Dynamic swap with cublas linear / optional improved precision with vram drawback
Browse files- float8_quantize.py +75 -8
- flux_pipeline.py +7 -1
- modules/flux_model.py +2 -7
- util.py +9 -0
float8_quantize.py
CHANGED
@@ -1,3 +1,4 @@
|
|
|
|
1 |
import torch
|
2 |
import torch.nn as nn
|
3 |
from torchao.float8.float8_utils import (
|
@@ -10,7 +11,8 @@ import math
|
|
10 |
from torch.compiler import is_compiling
|
11 |
from torch import __version__
|
12 |
from torch.version import cuda
|
13 |
-
|
|
|
14 |
|
15 |
IS_TORCH_2_4 = __version__ < (2, 4, 9)
|
16 |
LT_TORCH_2_4 = __version__ < (2, 4)
|
@@ -42,7 +44,7 @@ class F8Linear(nn.Module):
|
|
42 |
float8_dtype=torch.float8_e4m3fn,
|
43 |
float_weight: torch.Tensor = None,
|
44 |
float_bias: torch.Tensor = None,
|
45 |
-
num_scale_trials: int =
|
46 |
input_float8_dtype=torch.float8_e5m2,
|
47 |
) -> None:
|
48 |
super().__init__()
|
@@ -183,6 +185,11 @@ class F8Linear(nn.Module):
|
|
183 |
1, dtype=self.weight.dtype, device=self.weight.device, requires_grad=False
|
184 |
)
|
185 |
|
|
|
|
|
|
|
|
|
|
|
186 |
def quantize_input(self, x: torch.Tensor):
|
187 |
if self.input_scale_initialized:
|
188 |
return to_fp8_saturated(x * self.input_scale, self.input_float8_dtype)
|
@@ -279,10 +286,12 @@ class F8Linear(nn.Module):
|
|
279 |
return f8_lin
|
280 |
|
281 |
|
|
|
282 |
def recursive_swap_linears(
|
283 |
model: nn.Module,
|
284 |
float8_dtype=torch.float8_e4m3fn,
|
285 |
input_float8_dtype=torch.float8_e5m2,
|
|
|
286 |
) -> None:
|
287 |
"""
|
288 |
Recursively swaps all nn.Linear modules in the given model with F8Linear modules.
|
@@ -300,6 +309,8 @@ def recursive_swap_linears(
|
|
300 |
all linear layers in the model will be using 8-bit quantization.
|
301 |
"""
|
302 |
for name, child in model.named_children():
|
|
|
|
|
303 |
if isinstance(child, nn.Linear) and not isinstance(
|
304 |
child, (F8Linear, CublasLinear)
|
305 |
):
|
@@ -315,7 +326,35 @@ def recursive_swap_linears(
|
|
315 |
)
|
316 |
del child
|
317 |
else:
|
318 |
-
recursive_swap_linears(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
319 |
|
320 |
|
321 |
@torch.inference_mode()
|
@@ -325,6 +364,10 @@ def quantize_flow_transformer_and_dispatch_float8(
|
|
325 |
float8_dtype=torch.float8_e4m3fn,
|
326 |
input_float8_dtype=torch.float8_e5m2,
|
327 |
offload_flow=False,
|
|
|
|
|
|
|
|
|
328 |
) -> nn.Module:
|
329 |
"""
|
330 |
Quantize the flux flow transformer model (original BFL codebase version) and dispatch to the given device.
|
@@ -334,19 +377,36 @@ def quantize_flow_transformer_and_dispatch_float8(
|
|
334 |
Allows for fast dispatch to gpu & quantize without causing OOM on gpus with limited memory.
