[Hardware][Ascend] Add Wan2.2 MXFP4 UOS, Smooth and W4A8 fallback (#7210)
Signed-off-by: HAAZZZEEEE <wang-taicheng@qq.com>
This commit is contained in:
HAAZZZEEEE authored and GitHub committed
1 parent
4fab34c8f4
commit
becfefd440
20 files changed
+2435
-184
No files matched your search
@@ -0,0 +1,294 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
|
||||
"""CPU coverage for MXFP4 checkpoint loading and the W4A8 dtype contract."""
|
||||
|
||||
import sys
|
||||
from types import ModuleType
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
from vllm.config.load import LoadConfig
|
||||
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
|
||||
|
||||
from vllm_omni.diffusion.data import OmniDiffusionConfig
|
||||
from vllm_omni.diffusion.forward_context import (
|
||||
ForwardContext,
|
||||
override_forward_context,
|
||||
set_forward_context_denoise_step_idx,
|
||||
)
|
||||
from vllm_omni.diffusion.model_loader.diffusers_loader import DiffusersPipelineLoader
|
||||
from vllm_omni.quantization.mxfp4_config import (
|
||||
DiffusionMXFP4Config,
|
||||
DiffusionMXFP4DualScaleMixedConfig,
|
||||
NPUMxfp4DualScaleLinearMethod,
|
||||
NPUMxfp4LinearMethod,
|
||||
NPUMxfp4OnlineLinearMethod,
|
||||
)
|
||||
|
||||
pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
|
||||
|
||||
|
||||
@pytest.fixture(autouse=True)
|
||||
def _patch_tp_state(monkeypatch):
|
||||
monkeypatch.setattr("vllm.model_executor.parameter.get_tensor_model_parallel_rank", lambda: 0)
|
||||
monkeypatch.setattr("vllm.model_executor.parameter.get_tensor_model_parallel_world_size", lambda: 1)
|
||||
|
||||
|
||||
class _CheckpointLayer(torch.nn.Module):
|
||||
def __init__(self, method, dtype=torch.bfloat16):
|
||||
super().__init__()
|
||||
method.create_weights(
|
||||
self,
|
||||
input_size_per_partition=512,
|
||||
output_partition_sizes=[2],
|
||||
input_size=512,
|
||||
output_size=2,
|
||||
params_dtype=dtype,
|
||||
weight_loader=default_weight_loader,
|
||||
)
|
||||
|
||||
def load_weights(self, weights):
|
||||
params = dict(self.named_parameters())
|
||||
loaded = set()
|
||||
for name, value in weights:
|
||||
if name in params:
|
||||
params[name].weight_loader(params[name], value)
|
||||
loaded.add(name)
|
||||
return loaded
|
||||
|
||||
|
||||
def _load_checkpoint(layer, config, weights, monkeypatch):
|
||||
od_config = OmniDiffusionConfig(model="", dtype=torch.bfloat16, quantization_config=config)
|
||||
loader = DiffusersPipelineLoader(LoadConfig(), od_config)
|
||||
loader.counter_before_loading_weights = 0.0
|
||||
monkeypatch.setattr(loader, "get_all_weights", lambda model: iter(weights.items()))
|
||||
loader.load_weights(layer)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("serialized", [False, True])
|
||||
def test_single_scale_requires_checkpoint_scale_only_when_serialized(serialized, monkeypatch):
|
||||
config = DiffusionMXFP4Config(is_checkpoint_mxfp4_serialized=serialized)
|
||||
# Exercise the shared implementation directly as well as the online MRO
|
||||
# below: checkpoint requirements must follow the configured storage mode.
|
||||
layer = _CheckpointLayer(NPUMxfp4LinearMethod(config))
|
||||
weights = {"weight": torch.ones((2, 512), dtype=torch.bfloat16)}
|
||||
if serialized:
|
||||
with pytest.raises(ValueError, match="Required weights.*weight_scale"):
|
||||
_load_checkpoint(layer, config, weights, monkeypatch)
|
||||
else:
|
||||
_load_checkpoint(layer, config, weights, monkeypatch)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("scale_alg", [0, 2])
|
||||
def test_online_mxfp4_loads_bf16_checkpoint_without_scale(scale_alg, monkeypatch):
|
||||
config = DiffusionMXFP4Config(mxfp4_scale_alg=scale_alg)
|
||||
method = NPUMxfp4OnlineLinearMethod(config)
|
||||
layer = _CheckpointLayer(method)
|
||||
assert set(dict(layer.named_parameters())) == {"weight"}
|
||||
assert layer.weight.is_meta
|
||||
|
||||
weight = torch.arange(1024, dtype=torch.bfloat16).reshape(2, 512)
|
||||
packed_weight = torch.zeros((2, 256), dtype=torch.uint8)
|
||||
generated_scale = torch.full((2, 16), 127, dtype=torch.uint8)
|
||||
quantized_inputs = []
|
||||
|
||||
def quantize(value, **kwargs):
|
||||
quantized_inputs.append(value.clone())
|
||||
expected = dict(dst_type="fp4", axis=-1, block_size=32, round_mode="rint", scale_alg=scale_alg)
|
||||
if scale_alg == 2:
|
||||
expected["dst_type_max"] = 7.25
|
||||
assert kwargs == expected
|
||||
return packed_weight, generated_scale
|
||||
|
||||
# Keep real creation, lazy loading, processing, and strict loader checks;
|
||||
# substitute only the device quantization operator for this CPU test.
|
||||
npu = ModuleType("torch_npu")
|
||||
npu.__dict__.update(float4_e2m1fn_x2="fp4", npu_dynamic_mx_quant=quantize)
|
||||
monkeypatch.setitem(sys.modules, "torch_npu", npu)
|
||||
_load_checkpoint(layer, config, {"weight": weight}, monkeypatch)
|
||||
|
||||
assert len(quantized_inputs) == 1
|
||||
torch.testing.assert_close(quantized_inputs[0], weight, rtol=0, atol=0)
|
||||
torch.testing.assert_close(layer.weight, packed_weight, rtol=0, atol=0)
|
||||
torch.testing.assert_close(layer.weight_scale, generated_scale.reshape(2, 8, 2), rtol=0, atol=0)
|
||||
assert not getattr(layer.weight_scale, "is_checkpoint_required", False)
|
||||
method.process_weights_after_loading(layer)
|
||||
assert len(quantized_inputs) == 1
|
||||
|
||||
|
||||
@pytest.mark.parametrize("require_smooth", [False, True])
|
||||
@pytest.mark.parametrize("has_smooth", [False, True])
|
||||
def test_single_scale_checkpoint_smooth_default_and_required(require_smooth, has_smooth, monkeypatch):
|
||||
config = DiffusionMXFP4Config(is_checkpoint_mxfp4_serialized=True, require_smooth_scale=require_smooth)
|
||||
layer = _CheckpointLayer(NPUMxfp4LinearMethod(config))
|
||||
weights = {name: torch.ones_like(param) for name, param in layer.named_parameters()}
|
||||
scale = torch.linspace(0.5, 3.0, 512)
|
||||
if has_smooth:
|
||||
weights["mul_scale"] = scale
|
||||
else:
|
||||
weights.pop("mul_scale")
|
||||
|
||||
if require_smooth and not has_smooth:
|
||||
with pytest.raises(ValueError, match="mul_scale"):
|
||||
_load_checkpoint(layer, config, weights, monkeypatch)
|
||||
else:
|
||||
_load_checkpoint(layer, config, weights, monkeypatch)
|
||||
torch.testing.assert_close(layer.mul_scale, scale if has_smooth else torch.ones(512), rtol=0, atol=0)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("bad_scale", ["shape", "name"])
|
||||
def test_declared_single_scale_smooth_rejects_malformed_checkpoint(bad_scale, monkeypatch):
|
||||
config = DiffusionMXFP4Config(is_checkpoint_mxfp4_serialized=True, require_smooth_scale=True)
|
||||
layer = _CheckpointLayer(NPUMxfp4LinearMethod(config))
|
||||
weights = {name: torch.ones_like(param) for name, param in layer.named_parameters()}
|
||||
if bad_scale == "shape":
|
||||
weights["mul_scale"] = torch.ones(511)
|
||||
with pytest.raises(AssertionError):
|
||||
_load_checkpoint(layer, config, weights, monkeypatch)
|
||||
else:
|
||||
weights["multiply_scale"] = weights.pop("mul_scale")
|
||||
with pytest.raises(ValueError, match="mul_scale"):
|
||||
_load_checkpoint(layer, config, weights, monkeypatch)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("required", [False, True])
|
||||
@pytest.mark.parametrize("bad_value", [float("nan"), float("inf"), 0.0, -1.0])
|
||||
def test_single_scale_rejects_invalid_smooth_values(required, bad_value, monkeypatch):
|
||||
config = DiffusionMXFP4Config(is_checkpoint_mxfp4_serialized=True, require_smooth_scale=required)
|
||||
layer = _CheckpointLayer(NPUMxfp4LinearMethod(config))
|
||||
weights = {name: torch.ones_like(param) for name, param in layer.named_parameters()}
|
||||
weights["mul_scale"][7] = bad_value
|
||||
with pytest.raises(ValueError, match="finite, strictly positive"):
|
||||
_load_checkpoint(layer, config, weights, monkeypatch)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("bad_scale", [torch.ones(512, dtype=torch.int32), torch.ones(1, 512)])
|
||||
def test_single_scale_rejects_invalid_smooth_type_or_rank(bad_scale, monkeypatch):
|
||||
config = DiffusionMXFP4Config(is_checkpoint_mxfp4_serialized=True)
|
||||
layer = _CheckpointLayer(NPUMxfp4LinearMethod(config))
|
||||
weights = {name: torch.ones_like(param) for name, param in layer.named_parameters()}
|
||||
weights["mul_scale"] = bad_scale
|
||||
with pytest.raises(ValueError, match="one-dimensional floating-point"):
|
||||
_load_checkpoint(layer, config, weights, monkeypatch)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("bad_value", [1e10, 1e-10])
|
||||
def test_single_scale_rejects_smooth_overflow_or_underflow_after_cast(bad_value, monkeypatch):
|
||||
config = DiffusionMXFP4Config(is_checkpoint_mxfp4_serialized=True)
|
||||
method = NPUMxfp4LinearMethod(config)
|
||||
layer = _CheckpointLayer(method, torch.float16)
|
||||
layer.mul_scale.data.fill_(bad_value)
