File size: 20,979 Bytes
b78342b | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 | from pathlib import Path
import sys
from types import ModuleType, SimpleNamespace
import torch
import pytest
COMFY = Path(__file__).resolve().parents[2] / "ComfyUI"
if str(COMFY) not in sys.path:
sys.path.insert(0, str(COMFY))
from comfyui_nodes import (
LoadMiniMaxH3AVLatentCache,
MiniMaxH3SolAttentionSM120,
MiniMaxH3SparkAttentionSM120,
SaveMiniMaxH3AVLatentCache,
_RunState,
_make_spark_layout,
)
def test_h3_av_latent_cache_roundtrip(tmp_path, monkeypatch):
import sys
class NestedTensor:
def __init__(self, tensors):
self.tensors = list(tensors)
self.is_nested = True
def unbind(self):
return self.tensors
fake_comfy = SimpleNamespace(nested_tensor=SimpleNamespace(NestedTensor=NestedTensor))
monkeypatch.setitem(sys.modules, "comfy", fake_comfy)
monkeypatch.setitem(sys.modules, "comfy.nested_tensor", fake_comfy.nested_tensor)
video = torch.randn(1, 24, 3, 4, 5, dtype=torch.bfloat16)
audio = torch.randn(1, 32, 2, 7, dtype=torch.bfloat16)
cache = tmp_path / "av.safetensors"
latent = {"samples": NestedTensor((video, audio))}
result = SaveMiniMaxH3AVLatentCache().save(latent, str(cache))
assert result["result"][0] is latent
loaded = LoadMiniMaxH3AVLatentCache().load(str(cache))[0]["samples"].unbind()
torch.testing.assert_close(loaded[0], video)
torch.testing.assert_close(loaded[1], audio)
def test_comfy_layout_moves_target_video_first_and_roundtrips():
positions = torch.zeros(11, 3, dtype=torch.float64)
positions[3:11] = torch.cartesian_prod(
torch.arange(2), torch.arange(2), torch.arange(2)
)
layout = SimpleNamespace(
segments=[(0, 3, "text"), (3, 11, "video")],
position_ids=positions,
)
spark = _make_spark_layout(layout, 11, torch.device("cpu"))
assert spark.grid == (2, 2, 2)
assert spark.video_tokens == 8
assert spark.permutation.tolist() == list(range(3, 11)) + list(range(3))
values = torch.arange(11)
restored = values[spark.permutation][spark.inverse_permutation]
torch.testing.assert_close(restored, values)
def test_run_state_uses_comfy_model_evaluation_count_and_resets():
state = _RunState.create(20, 0.1, 0.1)
assert state.steps == 20
assert state.warmup_evaluations == 2
state.begin_evaluation({"sigmas": torch.tensor([1.0])})
assert state.controller.evaluation_index == 0 and state.is_warmup
state.begin_evaluation({"sigmas": torch.tensor([0.9])})
assert state.controller.evaluation_index == 1 and state.is_warmup
state.begin_evaluation({"sigmas": torch.tensor([0.8])})
assert state.controller.evaluation_index == 2 and not state.is_warmup
state.begin_evaluation({"sigmas": torch.tensor([1.0])})
assert state.controller.evaluation_index == 0
def test_spark_warmup_steps_mode_and_limits():
fixed = _RunState.create(20, 0.1, 0.1, warmup_mode="warmup_steps", warmup_steps=4)
assert fixed.warmup_evaluations == 4
assert _RunState.create(3, 0.1, 0.1, warmup_mode="warmup_steps", warmup_steps=4).warmup_evaluations == 3
assert _RunState.create(20, 0.1, 0.1, warmup_mode="warmup_steps", warmup_steps=0).warmup_evaluations == 0
assert _RunState.create(20, 0.1, 0.1).warmup_evaluations == 2
assert _RunState.create(20, 0.2, 0.2).warmup_evaluations == 4
for index in range(5):
fixed.begin_evaluation({"sigmas": torch.tensor([1.0 - index * 0.1])})
assert fixed.is_warmup == (index < 4)
inputs = MiniMaxH3SparkAttentionSM120.INPUT_TYPES()["required"]
assert inputs["warmup_mode"][0] == ["warmup_ratio", "warmup_steps"]
assert inputs["warmup_ratio"][1]["default"] == 0.2
