[Kernel] Add B12X dense linear backends (#52016)

Signed-off-by: Luke Alonso <lalonso@gmail.com>
This commit is contained in:
Luke Alonso
2026-08-13 19:48:33 -07:00
committed by GitHub
parent fe4c5dcd4c
commit 3c79b1a8bf
17 changed files with 2753 additions and 3 deletions
+24
View File
@@ -0,0 +1,24 @@
# B12X Linear Backend
[B12X](https://pypi.org/project/b12x/) provides optional CUDA kernels for
NVIDIA SM120 and SM121 GPUs. Install the dependency with:
```bash
uv pip install "vllm[b12x]"
```
B12X participates in automatic kernel selection after established optimized
backends and before emulation. Select it explicitly with:
```bash
vllm serve <model> --linear-backend b12x
```
## Supported Configurations
| Backend | Supported configurations |
| ------- | ------------------------ |
| Linear | Per-tensor FP8, 128x128 block FP8, MXFP8, NVFP4, and MXFP4 |
Dense W4A16 layers are not handled by B12X and continue to use another
compatible backend such as Marlin.
+1
View File
@@ -1291,6 +1291,7 @@ setup(
# only; also needs system GStreamer + libv4l (see docs).
"deepstream": ["nvidia-deepstream-videodecode-cu13>=9.0.2"],
"flashinfer": [], # Kept for backwards compatibility
"b12x": ["b12x==1.2.4"],
# Optional deps for Helion kernel development
# NOTE: When updating helion version, also update CI files:
# - .buildkite/test_areas/kernels.yaml
@@ -14,6 +14,10 @@ from tests.kernels.quant_utils import (
)
from tests.kernels.utils import fp8_ulp_distance
from vllm.config import VllmConfig
from vllm.model_executor.kernels.linear.scaled_mm.b12x_block import (
B12xFp8BlockScaledMMKernel,
_run_b12x_fp8_block_scaled_mm,
)
from vllm.model_executor.kernels.linear.scaled_mm.cutlass import cutlass_scaled_mm
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
per_token_group_quant_fp8,
@@ -353,3 +357,54 @@ def test_w8a8_block_fp8_flashinfer_matmul(M, N, K, block_size, out_dtype, seed):
torch.abs(out.to(torch.bfloat16) - ref_out.to(torch.bfloat16))
) / torch.mean(torch.abs(ref_out.to(torch.bfloat16)))
assert rel_diff < 0.001
@pytest.mark.parametrize(
"M,N,K",
[(1, 128, 256), (8, 256, 512), (129, 256, 256), (2, 4096, 4096)],
)
@torch.inference_mode()
def test_w8a8_block_fp8_b12x_matmul(M, N, K):
supported, reason = B12xFp8BlockScaledMMKernel.is_supported()
if not supported:
pytest.skip(reason)
torch.manual_seed(M)
fp8_max = torch.finfo(torch.float8_e4m3fn).max
A_bf16 = (torch.rand(M, K, dtype=torch.bfloat16) - 0.5) * 2 * fp8_max
B_bf16 = (torch.rand(N, K, dtype=torch.bfloat16) - 0.5) * 2 * fp8_max
A_fp8, As = per_token_group_quant_fp8(A_bf16, 128, use_ue8m0=False)
B_fp8, Bs = per_block_cast_to_fp8(
B_bf16,
block_size=[128, 128],
use_ue8m0=False,
)
As = As.float()
Bs = Bs.float()
ref_out = native_w8a8_block_matmul(
A_fp8,
B_fp8,
As,
Bs,
[128, 128],
torch.bfloat16,
)
out = _run_b12x_fp8_block_scaled_mm(
A_fp8,
B_fp8,
As,
Bs,
torch.bfloat16,
)
rel_diff = torch.mean(torch.abs(out.float() - ref_out.float())) / torch.mean(
torch.abs(ref_out.float())
)
cosine = torch.nn.functional.cosine_similarity(
out.float().flatten(),
ref_out.float().flatten(),
dim=0,
)
assert rel_diff < 0.002
assert cosine >= 0.9999
@@ -0,0 +1,199 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
import types
import torch
from vllm.model_executor.kernels.linear import (
_LINEAR_BACKEND_KERNEL_MAP,
_POSSIBLE_MXFP4_KERNELS,
init_mxfp4_linear_kernel,
)
from vllm.model_executor.kernels.linear.mxfp4.b12x import (
B12xMxFp4LinearKernel,
warmup_b12x_mxfp4_linear,
)
from vllm.model_executor.layers.quantization.utils.quant_utils import (
kMxfp4Dynamic,
)
from vllm.platforms import PlatformEnum
def test_b12x_backend_maps_mxfp4_kernel() -> None:
assert B12xMxFp4LinearKernel in _LINEAR_BACKEND_KERNEL_MAP["b12x"]
assert B12xMxFp4LinearKernel in _POSSIBLE_MXFP4_KERNELS[PlatformEnum.CUDA]
def test_b12x_mxfp4_fallback_priority() -> None:
kernels = _POSSIBLE_MXFP4_KERNELS[PlatformEnum.CUDA]
names = [kernel.__name__ for kernel in kernels]
assert (
names.index("HummingMxFp4LinearKernel")
< names.index("B12xMxFp4LinearKernel")
< names.index("EmulationMxfp4LinearKernel")
)
def test_b12x_mxfp4_explicit_backend_selects_native_kernel(monkeypatch) -> None:
import vllm.model_executor.kernels.linear as linear_mod
monkeypatch.setattr(linear_mod.current_platform, "_enum", PlatformEnum.CUDA)
monkeypatch.setattr(linear_mod, "_get_linear_backend", lambda: "b12x")
monkeypatch.setattr(
B12xMxFp4LinearKernel,
"is_supported",
classmethod(lambda cls, compute_capability=None: (True, None)),
)
monkeypatch.setattr(
B12xMxFp4LinearKernel,
"can_implement",
classmethod(lambda cls, config: (True, None)),
)
kernel = init_mxfp4_linear_kernel(activation_quant_key=kMxfp4Dynamic)
assert isinstance(kernel, B12xMxFp4LinearKernel)
def test_b12x_mxfp4_requires_dynamic_activations() -> None:
config = types.SimpleNamespace(activation_quant_key=kMxfp4Dynamic)
can_implement, reason = B12xMxFp4LinearKernel.can_implement(config)
assert can_implement
assert reason is None
config.activation_quant_key = None
can_implement, reason = B12xMxFp4LinearKernel.can_implement(config)
assert not can_implement
assert reason == "B12X MXFP4 GEMM requires dynamic MXFP4 activations"
def test_b12x_mxfp4_processes_scale_and_preserves_loader(monkeypatch) -> None:
import vllm.model_executor.kernels.linear.mxfp4.b12x as b12x_mod
scale = torch.empty((48, 8), dtype=torch.uint8)
swizzled_scale = torch.empty((128, 8), dtype=torch.uint8)
monkeypatch.setattr(
b12x_mod,
"_import_b12x_intrinsics",
lambda: types.SimpleNamespace(swizzle_block_scale=lambda value: swizzled_scale),
)
layer = torch.nn.Module()
layer.prefix = "model.layers.0.mlp.shared_expert.down_proj"
layer.weight_scale = torch.nn.Parameter(scale, requires_grad=False)
weight_loader = object()
layer.weight_scale.weight_loader = weight_loader
kernel = object.__new__(B12xMxFp4LinearKernel)
kernel.process_weights_after_loading(layer)
assert layer.weight_scale.data_ptr() == swizzled_scale.data_ptr()
assert layer.weight_scale.weight_loader is weight_loader
assert layer.b12x_mxfp4_linear
def test_b12x_mxfp4_apply_calls_native_blockscaled_gemm(monkeypatch) -> None:
import vllm.model_executor.kernels.linear.mxfp4.b12x as b12x_mod
import vllm.utils.flashinfer as flashinfer_utils
calls: list[tuple] = []
x_packed = torch.empty((6, 64), dtype=torch.uint8)
x_scale_storage = torch.empty((128, 4), dtype=torch.uint8)
def mm_mxfp4(*args, **kwargs):
calls.append((args, kwargs))
return torch.full((6, 48), 3.0, dtype=torch.bfloat16)
monkeypatch.setattr(
flashinfer_utils,
"flashinfer_mxfp4_quantize",
lambda *args, **kwargs: (x_packed, x_scale_storage),
)
monkeypatch.setattr(
b12x_mod,
"_import_b12x_blockscaled",
lambda: types.SimpleNamespace(mm_mxfp4=mm_mxfp4),
)
layer = torch.nn.Module()
layer.output_size_per_partition = 48
layer.weight = torch.empty((48, 64), dtype=torch.uint8)
layer.weight_scale = torch.empty((128, 4), dtype=torch.uint8)
x = torch.empty((2, 3, 128), dtype=torch.bfloat16)
bias = torch.ones(48, dtype=torch.bfloat16)
kernel = object.__new__(B12xMxFp4LinearKernel)
output = kernel.apply_weights(layer, x, bias)
assert output.shape == (2, 3, 48)
torch.testing.assert_close(output, torch.full_like(output, 4.0))
assert len(calls) == 1
args, kwargs = calls[0]
assert args == (
x_packed,
x_scale_storage,
layer.weight,
layer.weight_scale,
)
assert kwargs == {"out_dtype": torch.bfloat16}
def test_warmup_b12x_mxfp4_dedupes_weight_signatures(monkeypatch) -> None:
import vllm.model_executor.kernels.linear.mxfp4.b12x as b12x_mod
calls = []
def apply(source, weight, weight_scale, bias):
calls.append((source.shape, weight, weight_scale, bias))
return source.new_empty((source.shape[0], weight.shape[0]))
platform = types.SimpleNamespace(
is_cuda=lambda: True,