|
335 |
|
336 |
After dispatching, if offload_flow is True, offloads the model to cpu.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
337 |
"""
|
338 |
for module in flow_model.double_blocks:
|
339 |
module.to(device)
|
340 |
module.eval()
|
341 |
recursive_swap_linears(
|
342 |
-
module,
|
|
|
|
|
|
|
343 |
)
|
344 |
torch.cuda.empty_cache()
|
345 |
for module in flow_model.single_blocks:
|
346 |
module.to(device)
|
347 |
module.eval()
|
348 |
recursive_swap_linears(
|
349 |
-
module,
|
|
|
|
|
|
|
350 |
)
|
351 |
torch.cuda.empty_cache()
|
352 |
to_gpu_extras = [
|
@@ -364,8 +424,10 @@ def quantize_flow_transformer_and_dispatch_float8(
|
|
364 |
continue
|
365 |
m_extra.to(device)
|
366 |
m_extra.eval()
|
367 |
-
if
|
368 |
-
m_extra,
|
|
|
|
|
369 |
):
|
370 |
setattr(
|
371 |
flow_model,
|
@@ -377,13 +439,18 @@ def quantize_flow_transformer_and_dispatch_float8(
|
|
377 |
),
|
378 |
)
|
379 |
del m_extra
|
380 |
-
elif module != "final_layer":
|
381 |
recursive_swap_linears(
|
382 |
m_extra,
|
383 |
float8_dtype=float8_dtype,
|
384 |
input_float8_dtype=input_float8_dtype,
|
|
|
385 |
)
|
386 |
torch.cuda.empty_cache()
|
|
|
|
|
|
|
|
|
387 |
if offload_flow:
|
388 |
flow_model.to("cpu")
|
389 |
torch.cuda.empty_cache()
|
|
|
1 |
+
from loguru import logger
|
2 |
import torch
|
3 |
import torch.nn as nn
|
4 |
from torchao.float8.float8_utils import (
|
|
|
11 |
from torch.compiler import is_compiling
|
12 |
from torch import __version__
|
13 |
from torch.version import cuda
|
14 |
+
|
15 |
+
from modules.flux_model import Modulation
|
16 |
|
17 |
IS_TORCH_2_4 = __version__ < (2, 4, 9)
|
18 |
LT_TORCH_2_4 = __version__ < (2, 4)
|
|
|
44 |
float8_dtype=torch.float8_e4m3fn,
|
45 |
float_weight: torch.Tensor = None,
|
46 |
float_bias: torch.Tensor = None,
|
47 |
+
num_scale_trials: int = 12,
|
48 |
input_float8_dtype=torch.float8_e5m2,
|
49 |
) -> None:
|
50 |
super().__init__()
|
|
|
185 |
1, dtype=self.weight.dtype, device=self.weight.device, requires_grad=False
|
186 |
)
|
187 |
|
188 |
+
def set_weight_tensor(self, tensor: torch.Tensor):
|
189 |
+
self.weight.data = tensor
|
190 |
+
self.weight_initialized = False
|
191 |
+
self.quantize_weight()
|
192 |
+
|
193 |
def quantize_input(self, x: torch.Tensor):
|
194 |
if self.input_scale_initialized:
|
195 |
return to_fp8_saturated(x * self.input_scale, self.input_float8_dtype)
|
|
|
286 |
return f8_lin
|
287 |
|
288 |
|
289 |
+
@torch.inference_mode()
|
290 |
def recursive_swap_linears(
|
291 |
model: nn.Module,
|
292 |
float8_dtype=torch.float8_e4m3fn,
|
293 |
input_float8_dtype=torch.float8_e5m2,
|
294 |
+
quantize_modulation: bool = True,
|
295 |
) -> None:
|
296 |
"""
|
297 |
Recursively swaps all nn.Linear modules in the given model with F8Linear modules.
|
|
|
309 |
all linear layers in the model will be using 8-bit quantization.
