{{ message }}
Repository navigation
Expand file tree
/
Copy pathquant_ops.py
More file actions
284 lines (230 loc) · 9.91 KB
/
Copy pathquant_ops.py
File metadata and controls
284 lines (230 loc) · 9.91 KB
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
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
import torch
import logging
from comfy.cli_args import args
try:
import comfy_kitchen as ck
from comfy_kitchen.tensor import (
QuantizedTensor,
QuantizedLayout,
TensorCoreFP8Layout as _CKFp8Layout,
TensorCoreNVFP4Layout as _CKNvfp4Layout,
TensorCoreConvRotW4A4Layout as _CKTensorCoreConvRotW4A4Layout,
TensorWiseINT8Layout as _CKTensorWiseINT8Layout,
AsymW4A8Int8Layout as _CKAsymW4A8Int8Layout,
register_layout_op,
register_layout_class,
get_layout_class,
)
_CK_AVAILABLE = True
if torch.version.cuda is None:
ck.registry.disable("cuda")
else:
cuda_version = tuple(map(int, str(torch.version.cuda).split('.')))
if cuda_version < (13,):
ck.registry.disable("cuda")
logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.\nWARNING WARNING WARNING\nIf you are on nvidia 20 series and above it is required that you update your pytorch to cu130 or higher.\n")
# comfy-kitchen picks its accelerated backend on import: the HIP backend registers
# itself on a supported AMD device and takes dispatch priority there, CUDA on NVIDIA.
# Triton is an opt-in override, off by default on every platform.
if args.enable_triton_backend and not args.disable_triton_backend:
try:
import triton
logging.info("Found triton %s. Enabling comfy-kitchen triton backend.", triton.__version__)
except ImportError as e:
logging.error(f"Failed to import triton, Error: {e}, the comfy-kitchen triton backend will not be available.")
ck.registry.disable("triton")
else:
ck.registry.disable("triton")
for k, v in ck.list_backends().items():
logging.info(f"Found comfy_kitchen backend {k}: {v}")
except ImportError as e:
logging.error(f"Failed to import comfy_kitchen, Error: {e}, fp8 and fp4 support will not be available.")
_CK_AVAILABLE = False
class QuantizedTensor:
pass
class _CKFp8Layout:
pass
class _CKNvfp4Layout:
pass
class _CKTensorWiseINT8Layout:
pass
class _CKTensorCoreConvRotW4A4Layout:
pass
class _CKAsymW4A8Int8Layout:
pass
def register_layout_class(name, cls):
pass
def get_layout_class(name):
return None
_CK_MXFP8_AVAILABLE = False
if _CK_AVAILABLE:
try:
from comfy_kitchen.tensor import TensorCoreMXFP8Layout as _CKMxfp8Layout
_CK_MXFP8_AVAILABLE = True
except ImportError:
logging.warning("comfy_kitchen does not support MXFP8, please update comfy_kitchen.")
if not _CK_MXFP8_AVAILABLE:
class _CKMxfp8Layout:
pass
import comfy.float
# ==============================================================================
# FP8 Layouts with Comfy-Specific Extensions
# ==============================================================================
class _TensorCoreFP8LayoutBase(_CKFp8Layout):
FP8_DTYPE = None # Must be overridden in subclass
@classmethod
def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False):
if cls.FP8_DTYPE is None:
raise NotImplementedError(f"{cls.__name__} must define FP8_DTYPE")
orig_dtype = tensor.dtype
orig_shape = tuple(tensor.shape)
if isinstance(scale, str) and scale == "recalculate":
scale = torch.amax(tensor.abs()).to(dtype=torch.float32) / torch.finfo(cls.FP8_DTYPE).max
if tensor.dtype not in [torch.float32, torch.bfloat16]: # Prevent scale from being too small
tensor_info = torch.finfo(tensor.dtype)
scale = (1.0 / torch.clamp((1.0 / scale), min=tensor_info.min, max=tensor_info.max))
if scale is None:
scale = torch.ones((), device=tensor.device, dtype=torch.float32)
if not isinstance(scale, torch.Tensor):
scale = torch.tensor(scale, device=tensor.device, dtype=torch.float32)
if stochastic_rounding > 0:
if inplace_ops:
tensor *= (1.0 / scale).to(tensor.dtype)
else:
tensor = tensor * (1.0 / scale).to(tensor.dtype)
qdata = comfy.float.stochastic_rounding(tensor, dtype=cls.FP8_DTYPE, seed=stochastic_rounding)
else:
qdata = ck.quantize_per_tensor_fp8(tensor, scale, cls.FP8_DTYPE)
params = cls.Params(scale=scale.float(), orig_dtype=orig_dtype, orig_shape=orig_shape)
return qdata, params
class TensorCoreMXFP8Layout(_CKMxfp8Layout):
@classmethod
def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False):
if tensor.dim() != 2:
raise ValueError(f"MXFP8 requires 2D tensor, got {tensor.dim()}D")
orig_dtype = tensor.dtype
orig_shape = tuple(tensor.shape)
padded_shape = cls.get_padded_shape(orig_shape)
needs_padding = padded_shape != orig_shape
if stochastic_rounding > 0:
qdata, block_scale = comfy.float.stochastic_round_quantize_mxfp8_by_block(tensor, pad_32x=needs_padding, seed=stochastic_rounding)
else:
qdata, block_scale = ck.quantize_mxfp8(tensor, pad_32x=needs_padding)
params = cls.Params(
scale=block_scale,
orig_dtype=orig_dtype,
orig_shape=orig_shape,
)
return qdata, params
class TensorCoreNVFP4Layout(_CKNvfp4Layout):
@classmethod
def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False):
if tensor.dim() != 2:
raise ValueError(f"NVFP4 requires 2D tensor, got {tensor.dim()}D")
orig_dtype = tensor.dtype
orig_shape = tuple(tensor.shape)
if scale is None or (isinstance(scale, str) and scale == "recalculate"):
scale = torch.amax(tensor.abs()) / (ck.float_utils.F8_E4M3_MAX * ck.float_utils.F4_E2M1_MAX)
if not isinstance(scale, torch.Tensor):
scale = torch.tensor(scale)
scale = scale.to(device=tensor.device, dtype=torch.float32)
padded_shape = cls.get_padded_shape(orig_shape)
needs_padding = padded_shape != orig_shape
if stochastic_rounding > 0:
qdata, block_scale = comfy.float.stochastic_round_quantize_nvfp4_by_block(tensor, scale, pad_16x=needs_padding, seed=stochastic_rounding)
else:
qdata, block_scale = ck.quantize_nvfp4(tensor, scale, pad_16x=needs_padding)
params = cls.Params(
scale=scale,
orig_dtype=orig_dtype,
orig_shape=orig_shape,
block_scale=block_scale,
)
return qdata, params
class TensorCoreFP8E4M3Layout(_TensorCoreFP8LayoutBase):
FP8_DTYPE = torch.float8_e4m3fn
class TensorCoreFP8E5M2Layout(_TensorCoreFP8LayoutBase):
FP8_DTYPE = torch.float8_e5m2
# Backward compatibility alias - default to E4M3
TensorCoreFP8Layout = TensorCoreFP8E4M3Layout
TensorWiseINT8Layout = _CKTensorWiseINT8Layout
TensorCoreConvRotW4A4Layout = _CKTensorCoreConvRotW4A4Layout
AsymW4A8Int8Layout = _CKAsymW4A8Int8Layout
# ==============================================================================
# Registry
# ==============================================================================
register_layout_class("TensorCoreFP8Layout", TensorCoreFP8Layout)
register_layout_class("TensorCoreFP8E4M3Layout", TensorCoreFP8E4M3Layout)
register_layout_class("TensorCoreFP8E5M2Layout", TensorCoreFP8E5M2Layout)
register_layout_class("TensorCoreNVFP4Layout", TensorCoreNVFP4Layout)
register_layout_class("TensorWiseINT8Layout", _CKTensorWiseINT8Layout)
register_layout_class("TensorCoreConvRotW4A4Layout", _CKTensorCoreConvRotW4A4Layout)
if _CK_MXFP8_AVAILABLE:
register_layout_class("TensorCoreMXFP8Layout", TensorCoreMXFP8Layout)
register_layout_class("AsymW4A8Int8Layout", _CKAsymW4A8Int8Layout)
QUANT_ALGOS = {
"float8_e4m3fn": {
"storage_t": torch.float8_e4m3fn,
"parameters": {"weight_scale", "input_scale"},
"comfy_tensor_layout": "TensorCoreFP8E4M3Layout",
},
"float8_e5m2": {
"storage_t": torch.float8_e5m2,
"parameters": {"weight_scale", "input_scale"},
"comfy_tensor_layout": "TensorCoreFP8E5M2Layout",
},
"nvfp4": {
"storage_t": torch.uint8,
"parameters": {"weight_scale", "weight_scale_2", "input_scale", "pre_quant_scale"},
"comfy_tensor_layout": "TensorCoreNVFP4Layout",
"group_size": 16,
},
}
if _CK_MXFP8_AVAILABLE:
QUANT_ALGOS["mxfp8"] = {
"storage_t": torch.float8_e4m3fn,
"parameters": {"weight_scale", "input_scale"},
"comfy_tensor_layout": "TensorCoreMXFP8Layout",
"group_size": 32,
}
QUANT_ALGOS["int8_tensorwise"] = {
"storage_t": torch.int8,
"parameters": {"weight_scale"},
"comfy_tensor_layout": "TensorWiseINT8Layout",
"quantize_input": False,
}
QUANT_ALGOS["convrot_w4a4"] = {
"storage_t": torch.int8,
"parameters": {"weight_scale"},
"comfy_tensor_layout": "TensorCoreConvRotW4A4Layout",
"quantize_input": False,
}
QUANT_ALGOS["asym_w4a8_int8"] = {
"storage_t": torch.int8,
"parameters": {"weight_scale"},
"comfy_tensor_layout": "AsymW4A8Int8Layout",
"quantize_input": False,
}
# Same layout class and kernels, 6-bit uniform codes: weight is int8 [N, 3K/4], no codebook.
QUANT_ALGOS["w6a8_int8"] = {
"storage_t": torch.int8,
"parameters": {"weight_scale"},
"comfy_tensor_layout": "AsymW4A8Int8Layout",
"quantize_input": False,
}
# ==============================================================================
# Re-exports for backward compatibility
# ==============================================================================
__all__ = [
"QuantizedTensor",
"QuantizedLayout",
"TensorCoreFP8Layout",
"TensorCoreFP8E4M3Layout",
"TensorCoreFP8E5M2Layout",
"TensorCoreNVFP4Layout",
"TensorCoreConvRotW4A4Layout",
"TensorWiseINT8Layout",
"AsymW4A8Int8Layout",
"QUANT_ALGOS",
"register_layout_op",
]
You can’t perform that action at this time.