Spaces:
Running on Zero
Running on Zero
File size: 15,079 Bytes
b701455 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 | """Tests for FP8 quantization and torch.compile fixes.
Validates:
- FP8 quantization enables comfy_cast_weights so runtime forward casts FP8→input dtype
- torch.compile uses safe default mode (max-autotune-no-cudagraphs, not reduce-overhead)
- FP8 + torch.compile combination works without crashes
"""
import inspect
import pytest
import torch
class TestFP8Quantization:
"""Tests for FP8 weight quantization and runtime casting."""
def test_fp8_enables_comfy_cast_weights(self):
"""After FP8 quantization, CastWeightBiasOp modules must have comfy_cast_weights=True."""
from src.cond.cast import CastWeightBiasOp, disable_weight_init
linear = disable_weight_init.Linear(4, 4, bias=False)
linear.weight.data = torch.randn(4, 4, dtype=torch.float16)
assert linear.comfy_cast_weights is False, "Should start with comfy_cast_weights=False"
if not hasattr(torch, "float8_e4m3fn"):
pytest.skip("FP8 dtype not available in this PyTorch build")
linear.weight.data = linear.weight.data.to(torch.float8_e4m3fn)
assert isinstance(linear, CastWeightBiasOp)
linear.comfy_cast_weights = True
assert linear.comfy_cast_weights is True
def test_fp8_forward_no_dtype_mismatch(self):
"""FP8 weights with comfy_cast_weights=True must not cause dtype mismatch."""
from src.cond.cast import disable_weight_init
if not hasattr(torch, "float8_e4m3fn"):
pytest.skip("FP8 dtype not available in this PyTorch build")
linear = disable_weight_init.Linear(8, 8, bias=False)
linear.weight.data = torch.randn(8, 8, dtype=torch.float16).to(torch.float8_e4m3fn)
linear.comfy_cast_weights = True
inp = torch.randn(2, 8, dtype=torch.float16)
result = linear(inp)
assert result.dtype == torch.float16, f"Expected float16 output, got {result.dtype}"
assert result.shape == (2, 8)
def test_fp8_forward_with_bias(self):
"""FP8 weights + float16 bias should work with comfy_cast_weights=True."""
from src.cond.cast import disable_weight_init
if not hasattr(torch, "float8_e4m3fn"):
pytest.skip("FP8 dtype not available in this PyTorch build")
linear = disable_weight_init.Linear(8, 8, bias=True)
linear.weight.data = torch.randn(8, 8, dtype=torch.float16).to(torch.float8_e4m3fn)
# bias stays in float16 (apply_fp8 only quantizes ndim>=2)
assert linear.bias.dtype in (torch.float16, torch.float32)
linear.comfy_cast_weights = True
inp = torch.randn(2, 8, dtype=torch.float16)
result = linear(inp)
assert result.shape == (2, 8)
def test_fp8_without_comfy_cast_raises(self):
"""FP8 weights WITHOUT comfy_cast_weights should raise RuntimeError (the original bug)."""
from src.cond.cast import disable_weight_init
if not hasattr(torch, "float8_e4m3fn"):
pytest.skip("FP8 dtype not available in this PyTorch build")
linear = disable_weight_init.Linear(8, 8, bias=False)
linear.weight.data = torch.randn(8, 8, dtype=torch.float16).to(torch.float8_e4m3fn)
# Intentionally do NOT set comfy_cast_weights = True
assert linear.comfy_cast_weights is False
inp = torch.randn(2, 8, dtype=torch.float16)
with pytest.raises(RuntimeError, match="have the same dtype"):
linear(inp)
def test_fp8_conv2d_forward(self):
"""FP8 Conv2d weights with comfy_cast_weights should work."""
from src.cond.cast import disable_weight_init
if not hasattr(torch, "float8_e4m3fn"):
pytest.skip("FP8 dtype not available in this PyTorch build")
conv = disable_weight_init.Conv2d(4, 8, 3, padding=1, bias=False)
conv.weight.data = torch.randn(8, 4, 3, 3, dtype=torch.float16).to(torch.float8_e4m3fn)
conv.comfy_cast_weights = True
inp = torch.randn(1, 4, 16, 16, dtype=torch.float16)
result = conv(inp)
assert result.shape == (1, 8, 16, 16)
class TestTorchCompileMode:
"""Tests for torch.compile default mode safety."""
def test_compile_model_default_mode(self):
"""Device.compile_model should default to max-autotune-no-cudagraphs."""
from src.Device import Device
sig = inspect.signature(Device.compile_model)
default_mode = sig.parameters["mode"].default
assert default_mode == "max-autotune-no-cudagraphs", (
f"Default compile mode should be 'max-autotune-no-cudagraphs', got '{default_mode}'"
)
def test_apply_torch_compile_default_mode(self):
"""AbstractModel.apply_torch_compile should default to max-autotune-no-cudagraphs."""