assert list(inputs).index("warmup_mode") < list(inputs).index("warmup_ratio") < list(inputs).index("warmup_steps")
assert MiniMaxH3SparkAttentionSM120.INPUT_TYPES()["required"]["topk_ratio"][1]["default"] == 0.2
assert MiniMaxH3SparkAttentionSM120.INPUT_TYPES()["required"]["min_tokens"][1]["default"] == 12288
assert "warmup_mode" not in MiniMaxH3SolAttentionSM120.INPUT_TYPES()["required"]
with pytest.raises(ValueError, match="warmup_mode"):
_RunState.create(20, 0.1, 0.1, warmup_mode="invalid")
with pytest.raises(ValueError, match="warmup_steps"):
_RunState.create(20, 0.1, 0.1, warmup_steps=-1)
with pytest.raises(ValueError, match="warmup_ratio"):
_RunState.create(20, 20.0, 0.1)
def test_spark_ablation_modes_are_orthogonal():
full = _RunState.create(20, 0.2, 0.1, "full").controller
optimized = _RunState.create(20, 0.2, 0.1, "full_group841").controller
reuse = _RunState.create(20, 0.2, 0.1, "full_reuse2").controller
assert full.config.topk_ratio == 0.1
assert full.config.topk_mode == "topk_ratio"
assert full.config.topk_blocks == 228
fixed = _RunState.create(20, 0.2, 0.1, topk_mode="topk_blocks", topk_blocks=12)
assert fixed.controller.config.topk_mode == "topk_blocks"
assert fixed.controller.config.topk_blocks == 12
assert MiniMaxH3SparkAttentionSM120.INPUT_TYPES()["required"]["topk_mode"][0] == [
"topk_ratio", "topk_blocks",
]
assert MiniMaxH3SparkAttentionSM120.INPUT_TYPES()["required"]["topk_mode"][1]["default"] == "topk_ratio"
assert MiniMaxH3SparkAttentionSM120.INPUT_TYPES()["required"]["topk_blocks"][1]["default"] == 228
assert list(MiniMaxH3SparkAttentionSM120.INPUT_TYPES()["required"]).index("topk_mode") < list(MiniMaxH3SparkAttentionSM120.INPUT_TYPES()["required"]).index("topk_ratio") < list(MiniMaxH3SparkAttentionSM120.INPUT_TYPES()["required"]).index("topk_blocks")
assert "topk_blocks" not in MiniMaxH3SolAttentionSM120.INPUT_TYPES()["required"]
assert full.config.landmark_tree_v2_group_size == 1
assert full.config.landmark_tree_v2_layout_reuse == "q_from_k"
independent = _RunState.create(
20, 0.2, 0.1, reblock_layout="independent"
).controller
assert independent.config.landmark_tree_v2_layout_reuse == "independent"
assert optimized.config.landmark_tree_v2_group_size == (8, 4, 1)
assert reuse.reblock_reuse_layers == 2
assert reuse.config.landmark_tree_v2_group_size == 1
assert full.config.video_tail_mode == "dense"
assert full.config.landmark_tree_v2_midpoint_direction_mode == "legacy"
assert "midpoint_direction_mode" not in MiniMaxH3SparkAttentionSM120.INPUT_TYPES()["optional"]
assert "midpoint_direction_mode" not in MiniMaxH3SolAttentionSM120.INPUT_TYPES()["optional"]
assert full.config.global_anchor_dtype == "float32"
assert "global_anchor_dtype" not in MiniMaxH3SparkAttentionSM120.INPUT_TYPES()["optional"]
assert "global_anchor_dtype" not in MiniMaxH3SolAttentionSM120.INPUT_TYPES()["optional"]
reblock_layout = MiniMaxH3SparkAttentionSM120.INPUT_TYPES()["optional"]["reblock_layout"]
assert reblock_layout[0] == ["q_reuse_k", "independent"]
assert reblock_layout[1]["default"] == "q_reuse_k"
assert "reblock_layout" not in MiniMaxH3SolAttentionSM120.INPUT_TYPES()["optional"]
fused_midpoint = _RunState.create(20, 0.2, 0.1, midpoint_direction_mode="fused")
assert fused_midpoint.controller.config.landmark_tree_v2_midpoint_direction_mode == "fused"
with pytest.raises(ValueError, match="midpoint_direction_mode"):
_RunState.create(20, 0.2, 0.1, midpoint_direction_mode="invalid")
for legacy in ("full_target189", "topk10_base", "reblock_only", "reweight_only"):