|
||||
# Validation must fail before any device packing or GEMM.
|
||||
monkeypatch.setitem(sys.modules, "torch_npu", ModuleType("torch_npu"))
|
||||
with pytest.raises(ValueError, match="finite, strictly positive"):
|
||||
method.process_weights_after_loading(layer)
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
("mode", "missing"),
|
||||
[
|
||||
("single", "weight"),
|
||||
("single", "weight_scale"),
|
||||
("dual", "weight"),
|
||||
("dual", "weight_scale"),
|
||||
("dual", "weight_dual_scale"),
|
||||
("dual", "mul_scale"),
|
||||
],
|
||||
)
|
||||
def test_offline_mxfp4_required_tensors_do_not_use_legacy_scale_tolerance(mode, missing, monkeypatch):
|
||||
if mode == "single":
|
||||
config = DiffusionMXFP4Config(is_checkpoint_mxfp4_serialized=True)
|
||||
method = NPUMxfp4LinearMethod(config)
|
||||
else:
|
||||
config = DiffusionMXFP4DualScaleMixedConfig(is_checkpoint_serialized=True)
|
||||
method = NPUMxfp4DualScaleLinearMethod(config)
|
||||
layer = _CheckpointLayer(method)
|
||||
weights = {name: torch.ones_like(param) for name, param in layer.named_parameters() if name != missing}
|
||||
with pytest.raises(ValueError, match=missing):
|
||||
_load_checkpoint(layer, config, weights, monkeypatch)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("rank", [0, 1])
|
||||
def test_single_scale_smooth_uses_real_row_parallel_loader(rank):
|
||||
from vllm.model_executor.layers.linear import RowParallelLinear
|
||||
|
||||
layer = _CheckpointLayer(NPUMxfp4LinearMethod(DiffusionMXFP4Config(is_checkpoint_mxfp4_serialized=True)))
|
||||
layer.tp_rank = rank
|
||||
checkpoint_scale = torch.arange(1024, dtype=torch.float32)
|
||||
RowParallelLinear.weight_loader(layer, layer.mul_scale, checkpoint_scale)
|
||||
torch.testing.assert_close(layer.mul_scale, checkpoint_scale[rank * 512 : (rank + 1) * 512], rtol=0, atol=0)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("dtype", [torch.bfloat16, torch.float16])
|
||||
@pytest.mark.parametrize("fallback", [False, True])
|
||||
@pytest.mark.parametrize("with_bias", [False, True])
|
||||
@pytest.mark.parametrize("scale_alg", [0, 2])
|
||||
def test_single_scale_smooth_precedes_quantization_and_preserves_output(
|
||||
dtype, fallback, with_bias, scale_alg, monkeypatch
|
||||
):
|
||||
config = DiffusionMXFP4Config(
|
||||
is_checkpoint_mxfp4_serialized=True, w4a8_fallback_steps=[0], mxfp4_scale_alg=scale_alg
|
||||
)
|
||||
method = NPUMxfp4LinearMethod(config)
|
||||
layer = _CheckpointLayer(method, dtype)
|
||||
weights = {
|
||||
"weight": torch.linspace(-0.5, 1.5, 1024, dtype=dtype).reshape(2, 512),
|
||||
"weight_scale": torch.full((2, 16), 127, dtype=torch.uint8),
|
||||
"mul_scale": torch.linspace(0.5, 3.0, 512),
|
||||
}
|
||||
_load_checkpoint(layer, config, weights, monkeypatch)
|
||||
x = torch.linspace(-1.0, 2.0, 3072, dtype=dtype).reshape(2, 3, 512)
|
||||
bias = torch.tensor([0.5, -1.0], dtype=dtype) if with_bias else None
|
||||
quantized_inputs = []
|
||||
|
||||
# Only the two device operators are substituted. Exercise real apply(),
|
||||
# step selection, Smooth multiplication, bias handling, and reshape logic.
|
||||
npu = ModuleType("torch_npu")
|
||||
|
||||
def quantize(value, **kwargs):
|
||||
quantized_inputs.append(value.clone())
|
||||
expected = dict(
|
||||
dst_type=torch.float8_e4m3fn if fallback else "fp4",
|
||||
axis=-1,
|
||||
block_size=32,
|
||||
round_mode="rint",
|
||||
scale_alg=0 if fallback else scale_alg,
|
||||
)
|
||||
if not fallback and scale_alg == 2:
|
||||
expected["dst_type_max"] = 7.25
|
||||
assert kwargs == expected
|
||||
return value, torch.ones((6, 16), dtype=torch.uint8)
|
||||
|
||||
def matmul(x_q, weight, weight_scale, *, bias, output_dtype, **kwargs):
|
||||
if fallback:
|
||||
assert output_dtype == torch.bfloat16
|
||||
if bias is not None:
|
||||
assert bias.dtype == output_dtype
|
||||
assert bias.shape == (1, 2)
|
||||
result = x_q.float() @ weight.float()
|
||||
if bias is not None:
|
||||
result += bias.float()
|
||||
return result.to(output_dtype)
|
||||
|
||||
npu.__dict__.update(
|
||||
float8_e4m3fn=torch.float8_e4m3fn,
|
||||
float8_e8m0fnu=torch.uint8,
|
||||
float4_e2m1fn_x2="fp4",
|
||||
npu_dynamic_mx_quant=quantize,
|
||||
npu_quant_matmul=matmul,
|
||||
)
|
||||
monkeypatch.setitem(sys.modules, "torch_npu", npu)
|
||||
with override_forward_context(ForwardContext()):
|
||||
set_forward_context_denoise_step_idx(0 if fallback else 1)
|
||||
result = method.apply(layer, x, bias)
|
||||
|
||||
expected_input = x.reshape(-1, 512) * weights["mul_scale"].to(dtype)
|
||||
assert len(quantized_inputs) == 1
|
||||
torch.testing.assert_close(quantized_inputs[0], expected_input, rtol=0, atol=0)
|
||||
expected = expected_input.float() @ weights["weight"].float().T
|
||||
if bias is not None:
|
||||
expected += bias.float()
|
||||
if fallback:
|
||||
expected = expected.bfloat16()
|
||||
torch.testing.assert_close(result, expected.to(dtype).reshape(2, 3, 2), rtol=0, atol=0)
|
||||
assert result.dtype == dtype
|
||||
@@ -1,5 +1,5 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
|
||||
"""Tests for MXFP4 quantization configs and the MXFP4 DualScale + BF16 mixed config."""
