img_mod2 = self.img_mod(vec)
|
255 |
-
txt_mod1, txt_mod2 = self.txt_mod(vec)
|
256 |
-
|
257 |
-
# prepare image for attention
|
258 |
-
img_modulated = self.img_norm1(img)
|
259 |
-
img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
|
260 |
-
img_qkv = self.img_attn.qkv(img_modulated)
|
261 |
-
img_q, img_k, img_v = self.rearrange_for_norm(img_qkv)
|
262 |
-
img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
|
263 |
-
|
264 |
-
# prepare txt for attention
|
265 |
-
txt_modulated = self.txt_norm1(txt)
|
266 |
-
txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
|
267 |
-
txt_qkv = self.txt_attn.qkv(txt_modulated)
|
268 |
-
txt_q, txt_k, txt_v = self.rearrange_for_norm(txt_qkv)
|
269 |
-
txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
|
270 |
-
|
271 |
-
q = torch.cat((txt_q, img_q), dim=2)
|
272 |
-
k = torch.cat((txt_k, img_k), dim=2)
|
273 |
-
v = torch.cat((txt_v, img_v), dim=2)
|
274 |
-
|
275 |
-
attn = attention(q, k, v, pe=pe)
|
276 |
-
txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
|
277 |
-
# calculate the img bloks
|
278 |
-
img = img + img_mod1.gate * self.img_attn.proj(img_attn)
|
279 |
-
img = img + img_mod2.gate * self.img_mlp(
|
280 |
-
(1 + img_mod2.scale) * self.img_norm2(img) + img_mod2.shift
|
281 |
-
).clamp(min=-384 * 2, max=384 * 2)
|
282 |
-
|
283 |
-
# calculate the txt bloks
|
284 |
-
txt = txt + txt_mod1.gate * self.txt_attn.proj(txt_attn)
|
285 |
-
txt = txt + txt_mod2.gate * self.txt_mlp(
|
286 |
-
(1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift
|
287 |
-
).clamp(min=-384 * 2, max=384 * 2)
|
288 |
-
|
289 |
-
return img, txt
|
290 |
-
|
291 |
-
|
292 |
-
class SingleStreamBlock(nn.Module):
|
293 |
-
"""
|
294 |
-
A DiT block with parallel linear layers as described in
|
295 |
-
https://arxiv.org/abs/2302.05442 and adapted modulation interface.
|
296 |
-
"""
|
297 |
-
|
298 |
-
def __init__(
|
299 |
-
self,
|
300 |
-
hidden_size: int,
|
301 |
-
num_heads: int,
|
302 |
-
mlp_ratio: float = 4.0,
|
303 |
-
qk_scale: float | None = None,
|
304 |
-
dtype: torch.dtype = torch.float16,
|
305 |
-
):
|
306 |
-
super().__init__()
|
307 |
-
self.dtype = dtype
|
308 |
-
self.hidden_dim = hidden_size
|
309 |
-
self.num_heads = num_heads
|
310 |
-
head_dim = hidden_size // num_heads
|
311 |
-
self.scale = qk_scale or head_dim**-0.5
|
312 |
-
|
313 |
-
self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
|
314 |
-
# qkv and mlp_in
|
315 |
-
self.linear1 = F8Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim)
|
316 |
-
# proj and mlp_out
|
317 |
-
self.linear2 = F8Linear(hidden_size + self.mlp_hidden_dim, hidden_size)
|
318 |
-
|
319 |
-
self.norm = QKNorm(head_dim)
|
320 |
-
|
321 |
-
self.hidden_size = hidden_size
|