is_device_capability_family=lambda family: family == 120,
)
monkeypatch.setattr(b12x_mod, "current_platform", platform)
monkeypatch.setattr(
b12x_mod,
"_import_b12x_blockscaled",
lambda: types.SimpleNamespace(),
)
monkeypatch.setattr(b12x_mod, "_apply_b12x_mxfp4_linear", apply)
def layer(n: int):
return types.SimpleNamespace(
b12x_mxfp4_linear=True,
weight=torch.empty((n, 64), dtype=torch.uint8),
weight_scale=torch.empty((128, 4), dtype=torch.uint8),
)
layer_a = layer(48)
layer_b = layer(48)
layer_c = layer(96)
model = types.SimpleNamespace(
modules=lambda: iter([layer_a, layer_b, layer_c, types.SimpleNamespace()])
)
warmed = warmup_b12x_mxfp4_linear(
model,
max_tokens=8,
cudagraph_capture_sizes=[1, 2],
)
assert warmed == 6
assert [call[0] for call in calls] == [
torch.Size([1, 128]),
torch.Size([2, 128]),
torch.Size([8, 128]),
torch.Size([1, 128]),
torch.Size([2, 128]),
torch.Size([8, 128]),
]
assert calls[0][1] is layer_a.weight
assert calls[3][1] is layer_c.weight
assert all(call[3] is None for call in calls)
@@ -0,0 +1,935 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
import types
from dataclasses import dataclass
import pytest
import torch
from vllm.model_executor.kernels.linear import (
_LINEAR_BACKEND_KERNEL_MAP,
_POSSIBLE_FP8_BLOCK_KERNELS,
_POSSIBLE_FP8_KERNELS,
_POSSIBLE_MXFP8_KERNELS,
init_fp8_linear_kernel,
init_mxfp8_linear_kernel,
)
from vllm.model_executor.kernels.linear.mxfp8.b12x import (
B12xMxfp8LinearKernel,
_b12x_mxfp8_expected_m,
warmup_b12x_mxfp8_linear,
)
from vllm.model_executor.kernels.linear.mxfp8.Mxfp8LinearKernel import (
Mxfp8LinearLayerConfig,
)
from vllm.model_executor.kernels.linear.scaled_mm.b12x_block import (
B12xFp8BlockScaledMMKernel,
_run_b12x_fp8_block_scaled_mm,
warmup_b12x_block_fp8_linear,
)
from vllm.model_executor.kernels.linear.scaled_mm.b12x_tensor import (
B12xTensorFP8ScaledMMLinearKernel,
warmup_b12x_tensor_fp8_linear,
)
from vllm.model_executor.kernels.linear.scaled_mm.ScaledMMLinearKernel import (
FP8ScaledMMLinearLayerConfig,
)
from vllm.model_executor.layers.linear import (
ReplicatedLinear,
UnquantizedLinearMethod,
)
from vllm.model_executor.layers.quantization.utils.quant_utils import (
kFp8Dynamic128Sym,
kFp8Static128BlockSym,
kFp8StaticTensorSym,
)
from vllm.platforms import PlatformEnum
from vllm.utils.b12x import b12x_warmup_token_counts
def test_b12x_backend_maps_mxfp8_kernel() -> None:
assert B12xMxfp8LinearKernel in _LINEAR_BACKEND_KERNEL_MAP["b12x"]
assert B12xMxfp8LinearKernel in _POSSIBLE_MXFP8_KERNELS[PlatformEnum.CUDA]
def test_b12x_backend_maps_tensor_fp8_kernel() -> None:
assert B12xTensorFP8ScaledMMLinearKernel in _LINEAR_BACKEND_KERNEL_MAP["b12x"]
assert B12xTensorFP8ScaledMMLinearKernel in _POSSIBLE_FP8_KERNELS[PlatformEnum.CUDA]
@pytest.mark.parametrize(
("kernels", "before", "b12x", "after"),
[
(
_POSSIBLE_FP8_KERNELS[PlatformEnum.CUDA],
"CutlassFP8ScaledMMLinearKernel",
"B12xTensorFP8ScaledMMLinearKernel",
"PerTensorTorchFP8ScaledMMLinearKernel",
),
(
_POSSIBLE_FP8_BLOCK_KERNELS[PlatformEnum.CUDA],
"CutlassFp8BlockScaledMMKernel",
"B12xFp8BlockScaledMMKernel",
"MarlinFP8ScaledMMLinearKernel",
),
(
_POSSIBLE_MXFP8_KERNELS[PlatformEnum.CUDA],
"MarlinMxfp8LinearKernel",
"B12xMxfp8LinearKernel",
"EmulationMxfp8LinearKernel",
),
],
)
def test_b12x_fp8_fallback_priority(
kernels: list[type],
before: str,
b12x: str,
after: str,
) -> None:
names = [kernel.__name__ for kernel in kernels]
assert names.index(before) < names.index(b12x) < names.index(after)
@torch.inference_mode()
def test_b12x_backend_does_not_intercept_unquantized_bf16(
default_vllm_config,
dist_init,
) -> None:
default_vllm_config.kernel_config.linear_backend = "b12x"
layer = ReplicatedLinear(
128,
64,
bias=False,
params_dtype=torch.bfloat16,
quant_config=None,
prefix="bf16_linear",
).cuda()
layer.weight.data.normal_()
x = torch.randn(8, 128, device="cuda", dtype=torch.bfloat16)
output, output_bias = layer(x)
expected = torch.nn.functional.linear(x, layer.weight)
assert isinstance(layer.quant_method, UnquantizedLinearMethod)
assert output_bias is None
torch.testing.assert_close(output, expected)
def test_b12x_explicit_backend_selects_per_tensor_fp8(
monkeypatch,
default_vllm_config,
) -> None:
import vllm.model_executor.kernels.linear as linear_mod
monkeypatch.setattr(linear_mod.current_platform, "_enum", PlatformEnum.CUDA)
monkeypatch.setattr(linear_mod, "_get_linear_backend", lambda: "b12x")
monkeypatch.setattr(
B12xTensorFP8ScaledMMLinearKernel,
"is_supported",
classmethod(lambda cls, compute_capability=None: (True, None)),
)
monkeypatch.setattr(
B12xTensorFP8ScaledMMLinearKernel,
"can_implement",
classmethod(lambda cls, config: (True, None)),
)
kernel = init_fp8_linear_kernel(
activation_quant_key=kFp8StaticTensorSym,
weight_quant_key=kFp8StaticTensorSym,
input_dtype=torch.bfloat16,
out_dtype=torch.bfloat16,
weight_shape=(2048, 2048),
)
assert isinstance(kernel, B12xTensorFP8ScaledMMLinearKernel)
def test_b12x_tensor_fp8_can_implement_supported_config() -> None:
config = FP8ScaledMMLinearLayerConfig(
activation_quant_key=kFp8StaticTensorSym,
weight_quant_key=kFp8StaticTensorSym,
weight_shape=(64, 128),
input_dtype=torch.bfloat16,
out_dtype=torch.bfloat16,
)
can_implement, reason = B12xTensorFP8ScaledMMLinearKernel.can_implement(config)
assert can_implement
assert reason is None
def test_b12x_explicit_backend_selects_block_fp8(
monkeypatch,
default_vllm_config,
) -> None:
import vllm.model_executor.kernels.linear as linear_mod
monkeypatch.setattr(linear_mod.current_platform, "_enum", PlatformEnum.CUDA)
monkeypatch.setattr(linear_mod, "_get_linear_backend", lambda: "b12x")
monkeypatch.setattr(
B12xFp8BlockScaledMMKernel,
"is_supported",
classmethod(lambda cls, compute_capability=None: (True, None)),
)
kernel = init_fp8_linear_kernel(
activation_quant_key=kFp8Dynamic128Sym,
weight_quant_key=kFp8Static128BlockSym,
input_dtype=torch.bfloat16,
out_dtype=torch.bfloat16,
weight_shape=(2048, 2048),
)
assert isinstance(kernel, B12xFp8BlockScaledMMKernel)
def test_b12x_block_fp8_checks_runtime_support(monkeypatch) -> None:
import vllm.model_executor.kernels.linear.scaled_mm.b12x_block as b12x_mod
platform = types.SimpleNamespace(
is_cuda=lambda: True,
is_device_capability_family=lambda family: family == 120,
)
monkeypatch.setattr(b12x_mod, "current_platform", platform)
monkeypatch.setattr(
b12x_mod,
"_import_b12x_blockscaled",
lambda: types.SimpleNamespace(is_supported=lambda: False),
)
supported, reason = B12xFp8BlockScaledMMKernel.is_supported()
assert not supported
assert reason == "B12X regular block-FP8 GEMM is not supported"
def test_b12x_block_fp8_requires_matching_supported_dtypes() -> None:
def config(input_dtype: torch.dtype, out_dtype: torch.dtype):
return FP8ScaledMMLinearLayerConfig(
activation_quant_key=kFp8Dynamic128Sym,
weight_quant_key=kFp8Static128BlockSym,
weight_shape=(256, 128),
input_dtype=input_dtype,
out_dtype=out_dtype,
)
can_implement, reason = B12xFp8BlockScaledMMKernel.can_implement(
config(torch.float32, torch.float32)
)
assert not can_implement
assert reason == "Supports only bf16/fp16 input dtype"
can_implement, reason = B12xFp8BlockScaledMMKernel.can_implement(
config(torch.bfloat16, torch.float16)
)
assert not can_implement
assert reason == "Input and output dtype must match"
can_implement, reason = B12xFp8BlockScaledMMKernel.can_implement(
config(torch.float16, torch.float16)
)
assert can_implement
assert reason is None
def test_b12x_block_fp8_requires_aligned_features() -> None:
def can_implement(weight_shape: tuple[int, int]):
config = FP8ScaledMMLinearLayerConfig(
activation_quant_key=kFp8Dynamic128Sym,
weight_quant_key=kFp8Static128BlockSym,
weight_shape=weight_shape,