|
||||
|
||||
import pytest
|
||||
@@ -27,12 +27,42 @@ def test_mxfp4_config_get_name():
|
||||
assert DiffusionMXFP4Config.get_name() == "mxfp4"
|
||||
|
||||
|
||||
@pytest.mark.parametrize("value", [True, False, None, "2", 2.0, 1, -1, 3])
|
||||
def test_mxfp4_scale_alg_rejects_invalid_values(value):
|
||||
from vllm_omni.quantization import build_quant_config
|
||||
|
||||
with pytest.raises(ValueError, match="mxfp4_scale_alg"):
|
||||
build_quant_config({"method": "mxfp4", "mxfp4_scale_alg": value})
|
||||
|
||||
|
||||
@pytest.mark.parametrize("runtime_alg", [0, 2])
|
||||
@pytest.mark.parametrize("serialized", [False, True])
|
||||
def test_runtime_algorithm_survives_offline_storage_rebuild(runtime_alg, serialized):
|
||||
from vllm_omni.quantization.factory import resolve_quant_config_from_disk
|
||||
from vllm_omni.quantization.mxfp4_config import DiffusionMXFP4Config
|
||||
|
||||
active = DiffusionMXFP4Config(is_checkpoint_mxfp4_serialized=serialized, mxfp4_scale_alg=runtime_alg)
|
||||
resolved = resolve_quant_config_from_disk(
|
||||
active,
|
||||
{
|
||||
"quant_method": "mxfp4",
|
||||
"is_checkpoint_mxfp4_serialized": True,
|
||||
"mxfp4_scale_alg": 2,
|
||||
"ignored_layers": ["proj_out"],
|
||||
},
|
||||
)
|
||||
assert resolved.is_checkpoint_mxfp4_serialized
|
||||
assert resolved.mxfp4_scale_alg == runtime_alg
|
||||
assert resolved.w4a8_fallback_layers == resolved.w4a8_fallback_steps == []
|
||||
|
||||
|
||||
def test_mxfp4_config_from_config_defaults():
|
||||
from vllm_omni.quantization.mxfp4_config import DiffusionMXFP4Config
|
||||
|
||||
cfg = DiffusionMXFP4Config.from_config({})
|
||||
assert cfg.is_checkpoint_mxfp4_serialized is False
|
||||
assert cfg.ignored_layers == []
|
||||
assert cfg.w4a8_fallback_layers == []
|
||||
|
||||
|
||||
def test_mxfp4_config_from_config_serialized():
|
||||
@@ -42,6 +72,98 @@ def test_mxfp4_config_from_config_serialized():
|
||||
assert cfg.is_checkpoint_mxfp4_serialized is True
|
||||
|
||||
|
||||
def test_mxfp4_config_requires_offline_smooth():
|
||||
from vllm_omni.quantization import build_quant_config
|
||||
from vllm_omni.quantization.mxfp4_config import DiffusionMXFP4Config
|
||||
|
||||
values = {"is_checkpoint_mxfp4_serialized": True, "require_smooth_scale": True}
|
||||
assert DiffusionMXFP4Config.from_config(values).require_smooth_scale
|
||||
assert build_quant_config({"method": "mxfp4", **values}).require_smooth_scale
|
||||
with pytest.raises(ValueError, match="offline"):
|
||||
build_quant_config({"method": "mxfp4", "require_smooth_scale": True})
|
||||
with pytest.raises(ValueError, match="boolean"):
|
||||
DiffusionMXFP4Config.from_config({**values, "require_smooth_scale": "false"})
|
||||
|
||||
|
||||
@pytest.mark.parametrize("serialized", [False, True])
|
||||
def test_mxfp4_disk_smooth_requirement_preserves_step_policy(serialized):
|
||||
from vllm_omni.quantization.factory import resolve_quant_config_from_disk
|
||||
from vllm_omni.quantization.mxfp4_config import DiffusionMXFP4Config
|
||||
|
||||
active = DiffusionMXFP4Config(is_checkpoint_mxfp4_serialized=serialized, w4a8_fallback_steps=[1])
|
||||
resolved = resolve_quant_config_from_disk(
|
||||
active,
|
||||
{
|
||||
"quant_method": "mxfp4",
|
||||
"is_checkpoint_mxfp4_serialized": True,
|
||||
"require_smooth_scale": True,
|
||||
},
|
||||
)
|
||||
assert resolved.require_smooth_scale
|
||||
assert resolved.is_checkpoint_mxfp4_serialized
|
||||
assert resolved.w4a8_fallback_steps == [1]
|
||||
|
||||
|
||||
@pytest.mark.parametrize("disk_requirement", [None, False, True])
|
||||
def test_mxfp4_expert_rebuild_cannot_weaken_explicit_smooth_requirement(disk_requirement):
|
||||
from vllm_omni.quantization.factory import resolve_quant_config_from_disk
|
||||
from vllm_omni.quantization.mxfp4_config import DiffusionMXFP4Config
|
||||
|
||||
active = DiffusionMXFP4Config(is_checkpoint_mxfp4_serialized=True, require_smooth_scale=True)
|
||||
disk = {
|
||||
"quant_method": "mxfp4",
|
||||
"is_checkpoint_mxfp4_serialized": True,
|
||||
"ignored_layers": ["blocks.0.attn2.to_q"],
|
||||
}
|
||||
if disk_requirement is not None:
|
||||
disk["require_smooth_scale"] = disk_requirement
|
||||
resolved = resolve_quant_config_from_disk(active, disk)
|
||||
assert resolved.require_smooth_scale
|
||||
assert resolved.ignored_layers == disk["ignored_layers"]
|
||||
|
||||
|
||||
@pytest.mark.parametrize("disk_qc", ["mxfp4", {"quant_method": "mxfp4"}])
|
||||
def test_mxfp4_method_only_disk_config_preserves_caller_layer_policy(disk_qc):
|
||||
from vllm_omni.quantization.factory import resolve_quant_config_from_disk
|
||||
from vllm_omni.quantization.mxfp4_config import DiffusionMXFP4Config
|
||||
|
||||
active = DiffusionMXFP4Config(
|
||||
ignored_layers=["blocks.0.attn2.to_q"],
|
||||
w4a8_fallback_steps=[1],
|
||||
w4a8_fallback_layers=["blocks.1.attn1.to_qkv"],
|
||||
)
|
||||
|
||||
resolved = resolve_quant_config_from_disk(active, disk_qc)
|
||||
|
||||
assert resolved is active
|
||||
assert resolved.ignored_layers == ["blocks.0.attn2.to_q"]
|
||||
assert resolved.w4a8_fallback_steps == [1]
|
||||
assert resolved.w4a8_fallback_layers == ["blocks.1.attn1.to_qkv"]
|
||||
|
||||
|
||||
def test_quantization_disk_string_rejects_mismatched_active_method():
|
||||
from vllm_omni.quantization import build_quant_config
|
||||
from vllm_omni.quantization.factory import resolve_quant_config_from_disk
|
||||
|
||||
active = build_quant_config("mxfp4")
|
||||
|
||||
with pytest.raises(ValueError, match="mxfp4_dualscale.*active quantization config is 'mxfp4'"):
|
||||
resolve_quant_config_from_disk(active, "mxfp4_dualscale")
|
||||
|
||||
|
||||
def test_quantization_disk_string_accepts_equivalent_method_alias():
|
||||
from vllm_omni.quantization.factory import resolve_quant_config_from_disk
|
||||
|
||||
class ActiveAliasConfig:
|
||||
@staticmethod
|
||||
def get_name() -> str:
|
||||
return "inc"
|
||||
|
||||
active = ActiveAliasConfig()
|
||||
|
||||
assert resolve_quant_config_from_disk(active, "auto-round") is active
|
||||
|
||||
|
||||
def test_mxfp4_config_from_config_ignored_layers():
|
||||
from vllm_omni.quantization.mxfp4_config import DiffusionMXFP4Config
|
||||
|
||||
@@ -57,6 +179,202 @@ def test_mxfp4_config_from_config_modules_to_not_convert_fallback():
|
||||
assert cfg.ignored_layers == ["proj_out"]
|
||||
|
||||
|
||||
@pytest.mark.parametrize("serialized", [False, True])
|
||||
@pytest.mark.parametrize(
|
||||
"policy",
|
||||
[
|
||||
{"w4a8_fallback_steps": [0]},
|
||||
{"w4a8_fallback_layers": ["blocks.10.attn1.to_qkv"]},
|
||||
{"w4a8_fallback_steps": [1], "w4a8_fallback_layers": ["blocks.10.attn1.to_qkv"]},