from src.Core.AbstractModel import AbstractModel
sig = inspect.signature(AbstractModel.apply_torch_compile)
default_mode = sig.parameters["mode"].default
assert default_mode == "max-autotune-no-cudagraphs", (
f"Default compile mode should be 'max-autotune-no-cudagraphs', got '{default_mode}'"
)
def test_compile_model_not_reduce_overhead(self):
"""Ensure default is NOT reduce-overhead (causes CUDA graph assertion errors)."""
from src.Device import Device
sig = inspect.signature(Device.compile_model)
default_mode = sig.parameters["mode"].default
assert default_mode != "reduce-overhead", (
"reduce-overhead causes CUDA graph assertion errors with dynamic model state"
)
class TestFP8AndCompileCombined:
"""Tests for FP8 + torch.compile compatibility."""
def test_fp8_compile_forward(self):
"""FP8 quantized modules should work when torch.compiled."""
from src.cond.cast import disable_weight_init
if not hasattr(torch, "float8_e4m3fn"):
pytest.skip("FP8 dtype not available")
if not hasattr(torch, "compile"):
pytest.skip("torch.compile not available")
linear = disable_weight_init.Linear(16, 16, bias=False)
linear.weight.data = torch.randn(16, 16, dtype=torch.float16).to(torch.float8_e4m3fn)
linear.comfy_cast_weights = True
try:
compiled = torch.compile(linear, mode="max-autotune-no-cudagraphs")
inp = torch.randn(2, 16, dtype=torch.float16)
out = compiled(inp)
assert out.shape == (2, 16)
except Exception as e:
# torch.compile may not work on all platforms (e.g., CPU-only, Windows)
if "inductor" in str(e).lower() or "compile" in str(e).lower():
pytest.skip(f"torch.compile not functional in this environment: {e}")
raise
def test_apply_fp8_falls_back_to_top_level_model(caplog, monkeypatch):
"""Models without a 'diffusion_model' submodule (e.g., Flux2) should have FP8 quantization
applied to the top-level module rather than emitting a warning."""
import logging
import torch
from src.Core.AbstractModel import AbstractModel, ModelCapabilities
class DummyModel(AbstractModel):
def _create_capabilities(self):
return ModelCapabilities()
def load(self, model_path=None):
self.model = torch.nn.Sequential(torch.nn.Linear(8, 8, bias=False))
self._loaded = True
return self
def encode_prompt(self, prompt, negative_prompt="", clip_skip=-2):
return None, None
def generate(self, ctx, positive, negative, *args, **kwargs):
raise NotImplementedError
def decode(self, latents):
raise NotImplementedError
dummy = DummyModel()
dummy.load()
caplog.set_level(logging.INFO)
# Force FP8 support path and spy on cast_to_fp8 calls
# Note: Device functions live in src.Device.Device module
monkeypatch.setattr('src.Device.Device.is_fp8_supported', lambda *args, **kwargs: True)
called = {'count': 0}
def fake_cast(tensor, scale=1.0):
called['count'] += 1
return tensor
monkeypatch.setattr('src.Device.Device.cast_to_fp8', fake_cast)
dummy.apply_fp8()
assert "No diffusion_model found for FP8 quantization" not in caplog.text
assert called['count'] > 0, "Expected cast_to_fp8 to be invoked on top-level model modules"
def test_apply_torch_compile_falls_back_to_top_level_model(caplog, monkeypatch):
"""If a model has no 'diffusion_model' attribute, torch.compile should be
applied to the top-level module instead of logging a warning."""
import logging
import torch
from src.Core.AbstractModel import AbstractModel, ModelCapabilities
if not hasattr(torch, 'compile'):
pytest.skip("torch.compile not available in this environment")
class DummyModel(AbstractModel):
def _create_capabilities(self):
return ModelCapabilities()
def load(self, model_path=None):
self.model = torch.nn.Sequential(torch.nn.Linear(4, 4, bias=False))
self._loaded = True
return self
def encode_prompt(self, prompt, negative_prompt="", clip_skip=-2):
return None, None
def generate(self, ctx, positive, negative, *args, **kwargs):
raise NotImplementedError
def decode(self, latents):
raise NotImplementedError
dummy = DummyModel()
dummy.load()
caplog.set_level(logging.INFO)
# Spy on Device.compile_model
compiled_called = {'count': 0}
def fake_compile(model_obj, mode='max-autotune-no-cudagraphs'):
compiled_called['count'] += 1
return model_obj # Return same object for simplicity
monkeypatch.setattr('src.Device.Device.compile_model', fake_compile)
dummy.apply_torch_compile()
assert "No diffusion_model found for torch.compile" not in caplog.text
assert compiled_called['count'] > 0, "Expected Device.compile_model to be invoked on the top-level module"
def test_apply_torch_compile_registers_wrapper_when_compile_returns_callable(monkeypatch):
"""If Device.compile_model returns a callable (not nn.Module), AbstractModel.apply_torch_compile
should attach the compiled callable to the module.forward while preserving the module object.