input_dtype=torch.bfloat16,
out_dtype=torch.bfloat16,
)
return B12xFp8BlockScaledMMKernel.can_implement(config)
assert can_implement((256, 192)) == (
False,
"Input features must be a positive multiple of 128",
)
assert can_implement((192, 256)) == (
False,
"Output features must be a positive multiple of 128",
)
def test_b12x_warmup_token_counts_cover_serving_regimes() -> None:
assert b12x_warmup_token_counts(
max_tokens=2048,
cudagraph_capture_sizes=[1, 2, 8, 128],
) == (1, 2, 8, 128, 2048)
def test_warmup_b12x_block_fp8_dedupes_weight_signatures(monkeypatch) -> None:
import vllm.model_executor.kernels.linear.scaled_mm.b12x_block as b12x_mod
calls = []
def run(a, weight, a_scale, weight_scale, out_dtype):
calls.append((a.shape, weight, a_scale.shape, weight_scale, out_dtype))
return torch.empty((a.shape[0], weight.shape[0]), dtype=out_dtype)
platform = types.SimpleNamespace(
is_cuda=lambda: True,
is_device_capability_family=lambda family: family == 120,
)
monkeypatch.setattr(b12x_mod, "current_platform", platform)
monkeypatch.setattr(
b12x_mod,
"_import_b12x_blockscaled",
lambda: types.SimpleNamespace(),
)
monkeypatch.setattr(b12x_mod, "_run_b12x_fp8_block_scaled_mm", run)
def layer(in_features: int, out_features: int):
return types.SimpleNamespace(
b12x_block_fp8_linear=True,
weight=torch.empty((out_features, in_features), dtype=torch.float8_e4m3fn),
weight_scale_inv=torch.empty(
(out_features // 128, in_features // 128), dtype=torch.float32
),
)
layer_a = layer(128, 256)
layer_b = layer(256, 128)
modules = [
layer_a,
layer_a,
layer_b,
types.SimpleNamespace(),
]
model = types.SimpleNamespace(modules=lambda: iter(modules))
warmed = warmup_b12x_block_fp8_linear(
model,
max_tokens=32,
cudagraph_capture_sizes=[2, 8],
output_dtype=torch.bfloat16,
)
assert warmed == 8
assert [call[0][0] for call in calls] == [1, 2, 8, 32] * 2
assert [call[2][0] for call in calls] == [1, 2, 8, 32] * 2
assert calls[0][1] is layer_a.weight
assert calls[4][1] is layer_b.weight
assert calls[0][3] is layer_a.weight_scale_inv
assert calls[4][3] is layer_b.weight_scale_inv
assert all(call[4] == torch.bfloat16 for call in calls)
def test_b12x_tensor_fp8_process_weights_packs_modelopt_layout(
monkeypatch,
) -> None:
import vllm.model_executor.kernels.linear.scaled_mm.b12x_tensor as b12x_mod
calls = []
packed = types.SimpleNamespace(out_features=64)
def pack(weight: torch.Tensor, output_scale: torch.Tensor):
calls.append((weight, output_scale))
return packed
monkeypatch.setattr(
b12x_mod,
"_import_b12x_tensor_fp8",
lambda: types.SimpleNamespace(pack_weight=pack),
)
layer = torch.nn.Module()
layer.prefix = "model.layers.0.self_attn.qkv_proj"
original_weight = (
torch.randn((128, 64), dtype=torch.float32).clamp(-4, 4).to(torch.float8_e4m3fn)
)
layer.weight = torch.nn.Parameter(original_weight, requires_grad=False)
layer.weight_scale = torch.nn.Parameter(torch.tensor(0.25), requires_grad=False)
layer.input_scale = torch.nn.Parameter(torch.tensor(0.5), requires_grad=False)
weight_loader = object()
scale_loader = object()
layer.weight.weight_loader = weight_loader
layer.weight_scale.weight_loader = scale_loader
kernel = object.__new__(B12xTensorFP8ScaledMMLinearKernel)
kernel.config = types.SimpleNamespace(weight_shape=(64, 128))
kernel.layer_param_names = (
"weight",
"weight_scale",
"input_scale",
"input_scale_ub",
)
kernel.process_weights_after_loading(layer)
assert layer.b12x_tensor_fp8_packed_weight is packed
assert len(calls) == 1
weight, output_scale = calls[0]
torch.testing.assert_close(weight, original_weight.T.contiguous())
torch.testing.assert_close(output_scale, torch.tensor([0.125]))
assert layer.weight.numel() == 0
assert layer.weight_scale.numel() == 0
assert layer.weight.weight_loader is weight_loader
assert layer.weight_scale.weight_loader is scale_loader
torch.testing.assert_close(layer.input_scale, torch.tensor(0.5))
def test_warmup_b12x_tensor_fp8_dedupes_weight_signatures(monkeypatch) -> None:
import vllm.model_executor.kernels.linear.scaled_mm.b12x_tensor as b12x_mod
calls = []
def prewarm(packed_weight, token_counts, *, out_dtype, stream):
del stream
calls.append((packed_weight, tuple(token_counts), out_dtype))
return len(tuple(token_counts))
platform = types.SimpleNamespace(
is_cuda=lambda: True,
is_device_capability_family=lambda family: family == 120,
)
monkeypatch.setattr(b12x_mod, "current_platform", platform)
monkeypatch.setattr(
b12x_mod,
"_import_b12x_tensor_fp8",
lambda: types.SimpleNamespace(prewarm=prewarm),
)
monkeypatch.setattr(
b12x_mod,
"current_stream",
lambda: types.SimpleNamespace(cuda_stream=object()),
)
def packed(in_features: int, padded_in_features: int, out_features: int):
return types.SimpleNamespace(
in_features=in_features,
padded_in_features=padded_in_features,
out_features=out_features,
values=torch.empty(1),
)
packed_a = packed(128, 128, 256)
packed_b = packed(160, 256, 512)
modules = [
types.SimpleNamespace(b12x_tensor_fp8_packed_weight=packed_a),
types.SimpleNamespace(b12x_tensor_fp8_packed_weight=packed_a),
types.SimpleNamespace(b12x_tensor_fp8_packed_weight=packed_b),
types.SimpleNamespace(),
]
model = types.SimpleNamespace(modules=lambda: iter(modules))
warmed = warmup_b12x_tensor_fp8_linear(
model,
max_tokens=2048,
cudagraph_capture_sizes=[1, 2],
)
assert warmed == 6
assert calls == [
(packed_a, (1, 2, 2048), torch.bfloat16),
(packed_b, (1, 2, 2048), torch.bfloat16),
]
def test_b12x_tensor_fp8_apply_quantizes_and_uses_packed_weight(
monkeypatch,
) -> None:
import vllm.model_executor.kernels.linear.scaled_mm.b12x_tensor as b12x_mod
calls = []
def mm(
source: torch.Tensor,
packed_weight,
*,
bias: torch.Tensor | None = None,
out_dtype: torch.dtype,
expected_m: int,
stream: object = None,
) -> torch.Tensor:
del stream
calls.append((source, packed_weight, bias, out_dtype, expected_m))
return torch.full(
(source.shape[0], packed_weight.out_features),
3.0,
dtype=out_dtype,
)
monkeypatch.setattr(
b12x_mod,
"_import_b12x_tensor_fp8",
lambda: types.SimpleNamespace(mm=mm),
)
monkeypatch.setattr(
b12x_mod,
"current_stream",
lambda: types.SimpleNamespace(cuda_stream=object()),
)
monkeypatch.setattr(torch.compiler, "is_compiling", lambda: False)
layer = torch.nn.Module()
packed = types.SimpleNamespace(out_features=48)
layer.b12x_tensor_fp8_packed_weight = packed
layer.weight = torch.nn.Parameter(
torch.empty((128, 48), dtype=torch.float8_e4m3fn),
requires_grad=False,
)
layer.weight_scale = torch.nn.Parameter(torch.tensor(0.25), requires_grad=False)
layer.input_scale = torch.nn.Parameter(torch.tensor(0.5), requires_grad=False)
x = torch.empty((2, 3, 128), dtype=torch.bfloat16)
x_q = torch.empty((6, 128), dtype=torch.float8_e4m3fn)
bias = torch.empty((48,), dtype=torch.bfloat16)
kernel = object.__new__(B12xTensorFP8ScaledMMLinearKernel)
kernel.config = types.SimpleNamespace(out_dtype=torch.bfloat16)
kernel.layer_param_names = (
"weight",
"weight_scale",
"input_scale",
"input_scale_ub",
)
kernel.quant_fp8 = lambda source, scale, scale_ub: (x_q, scale)
output = kernel.apply_weights(layer, x, bias)
assert output.shape == (2, 3, 48)
assert output.dtype == torch.bfloat16
assert len(calls) == 1
source, called_packed, called_bias, out_dtype, expected_m = calls[0]
assert source.data_ptr() == x_q.data_ptr()
assert called_packed is packed
assert called_bias is bias
assert out_dtype == torch.bfloat16
assert expected_m == 6
def test_b12x_mxfp8_explicit_backend_selects_kernel(monkeypatch) -> None:
import vllm.model_executor.kernels.linear as linear_mod
monkeypatch.setattr(linear_mod.current_platform, "_enum", PlatformEnum.CUDA)
monkeypatch.setattr(linear_mod, "_get_linear_backend", lambda: "b12x")
monkeypatch.setattr(
B12xMxfp8LinearKernel,
"is_supported",