322 |
-
self.pre_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
323 |
-
|
324 |
-
self.mlp_act = nn.GELU(approximate="tanh")
|
325 |
-
self.modulation = Modulation(hidden_size, double=False)
|
326 |
-
|
327 |
-
self.K = 3
|
328 |
-
self.H = self.num_heads
|
329 |
-
self.KH = self.K * self.H
|
330 |
-
|
331 |
-
def forward(self, x: Tensor, vec: Tensor, pe: Tensor) -> Tensor:
|
332 |
-
mod = self.modulation(vec)[0]
|
333 |
-
pre_norm = self.pre_norm(x)
|
334 |
-
x_mod = (1 + mod.scale) * pre_norm + mod.shift
|
335 |
-
qkv, mlp = torch.split(
|
336 |
-
self.linear1(x_mod),
|
337 |
-
[3 * self.hidden_size, self.mlp_hidden_dim],
|
338 |
-
dim=-1,
|
339 |
-
)
|
340 |
-
B, L, D = qkv.shape
|
341 |
-
q, k, v = qkv.reshape(B, L, self.K, self.H, D // self.KH).permute(2, 0, 3, 1, 4)
|
342 |
-
q, k = self.norm(q, k, v)
|
343 |
-
attn = attention(q, k, v, pe=pe)
|
344 |
-
output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2)).clamp(
|
345 |
-
min=-384 * 4, max=384 * 4
|
346 |
-
)
|
347 |
-
return x + mod.gate * output
|
348 |
-
|
349 |
-
|
350 |
-
class LastLayer(nn.Module):
|
351 |
-
def __init__(self, hidden_size: int, patch_size: int, out_channels: int):
|
352 |
-
super().__init__()
|
353 |
-
self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
354 |
-
self.linear = CublasLinear(
|
355 |
-
hidden_size, patch_size * patch_size * out_channels, bias=True
|
356 |
-
)
|
357 |
-
self.adaLN_modulation = nn.Sequential(
|
358 |
-
nn.SiLU(), CublasLinear(hidden_size, 2 * hidden_size, bias=True)
|
359 |
-
)
|
360 |
-
|
361 |
-
def forward(self, x: Tensor, vec: Tensor) -> Tensor:
|
362 |
-
shift, scale = self.adaLN_modulation(vec).chunk(2, dim=1)
|
363 |
-
x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
|
364 |
-
x = self.linear(x)
|
365 |
-
return x
|
366 |
-
|
367 |
-
|
368 |
-
class Flux(nn.Module):
|
369 |
-
"""
|
370 |
-
Transformer model for flow matching on sequences.
|
371 |
-
"""
|
372 |
-
|
373 |
-
def __init__(self, params: FluxParams, dtype: torch.dtype = torch.float16):
|
374 |
-
super().__init__()
|
375 |
-
|
376 |
-
self.dtype = dtype
|
377 |
-
self.params = params
|
378 |
-
self.in_channels = params.in_channels
|
379 |
-
self.out_channels = self.in_channels
|
380 |
-
if params.hidden_size % params.num_heads != 0:
|
381 |
-
raise ValueError(
|
382 |
-
f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}"
|
383 |
-
)
|
384 |
-
pe_dim = params.hidden_size // params.num_heads
|
385 |
-
if sum(params.axes_dim) != pe_dim:
|
386 |
-
raise ValueError(
|
387 |
-
f"Got {params.axes_dim} but expected positional dim {pe_dim}"
|
388 |
-
)
|
389 |
-
self.hidden_size = params.hidden_size
|
390 |
-
self.num_heads = params.num_heads
|
391 |
-
self.pe_embedder = EmbedND(
|
392 |