"""
import torch
import logging
from src.Core.AbstractModel import AbstractModel, ModelCapabilities
from src.Model.ModelPatcher import ModelPatcher
import torch.nn as nn
if not hasattr(torch, 'compile'):
pytest.skip("torch.compile not available")
class DummyModel(AbstractModel):
def _create_capabilities(self):
return ModelCapabilities()
def load(self, model_path=None):
base = nn.Sequential(nn.Linear(4, 4, bias=False))
mp = ModelPatcher(base, load_device=torch.device('cpu'), offload_device=torch.device('cpu'))
self.model = mp
self._loaded = True
return self
def encode_prompt(self, prompt, negative_prompt="", clip_skip=-2):
return None, None
def generate(self, ctx, positive, negative, *args, **kwargs):
raise NotImplementedError
def decode(self, latents):
raise NotImplementedError
dummy = DummyModel()
dummy.load()
compiled_called = {'count': 0}
def fake_compiled(*args, **kwargs):
compiled_called['count'] += 1
for a in args:
if hasattr(a, 'shape'):
return a
return torch.zeros(1)
monkeypatch.setattr('src.Device.Device.compile_model', lambda model_obj, mode='max-autotune-no-cudagraphs': fake_compiled)
dummy.apply_torch_compile()
assert isinstance(dummy.model, ModelPatcher)
assert hasattr(dummy.model, 'model') and isinstance(dummy.model.model, nn.Module)
assert hasattr(dummy.model.model, '_compiled_fn')
inp = torch.randn(1, 4)
_ = dummy.model.model(inp)
assert compiled_called['count'] > 0
# Ensure latent_format access still works
from src.Utilities import Latent as LatentUtil
dummy.model.model.latent_format = LatentUtil.SD15()
from src.Utilities.Latent import fix_empty_latent_channels
out = fix_empty_latent_channels(dummy.model, torch.zeros((1, 4, 64, 64)))
assert isinstance(out, torch.Tensor)
assert out.shape[1] == dummy.model.model.latent_format.latent_channels
def test_apply_torch_compile_attaches_compiled_forward_for_flux_like_module(monkeypatch):
"""When compiling a Flux-like module (top-level with apply_model/forward signature),
the compiled callable should be attached to the module.forward and model.apply_model
should continue to accept Flux-style kwargs like `c_crossattn`.
"""
import torch
import torch.nn as nn
from src.Core.AbstractModel import AbstractModel, ModelCapabilities
from src.Model.ModelPatcher import ModelPatcher
class FakeFluxModule(nn.Module):
def forward(self, img, txt=None, timesteps=None, y=None, guidance=None, control=None, transformer_options=None, attn_mask=None, img_h=None, img_w=None):
return img
def apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options=None, **kwargs):
# Map to forward signature used by Flux-like models
return self.forward(img=x, txt=c_crossattn, timesteps=t, y=kwargs.get('y'),
guidance=kwargs.get('guidance'), control=control, transformer_options=transformer_options,
attn_mask=kwargs.get('attention_mask'), img_h=(transformer_options or {}).get('img_h'), img_w=(transformer_options or {}).get('img_w'))
class DummyModel(AbstractModel):
def _create_capabilities(self):
return ModelCapabilities(is_flux=True, is_flux2=True)
def load(self, model_path=None):
base = FakeFluxModule()
mp = ModelPatcher(base, load_device=torch.device('cpu'), offload_device=torch.device('cpu'))
self.model = mp
self._loaded = True
return self
def encode_prompt(self, prompt, negative_prompt="", clip_skip=-2):
return None, None
def generate(self, ctx, positive, negative, *args, **kwargs):
raise NotImplementedError
def decode(self, latents):
raise NotImplementedError
dummy = DummyModel()
dummy.load()
compiled_called = {'count': 0}
def fake_compiled(img, txt=None, timesteps=None, y=None, **kwargs):
compiled_called['count'] += 1
return img
monkeypatch.setattr('src.Device.Device.compile_model', lambda model_obj, mode='max-autotune-no-cudagraphs': fake_compiled)
dummy.apply_torch_compile()
# After compile, calling apply_model should route through compiled forward without error
out = dummy.model.model.apply_model(torch.randn(1, 128, 8, 8), torch.tensor([1.0]), c_crossattn=torch.randn(1, 10, 768), transformer_options={'img_h':128,'img_w':128})
assert compiled_called['count'] > 0
assert out.shape[0] == 1
|