with pytest.raises(ValueError, match="only comfy-kitchen global modes"):
_RunState.create(20, 0.2, 0.1, legacy)
with pytest.raises(ValueError, match="topk_mode"):
_RunState.create(20, 0.2, 0.1, topk_mode="invalid")
with pytest.raises(ValueError, match="topk_blocks"):
_RunState.create(20, 0.2, 0.1, topk_blocks=0)
with pytest.raises(ValueError, match="reblock_layout"):
_RunState.create(20, 0.2, 0.1, reblock_layout="invalid")
def test_comfy_video_tail_rejects_unimplemented_pad():
from comfyui_backend import ComfySparkConfig
assert "video_tail_mode" not in MiniMaxH3SparkAttentionSM120.INPUT_TYPES()["optional"]
assert "ablation_mode" not in MiniMaxH3SparkAttentionSM120.INPUT_TYPES()["optional"]
for mode in ("pad", "invalid"):
with pytest.raises(ValueError, match="only supports 'dense'"):
ComfySparkConfig(video_tail_mode=mode)
with pytest.raises(ValueError, match="only supports 'dense'"):
_RunState.create(20, .2, .1, video_tail_mode=mode)
def test_spark_tail_granularity_is_independent():
for grain in ("block", "query"):
state = _RunState.create(20, .2, .1, tail_granularity=grain)
assert state.controller.config.tail_granularity == grain
assert state.controller.config.global_anchor_dtype == "float32"
assert _RunState.create(20, .2, .1).controller.config.tail_granularity == "query"
with pytest.raises(ValueError, match="tail_granularity"):
_RunState.create(20, .2, .1, tail_granularity="invalid")
assert "tail_granularity" in MiniMaxH3SparkAttentionSM120.INPUT_TYPES()["optional"]
def test_spark_local_exact_override_validation():
from comfyui_backend import ComfySparkConfig
assert ComfySparkConfig().force_local_blocks is None
assert ComfySparkConfig(force_local_blocks=True).force_local_blocks is True
assert ComfySparkConfig(force_local_blocks=False).force_local_blocks is False
with pytest.raises(ValueError, match="force_local_blocks"):
ComfySparkConfig(force_local_blocks=1)
with pytest.raises(ValueError, match="layout_reuse"):
ComfySparkConfig(landmark_tree_v2_layout_reuse="k_from_q")
assert "tail_granularity" not in MiniMaxH3SolAttentionSM120.INPUT_TYPES()["optional"]
@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required")
@pytest.mark.parametrize("group_size", [1, (8, 4, 1)])
def test_comfy_q_reuse_k_aliases_independent_key_layout_bitwise(
group_size, monkeypatch
):
import comfyui_reblock_plan
from comfyui_backend import (
ComfyPackedLayout,
ComfySparkConfig,
ComfySparkController,
comfy_kitchen_spark_attention,
)
from comfyui_reblock_plan import build_comfy_reblock_permutations
device = torch.device("cuda")
video_tokens = 2048
tokens = video_tokens + 64
heads = 2
ids = torch.arange(tokens, device=device)
layout = ComfyPackedLayout(
permutation=ids,
inverse_permutation=ids.clone(),
grid=(1, 32, 64),
video_tokens=video_tokens,
sequence_length=tokens,
video_positions=torch.zeros(video_tokens, 3, device=device),
)
generator = torch.Generator(device=device).manual_seed(20261008)
query = torch.randn(
1, tokens, heads, 128,
generator=generator, device=device, dtype=torch.bfloat16,
)
key = torch.randn(
query.shape,
generator=generator, device=device, dtype=torch.bfloat16,
)
value = torch.randn(
query.shape,
generator=generator, device=device, dtype=torch.bfloat16,
)
independent_controller = ComfySparkController(ComfySparkConfig(
landmark_tree_v2_layout_reuse="independent",
landmark_tree_v2_group_size=group_size,
))
shared_controller = ComfySparkController(ComfySparkConfig(
landmark_tree_v2_layout_reuse="q_from_k",
landmark_tree_v2_group_size=group_size,
))