|
||||
],
|
||||
)
|
||||
def test_dualscale_rejects_w4a8_configuration(serialized, policy):
|
||||
from vllm_omni.quantization import build_quant_config
|
||||
from vllm_omni.quantization.factory import resolve_quant_config_from_disk
|
||||
from vllm_omni.quantization.mxfp4_config import DiffusionMXFP4DualScaleMixedConfig
|
||||
|
||||
config = {"quant_method": "mxfp4_dualscale", "is_checkpoint_serialized": serialized, **policy}
|
||||
with pytest.raises(ValueError, match="does not support W4A8"):
|
||||
DiffusionMXFP4DualScaleMixedConfig(is_checkpoint_serialized=serialized, **policy)
|
||||
with pytest.raises(ValueError, match="does not support W4A8"):
|
||||
DiffusionMXFP4DualScaleMixedConfig.from_config(config)
|
||||
with pytest.raises(ValueError, match="does not support W4A8"):
|
||||
build_quant_config({"transformer": config})
|
||||
with pytest.raises(ValueError, match="does not support W4A8"):
|
||||
resolve_quant_config_from_disk(None, config)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("method", ["mxfp4", "mxfp4_dualscale"])
|
||||
@pytest.mark.parametrize("steps", [[-1], [True], [1.5], ["2"], "0,2", (0, 2)])
|
||||
def test_mxfp4_rejects_invalid_fallback_steps(method, steps):
|
||||
from vllm_omni.quantization import build_quant_config
|
||||
|
||||
with pytest.raises(ValueError, match="non-negative integer"):
|
||||
build_quant_config({"method": method, "w4a8_fallback_steps": steps})
|
||||
|
||||
|
||||
def test_mxfp4_fallback_steps_follow_request_context():
|
||||
from vllm_omni.diffusion.forward_context import (
|
||||
ForwardContext,
|
||||
override_forward_context,
|
||||
set_forward_context_denoise_step_idx,
|
||||
)
|
||||
from vllm_omni.quantization import build_quant_config
|
||||
from vllm_omni.quantization.mxfp4_config import _is_w4a8_fallback_step
|
||||
|
||||
steps = [2, 0, 2]
|
||||
cfg = build_quant_config({"method": "mxfp4", "w4a8_fallback_steps": steps})
|
||||
steps.append(1)
|
||||
assert cfg.w4a8_fallback_steps == [0, 2]
|
||||
with override_forward_context(None):
|
||||
assert not _is_w4a8_fallback_step([])
|
||||
with pytest.raises(RuntimeError, match="publishes its denoise step"):
|
||||
_is_w4a8_fallback_step(cfg.w4a8_fallback_steps)
|
||||
for _request in range(2):
|
||||
with override_forward_context(ForwardContext()):
|
||||
for step, expected in [(0, True), (1, False), (2, True), (3, False)]:
|
||||
set_forward_context_denoise_step_idx(step)
|
||||
# Both CFG passes and both experts read the same index.
|
||||
assert [_is_w4a8_fallback_step(cfg.w4a8_fallback_steps) for _cfg_pass in range(2)] == [
|
||||
expected,
|
||||
expected,
|
||||
]
|
||||
with pytest.raises(RuntimeError, match="publishes its denoise step"):
|
||||
_is_w4a8_fallback_step(cfg.w4a8_fallback_steps)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("method,flag", [("mxfp4", "is_checkpoint_mxfp4_serialized")])
|
||||
@pytest.mark.parametrize("steps", [[], [0, 2]])
|
||||
@pytest.mark.parametrize("already_serialized", [False, True])
|
||||
def test_mxfp4_disk_rebuild_preserves_runtime_fallback(method, flag, steps, already_serialized):
|
||||
from vllm_omni.quantization import build_quant_config
|
||||
from vllm_omni.quantization.factory import resolve_quant_config_from_disk
|
||||
|
||||
layers = ["blocks.10.attn1.to_qkv"] if steps else []
|
||||
active = build_quant_config(
|
||||
{
|
||||
"method": method,
|
||||
flag: already_serialized,
|
||||
"w4a8_fallback_steps": steps,
|
||||
"w4a8_fallback_layers": layers,
|
||||
}
|
||||
)
|
||||
# Separate experts may have distinct high precision layers. Both must
|
||||
# retain the requested step policy during online/offline reconciliation.
|
||||
for ignored in [["blocks.0.attn1.to_q"], ["blocks.1.attn1.to_q"]]:
|
||||
resolved = resolve_quant_config_from_disk(
|
||||
active,
|
||||
{
|
||||
"quant_method": method,
|
||||
flag: True,
|
||||
"ignored_layers": ignored,
|
||||
"w4a8_fallback_steps": [9],
|
||||
"w4a8_fallback_layers": ["blocks.9.ffn.net_2"],
|
||||
},
|
||||
)
|
||||
assert getattr(resolved, flag)
|
||||
assert resolved.ignored_layers == ignored
|
||||
assert resolved.w4a8_fallback_steps == steps
|
||||
assert resolved.w4a8_fallback_layers == layers
|
||||
assert active.ignored_layers == []
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"runtime_policy",
|
||||
[{}, {"w4a8_fallback_steps": [1]}, {"w4a8_fallback_layers": ["blocks.10.attn1.to_qkv"]}],
|
||||
)
|
||||
def test_mxfp4_explicit_omission_disables_saved_policies_for_both_experts(runtime_policy):
|
||||
from vllm_omni.quantization import build_quant_config
|
||||
from vllm_omni.quantization.factory import resolve_quant_config_from_disk
|
||||
|
||||
active = build_quant_config({"method": "mxfp4", **runtime_policy})
|
||||
resolved_experts = []
|
||||
for ignored in (["blocks.0.attn2.to_q"], ["blocks.1.ffn.net_2"]):
|
||||
disk = {
|
||||
"quant_method": "mxfp4",
|
||||
"is_checkpoint_mxfp4_serialized": True,
|
||||
"ignored_layers": ignored,
|
||||
"w4a8_fallback_steps": [9],
|
||||
"w4a8_fallback_layers": ["blocks.9.ffn.net_2"],
|
||||
}
|
||||
resolved = resolve_quant_config_from_disk(active, disk)
|
||||
assert resolved.w4a8_fallback_steps == runtime_policy.get("w4a8_fallback_steps", [])
|
||||
assert resolved.w4a8_fallback_layers == runtime_policy.get("w4a8_fallback_layers", [])
|
||||
assert resolved.ignored_layers == ignored
|
||||
detected = resolve_quant_config_from_disk(None, disk)
|
||||
assert detected.w4a8_fallback_steps == [9]
|
||||
assert detected.w4a8_fallback_layers == ["blocks.9.ffn.net_2"]
|
||||
resolved_experts.append(resolved)
|
||||
assert resolved_experts[0] is not resolved_experts[1]
|
||||
assert active.is_checkpoint_mxfp4_serialized is False
|
||||
assert active.ignored_layers == []
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"method,flag", [("mxfp4", "is_checkpoint_mxfp4_serialized"), ("mxfp4_dualscale", "is_checkpoint_serialized")]
|
||||
)
|
||||
def test_serialized_mxfp4_disk_omission_clears_active_ignored_layers(method, flag):
|
||||
from vllm_omni.quantization import build_quant_config
|
||||
from vllm_omni.quantization.factory import resolve_quant_config_from_disk
|
||||
|
||||
active_spec = {
|
||||
"method": method,
|
||||
flag: True,
|
||||
"ignored_layers": ["blocks.0.attn2.to_q"],
|
||||
"w4a8_fallback_steps": [1] if method == "mxfp4" else [],
|
||||
"w4a8_fallback_layers": ["blocks.1.attn1.to_qkv"] if method == "mxfp4" else [],
|
||||
}
|
||||
if method == "mxfp4":
|
||||
active_spec["require_smooth_scale"] = True
|
||||
active = build_quant_config(active_spec)
|
||||
|
||||
resolved = resolve_quant_config_from_disk(active, {"quant_method": method, flag: True})