classmethod(lambda cls, compute_capability=None: (True, None)),
)
monkeypatch.setattr(
B12xMxfp8LinearKernel,
"can_implement",
classmethod(lambda cls, c: (True, None)),
)
kernel = init_mxfp8_linear_kernel()
assert isinstance(kernel, B12xMxfp8LinearKernel)
def test_b12x_mxfp8_can_implement_supported_config() -> None:
can_implement, reason = B12xMxfp8LinearKernel.can_implement(
Mxfp8LinearLayerConfig()
)
assert can_implement
assert reason is None
def test_b12x_mxfp8_expected_m_uses_live_m() -> None:
assert _b12x_mxfp8_expected_m(0) == 1
assert _b12x_mxfp8_expected_m(1) == 1
assert _b12x_mxfp8_expected_m(2) == 2
assert _b12x_mxfp8_expected_m(8) == 8
assert _b12x_mxfp8_expected_m(9) == 9
assert _b12x_mxfp8_expected_m(128) == 128
assert _b12x_mxfp8_expected_m(129) == 129
assert _b12x_mxfp8_expected_m(2048) == 2048
def test_warmup_b12x_mxfp8_linear_dedupes_weight_signatures(
monkeypatch,
) -> None:
import vllm.model_executor.kernels.linear.mxfp8.b12x as b12x_mod
calls = []
def mm(
source: torch.Tensor,
packed_weight,
*,
bias: torch.Tensor | None = None,
expected_m: int | None = None,
stream: object = None,
) -> torch.Tensor:
del stream
calls.append((source.shape, packed_weight, bias, expected_m))
return source.new_empty((source.shape[0], packed_weight.out_features))
platform = types.SimpleNamespace(
is_cuda=lambda: True,
is_device_capability_family=lambda family: family == 120,
)
monkeypatch.setattr(b12x_mod, "current_platform", platform)
monkeypatch.setattr(
b12x_mod,
"_import_b12x_mxfp8",
lambda: types.SimpleNamespace(mm=mm),
)
monkeypatch.setattr(
b12x_mod,
"current_stream",
lambda: types.SimpleNamespace(cuda_stream=object()),
)
def packed(in_features: int, padded_in_features: int, out_features: int):
return types.SimpleNamespace(
in_features=in_features,
padded_in_features=padded_in_features,
out_features=out_features,
weight=types.SimpleNamespace(values=torch.empty(1)),
)
packed_a = packed(128, 128, 256)
packed_b = packed(128, 128, 512)
modules = [
types.SimpleNamespace(b12x_mxfp8_packed_weight=packed_a),
types.SimpleNamespace(b12x_mxfp8_packed_weight=packed_a),
types.SimpleNamespace(b12x_mxfp8_packed_weight=packed_b),
types.SimpleNamespace(),
]
model = types.SimpleNamespace(modules=lambda: iter(modules))
warmed = warmup_b12x_mxfp8_linear(
model,
max_tokens=2048,
cudagraph_capture_sizes=[1, 2],
)
assert warmed == 6
assert [call[0] for call in calls] == [
torch.Size([1, 128]),
torch.Size([2, 128]),
torch.Size([2048, 128]),
torch.Size([1, 128]),
torch.Size([2, 128]),
torch.Size([2048, 128]),
]
assert [call[3] for call in calls] == [1, 2, 2048, 1, 2, 2048]
assert calls[0][1] is packed_a
assert calls[3][1] is packed_b
def test_b12x_mxfp8_support_check_reports_missing_import(monkeypatch) -> None:
import vllm.model_executor.kernels.linear.mxfp8.b12x as b12x_mod
monkeypatch.setattr(b12x_mod.current_platform, "is_cuda", lambda: True)
monkeypatch.setattr(
b12x_mod.current_platform,
"is_device_capability_family",
lambda family: family == 120,
)
monkeypatch.setattr(b12x_mod, "_import_b12x_mxfp8", lambda: None)
is_supported, reason = B12xMxfp8LinearKernel.is_supported()
assert not is_supported
assert reason == "Install the B12X backend with `pip install vllm[b12x]`"
def test_b12x_mxfp8_support_respects_runtime_probe(monkeypatch) -> None:
import vllm.model_executor.kernels.linear.mxfp8.b12x as b12x_mod
monkeypatch.setattr(b12x_mod.current_platform, "is_cuda", lambda: True)
monkeypatch.setattr(
b12x_mod.current_platform,
"is_device_capability_family",
lambda family: family == 120,
)
monkeypatch.setattr(
b12x_mod,
"_import_b12x_mxfp8",
lambda: types.SimpleNamespace(is_supported=lambda: False),
)
is_supported, reason = B12xMxfp8LinearKernel.is_supported()
assert not is_supported
assert reason == "b12x.gemm.mxfp8_linear is not supported"
def test_b12x_mxfp8_process_weights_packs_modelopt_layout(monkeypatch) -> None:
import vllm.model_executor.kernels.linear.mxfp8.b12x as b12x_mod
calls = []
packed = types.SimpleNamespace(out_features=48)
def pack(weight: torch.Tensor, weight_scale: torch.Tensor):
calls.append((weight, weight_scale))
return packed
monkeypatch.setattr(
b12x_mod,
"_import_b12x_mxfp8",
lambda: types.SimpleNamespace(pack_weight=pack),
)
layer = torch.nn.Module()
layer.prefix = "model.layers.0.self_attn.qkv_proj"
layer.weight = torch.nn.Parameter(
torch.empty((48, 128), dtype=torch.float8_e4m3fn),
requires_grad=False,
)
layer.weight_scale = torch.nn.Parameter(
torch.empty((64, 8), dtype=torch.uint8),
requires_grad=False,
)
weight_loader = object()
scale_loader = object()
layer.weight.weight_loader = weight_loader
layer.weight_scale.weight_loader = scale_loader
kernel = object.__new__(B12xMxfp8LinearKernel)
kernel.process_weights_after_loading(layer)
assert layer.b12x_mxfp8_packed_weight is packed
assert len(calls) == 1
weight, weight_scale = calls[0]
assert weight.shape == (48, 128)
assert weight_scale.shape == (48, 4)
assert weight.dtype == torch.float8_e4m3fn
assert weight_scale.dtype == torch.uint8
assert layer.weight.numel() == 0
assert layer.weight_scale.numel() == 0
assert layer.weight.weight_loader is weight_loader
assert layer.weight_scale.weight_loader is scale_loader
def test_b12x_mxfp8_reload_reuses_packed_tensor_addresses(monkeypatch) -> None:
import vllm.model_executor.kernels.linear.mxfp8.b12x as b12x_mod
@dataclass(frozen=True)
class PackedWeight:
values: torch.Tensor
scales: torch.Tensor
out_features: int
def pack(weight: torch.Tensor, weight_scale: torch.Tensor) -> PackedWeight:
return PackedWeight(
values=weight.clone(),
scales=weight_scale.clone(),
out_features=int(weight.shape[0]),
)
monkeypatch.setattr(
b12x_mod,
"_import_b12x_mxfp8",
lambda: types.SimpleNamespace(pack_weight=pack),
)
layer = torch.nn.Module()
layer.prefix = "model.layers.0.mlp.down_proj"
layer.weight = torch.nn.Parameter(
torch.zeros((48, 128), dtype=torch.float8_e4m3fn),
requires_grad=False,
)
layer.weight_scale = torch.nn.Parameter(
torch.zeros((48, 4), dtype=torch.uint8),
requires_grad=False,
)
kernel = object.__new__(B12xMxfp8LinearKernel)
kernel.process_weights_after_loading(layer)
packed = layer.b12x_mxfp8_packed_weight
values_ptr = packed.values.data_ptr()
scales_ptr = packed.scales.data_ptr()
layer.weight = torch.nn.Parameter(
torch.ones((48, 128), dtype=torch.float8_e4m3fn),
requires_grad=False,
)
layer.weight_scale = torch.nn.Parameter(
torch.full((48, 4), 3, dtype=torch.uint8),
requires_grad=False,
)
kernel.process_weights_after_loading(layer)
assert layer.b12x_mxfp8_packed_weight is packed
assert packed.values.data_ptr() == values_ptr
assert packed.scales.data_ptr() == scales_ptr
torch.testing.assert_close(
packed.values,
torch.ones((48, 128), dtype=torch.float8_e4m3fn),
)
torch.testing.assert_close(
packed.scales,
torch.full((48, 4), 3, dtype=torch.uint8),
)
assert layer.weight.numel() == 0
assert layer.weight_scale.numel() == 0
def test_b12x_block_fp8_process_weights_keeps_native_block_layout() -> None:
layer = torch.nn.Module()
layer.weight = torch.nn.Parameter(
torch.empty((128, 128), dtype=torch.float8_e4m3fn),
requires_grad=False,
)
layer.weight_scale_inv = torch.nn.Parameter(
torch.empty((1, 1), dtype=torch.float32),
requires_grad=False,
)
layer.weight_block_size = [128, 128]
weight_loader = object()
scale_loader = object()
layer.weight.weight_loader = weight_loader
layer.weight_scale_inv.weight_loader = scale_loader
kernel = object.__new__(B12xFp8BlockScaledMMKernel)
kernel.process_weights_after_loading(layer)
assert layer.b12x_block_fp8_linear
assert layer.weight.shape == (128, 128)
assert layer.weight.dtype == torch.float8_e4m3fn
assert layer.weight_scale_inv.shape == (1, 1)
assert layer.weight_scale_inv.dtype == torch.float32