-
dim=pe_dim,
|
393 |
-
theta=params.theta,
|
394 |
-
axes_dim=params.axes_dim,
|
395 |
-
dtype=self.dtype,
|
396 |
-
)
|
397 |
-
self.img_in = F8Linear(self.in_channels, self.hidden_size, bias=True)
|
398 |
-
self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size)
|
399 |
-
self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size)
|
400 |
-
self.guidance_in = (
|
401 |
-
MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size)
|
402 |
-
if params.guidance_embed
|
403 |
-
else nn.Identity()
|
404 |
-
)
|
405 |
-
self.txt_in = F8Linear(params.context_in_dim, self.hidden_size)
|
406 |
-
|
407 |
-
self.double_blocks = nn.ModuleList(
|
408 |
-
[
|
409 |
-
DoubleStreamBlock(
|
410 |
-
self.hidden_size,
|
411 |
-
self.num_heads,
|
412 |
-
mlp_ratio=params.mlp_ratio,
|
413 |
-
qkv_bias=params.qkv_bias,
|
414 |
-
dtype=self.dtype,
|
415 |
-
)
|
416 |
-
for _ in range(params.depth)
|
417 |
-
]
|
418 |
-
)
|
419 |
-
|
420 |
-
self.single_blocks = nn.ModuleList(
|
421 |
-
[
|
422 |
-
SingleStreamBlock(
|
423 |
-
self.hidden_size,
|
424 |
-
self.num_heads,
|
425 |
-
mlp_ratio=params.mlp_ratio,
|
426 |
-
dtype=self.dtype,
|
427 |
-
)
|
428 |
-
for _ in range(params.depth_single_blocks)
|
429 |
-
]
|
430 |
-
)
|
431 |
-
|
432 |
-
self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels)
|
433 |
-
|
434 |
-
def forward(
|
435 |
-
self,
|
436 |
-
img: Tensor,
|
437 |
-
img_ids: Tensor,
|
438 |
-
txt: Tensor,
|
439 |
-
txt_ids: Tensor,
|
440 |
-
timesteps: Tensor,
|
441 |
-
y: Tensor,
|
442 |
-
guidance: Tensor | None = None,
|
443 |
-
) -> Tensor:
|
444 |
-
if img.ndim != 3 or txt.ndim != 3:
|
445 |
-
raise ValueError("Input img and txt tensors must have 3 dimensions.")
|
446 |
-
|
447 |
-
# running on sequences img
|
448 |
-
img = self.img_in(img)
|
449 |
-
vec = self.time_in(timestep_embedding(timesteps, 256).type(self.dtype))
|
450 |
-
|
451 |
-
if self.params.guidance_embed:
|
452 |
-
if guidance is None:
|
453 |
-
raise ValueError(
|
454 |
-
"Didn't get guidance strength for guidance distilled model."
|
455 |
-
)
|
456 |
-
vec = vec + self.guidance_in(
|
457 |
-
timestep_embedding(guidance, 256).type(self.dtype)
|
458 |
-
)
|
459 |
-
vec = vec + self.vector_in(y)
|
460 |
-
|
461 |
-
txt = self.txt_in(txt)
|
462 |
-
|
463 |
-
ids = torch.cat((txt_ids, img_ids), dim=1)
|
464 |
-
pe = self.pe_embedder(ids)
|
465 |
-
|
466 |
-
# double stream blocks
|
467 |
-
for block in self.double_blocks:
|
468 |
-
img, txt = block(img=img, txt=txt, vec=vec, pe=pe)
|
469 |
-
|
470 |
-
img = torch.cat((txt, img), 1)
|
471 |
-
|
472 |
-
# single stream blocks
|
473 |
-
for block in self.single_blocks:
|
474 |
-
img = block(img, vec=vec, pe=pe)
|
475 |
-
|
476 |
-
img = img[:, txt.shape[1] :, ...]