|
||||
|
||||
assert resolved is not active
|
||||
assert resolved.ignored_layers == []
|
||||
assert resolved.w4a8_fallback_steps == active.w4a8_fallback_steps
|
||||
assert resolved.w4a8_fallback_layers == active.w4a8_fallback_layers
|
||||
if method == "mxfp4":
|
||||
assert resolved.require_smooth_scale
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"method,flag", [("mxfp4", "is_checkpoint_mxfp4_serialized"), ("mxfp4_dualscale", "is_checkpoint_serialized")]
|
||||
)
|
||||
@pytest.mark.parametrize("with_active", [False, True])
|
||||
def test_serialized_mxfp4_disk_uses_modules_to_not_convert(method, flag, with_active):
|
||||
from vllm_omni.quantization import build_quant_config
|
||||
from vllm_omni.quantization.factory import resolve_quant_config_from_disk
|
||||
|
||||
active = None
|
||||
disk_steps = [9] if method == "mxfp4" else []
|
||||
active_steps = [1] if method == "mxfp4" else []
|
||||
expected_steps = disk_steps
|
||||
if with_active:
|
||||
active = build_quant_config(
|
||||
{
|
||||
"method": method,
|
||||
flag: True,
|
||||
"ignored_layers": ["blocks.0.attn2.to_q"],
|
||||
"w4a8_fallback_steps": active_steps,
|
||||
}
|
||||
)
|
||||
expected_steps = active_steps
|
||||
|
||||
resolved = resolve_quant_config_from_disk(
|
||||
active,
|
||||
{
|
||||
"quant_method": method,
|
||||
flag: True,
|
||||
"modules_to_not_convert": ["blocks.2.ffn.net_2"],
|
||||
"w4a8_fallback_steps": disk_steps,
|
||||
},
|
||||
)
|
||||
|
||||
assert resolved.ignored_layers == ["blocks.2.ffn.net_2"]
|
||||
assert resolved.w4a8_fallback_steps == expected_steps
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# build_quant_config integration
|
||||
# ---------------------------------------------------------------------------
|
||||
@@ -594,18 +912,31 @@ def test_rocm_online_dispatch_returns_rocm_method(_rocm_platform, monkeypatch):
|
||||
assert isinstance(method, ROCmMxfp4OnlineLinearMethod)
|
||||
|
||||
|
||||
def test_rocm_ignored_layer_returns_unquantized(_rocm_platform, monkeypatch):
|
||||
@pytest.mark.parametrize("steps", [[], [0, 2]])
|
||||
def test_rocm_ignored_layer_returns_unquantized(_rocm_platform, monkeypatch, steps):
|
||||
"""A prefix in ignored_layers must return UnquantizedLinearMethod before the gfx950 probe."""
|
||||
from vllm.model_executor.layers.linear import UnquantizedLinearMethod
|
||||
|
||||
from vllm_omni.quantization.mxfp4_config import DiffusionMXFP4Config
|
||||
|
||||
_patch_gcn_arch(monkeypatch, "gfx950:sramecc+:xnack-")
|
||||
cfg = DiffusionMXFP4Config(is_checkpoint_mxfp4_serialized=False, ignored_layers=["proj_out"])
|
||||
cfg = DiffusionMXFP4Config(
|
||||
is_checkpoint_mxfp4_serialized=False, ignored_layers=["proj_out"], w4a8_fallback_steps=steps
|
||||
)
|
||||
|
||||
assert isinstance(cfg.get_quant_method(_fake_linear_layer(), "proj_out"), UnquantizedLinearMethod)
|
||||
|
||||
|
||||
def test_rocm_rejects_w4a8_step_fallback(_rocm_platform, monkeypatch):
|
||||
from vllm_omni.quantization.mxfp4_config import DiffusionMXFP4Config
|
||||
|
||||
_patch_gcn_arch(monkeypatch, "gfx950:sramecc+:xnack-")
|
||||
cfg = DiffusionMXFP4Config(w4a8_fallback_steps=[0, 2])
|
||||
|
||||
with pytest.raises(NotImplementedError, match="only supported on NPU"):
|
||||
cfg.get_quant_method(_fake_linear_layer(), "blocks.0.attn1.to_q")
|
||||
|
||||
|
||||
def test_rocm_non_gfx950_raises(_rocm_platform, monkeypatch):
|
||||
"""MXFP4 on ROCm requires gfx950; any other arch must raise."""
|
||||
from vllm_omni.quantization.mxfp4_config import DiffusionMXFP4Config
|
||||
@@ -652,3 +983,221 @@ def test_rocm_create_weights_row_parallel_tp2(_rocm_platform):
|
||||
assert layer.weight.shape == (_TP2_N, _TP2_K // _TP2)
|
||||
assert layer.input_size_per_partition == _TP2_K // _TP2
|
||||
assert layer.output_size_per_partition == _TP2_N
|
||||
|
||||
|
||||
@pytest.mark.parametrize("method", ["mxfp4", "mxfp4_dualscale"])
|
||||
@pytest.mark.parametrize("layers", ["blocks.1.attn1.to_qkv", [1], [True], [""], [" a"], ["blocks.*"], ["a..b"]])
|
||||
def test_mxfp4_rejects_invalid_fallback_layers(method, layers):
|
||||
from vllm_omni.quantization import build_quant_config
|
||||
|
||||
with pytest.raises(ValueError, match="exact runtime Linear paths"):
|
||||
build_quant_config({"method": method, "w4a8_fallback_layers": layers})
|
||||
|
||||
|
||||
@pytest.mark.parametrize("method", ["mxfp4"])
|
||||
def test_mxfp4_layer_policy_from_checkpoint_and_component_config(method):
|
||||
from vllm_omni.quantization import build_quant_config
|
||||
from vllm_omni.quantization.component_config import resolve_component_quant_config
|
||||
from vllm_omni.quantization.factory import resolve_quant_config_from_disk
|
||||
|
||||
names = ["blocks.10.attn1.to_qkv", "blocks.10.attn1.to_qkv"]
|
||||
components = build_quant_config(
|
||||
{"transformer": {"method": method, "w4a8_fallback_layers": names}, "transformer_2": {"method": method}}
|
||||
)
|
||||
names.append("blocks.11.attn1.to_qkv")
|
||||
assert resolve_component_quant_config(components, "transformer").w4a8_fallback_layers == [names[0]]
|
||||
assert resolve_component_quant_config(components, "transformer_2").w4a8_fallback_layers == []
|
||||
detected = resolve_quant_config_from_disk(None, {"quant_method": method, "w4a8_fallback_layers": [names[0]]})
|
||||
assert detected.w4a8_fallback_layers == [names[0]]
|
||||
assert type(detected).from_config({"w4a8_fallback_layers": [names[0]]}).w4a8_fallback_layers == [names[0]]
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"kind, serialized_flag",
|
||||
[("mxfp4", "is_checkpoint_mxfp4_serialized")],
|
||||
)
|
||||
@pytest.mark.parametrize("offline", [False, True])
|
||||
@pytest.mark.parametrize("scale_alg", [0, 2])
|
||||
def test_mxfp4_layer_and_step_union_routes_actual_methods(
|
||||
kind, serialized_flag, offline, scale_alg, monkeypatch, mocker
|
||||
):
|
||||
from vllm.model_executor.layers.linear import LinearBase, UnquantizedLinearMethod
|
||||
|
||||
from vllm_omni.diffusion.forward_context import ForwardContext, override_forward_context
|
||||
from vllm_omni.quantization import build_quant_config, mxfp4_config
|
||||
|
||||
monkeypatch.setattr(mxfp4_config.current_omni_platform, "is_npu", lambda: True)
|
||||
selected = "blocks.10.attn1.to_qkv"
|
||||
other = "blocks.11.attn2.to_q"
|
||||
preserved = "blocks.12.ffn.net_2"
|
||||
config = build_quant_config(
|
||||
{
|
||||
"method": kind,
|
||||
"mxfp4_scale_alg": scale_alg,
|
||||
serialized_flag: offline,
|
||||
"w4a8_fallback_layers": [selected, preserved],
|
||||
"ignored_layers": [preserved],
|
||||
"w4a8_fallback_steps": [1],
|
||||
}
|
||||
)