assert layer.weight.weight_loader is weight_loader
assert layer.weight_scale_inv.weight_loader is scale_loader
@pytest.mark.parametrize("scale_dtype", [torch.float8_e8m0fnu, torch.uint8])
def test_b12x_block_fp8_upcasts_e8m0_weight_scales(scale_dtype) -> None:
layer = torch.nn.Module()
layer.weight = torch.nn.Parameter(
torch.empty((128, 128), dtype=torch.float8_e4m3fn),
requires_grad=False,
)
scale_bytes = torch.tensor([[125]], dtype=torch.uint8)
layer.weight_scale_inv = torch.nn.Parameter(
scale_bytes.view(scale_dtype),
requires_grad=False,
)
layer.weight_block_size = [128, 128]
kernel = object.__new__(B12xFp8BlockScaledMMKernel)
kernel.process_weights_after_loading(layer)
assert layer.weight_scale_inv.dtype == torch.float32
torch.testing.assert_close(
layer.weight_scale_inv,
torch.tensor([[0.25]], dtype=torch.float32),
)
def test_b12x_mxfp8_apply_uses_packed_weight(monkeypatch) -> None:
import vllm.model_executor.kernels.linear.mxfp8.b12x as b12x_mod
calls = []
def mxfp8_linear(
source: torch.Tensor,
packed_weight,
*,
bias: torch.Tensor | None = None,
expected_m: int | None = None,
stream: object = None,
) -> torch.Tensor:
del stream
calls.append((source, packed_weight, bias, expected_m))
return source.new_full((source.shape[0], packed_weight.out_features), 3.0)
monkeypatch.setattr(
b12x_mod,
"_import_b12x_mxfp8",
lambda: types.SimpleNamespace(mm=mxfp8_linear),
)
layer = torch.nn.Module()
packed = types.SimpleNamespace(out_features=48)
layer.b12x_mxfp8_packed_weight = packed
x = torch.empty((2, 3, 128), dtype=torch.bfloat16)
bias = torch.empty((48,), dtype=torch.bfloat16)
kernel = object.__new__(B12xMxfp8LinearKernel)
output = kernel.apply_weights(layer, x, bias)
assert output.shape == (2, 3, 48)
assert output.dtype == x.dtype
assert len(calls) == 1
source, called_packed, called_bias, expected_m = calls[0]
assert source.shape == (6, 128)
assert called_packed is packed
assert called_bias is bias
assert expected_m == 6
def test_b12x_block_fp8_apply_uses_b12x_recipe_api(monkeypatch) -> None:
import vllm.model_executor.kernels.linear.scaled_mm.b12x_block as b12x_mod
calls = []
def mm_block_fp8(*args, **kwargs):
calls.append((args, kwargs))
return torch.full(
(args[0].shape[0], args[2].shape[0]),
13.0,
dtype=kwargs["out_dtype"],
)
monkeypatch.setattr(
b12x_mod,
"_import_b12x_blockscaled",
lambda: types.SimpleNamespace(mm_block_fp8=mm_block_fp8),
)
a = torch.empty((6, 128), dtype=torch.float8_e4m3fn)
weight = torch.empty((256, 128), dtype=torch.float8_e4m3fn)
a_scale = torch.empty((6, 1), dtype=torch.float32)
weight_scale = torch.empty((2, 1), dtype=torch.float32)
kernel = object.__new__(B12xFp8BlockScaledMMKernel)
kernel.config = types.SimpleNamespace(out_dtype=torch.bfloat16)
output = kernel.apply_block_scaled_mm(a, weight, a_scale, weight_scale)
assert output.shape == (6, 256)
assert output.dtype == torch.bfloat16
assert len(calls) == 1
assert calls[0] == (
(a, a_scale, weight, weight_scale),
{"out_dtype": torch.bfloat16},
)
torch.testing.assert_close(output, torch.full_like(output, 13.0))
def test_b12x_block_fp8_helper_uses_regular_compact_scale_api(monkeypatch) -> None:
import vllm.model_executor.kernels.linear.scaled_mm.b12x_block as b12x_mod
calls = []
def mm_block_fp8(*args, **kwargs):
calls.append((args, kwargs))
return torch.full((6, 256), 17.0, dtype=torch.bfloat16)
monkeypatch.setattr(
b12x_mod,
"_import_b12x_blockscaled",
lambda: types.SimpleNamespace(mm_block_fp8=mm_block_fp8),
)
a = torch.empty((6, 128), dtype=torch.float8_e4m3fn)
weight = torch.empty((256, 128), dtype=torch.float8_e4m3fn)
a_scale = torch.empty((6, 1), dtype=torch.float32)
weight_scale = torch.empty((2, 1), dtype=torch.float32)
output = _run_b12x_fp8_block_scaled_mm(
a,
weight,
a_scale,
weight_scale,
torch.bfloat16,
)
assert output.shape == (6, 256)
args, kwargs = calls[0]
assert args == (a, a_scale, weight, weight_scale)
assert kwargs == {
"out_dtype": torch.bfloat16,
}
torch.testing.assert_close(output, torch.full_like(output, 17.0))
@@ -0,0 +1,246 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
import types
import torch
from vllm.model_executor.kernels.linear import (
_LINEAR_BACKEND_KERNEL_MAP,
_POSSIBLE_NVFP4_KERNELS,
_resolve_backend_kernels,
init_nvfp4_linear_kernel,
)
from vllm.model_executor.kernels.linear.nvfp4.b12x import (
B12xNvFp4LinearKernel,
warmup_b12x_nvfp4_linear,
)
from vllm.model_executor.kernels.linear.nvfp4.marlin import (
MarlinNvFp4LinearKernel,
)
from vllm.platforms import PlatformEnum
def test_b12x_backend_maps_nvfp4_kernel() -> None:
assert B12xNvFp4LinearKernel in _LINEAR_BACKEND_KERNEL_MAP["b12x"]
assert B12xNvFp4LinearKernel in _POSSIBLE_NVFP4_KERNELS[PlatformEnum.CUDA]
def test_b12x_nvfp4_fallback_priority() -> None:
kernels = _POSSIBLE_NVFP4_KERNELS[PlatformEnum.CUDA]
names = [kernel.__name__ for kernel in kernels]
assert (
names.index("FbgemmNvFp4LinearKernel")
< names.index("B12xNvFp4LinearKernel")
< names.index("EmulationNvFp4LinearKernel")
)
def test_b12x_nvfp4_explicit_backend_selects_native_kernel(monkeypatch) -> None:
import vllm.model_executor.kernels.linear as linear_mod
monkeypatch.setattr(linear_mod.current_platform, "_enum", PlatformEnum.CUDA)
monkeypatch.setattr(linear_mod, "_get_linear_backend", lambda: "b12x")
monkeypatch.setattr(
B12xNvFp4LinearKernel,
"is_supported",
classmethod(lambda cls, compute_capability=None: (True, None)),
)
monkeypatch.setattr(
B12xNvFp4LinearKernel,
"can_implement",
classmethod(lambda cls, config: (True, None)),
)
kernel = init_nvfp4_linear_kernel()
assert isinstance(kernel, B12xNvFp4LinearKernel)
def test_b12x_nvfp4_can_implement_supported_config() -> None:
can_implement, reason = B12xNvFp4LinearKernel.can_implement(None)
assert can_implement
assert reason is None
def test_b12x_backend_preserves_w4a16_fallback(monkeypatch) -> None:
import vllm.model_executor.kernels.linear as linear_mod
monkeypatch.setattr(linear_mod.current_platform, "_enum", PlatformEnum.CUDA)
monkeypatch.setattr(linear_mod, "_get_linear_backend", lambda: "b12x")
monkeypatch.setattr(
MarlinNvFp4LinearKernel,
"is_supported",
classmethod(lambda cls, compute_capability=None: (True, None)),
)
kernel = init_nvfp4_linear_kernel(use_a16=True)
assert isinstance(kernel, MarlinNvFp4LinearKernel)
def test_backend_without_w4a16_kernel_preserves_fallback(monkeypatch) -> None:
import vllm.model_executor.kernels.linear as linear_mod
kernels = [MarlinNvFp4LinearKernel]
monkeypatch.setattr(linear_mod, "_get_linear_backend", lambda: "b12x")
assert _resolve_backend_kernels(kernels, "NVFP4") == kernels
def test_b12x_nvfp4_processes_scale_and_preserves_loader(monkeypatch) -> None:
import vllm.model_executor.kernels.linear.nvfp4.b12x as b12x_mod
scale = torch.empty((48, 8), dtype=torch.float8_e4m3fn)
swizzled_scale = torch.empty((128, 8), dtype=torch.float8_e4m3fn)
intrinsics = types.SimpleNamespace(swizzle_block_scale=lambda value: swizzled_scale)
monkeypatch.setattr(b12x_mod, "_import_b12x_intrinsics", lambda: intrinsics)
layer = torch.nn.Module()
layer.prefix = "model.layers.0.mlp.shared_expert.down_proj"
layer.weight_scale = torch.nn.Parameter(scale, requires_grad=False)
weight_loader = object()
layer.weight_scale.weight_loader = weight_loader
kernel = object.__new__(B12xNvFp4LinearKernel)
kernel.process_weights_after_loading(layer)
assert layer.weight_scale.data_ptr() == swizzled_scale.data_ptr()
assert layer.weight_scale.weight_loader is weight_loader
assert layer.b12x_nvfp4_linear
def test_b12x_nvfp4_apply_calls_native_blockscaled_gemm(monkeypatch) -> None:
import vllm.model_executor.kernels.linear.nvfp4.b12x as b12x_mod
calls: list[tuple] = []
quant_calls: list[tuple] = []
x_packed = torch.empty((6, 64), dtype=torch.uint8)