|
477 |
-
img = self.final_layer(img, vec) # (N, T, patch_size ** 2 * out_channels)
|
478 |
-
return img
|
479 |
-
|
480 |
-
@classmethod
|
481 |
-
def from_pretrained(cls, path: str, dtype: torch.dtype = torch.bfloat16) -> "Flux":
|
482 |
-
from util import load_config_from_path
|
483 |
-
from safetensors.torch import load_file
|
484 |
-
|
485 |
-
config = load_config_from_path(path)
|
486 |
-
with torch.device("meta"):
|
487 |
-
klass = cls(params=config.params, dtype=dtype).type(dtype)
|
488 |
-
|
489 |
-
ckpt = load_file(config.ckpt_path, device="cpu")
|
490 |
-
klass.load_state_dict(ckpt, assign=True)
|
491 |
-
return klass.to("cpu")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
util.py
CHANGED
@@ -6,7 +6,6 @@ import torch
|
|
6 |
from modules.autoencoder import AutoEncoder, AutoEncoderParams
|
7 |
from modules.conditioner import HFEmbedder
|
8 |
from modules.flux_model import Flux, FluxParams
|
9 |
-
from modules.flux_model_f8 import Flux as FluxF8
|
10 |
from safetensors.torch import load_file as load_sft
|
11 |
|
12 |
try:
|
@@ -68,7 +67,7 @@ class ModelSpec(BaseModel):
|
|
68 |
# Improved precision via not quanitzing the modulation linear layers
|
69 |
quantize_modulation: bool = True
|
70 |
# Improved precision via not quanitzing the flow embedder layers
|
71 |
-
quantize_flow_embedder_layers: bool =
|
72 |
|
73 |
model_config: ConfigDict = {
|
74 |
"arbitrary_types_allowed": True,
|
@@ -230,16 +229,14 @@ def print_load_warning(missing: list[str], unexpected: list[str]) -> None:
|
|
230 |
)
|
231 |
|
232 |
|
233 |
-
def load_flow_model(config: ModelSpec) -> Flux
|
234 |
ckpt_path = config.ckpt_path
|
235 |
FluxClass = Flux
|
236 |
-
if config.prequantized_flow:
|
237 |
-
FluxClass = FluxF8
|
238 |
|
239 |
with torch.device("meta"):
|
240 |
-
model = FluxClass(config
|
241 |
-
|
242 |
-
|
243 |
|
244 |
if ckpt_path is not None:
|
245 |
# load_sft doesn't support torch.device
|
@@ -290,7 +287,7 @@ def load_autoencoder(config: ModelSpec) -> AutoEncoder:
|
|
290 |
|
291 |
|
292 |
class LoadedModels(BaseModel):
|
293 |
-
flow: Flux
|
294 |
ae: AutoEncoder
|
295 |
clip: HFEmbedder
|
296 |
t5: HFEmbedder
|
|
|
6 |
from modules.autoencoder import AutoEncoder, AutoEncoderParams
|
7 |
from modules.conditioner import HFEmbedder
|
8 |
from modules.flux_model import Flux, FluxParams
|
|
|
9 |
from safetensors.torch import load_file as load_sft
|
10 |
|
11 |
try:
|
|
|
67 |
# Improved precision via not quanitzing the modulation linear layers
|
68 |
quantize_modulation: bool = True
|
69 |
# Improved precision via not quanitzing the flow embedder layers
|
70 |
+
quantize_flow_embedder_layers: bool = False
|
71 |
|
72 |
model_config: ConfigDict = {
|
73 |
"arbitrary_types_allowed": True,
|
|
|
229 |
)
|
230 |
|
231 |
|
232 |
+
def load_flow_model(config: ModelSpec) -> Flux:
|
233 |
ckpt_path = config.ckpt_path
|
234 |
FluxClass = Flux
|
|
|
|
|
235 |
|
236 |
with torch.device("meta"):
|
237 |
+
model = FluxClass(config, dtype=into_dtype(config.flow_dtype))
|
238 |
+
if not config.prequantized_flow:
|
239 |
+
model.type(into_dtype(config.flow_dtype))
|
240 |
|
241 |
if ckpt_path is not None:
|
242 |
# load_sft doesn't support torch.device
|
|
|
287 |
|
288 |
|
289 |
class LoadedModels(BaseModel):
|
290 |
+
flow: Flux
|
291 |
ae: AutoEncoder
|
292 |
clip: HFEmbedder
|
293 |
t5: HFEmbedder
|