|
||||
# Real config dispatch creates separate methods for selected and other layers.
|
||||
linear = mocker.Mock(spec=LinearBase)
|
||||
assert isinstance(config.get_quant_method(linear, preserved), UnquantizedLinearMethod)
|
||||
x = torch.ones(2, 64, dtype=torch.bfloat16)
|
||||
layer = torch.nn.Module()
|
||||
for name in ("weight", "weight_scale", "mul_scale"):
|
||||
layer.register_buffer(name, torch.ones(64, dtype=x.dtype))
|
||||
calls = []
|
||||
|
||||
def record_a8(*args):
|
||||
calls.append("A8")
|
||||
return x
|
||||
|
||||
def record_a4(*args):
|
||||
calls.append("A4")
|
||||
return x
|
||||
|
||||
monkeypatch.setattr(mxfp4_config, "_npu_w4a8_matmul", record_a8)
|
||||
for name in (selected, other):
|
||||
method = config.get_quant_method(linear, name)
|
||||
assert method.is_w4a8_fallback_layer == (name == selected)
|
||||
scales = (x, x)
|
||||
monkeypatch.setattr(method, "_quantize_activation", lambda *args: scales)
|
||||
monkeypatch.setattr(method, "_quant_matmul", record_a4)
|
||||
for step in (0, 1, 2):
|
||||
with override_forward_context(ForwardContext(denoise_step_idx=step)):
|
||||
method.apply(layer, x)
|
||||
assert calls == ["A8", "A8", "A8", "A4", "A8", "A4"]
|
||||
# A permanently selected layer works without a request context, even if steps are also configured.
|
||||
with override_forward_context(None):
|
||||
config.get_quant_method(linear, selected).apply(layer, x)
|
||||
assert calls[-1] == "A8"
|
||||
|
||||
|
||||
@pytest.mark.parametrize("online", [False, True])
|
||||
@pytest.mark.parametrize("selected", [False, True])
|
||||
@pytest.mark.parametrize("scale_alg", [0, 2])
|
||||
def test_single_scale_prepares_one_weight_for_both_activation_precisions(online, selected, scale_alg, monkeypatch):
|
||||
import sys
|
||||
from types import SimpleNamespace
|
||||
|
||||
from vllm_omni.diffusion.forward_context import ForwardContext, override_forward_context
|
||||
from vllm_omni.quantization import mxfp4_config
|
||||
|
||||
monkeypatch.setattr(torch.Tensor, "npu", lambda self: self, raising=False)
|
||||
prepared = []
|
||||
packed = []
|
||||
|
||||
def quantize(x, **kwargs):
|
||||
prepared.append(kwargs)
|
||||
return x.to(torch.int8), torch.full((x.shape[0], 2), 127, dtype=torch.uint8)
|
||||
|
||||
def pack(x, dtype):
|
||||
packed.append(x)
|
||||
return x.to(dtype)
|
||||
|
||||
fake_npu = SimpleNamespace(
|
||||
float4_e2m1fn_x2=torch.int8,
|
||||
float8_e8m0fnu=torch.uint8,
|
||||
float8_e4m3fn=torch.float8_e4m3fn,
|
||||
npu_dtype_cast=pack,
|
||||
npu_dynamic_mx_quant=quantize,
|
||||
)
|
||||
monkeypatch.setitem(sys.modules, "torch_npu", fake_npu)
|
||||
config = mxfp4_config.DiffusionMXFP4Config(
|
||||
is_checkpoint_mxfp4_serialized=not online,
|
||||
w4a8_fallback_layers=["blocks.0.attn1.to_qkv"] if selected else [],
|
||||
w4a8_fallback_steps=[1],
|
||||
mxfp4_scale_alg=scale_alg,
|
||||
)
|
||||
cls = mxfp4_config.NPUMxfp4OnlineLinearMethod if online else mxfp4_config.NPUMxfp4LinearMethod
|
||||
method = cls(config, prefix="blocks.0.attn1.to_qkv")
|
||||
layer = torch.nn.Linear(64, 2, bias=False, dtype=torch.bfloat16)
|
||||
layer.orig_dtype = torch.bfloat16
|
||||
if not online:
|
||||
layer.register_parameter("weight_scale", torch.nn.Parameter(torch.full((2, 2), 127, dtype=torch.uint8), False))
|
||||
layer.register_parameter("mul_scale", torch.nn.Parameter(torch.linspace(0.5, 2, 64), False))
|
||||
method.process_weights_after_loading(layer)
|
||||
pointers = (layer.weight.data_ptr(), layer.weight_scale.data_ptr())
|
||||
method.process_weights_after_loading(layer)
|
||||
assert len(prepared) == int(online)
|
||||
expected = dict(dst_type=torch.int8, axis=-1, block_size=32, round_mode="rint", scale_alg=scale_alg)
|
||||
if scale_alg == 2:
|
||||
expected["dst_type_max"] = 7.25
|
||||
assert prepared == ([expected] if online else [])
|
||||
assert len(packed) == int(not online)
|
||||
branches = []
|
||||
|
||||
def a4(x_q, x_scale, target, bias, dtype):
|
||||
assert (target.weight.data_ptr(), target.weight_scale.data_ptr()) == pointers
|
||||
branches.append("A4")
|
||||
return torch.zeros(x_q.shape[0], 2, dtype=dtype)
|
||||
|
||||
def a8(x, weight, scale, bias, dtype):
|
||||
assert (weight.data_ptr(), scale.data_ptr()) == pointers
|
||||
branches.append("A8")
|
||||
return torch.zeros(x.shape[0], 2, dtype=dtype)
|
||||
|
||||
monkeypatch.setattr(method, "_quant_matmul", a4)
|
||||
monkeypatch.setattr(mxfp4_config, "_npu_w4a8_matmul", a8)
|
||||
monkeypatch.setattr(method, "_quantize_activation", lambda x: (x, x))
|
||||
monkeypatch.setattr(fake_npu, "npu_dynamic_mx_quant", lambda *a, **kw: pytest.fail("weight requantization"))
|
||||
monkeypatch.setattr(fake_npu, "npu_dtype_cast", lambda *a, **kw: pytest.fail("weight repacking"))
|
||||
for step in (0, 1, 2):
|
||||
with override_forward_context(ForwardContext(denoise_step_idx=step)):
|
||||
method.apply(layer, torch.ones(2, 64, dtype=torch.bfloat16))
|
||||
assert branches == (["A8"] * 3 if selected else ["A4", "A8", "A4"])
|
||||
assert (layer.weight.data_ptr(), layer.weight_scale.data_ptr()) == pointers
|
||||
assert set(layer.state_dict()) == (
|
||||
{"weight", "weight_scale"} if online else {"weight", "weight_scale", "mul_scale"}
|
||||
)
|
||||
assert not hasattr(layer, "w4a8_weight")
|
||||
assert not hasattr(layer, "w4a8_weight_scale")
|
||||
|
||||
|
||||
@pytest.mark.parametrize("online", [False, True])
|
||||
def test_dualscale_without_fallback_keeps_original_a4_path(online, monkeypatch):
|
||||
import sys
|
||||
from types import SimpleNamespace
|
||||
|
||||
from vllm_omni.quantization import mxfp4_config
|
||||
|
||||
monkeypatch.setattr(torch.Tensor, "npu", lambda self: self, raising=False)
|
||||
original_ones = torch.ones
|
||||
monkeypatch.setattr(torch, "ones", lambda *args, **kw: original_ones(*args, **{**kw, "device": "cpu"}))
|
||||
calls = []
|
||||
|
||||
def quantize_dual(x, **kw):
|
||||
calls.append(kw.get("smooth_scale"))
|
||||
return x.to(torch.int8), torch.ones(x.shape[0], 1), torch.ones(x.shape[0], 16, dtype=torch.uint8)
|
||||
|
||||
fake_npu = SimpleNamespace(
|
||||
float4_e2m1fn_x2=torch.int8,
|
||||
npu_dtype_cast=lambda x, dtype: x.to(dtype),
|
||||
npu_format_cast=lambda x, *args, **kw: x,
|
||||
npu_dynamic_dual_level_mx_quant=quantize_dual,
|
||||
)
|
||||
monkeypatch.setitem(sys.modules, "torch_npu", fake_npu)
|
||||
config = mxfp4_config.DiffusionMXFP4DualScaleMixedConfig(
|
||||
is_checkpoint_serialized=not online, w4a8_fallback_steps=[], w4a8_fallback_layers=[]
|
||||
)
|
||||
cls = mxfp4_config.NPUMxfp4DualScaleOnlineLinearMethod if online else mxfp4_config.NPUMxfp4DualScaleLinearMethod
|
||||
method = cls(config)
|
||||
layer = torch.nn.Module()
|
||||
layer.input_size_per_partition = 512
|
||||
for name, value in {
|
||||
"weight": torch.ones(2, 512, dtype=torch.bfloat16),
|
||||
"weight_scale": torch.full((2, 16), 127, dtype=torch.uint8),
|
||||
"weight_dual_scale": torch.ones(2, 1, 1),
|
||||
"mul_scale": torch.ones(512),
|
||||
}.items():
|
||||
layer.register_parameter(name, torch.nn.Parameter(value, requires_grad=False))
|
||||
method.process_weights_after_loading(layer)
|
||||
method.process_weights_after_loading(layer)
|
||||
assert len(calls) == int(online)