x_scale_storage = torch.empty((128, 8), dtype=torch.float8_e4m3fn)
def quant(*args, **kwargs):
quant_calls.append((args, kwargs))
return x_packed, x_scale_storage
def mm_nvfp4(*args, **kwargs):
calls.append((args, kwargs))
return torch.full((6, 48), 3.0, dtype=torch.bfloat16)
monkeypatch.setattr(
b12x_mod,
"scaled_fp4_quant",
quant,
)
monkeypatch.setattr(
b12x_mod,
"_import_b12x_blockscaled",
lambda: types.SimpleNamespace(mm_nvfp4=mm_nvfp4),
)
layer = torch.nn.Module()
layer.output_size_per_partition = 48
layer.weight = torch.empty((48, 64), dtype=torch.uint8)
layer.weight_scale = torch.empty((128, 8), dtype=torch.float8_e4m3fn)
layer.input_global_scale_inv = torch.tensor(2.0)
layer.alpha = torch.tensor(0.25)
x = torch.empty((2, 3, 256), dtype=torch.bfloat16)[..., ::2]
bias = torch.ones(48, dtype=torch.bfloat16)
kernel = object.__new__(B12xNvFp4LinearKernel)
output = kernel.apply_weights(layer, x, bias)
assert output.shape == (2, 3, 48)
torch.testing.assert_close(output, torch.full_like(output, 4.0))
assert len(quant_calls) == 1
quant_args, quant_kwargs = quant_calls[0]
assert quant_args[0].shape == (6, 128)
assert quant_args[0].data_ptr() == x.data_ptr()
assert quant_args[1] is layer.input_global_scale_inv
assert not quant_args[0].is_contiguous()
assert quant_kwargs == {"is_sf_swizzled_layout": True}
assert len(calls) == 1
args, kwargs = calls[0]
assert args == (
x_packed,
x_scale_storage,
layer.weight,
layer.weight_scale,
layer.alpha,
)
assert kwargs == {"out_dtype": torch.bfloat16}
def test_warmup_b12x_nvfp4_dedupes_weight_signatures(monkeypatch) -> None:
import vllm.model_executor.kernels.linear.nvfp4.b12x as b12x_mod
calls = []
def apply(
source,
weight,
weight_scale,
input_global_scale_inv,
alpha,
bias,
):
calls.append(
(
source.shape,
weight,
weight_scale,
input_global_scale_inv,
alpha,
bias,
)
)
return source.new_empty((source.shape[0], weight.shape[0]))
platform = types.SimpleNamespace(
is_cuda=lambda: True,
is_device_capability_family=lambda family: family == 120,
)
monkeypatch.setattr(b12x_mod, "current_platform", platform)
monkeypatch.setattr(
b12x_mod,
"_import_b12x_blockscaled",
lambda: types.SimpleNamespace(),
)
monkeypatch.setattr(b12x_mod, "_apply_b12x_nvfp4_linear", apply)
def layer(n: int):
return types.SimpleNamespace(
b12x_nvfp4_linear=True,
weight=torch.empty((n, 64), dtype=torch.uint8),
weight_scale=torch.empty((128, 8), dtype=torch.float8_e4m3fn),
input_global_scale_inv=torch.tensor(2.0),
alpha=torch.tensor(0.25),
)
layer_a = layer(48)
layer_b = layer(48)
layer_c = layer(96)
model = types.SimpleNamespace(
modules=lambda: iter([layer_a, layer_b, layer_c, types.SimpleNamespace()])
)
warmed = warmup_b12x_nvfp4_linear(
model,
max_tokens=8,
cudagraph_capture_sizes=[1, 2],
)
assert warmed == 6
assert [call[0] for call in calls] == [
torch.Size([1, 128]),
torch.Size([2, 128]),
torch.Size([8, 128]),
torch.Size([1, 128]),
torch.Size([2, 128]),
torch.Size([8, 128]),
]
assert calls[0][1] is layer_a.weight
assert calls[3][1] is layer_c.weight
assert all(call[5] is None for call in calls)
+100
View File
@@ -0,0 +1,100 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import importlib
from types import SimpleNamespace
import pytest
import torch
from vllm.model_executor.warmup.b12x_warmup import b12x_warmup
@pytest.mark.parametrize(
("module_name", "warmup_name", "import_name"),
[
(
"vllm.model_executor.kernels.linear.scaled_mm.b12x_block",
"warmup_b12x_block_fp8_linear",
"_import_b12x_blockscaled",
),
(
"vllm.model_executor.kernels.linear.scaled_mm.b12x_tensor",
"warmup_b12x_tensor_fp8_linear",
"_import_b12x_tensor_fp8",
),
(
"vllm.model_executor.kernels.linear.mxfp8.b12x",
"warmup_b12x_mxfp8_linear",
"_import_b12x_mxfp8",
),
(
"vllm.model_executor.kernels.linear.mxfp4.b12x",
"warmup_b12x_mxfp4_linear",
"_import_b12x_blockscaled",
),
(
"vllm.model_executor.kernels.linear.nvfp4.b12x",
"warmup_b12x_nvfp4_linear",
"_import_b12x_blockscaled",
),
],
)
def test_b12x_linear_warmup_skips_unused_provider(
monkeypatch,
module_name: str,
warmup_name: str,
import_name: str,
) -> None:
module = importlib.import_module(module_name)
platform = SimpleNamespace(
is_cuda=lambda: True,
is_device_capability_family=lambda family: family == 120,
)
monkeypatch.setattr(module, "current_platform", platform)
def fail_import():
pytest.fail("unused B12X provider was imported")
monkeypatch.setattr(module, import_name, fail_import)
warmup = getattr(module, warmup_name)
assert warmup(torch.nn.Module(), max_tokens=128) == 0
def test_b12x_warmup_covers_linear_serving_shapes(monkeypatch) -> None:
import vllm.model_executor.warmup.b12x_warmup as warmup_mod
model = torch.nn.Module()
linear_calls: list[tuple[torch.nn.Module, dict]] = []
def linear_warmup(model, **kwargs):
linear_calls.append((model, kwargs))
return 0
monkeypatch.setattr(warmup_mod, "warmup_b12x_block_fp8_linear", linear_warmup)
monkeypatch.setattr(warmup_mod, "warmup_b12x_mxfp4_linear", linear_warmup)
monkeypatch.setattr(warmup_mod, "warmup_b12x_mxfp8_linear", linear_warmup)
monkeypatch.setattr(warmup_mod, "warmup_b12x_nvfp4_linear", linear_warmup)
monkeypatch.setattr(warmup_mod, "warmup_b12x_tensor_fp8_linear", linear_warmup)
worker = SimpleNamespace(
get_model=lambda: model,
scheduler_config=SimpleNamespace(
max_num_batched_tokens=256,
max_num_scheduled_tokens=320,
),
model_config=SimpleNamespace(dtype=torch.float16),
vllm_config=SimpleNamespace(
compilation_config=SimpleNamespace(compile_sizes=[32, "dynamic", 64])
),
)
b12x_warmup(worker, [8, 16])
expected_linear_kwargs = {
"max_tokens": 256,
"cudagraph_capture_sizes": [8, 16],
"output_dtype": torch.float16,
}
assert linear_calls == [(model, expected_linear_kwargs)] * 5
+2
View File
@@ -147,6 +147,7 @@ LinearBackend = Literal[
"flashinfer_trtllm",
"flashinfer_cudnn",
"flashinfer_b12x",
"b12x",
"marlin",
"humming",
"triton",
@@ -223,6 +224,7 @@ class KernelConfig:
- "flashinfer_trtllm": Use FlashInfer with TensorRT-LLM kernels
- "flashinfer_cudnn": Use FlashInfer with cuDNN kernels
- "flashinfer_b12x": Use FlashInfer b12x CuteDSL NVFP4 GEMM (SM120+)
- "b12x": Use native B12X FP8 and FP4 linear kernels on SM12x
- "marlin": Use Marlin kernels
- "triton": Use Triton-based kernels
- "deep_gemm": Use DeepGEMM kernels
@@ -77,6 +77,9 @@ from vllm.model_executor.kernels.linear.mxfp4 import (
from vllm.model_executor.kernels.linear.mxfp4.aiter import (
AiterMxfp4LinearKernel,
)
from vllm.model_executor.kernels.linear.mxfp4.b12x import (
B12xMxFp4LinearKernel,
)
from vllm.model_executor.kernels.linear.mxfp4.emulation import (
EmulationMxfp4LinearKernel,
)
@@ -103,6 +106,9 @@ from vllm.model_executor.kernels.linear.mxfp8 import (
Mxfp8LinearKernel,
Mxfp8LinearLayerConfig,
)
from vllm.model_executor.kernels.linear.mxfp8.b12x import (
B12xMxfp8LinearKernel,
)
from vllm.model_executor.kernels.linear.mxfp8.emulation import (
EmulationMxfp8LinearKernel,
)
@@ -126,6 +132,9 @@ from vllm.model_executor.kernels.linear.nvfp4 import (
NvFp4LinearKernel,
NvFp4LinearLayerConfig,
)
from vllm.model_executor.kernels.linear.nvfp4.b12x import (
B12xNvFp4LinearKernel,
)
from vllm.model_executor.kernels.linear.nvfp4.cutlass import (
CutlassNvFp4LinearKernel,
)
@@ -163,6 +172,12 @@ from vllm.model_executor.kernels.linear.scaled_mm.aiter import (
AiterPerTokenFp8ScaledMMLinearKernel,
AiterPreshuffledPerTokenFp8ScaledMMLinearKernel,
)
from vllm.model_executor.kernels.linear.scaled_mm.b12x_block import (
B12xFp8BlockScaledMMKernel,
)
from vllm.model_executor.kernels.linear.scaled_mm.b12x_tensor import (
B12xTensorFP8ScaledMMLinearKernel,
)
from vllm.model_executor.kernels.linear.scaled_mm.cpu import (
CPUFp8BlockScaledMMKernel,
CPUInt8ScaledMMLinearKernel,
@@ -228,6 +243,13 @@ def _get_linear_backend() -> str:
# set are considered candidates. If none can implement the layer config,
# an error is raised to respect the user's explicit intent.
_LINEAR_BACKEND_KERNEL_MAP: dict[str, set[type]] = {
"b12x": {
B12xFp8BlockScaledMMKernel,
B12xMxFp4LinearKernel,
B12xMxfp8LinearKernel,
B12xNvFp4LinearKernel,
B12xTensorFP8ScaledMMLinearKernel,
},
"cutlass": {
CutlassInt8ScaledMMLinearKernel,
CutlassFP8ScaledMMLinearKernel,
@@ -377,6 +399,7 @@ _POSSIBLE_FP8_KERNELS: dict[PlatformEnum, list[type[FP8ScaledMMLinearKernel]]] =
MarlinFP8ScaledMMLinearKernel,
FlashInferFP8ScaledMMLinearKernel,
CutlassFP8ScaledMMLinearKernel,
B12xTensorFP8ScaledMMLinearKernel,
PerTensorTorchFP8ScaledMMLinearKernel,
ChannelWiseTorchFP8ScaledMMLinearKernel,
HummingFP8ScaledMMLinearKernel,
@@ -411,6 +434,7 @@ _POSSIBLE_FP8_BLOCK_KERNELS: dict[
FlashInferFp8DeepGEMMDynamicBlockScaledKernel,
DeepGemmFp8BlockScaledMMKernel,
CutlassFp8BlockScaledMMKernel,
B12xFp8BlockScaledMMKernel,
MarlinFP8ScaledMMLinearKernel,
TritonFp8BlockScaledMMKernel,
HummingFP8ScaledMMLinearKernel,
@@ -482,6 +506,7 @@ _POSSIBLE_MXFP8_KERNELS: dict[PlatformEnum, list[type[Mxfp8LinearKernel]]] = {
FlashInferCutedslMxfp8LinearKernel,
FlashInferCutlassMxfp8LinearKernel,
MarlinMxfp8LinearKernel,
B12xMxfp8LinearKernel,
EmulationMxfp8LinearKernel,
HummingMxfp8LinearKernel,
],
@@ -507,6 +532,7 @@ _POSSIBLE_NVFP4_KERNELS: dict[PlatformEnum, list[type[NvFp4LinearKernel]]] = {
FlashInferTrtllmNvFp4LinearKernel,
FlashInferCudnnNvFp4LinearKernel,
FbgemmNvFp4LinearKernel,
B12xNvFp4LinearKernel,
EmulationNvFp4LinearKernel,
HummingNvFp4LinearKernel,
],
@@ -529,6 +555,7 @@ _POSSIBLE_MXFP4_KERNELS: dict[PlatformEnum, list[type[MxFp4LinearKernel]]] = {
FlashInferMxFp4LinearKernel,
MarlinMxFp4LinearKernel,
HummingMxFp4LinearKernel,
B12xMxFp4LinearKernel,
EmulationMxfp4LinearKernel,
],
PlatformEnum.ROCM: [
@@ -1175,6 +1202,9 @@ __all__ = [
"init_mxfp8_linear_kernel",
"Mxfp8LinearKernel",
"Mxfp8LinearLayerConfig",
"B12xMxfp8LinearKernel",
"B12xMxFp4LinearKernel",
"B12xNvFp4LinearKernel",
"init_mxfp4_linear_kernel",
"MxFp4LinearKernel",
"MxFp4LinearLayerConfig",
@@ -1203,4 +1233,6 @@ __all__ = [
"_KernelT",
"DeepGemmFp8BlockScaledMMKernel",
"FlashInferFp8DeepGEMMDynamicBlockScaledKernel",
"B12xFp8BlockScaledMMKernel",
"B12xTensorFP8ScaledMMLinearKernel",
]
@@ -0,0 +1,170 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
from collections.abc import Iterable
from typing import Any
import torch
from vllm.model_executor.layers.quantization.utils.quant_utils import (
kMxfp4Dynamic,
)
from vllm.model_executor.utils import replace_parameter
from vllm.platforms import current_platform
from vllm.utils.b12x import (
b12x_warmup_token_counts,
)
from vllm.utils.b12x import (
get_b12x_blockscaled as _import_b12x_blockscaled,
)
from vllm.utils.b12x import get_b12x_intrinsics as _import_b12x_intrinsics
from .base import MxFp4LinearKernel, MxFp4LinearLayerConfig
def _apply_b12x_mxfp4_linear(
x: torch.Tensor,
weight: torch.Tensor,
weight_scale_storage: torch.Tensor,
bias: torch.Tensor | None,
) -> torch.Tensor:
from vllm.utils.flashinfer import flashinfer_mxfp4_quantize
blockscaled = _import_b12x_blockscaled()
assert blockscaled is not None
output_size = int(weight.shape[0])
output_shape = [*x.shape[:-1], output_size]
x_2d = x.reshape(-1, x.shape[-1]).contiguous()
x_packed, x_scale_swizzled = flashinfer_mxfp4_quantize(x_2d, backend="cute-dsl")
output = blockscaled.mm_mxfp4(
x_packed,
x_scale_swizzled,
weight,
weight_scale_storage,
out_dtype=x.dtype,
)
if bias is not None:
output = output + bias
return output.view(*output_shape)
def warmup_b12x_mxfp4_linear(
model: torch.nn.Module,
*,
max_tokens: int,
cudagraph_capture_sizes: Iterable[int] = (),
output_dtype: torch.dtype = torch.bfloat16,
) -> int:
if not current_platform.is_cuda():
return 0
if not current_platform.is_device_capability_family(120):
return 0
if output_dtype not in (torch.bfloat16, torch.float16):
output_dtype = torch.bfloat16
layer_map: dict[tuple[Any, ...], torch.nn.Module] = {}
for layer in model.modules():
if not getattr(layer, "b12x_mxfp4_linear", False):
continue
weight = layer.weight
weight_scale = layer.weight_scale
n, packed_k = map(int, weight.shape)
signature = (
weight.device,
n,
packed_k * 2,
weight.dtype,
weight_scale.dtype,
output_dtype,
)
layer_map.setdefault(signature, layer)
if not layer_map:
return 0
if _import_b12x_blockscaled() is None:
return 0
token_counts = b12x_warmup_token_counts(
max_tokens=max_tokens,
cudagraph_capture_sizes=cudagraph_capture_sizes,
)
warmed = 0
last_device: torch.device | None = None
with torch.inference_mode():
for signature, layer in layer_map.items():
weight = layer.weight
weight_scale = layer.weight_scale
k = signature[2]
last_device = weight.device
for tokens in token_counts:
source = torch.zeros(
(tokens, k),
dtype=output_dtype,
device=weight.device,
)
_apply_b12x_mxfp4_linear(
source,
weight,
weight_scale,
None,
)
warmed += 1
if warmed > 0 and last_device is not None and last_device.type == "cuda":
torch.accelerator.synchronize(last_device)
return warmed
class B12xMxFp4LinearKernel(MxFp4LinearKernel):
"""MXFP4 linear through the native B12X SM120 dense GEMM."""
@classmethod
def is_supported(
cls, compute_capability: int | None = None
) -> tuple[bool, str | None]:
del compute_capability
if not current_platform.is_cuda():
return False, "B12X MXFP4 kernels are only available on CUDA"
if not current_platform.is_device_capability_family(120):
return False, "B12X MXFP4 kernels require a Blackwell 12x device"
blockscaled = _import_b12x_blockscaled()
if blockscaled is None or _import_b12x_intrinsics() is None:
return False, "Install the B12X backend with `pip install vllm[b12x]`"
if not blockscaled.is_supported():
return False, "b12x native MXFP4 GEMM is not supported"
return True, None
@classmethod
def can_implement(cls, config: MxFp4LinearLayerConfig) -> tuple[bool, str | None]:
if config.activation_quant_key != kMxfp4Dynamic:
return False, "B12X MXFP4 GEMM requires dynamic MXFP4 activations"
return True, None
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
intrinsics = _import_b12x_intrinsics()
assert intrinsics is not None
replace_parameter(
layer,
"weight_scale",
intrinsics.swizzle_block_scale(layer.weight_scale.data),
)
layer.b12x_mxfp4_linear = True
def apply_weights(
self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
return _apply_b12x_mxfp4_linear(
x,
layer.weight,
layer.weight_scale,
bias,
)
__all__ = ["B12xMxFp4LinearKernel", "warmup_b12x_mxfp4_linear"]
@@ -0,0 +1,193 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
from collections.abc import Iterable
from typing import Any
import torch
from vllm.model_executor.layers.quantization.utils.mxfp8_utils import (
MXFP8_BLOCK_SIZE,
MXFP8_SCALE_DTYPE,
MXFP8_VALUE_DTYPE,
)
from vllm.model_executor.utils import replace_parameter
from vllm.platforms import current_platform
from vllm.utils.b12x import (
b12x_warmup_token_counts,
reuse_packed_weight_storage,
)
from vllm.utils.b12x import (
get_b12x_mxfp8_linear as _import_b12x_mxfp8,
)
from vllm.utils.torch_utils import current_stream
from .Mxfp8LinearKernel import Mxfp8LinearKernel, Mxfp8LinearLayerConfig
def _b12x_mxfp8_expected_m(tokens: int) -> int:
return max(1, int(tokens))
def _apply_b12x_mxfp8_packed_linear(
layer: torch.nn.Module,
x: torch.Tensor,
bias: torch.Tensor | None,
) -> torch.Tensor:
packed_weight = getattr(layer, "b12x_mxfp8_packed_weight", None)
if packed_weight is None:
raise RuntimeError(
"b12x MXFP8 packed weights are missing; "
"process_weights_after_loading did not run for this layer"
)
input_2d = x.reshape(-1, x.shape[-1]).contiguous()
output_shape = [*x.shape[:-1], int(packed_weight.out_features)]
mxfp8 = _import_b12x_mxfp8()
assert mxfp8 is not None
output = mxfp8.mm(
input_2d,
packed_weight,
bias=bias,
expected_m=_b12x_mxfp8_expected_m(int(input_2d.shape[0])),
)
return output.view(*output_shape)
def warmup_b12x_mxfp8_linear(
model: torch.nn.Module,
*,
max_tokens: int,
cudagraph_capture_sizes: Iterable[int] = (),
output_dtype: torch.dtype = torch.bfloat16,
) -> int:
if not current_platform.is_cuda():
return 0
if not current_platform.is_device_capability_family(120):
return 0
if output_dtype not in (torch.bfloat16, torch.float16):
output_dtype = torch.bfloat16
layer_map: dict[tuple[Any, ...], Any] = {}
for layer in model.modules():
packed_weight = getattr(layer, "b12x_mxfp8_packed_weight", None)
if packed_weight is None:
continue
device = torch.device(packed_weight.weight.values.device)
signature = (
device,
int(packed_weight.in_features),
int(packed_weight.padded_in_features),
int(packed_weight.out_features),
output_dtype,
)
layer_map.setdefault(signature, packed_weight)
if not layer_map:
return 0
mxfp8 = _import_b12x_mxfp8()
if mxfp8 is None:
return 0
token_counts = b12x_warmup_token_counts(
max_tokens=max_tokens,
cudagraph_capture_sizes=cudagraph_capture_sizes,
)
warmed = 0
last_device: torch.device | None = None
with torch.inference_mode():
for signature, packed_weight in layer_map.items():
device = signature[0]
last_device = device
for tokens in token_counts:
source = torch.zeros(
(tokens, int(packed_weight.in_features)),
dtype=output_dtype,
device=device,
)
mxfp8.mm(
source,
packed_weight,
expected_m=_b12x_mxfp8_expected_m(tokens),
stream=current_stream().cuda_stream,
)
warmed += 1
if warmed > 0 and last_device is not None and last_device.type == "cuda":
torch.accelerator.synchronize(last_device)
return warmed
class B12xMxfp8LinearKernel(Mxfp8LinearKernel):
"""ModelOpt MXFP8 linear through the native b12x SM120 dense GEMM path."""