|
310 |
"""
|
311 |
for name, child in model.named_children():
|
312 |
+
if isinstance(child, Modulation) and not quantize_modulation:
|
313 |
+
continue
|
314 |
if isinstance(child, nn.Linear) and not isinstance(
|
315 |
child, (F8Linear, CublasLinear)
|
316 |
):
|
|
|
326 |
)
|
327 |
del child
|
328 |
else:
|
329 |
+
recursive_swap_linears(
|
330 |
+
child,
|
331 |
+
float8_dtype=float8_dtype,
|
332 |
+
input_float8_dtype=input_float8_dtype,
|
333 |
+
quantize_modulation=quantize_modulation,
|
334 |
+
)
|
335 |
+
|
336 |
+
|
337 |
+
@torch.inference_mode()
|
338 |
+
def swap_to_cublaslinear(model: nn.Module):
|
339 |
+
if not isinstance(CublasLinear, torch.nn.Module):
|
340 |
+
return
|
341 |
+
for name, child in model.named_children():
|
342 |
+
if isinstance(child, nn.Linear) and not isinstance(
|
343 |
+
child, (F8Linear, CublasLinear)
|
344 |
+
):
|
345 |
+
cublas_lin = CublasLinear(
|
346 |
+
child.in_features,
|
347 |
+
child.out_features,
|
348 |
+
bias=child.bias is not None,
|
349 |
+
dtype=child.weight.dtype,
|
350 |
+
device=child.weight.device,
|
351 |
+
)
|
352 |
+
cublas_lin.weight.data = child.weight.clone().detach()
|
353 |
+
cublas_lin.bias.data = child.bias.clone().detach()
|
354 |
+
setattr(model, name, cublas_lin)
|
355 |
+
del child
|
356 |
+
else:
|
357 |
+
swap_to_cublaslinear(child)
|
358 |
|
359 |
|
360 |
@torch.inference_mode()
|
|
|
364 |
float8_dtype=torch.float8_e4m3fn,
|
365 |
input_float8_dtype=torch.float8_e5m2,
|
366 |
offload_flow=False,
|
367 |
+
swap_linears_with_cublaslinear=True,
|
368 |
+
flow_dtype=torch.float16,
|
369 |
+
quantize_modulation: bool = True,
|
370 |
+
quantize_flow_embedder_layers: bool = True,
|
371 |
) -> nn.Module:
|
372 |
"""
|
373 |
Quantize the flux flow transformer model (original BFL codebase version) and dispatch to the given device.
|
|
|
377 |
Allows for fast dispatch to gpu & quantize without causing OOM on gpus with limited memory.
|
378 |
|
379 |
After dispatching, if offload_flow is True, offloads the model to cpu.
|
380 |
+
|
381 |
+
if swap_linears_with_cublaslinear is true, and flow_dtype == torch.float16, then swap all linears with cublaslinears for 2x performance boost on consumer GPUs.
|
382 |
+
Otherwise will skip the cublaslinear swap.
|
383 |
+
|
384 |
+
For added extra precision, you can set quantize_flow_embedder_layers to False,
|
385 |
+
this helps maintain the output quality of the flow transformer moreso than fully quantizing,
|
386 |
+
at the expense of ~512MB more VRAM usage.
|
387 |
+
|
388 |
+
For added extra precision, you can set quantize_modulation to False,
|
389 |
+
this helps maintain the output quality of the flow transformer moreso than fully quantizing,
|
390 |
+
at the expense of ~2GB more VRAM usage, but- has a much higher impact on image quality than the embedder layers.
|
391 |
"""
|
392 |
for module in flow_model.double_blocks:
|
393 |
module.to(device)
|
394 |
module.eval()
|
395 |
recursive_swap_linears(
|
396 |
+
module,
|
397 |
+
float8_dtype=float8_dtype,
|
398 |
+
input_float8_dtype=input_float8_dtype,
|
399 |
+
quantize_modulation=quantize_modulation,
|
400 |
)
|
401 |
torch.cuda.empty_cache()
|
402 |
for module in flow_model.single_blocks:
|
403 |
module.to(device)
|
404 |
module.eval()
|
405 |
recursive_swap_linears(
|
406 |
+
module,
|
407 |
+
float8_dtype=float8_dtype,
|
408 |
+
input_float8_dtype=input_float8_dtype,
|
409 |
+
quantize_modulation=quantize_modulation,
|
410 |
)
|
411 |
torch.cuda.empty_cache()
|
412 |
to_gpu_extras = [
|
|
|
424 |
continue
|
425 |
m_extra.to(device)
|
426 |
m_extra.eval()
|
427 |
+
if (
|
428 |
+
isinstance(m_extra, nn.Linear)
|
429 |
+
and not isinstance(m_extra, (F8Linear, CublasLinear))
|
430 |
+
and quantize_flow_embedder_layers
|
431 |
):
|
432 |
setattr(
|
433 |
flow_model,
|
|