# Three calls cover initial graph capture and subsequent graph replay.
for _ in range(3):
independent = build_comfy_reblock_permutations(
independent_controller, query, key, layout
)
shared = build_comfy_reblock_permutations(
shared_controller, query, key, layout
)
assert torch.equal(shared[0], independent[2])
assert torch.equal(shared[2], independent[2])
assert shared[0].data_ptr() == shared[2].data_ptr()
assert shared[1] is None
expected_layout = (independent[2], None, independent[2], None)
monkeypatch.setattr(
comfyui_reblock_plan,
"build_comfy_reblock_permutations",
lambda *_args, **_kwargs: expected_layout,
)
expected = comfy_kitchen_spark_attention(
independent_controller, query, key, value, layout, layer=1
)
reused_layout = (shared[0], None, shared[2], None)
monkeypatch.setattr(
comfyui_reblock_plan,
"build_comfy_reblock_permutations",
lambda *_args, **_kwargs: reused_layout,
)
actual = comfy_kitchen_spark_attention(
shared_controller, query, key, value, layout, layer=1
)
assert torch.equal(actual, expected)
@pytest.mark.parametrize("video_tokens,total_tokens,topk_mode,topk_blocks,expected_ratio", [
(64, 70, "topk_ratio", 128, 0.2),
(65, 70, "topk_blocks", 1, 1.0),
(640, 646, "topk_blocks", 3, 0.3),
(640, 646, "topk_blocks", 20, 1.0),
(64, 64, "topk_ratio", 128, 0.2),
])
def test_spark_only_target_video_queries_use_sparse_output(
monkeypatch, video_tokens, total_tokens, topk_mode, topk_blocks, expected_ratio
):
import comfyui_reblock_plan
from comfyui_backend import (
ComfyPackedLayout, ComfySparkConfig, ComfySparkController,
comfy_kitchen_spark_attention,
)
generator = torch.Generator().manual_seed(42)
q, k, v = (
torch.randn(1, total_tokens, 2, 128, generator=generator).to(torch.bfloat16)
for _ in range(3)
)
identity = torch.arange(video_tokens).view(1, 1, -1).expand(1, 2, -1).contiguous()
monkeypatch.setattr(
comfyui_reblock_plan, "build_comfy_reblock_permutations",
lambda controller, query, key, layout: (identity, None, identity, None),
)
call = {}
def fake_spark(query, key, value, **kwargs):
call.update(kwargs)
return torch.full_like(query, 7)
cuda_module = ModuleType("comfy_kitchen.backends.cuda")
cuda_module.spark_attn = fake_spark
monkeypatch.setitem(sys.modules, "comfy_kitchen.backends.cuda", cuda_module)
layout = ComfyPackedLayout(
permutation=torch.arange(total_tokens),
inverse_permutation=torch.arange(total_tokens),
grid=(1, 1, video_tokens),
video_tokens=video_tokens,
sequence_length=total_tokens,
video_positions=torch.zeros(video_tokens, 3),
)
controller = ComfySparkController(ComfySparkConfig(
topk_mode=topk_mode, topk_blocks=topk_blocks,
))
output = comfy_kitchen_spark_attention(controller, q, k, v, layout, layer=1)
assert call["topk_ratio"] == pytest.approx(expected_ratio)
assert call["sink_q"] == [video_tokens // 64, (total_tokens + 63) // 64]
assert call["sink_blocks"] == [video_tokens // 64, (total_tokens + 63) // 64]
torch.testing.assert_close(output[:, :video_tokens], torch.full_like(q[:, :video_tokens], 7))
if video_tokens < total_tokens:
torch.testing.assert_close(output[:, video_tokens:], torch.full_like(q[:, video_tokens:], 7))
assert "sink_query_mode" not in MiniMaxH3SparkAttentionSM120.INPUT_TYPES()["optional"]
class _FakeAttention:
heads = 2
head_dim = 128
def forward(self, *args, **kwargs):
raise AssertionError("not called during patch installation")
class MiniMaxH3Model:
def __init__(self):
self.blocks = [SimpleNamespace(attn=_FakeAttention()) for _ in range(3)]
class _FakePatcher:
def __init__(self, diffusion_model):
self.diffusion_model = diffusion_model
self.object_patches = {}
self.model_sampling = SimpleNamespace(percent_to_sigma=lambda percent: 1.0 - percent)
self.block_patches = {}
self.callbacks = []
def clone(self):
clone = _FakePatcher(self.diffusion_model)
clone.object_patches = dict(self.object_patches)
clone.block_patches = dict(self.block_patches)
clone.callbacks = list(self.callbacks)
return clone
def get_model_object(self, path):
obj = self
for part in path.split("."):
if part == "diffusion_model":
obj = self.diffusion_model
elif part == "model_sampling":
obj = self.model_sampling