|
||||
assert set(layer.state_dict()) == {"weight", "weight_scale", "weight_dual_scale", "mul_scale"}
|
||||
assert not hasattr(layer, "w4a8_weight")
|
||||
monkeypatch.setattr(method, "_quant_matmul", lambda *args: torch.zeros(2, 2, dtype=torch.bfloat16))
|
||||
method.apply(layer, torch.ones(2, 512, dtype=torch.bfloat16))
|
||||
assert len(calls) == 1 + int(online)
|
||||
torch.testing.assert_close(calls[-1], layer.mul_scale)
|
||||
@@ -0,0 +1,237 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
|
||||
"""Exercise Wan's real name mapping and vLLM's QKV parameter loaders on CPU."""
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
from vllm.config.load import LoadConfig
|
||||
from vllm.model_executor.layers.linear import QKVParallelLinear
|
||||
|
||||
from vllm_omni.diffusion.data import OmniDiffusionConfig
|
||||
from vllm_omni.diffusion.model_loader.diffusers_loader import DiffusersPipelineLoader
|
||||
from vllm_omni.diffusion.models.wan2_2.wan2_2_transformer import WanTransformer3DModel
|
||||
from vllm_omni.quantization.mxfp4_config import DiffusionMXFP4Config, DiffusionMXFP4DualScaleMixedConfig
|
||||
|
||||
pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
|
||||
|
||||
|
||||
@pytest.fixture(autouse=True)
|
||||
def _patch_tp_state(monkeypatch):
|
||||
monkeypatch.setattr("vllm.model_executor.parameter.get_tensor_model_parallel_rank", lambda: 0)
|
||||
monkeypatch.setattr("vllm.model_executor.parameter.get_tensor_model_parallel_world_size", lambda: 1)
|
||||
monkeypatch.setattr("vllm_omni.platforms.current_omni_platform.is_npu", lambda: True)
|
||||
prefix = "vllm_omni.diffusion.models.wan2_2.wan2_2_transformer"
|
||||
monkeypatch.setattr(f"{prefix}.get_tensor_model_parallel_rank", lambda: 0)
|
||||
monkeypatch.setattr(f"{prefix}.get_tensor_model_parallel_world_size", lambda: 1)
|
||||
|
||||
|
||||
class _WanCheckpoint(torch.nn.Module):
|
||||
load_weights = WanTransformer3DModel.load_weights
|
||||
_log_w4a8_fallback_load_summary = WanTransformer3DModel._log_w4a8_fallback_load_summary
|
||||
|
||||
def __init__(self, config, *, num_blocks=1):
|
||||
super().__init__()
|
||||
blocks = []
|
||||
for block_idx in range(num_blocks):
|
||||
block = torch.nn.Module()
|
||||
block.attn1 = torch.nn.Module()
|
||||
block.attn1.to_qkv = QKVParallelLinear(
|
||||
hidden_size=512,
|
||||
head_size=32,
|
||||
total_num_heads=4,
|
||||
bias=False,
|
||||
params_dtype=torch.bfloat16,
|
||||
quant_config=config,
|
||||
prefix=f"blocks.{block_idx}.attn1.to_qkv",
|
||||
disable_tp=True,
|
||||
)
|
||||
blocks.append(block)
|
||||
self.blocks = torch.nn.ModuleList(blocks)
|
||||
|
||||
|
||||
def _checkpoint(mode, fused):
|
||||
if mode == "single":
|
||||
config = DiffusionMXFP4Config(is_checkpoint_mxfp4_serialized=True, require_smooth_scale=True)
|
||||
elif mode == "dual":
|
||||
config = DiffusionMXFP4DualScaleMixedConfig(is_checkpoint_serialized=True)
|
||||
else:
|
||||
config = None
|
||||
model = _WanCheckpoint(config)
|
||||
weights = {}
|
||||
expected = {}
|
||||
for name, param in model.named_parameters():
|
||||
suffix = name.rsplit(".", 1)[1]
|
||||
output_dim = getattr(param, "output_dim", None)
|
||||
if output_dim is None:
|
||||
value = torch.full_like(param, 2.5)
|
||||
shards = [value] * 3
|
||||
else:
|
||||
parts = param.chunk(3, dim=output_dim)
|
||||
shards = [
|
||||
torch.full_like(part, (121 if suffix == "weight_scale" else 1) + i) for i, part in enumerate(parts)
|
||||
]
|
||||
value = torch.cat(shards, dim=output_dim)
|
||||
expected[name] = value
|
||||
if fused:
|
||||
weights[name] = value
|
||||
else:
|
||||
for shard_id, shard in zip(("q", "k", "v"), shards):
|
||||
weights[f"blocks.0.attn1.to_{shard_id}.{suffix}"] = shard
|
||||
return model, config, weights, expected
|
||||
|
||||
|
||||
def _load(model, config, weights, monkeypatch):
|
||||
loader = DiffusersPipelineLoader(
|
||||
LoadConfig(), OmniDiffusionConfig(model="", dtype=torch.bfloat16, quantization_config=config)
|
||||
)
|
||||
loader.counter_before_loading_weights = 0.0
|
||||
monkeypatch.setattr(loader, "get_all_weights", lambda model: iter(weights.items()))
|
||||
loader.load_weights(model)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("mode", ["single", "dual", "bf16"])
|
||||
@pytest.mark.parametrize("fused", [False, True])
|
||||
def test_wan_qkv_loads_actual_weights_and_scales(mode, fused, monkeypatch):
|
||||
model, config, weights, expected = _checkpoint(mode, fused)
|
||||
_load(model, config, weights, monkeypatch)
|
||||
for name, param in model.named_parameters():
|
||||
torch.testing.assert_close(param, expected[name], rtol=0, atol=0)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("mode", ["single", "dual", "bf16"])
|
||||
def test_wan_incomplete_qkv_weight_is_not_marked_fully_loaded(mode, monkeypatch):
|
||||
model, config, weights, _ = _checkpoint(mode, fused=False)
|
||||
weights.pop("blocks.0.attn1.to_k.weight")
|
||||
weights.pop("blocks.0.attn1.to_v.weight")
|
||||
with pytest.raises(ValueError, match="to_qkv.weight"):
|
||||
_load(model, config, weights, monkeypatch)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("mode", ["single", "dual"])
|
||||
def test_wan_incomplete_qkv_scale_is_not_marked_fully_loaded(mode, monkeypatch):
|
||||
model, config, weights, _ = _checkpoint(mode, fused=False)
|
||||
weights.pop("blocks.0.attn1.to_k.weight_scale")
|
||||
weights.pop("blocks.0.attn1.to_v.weight_scale")
|
||||
with pytest.raises(ValueError, match="to_qkv.weight_scale"):
|
||||
_load(model, config, weights, monkeypatch)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("mode", ["single", "dual"])
|
||||
@pytest.mark.parametrize("reverse", [False, True])
|
||||
def test_wan_split_qkv_rejects_different_smooth_scales(mode, reverse, monkeypatch):
|
||||
model, config, weights, _ = _checkpoint(mode, fused=False)
|
||||
weights["blocks.0.attn1.to_k.mul_scale"] = weights["blocks.0.attn1.to_k.mul_scale"] * 2
|
||||
if reverse:
|
||||
weights = dict(reversed(weights.items()))
|
||||
with pytest.raises(ValueError, match="must share the same Smooth tensor"):
|
||||
_load(model, config, weights, monkeypatch)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("name", ["to_q.missing", "to_qkv.missing", "to_qkv_extra.weight", "to_query.weight"])
|
||||
def test_wan_unknown_qkv_key_is_not_reported_as_loaded(name):
|
||||
model, _, _, _ = _checkpoint("single", fused=True)
|
||||
loaded = model.load_weights([(f"blocks.0.attn1.{name}", torch.ones(2))])
|
||||
assert loaded == set()
|
||||
|
||||
|
||||
@pytest.mark.parametrize("fused", [False, True])
|
||||
def test_wan_single_scale_rejects_relabelled_dualscale_checkpoint(fused, monkeypatch):
|
||||
model, config, weights, _ = _checkpoint("single", fused=fused)
|
||||
projection = "to_qkv" if fused else "to_q"
|
||||
weights[f"blocks.0.attn1.{projection}.weight_dual_scale"] = torch.ones(128, 1, 1)
|
||||
with pytest.raises(ValueError, match="cannot load DualScale tensor"):
|
||||
_load(model, config, weights, monkeypatch)
|
||||
|
||||
|
||||
def test_wan_layer_policy_validates_runtime_qkv_and_bf16_override():
|
||||
config = DiffusionMXFP4Config(w4a8_fallback_layers=["blocks.0.attn1.to_qkv"])
|
||||
model = _WanCheckpoint(config)