|
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)
|
452 |
+
elif swap_linears_with_cublaslinear and flow_dtype != torch.float16:
|
453 |
+
logger.warning("Skipping cublas linear swap because flow_dtype is not float16")
|
454 |
if offload_flow:
|
455 |
flow_model.to("cpu")
|
456 |
torch.cuda.empty_cache()
|
flux_pipeline.py
CHANGED
@@ -645,7 +645,13 @@ class FluxPipeline:
|
|
645 |
|
646 |
if not config.prequantized_flow:
|
647 |
flow_model = quantize_flow_transformer_and_dispatch_float8(
|
648 |
-
flow_model,
|
|
|
|
|
|
|
|
|
|
|
|
|
649 |
)
|
650 |
else:
|
651 |
flow_model.eval().requires_grad_(False)
|
|
|
645 |
|
646 |
if not config.prequantized_flow:
|
647 |
flow_model = quantize_flow_transformer_and_dispatch_float8(
|
648 |
+
flow_model,
|
649 |
+
flux_device,
|
650 |
+
offload_flow=config.offload_flow,
|
651 |
+
swap_linears_with_cublaslinear=flux_dtype == torch.float16,
|
652 |
+
flow_dtype=flux_dtype,
|
653 |
+
quantize_modulation=config.quantize_modulation,
|
654 |
+
quantize_flow_embedder_layers=config.quantize_flow_embedder_layers,
|
655 |
)
|
656 |
else:
|
657 |
flow_model.eval().requires_grad_(False)
|
modules/flux_model.py
CHANGED
@@ -14,11 +14,6 @@ from torch import Tensor, nn
|
|
14 |
from pydantic import BaseModel
|
15 |
from torch.nn import functional as F
|
16 |
|
17 |
-
try:
|
18 |
-
from cublas_ops import CublasLinear
|
19 |
-
except ImportError:
|
20 |
-
CublasLinear = nn.Linear
|
21 |
-
|
22 |
|
23 |
class FluxParams(BaseModel):
|
24 |
in_channels: int
|
@@ -350,11 +345,11 @@ class LastLayer(nn.Module):
|
|
350 |
def __init__(self, hidden_size: int, patch_size: int, out_channels: int):
|
351 |
super().__init__()
|
352 |
self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
353 |
-
self.linear =
|
354 |
hidden_size, patch_size * patch_size * out_channels, bias=True
|
355 |
)
|
356 |
self.adaLN_modulation = nn.Sequential(
|
357 |
-
nn.SiLU(),
|
358 |
)
|
359 |
|
360 |
def forward(self, x: Tensor, vec: Tensor) -> Tensor:
|
|
|
14 |
from pydantic import BaseModel
|
15 |
from torch.nn import functional as F
|
16 |
|
|
|
|
|
|
|
|
|
|
|
17 |
|
18 |
class FluxParams(BaseModel):
|
19 |
in_channels: int
|
|
|
345 |
def __init__(self, hidden_size: int, patch_size: int, out_channels: int):
|
346 |
super().__init__()
|
347 |
self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
348 |
+
self.linear = nn.Linear(
|
349 |
hidden_size, patch_size * patch_size * out_channels, bias=True
|
350 |
)
|
351 |
self.adaLN_modulation = nn.Sequential(
|
352 |
+
nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True)
|
353 |
)
|
354 |
|
355 |
def forward(self, x: Tensor, vec: Tensor) -> Tensor:
|
util.py
CHANGED
@@ -8,12 +8,16 @@ 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 |
try:
|
12 |
from enum import StrEnum
|
13 |
except:
|
14 |
from enum import Enum
|
|
|
15 |
class StrEnum(str, Enum):
|
16 |
pass
|
|
|
|
|
17 |
from pydantic import BaseModel, ConfigDict
|
18 |
from loguru import logger
|
19 |
|
@@ -61,6 +65,11 @@ class ModelSpec(BaseModel):
|
|
61 |
offload_flow: bool = False
|
62 |
prequantized_flow: bool = False
|
63 |
|
|
|
|
|
|
|
|
|
|
|
64 |
model_config: ConfigDict = {
|
65 |
"arbitrary_types_allowed": True,
|
66 |
"use_enum_values": True,
|
|
|
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:
|
13 |
from enum import StrEnum
|
14 |
except:
|
15 |
from enum import Enum
|
16 |
+
|
17 |
class StrEnum(str, Enum):
|
18 |
pass
|
19 |
+
|
20 |
+
|
21 |
from pydantic import BaseModel, ConfigDict
|
22 |
from loguru import logger
|
23 |
|
|
|
65 |
offload_flow: bool = False
|
66 |
prequantized_flow: bool = False
|
67 |
|
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 = True
|
72 |
+
|
73 |
model_config: ConfigDict = {
|
74 |
"arbitrary_types_allowed": True,
|
75 |
"use_enum_values": True,
|