elif part == "blocks":
obj = obj.blocks
elif part.isdigit():
obj = obj[int(part)]
else:
obj = getattr(obj, part)
return obj
def add_object_patch(self, path, value):
self.object_patches[path] = value
def set_model_patch_replace(self, patch, name, block_name, number, transformer_index=None):
self.block_patches[name, block_name, number, transformer_index] = patch
def add_callback_with_key(self, call_type, key, callback):
self.callbacks.append((call_type, key, callback))
def test_spark_node_uses_comfyui_native_block_patches_and_cleanup():
model = _FakePatcher(MiniMaxH3Model())
(patched,) = MiniMaxH3SparkAttentionSM120().patch(
model, True, 20, 0.1, 0.1, 1, 4096, True
)
assert patched is not model
assert sorted(patched.block_patches) == [
("dit", "double_block", i, None) for i in range(3)
]
assert patched.object_patches == {}
assert len(patched.callbacks) == 1
assert patched.callbacks[0][1] == "spark_h3_native_attention"
def test_sol_node_delegates_to_comfyui_official_patch(monkeypatch):
model = _FakePatcher(MiniMaxH3Model())
model.model_options = {"transformer_options": {}}
captured = {}
def official(model_arg, **kwargs):
captured.update(kwargs)
return model_arg.clone()
import comfy_extras.nodes_sparse_attention as official_module
monkeypatch.setattr(official_module, "apply_block_sparse_attention", official)
(patched,) = MiniMaxH3SolAttentionSM120().patch(
model, True, 20, 0.2, 1, 4096, True, 1.0
)
assert patched is not model
assert captured["tau"] == 1.0
assert captured["topk_ratio"] == 0.0
assert captured["start_percent"] == 0.2
assert captured["sink_conditioning"] == "exact_kv_and_rows"
assert captured["extra_tokens"] == 0
def test_comfyui_module_has_no_legacy_attention_interface():
source = (Path(__file__).resolve().parents[1] / "comfyui_nodes.py").read_text()
assert "_sol_attention" not in source
assert "_make_attention_forward" not in source
assert "import triton" not in source
assert "cutlass.cute" not in source
def test_comfyui_pipeline_does_not_import_diffusers_pipeline():
root = Path(__file__).resolve().parents[1]
node_source = (root / "comfyui_nodes.py").read_text()
backend_source = (root / "comfyui_backend.py").read_text()
assert "h3_sparse_attention.processor" not in node_source
assert "spark_attention_bthd" not in node_source
assert "spark_attention_bthd(" not in backend_source
assert "spark_integration" not in backend_source
assert "H3SparseAttentionConfig" not in backend_source
assert "_Controller" not in backend_source
@pytest.mark.skipif(
not torch.cuda.is_available() or torch.cuda.get_device_capability() != (12, 0),
reason="SM120 CUDA device required",
)
def test_comfy_native_chunked_qkv_matches_monolithic_projection():
import comfy.model_management
import comfy.quant_ops
from comfy.ldm.minimax.model import rope_rotation_table
from comfyui_nodes import _native_spark_qkv
torch.manual_seed(37)
heads, dim, hidden, tokens = 2, 128, 256, 73
class Attention(torch.nn.Module):
def __init__(self):
super().__init__()
self.heads = heads
self.head_dim = dim
self.qkv_proj = torch.nn.Linear(hidden, 3 * hidden, bias=False)
self.q_norm = torch.nn.RMSNorm(dim, eps=1e-5)
self.k_norm = torch.nn.RMSNorm(dim, eps=1e-5)
attn = Attention().cuda().to(torch.bfloat16).eval().requires_grad_(False)
x = torch.randn(tokens, hidden, device="cuda", dtype=torch.bfloat16)
permutation = torch.randperm(tokens, device="cuda")
rope = rope_rotation_table(
torch.randn(tokens, dim, device="cuda"), torch.bfloat16
)
with torch.no_grad():
actual = _native_spark_qkv(attn, x, rope, permutation)
projected = attn.qkv_proj(x.index_select(0, permutation))
q, k, v = projected.split(hidden, dim=-1)
q = q.view(1, tokens, heads, dim)
k = k.view(1, tokens, heads, dim)
v = v.view(1, tokens, heads, dim)
comfy.quant_ops.ck.rms_rope_split_half_(
q,
k,
rope.index_select(1, permutation),
comfy.model_management.cast_to(attn.q_norm.weight, device=x.device),
comfy.model_management.cast_to(attn.k_norm.weight, device=x.device),
epsilon=attn.q_norm.eps,
rot_dim=dim,
)
for got, expected in zip(actual, (q, k, v)):
torch.testing.assert_close(got, expected, rtol=0, atol=0)
|