|
||||
# Uses the same validation called at the end of WanTransformer3DModel construction.
|
||||
WanTransformer3DModel._validate_w4a8_fallback_layers(model, config)
|
||||
for bad in ("blocks.0.attn1.to_q", "blocks.0.attn1", "blocks.99.attn1.to_qkv", "transformer.blocks.0.attn1.to_qkv"):
|
||||
config.w4a8_fallback_layers = [bad]
|
||||
with pytest.raises(ValueError, match="exact runtime Linear paths"):
|
||||
WanTransformer3DModel._validate_w4a8_fallback_layers(model, config)
|
||||
|
||||
|
||||
def test_wan_layer_policy_defers_only_other_pp_stage_paths():
|
||||
from vllm.model_executor.models.utils import PPMissingLayer
|
||||
|
||||
config = DiffusionMXFP4Config(w4a8_fallback_layers=["blocks.0.attn1.to_qkv", "blocks.1.attn1.to_qkv"])
|
||||
model = _WanCheckpoint(config)
|
||||
model.blocks.append(PPMissingLayer())
|
||||
WanTransformer3DModel._validate_w4a8_fallback_layers(model, config)
|
||||
config.w4a8_fallback_layers = ["blocks.1"]
|
||||
with pytest.raises(ValueError, match="exact runtime Linear paths"):
|
||||
WanTransformer3DModel._validate_w4a8_fallback_layers(model, config)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("is_checkpoint_serialized", [False, True], ids=["online", "offline"])
|
||||
def test_loader_logs_wan_expert_summaries_after_weight_processing(is_checkpoint_serialized, mocker):
|
||||
selected = "blocks.0.attn1.to_qkv"
|
||||
bf16 = "blocks.1.attn1.to_qkv"
|
||||
transformer_config = DiffusionMXFP4Config(
|
||||
is_checkpoint_mxfp4_serialized=is_checkpoint_serialized,
|
||||
w4a8_fallback_layers=[selected],
|
||||
)
|
||||
transformer_2_config = DiffusionMXFP4Config(
|
||||
is_checkpoint_mxfp4_serialized=is_checkpoint_serialized,
|
||||
ignored_layers=[bf16],
|
||||
w4a8_fallback_layers=[selected, bf16],
|
||||
)
|
||||
pipeline = torch.nn.Module()
|
||||
pipeline.transformer = _WanCheckpoint(transformer_config)
|
||||
pipeline.transformer_2 = _WanCheckpoint(transformer_2_config, num_blocks=2)
|
||||
WanTransformer3DModel._validate_w4a8_fallback_layers(pipeline.transformer, transformer_config)
|
||||
WanTransformer3DModel._validate_w4a8_fallback_layers(pipeline.transformer_2, transformer_2_config)
|
||||
|
||||
loader = DiffusersPipelineLoader(
|
||||
LoadConfig(),
|
||||
OmniDiffusionConfig(model="", dtype=torch.bfloat16, quantization_config=transformer_config),
|
||||
)
|
||||
events = []
|
||||
summaries = []
|
||||
|
||||
def process_weights(model, _target_device):
|
||||
events.append("process")
|
||||
for module in model.modules():
|
||||
method = getattr(module, "quant_method", None)
|
||||
if getattr(method, "is_w4a8_fallback_layer", False):
|
||||
module._already_called_process_weights_after_loading = True
|
||||
|
||||
def capture_summary(message, *args):
|
||||
rendered = message % args
|
||||
summaries.append(rendered)
|
||||
events.append(f"summary:{args[0]}")
|
||||
|
||||
loader._init_from_load_format = lambda *_args, **_kwargs: pipeline # type: ignore[method-assign]
|
||||
loader.load_weights = lambda _model: events.append("load") # type: ignore[method-assign]
|
||||
loader._process_weights_after_loading = process_weights # type: ignore[method-assign]
|
||||
loader._apply_skip_softmax_calibration = lambda _model: None # type: ignore[method-assign]
|
||||
mocker.patch(
|
||||
"vllm_omni.diffusion.models.wan2_2.wan2_2_transformer.logger.info",
|
||||
side_effect=capture_summary,
|
||||
)
|
||||
|
||||
assert loader.load_model(load_device="cpu") is pipeline
|
||||
assert events == ["load", "process", "summary:transformer", "summary:transformer_2"]
|
||||
assert "component=transformer; selected=1 ['transformer.blocks.0.attn1.to_qkv']" in summaries[0]
|
||||
assert "BF16 overrides=0 []" in summaries[0]
|
||||
assert "weight/A8 processing state=ready (1/1 selected layers processed)" in summaries[0]
|
||||
assert "component=transformer_2; selected=1 ['transformer_2.blocks.0.attn1.to_qkv']" in summaries[1]
|
||||
assert "BF16 overrides=1 ['transformer_2.blocks.1.attn1.to_qkv']" in summaries[1]
|
||||
assert "weight/A8 processing state=ready (1/1 selected layers processed)" in summaries[1]
|
||||
|
||||
|
||||
def test_wan_load_summary_does_not_claim_unprocessed_layer_is_ready(mocker):
|
||||
selected = "blocks.0.attn1.to_qkv"
|
||||
config = DiffusionMXFP4Config(w4a8_fallback_layers=[selected])
|
||||
model = _WanCheckpoint(config)
|
||||
WanTransformer3DModel._validate_w4a8_fallback_layers(model, config)
|
||||
logger_info = mocker.patch("vllm_omni.diffusion.models.wan2_2.wan2_2_transformer.logger.info")
|
||||
|
||||
WanTransformer3DModel._log_w4a8_fallback_load_summary(model, "transformer")
|
||||
|
||||
message, *args = logger_info.call_args.args
|
||||
rendered = message % tuple(args)
|
||||
assert "weight/A8 processing state=not-ready (0/1 selected layers processed)" in rendered
|
||||
Reference in new issue
Block a user