@classmethod
def is_supported(
cls,
compute_capability: int | None = None,
) -> tuple[bool, str | None]:
del compute_capability
if not current_platform.is_cuda():
return False, "b12x MXFP8 kernels are only available on CUDA"
if not current_platform.is_device_capability_family(120):
return False, "b12x MXFP8 kernels require a Blackwell 12x device"
mxfp8 = _import_b12x_mxfp8()
if mxfp8 is None:
return False, "Install the B12X backend with `pip install vllm[b12x]`"
if not mxfp8.is_supported():
return False, "b12x.gemm.mxfp8_linear is not supported"
return True, None
@classmethod
def can_implement(cls, c: Mxfp8LinearLayerConfig) -> tuple[bool, str | None]:
del c
return True, None
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
weight = layer.weight.data
assert weight.dtype == MXFP8_VALUE_DTYPE, (
f"b12x MXFP8 requires {MXFP8_VALUE_DTYPE}, got {weight.dtype}"
)
assert weight.ndim == 2, f"b12x MXFP8 weight must be 2D, got {weight.ndim}D"
assert hasattr(layer, "weight_scale"), "b12x MXFP8 linear requires weight_scale"
out_features, in_features = map(int, weight.shape)
assert in_features % MXFP8_BLOCK_SIZE == 0, (
"b12x MXFP8 requires input features divisible by "
f"{MXFP8_BLOCK_SIZE}, got {in_features}"
)
weight_scale = layer.weight_scale.data
assert weight_scale.dtype == MXFP8_SCALE_DTYPE, (
f"b12x MXFP8 requires {MXFP8_SCALE_DTYPE} weight_scale, "
f"got {weight_scale.dtype}"
)
assert weight_scale.ndim == 2, (
f"b12x MXFP8 weight_scale must be 2D, got {weight_scale.ndim}D"
)
mxfp8 = _import_b12x_mxfp8()
assert mxfp8 is not None
scale_k = in_features // MXFP8_BLOCK_SIZE
packed_weight = mxfp8.pack_weight(
weight[:out_features, :in_features].detach(),
weight_scale[:out_features, :scale_k].detach(),
)
layer.b12x_mxfp8_packed_weight = reuse_packed_weight_storage(
getattr(layer, "b12x_mxfp8_packed_weight", None),
packed_weight,
)
replace_parameter(layer, "weight", weight.new_empty((0,)))
replace_parameter(layer, "weight_scale", weight_scale.new_empty((0,)))
def apply_weights(
self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
return _apply_b12x_mxfp8_packed_linear(layer, x, bias)
@@ -0,0 +1,176 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
from collections.abc import Iterable
from typing import Any
import torch
from vllm._custom_ops import scaled_fp4_quant
from vllm.model_executor.utils import replace_parameter
from vllm.platforms import current_platform
from vllm.utils.b12x import (
b12x_warmup_token_counts,
)
from vllm.utils.b12x import (
get_b12x_blockscaled as _import_b12x_blockscaled,
)
from vllm.utils.b12x import get_b12x_intrinsics as _import_b12x_intrinsics
from .base import NvFp4LinearKernel, NvFp4LinearLayerConfig
def _apply_b12x_nvfp4_linear(
x: torch.Tensor,
weight: torch.Tensor,
weight_scale_storage: torch.Tensor,
input_global_scale_inv: torch.Tensor,
alpha: torch.Tensor,
bias: torch.Tensor | None,
) -> torch.Tensor:
blockscaled = _import_b12x_blockscaled()
assert blockscaled is not None
output_size = int(weight.shape[0])
output_shape = [*x.shape[:-1], output_size]
x_2d = x.reshape(-1, x.shape[-1])
x_packed, x_scale_swizzled = scaled_fp4_quant(
x_2d,
input_global_scale_inv,
is_sf_swizzled_layout=True,
)
output = blockscaled.mm_nvfp4(
x_packed,
x_scale_swizzled,
weight,
weight_scale_storage,
alpha,
out_dtype=x.dtype,
)
if bias is not None:
output = output + bias
return output.view(*output_shape)
def warmup_b12x_nvfp4_linear(
model: torch.nn.Module,
*,
max_tokens: int,
cudagraph_capture_sizes: Iterable[int] = (),
output_dtype: torch.dtype = torch.bfloat16,
) -> int:
if not current_platform.is_cuda():
return 0
if not current_platform.is_device_capability_family(120):
return 0
if output_dtype not in (torch.bfloat16, torch.float16):
output_dtype = torch.bfloat16
layer_map: dict[tuple[Any, ...], torch.nn.Module] = {}
for layer in model.modules():
if not getattr(layer, "b12x_nvfp4_linear", False):
continue
weight = layer.weight
weight_scale = layer.weight_scale
n, packed_k = map(int, weight.shape)
signature = (
weight.device,
n,
packed_k * 2,
weight.dtype,
weight_scale.dtype,
output_dtype,
)
layer_map.setdefault(signature, layer)
if not layer_map:
return 0
if _import_b12x_blockscaled() is None:
return 0
token_counts = b12x_warmup_token_counts(
max_tokens=max_tokens,
cudagraph_capture_sizes=cudagraph_capture_sizes,
)
warmed = 0
last_device: torch.device | None = None
with torch.inference_mode():
for signature, layer in layer_map.items():
weight = layer.weight
weight_scale = layer.weight_scale
k = signature[2]
last_device = weight.device
for tokens in token_counts:
source = torch.zeros(
(tokens, k),
dtype=output_dtype,
device=weight.device,
)
_apply_b12x_nvfp4_linear(
source,
weight,
weight_scale,
layer.input_global_scale_inv,
layer.alpha,
None,
)
warmed += 1
if warmed > 0 and last_device is not None and last_device.type == "cuda":
torch.accelerator.synchronize(last_device)
return warmed
class B12xNvFp4LinearKernel(NvFp4LinearKernel):
"""ModelOpt NVFP4 linear through the native B12X SM120 dense GEMM."""
@classmethod
def is_supported(
cls, compute_capability: int | None = None
) -> tuple[bool, str | None]:
del compute_capability
if not current_platform.is_cuda():
return False, "B12X NVFP4 kernels are only available on CUDA"
if not current_platform.is_device_capability_family(120):
return False, "B12X NVFP4 kernels require a Blackwell 12x device"
blockscaled = _import_b12x_blockscaled()
if blockscaled is None or _import_b12x_intrinsics() is None:
return False, "Install the B12X backend with `pip install vllm[b12x]`"
if not blockscaled.is_supported():
return False, "b12x native NVFP4 GEMM is not supported"
return True, None
@classmethod
def can_implement(cls, config: NvFp4LinearLayerConfig) -> tuple[bool, str | None]:
del config
return True, None
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
intrinsics = _import_b12x_intrinsics()
assert intrinsics is not None
replace_parameter(
layer,
"weight_scale",
intrinsics.swizzle_block_scale(layer.weight_scale.data),
)
layer.b12x_nvfp4_linear = True
def apply_weights(
self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
return _apply_b12x_nvfp4_linear(
x,
layer.weight,
layer.weight_scale,
layer.input_global_scale_inv,
layer.alpha,
bias,
)
__all__ = ["B12xNvFp4LinearKernel", "warmup_b12x_nvfp4_linear"]
@@ -0,0 +1,214 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
from collections.abc import Iterable
from typing import Any
import torch
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
_upcast_e8m0_to_fp32,
)
from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
from vllm.model_executor.utils import replace_parameter
from vllm.platforms import current_platform
from vllm.utils.b12x import (
b12x_warmup_token_counts,
)
from vllm.utils.b12x import (
get_b12x_blockscaled as _import_b12x_blockscaled,
)
from .BlockScaledMMLinearKernel import (
Fp8BlockScaledMMLinearKernel,
FP8ScaledMMLinearLayerConfig,
)
def _run_b12x_fp8_block_scaled_mm(
a: torch.Tensor,
weight: torch.Tensor,
a_scale: torch.Tensor,
weight_scale: torch.Tensor,
out_dtype: torch.dtype,
) -> torch.Tensor:
blockscaled = _import_b12x_blockscaled()
assert blockscaled is not None
return blockscaled.mm_block_fp8(
a,
a_scale,
weight,
weight_scale,
out_dtype=out_dtype,
)
def warmup_b12x_block_fp8_linear(
model: torch.nn.Module,
*,
max_tokens: int,
cudagraph_capture_sizes: Iterable[int] = (),
output_dtype: torch.dtype = torch.bfloat16,
) -> int:
if not current_platform.is_cuda():
return 0
if not current_platform.is_device_capability_family(120):
return 0
if output_dtype not in (torch.bfloat16, torch.float16):
output_dtype = torch.bfloat16
layer_map: dict[tuple[Any, ...], torch.nn.Module] = {}
for layer in model.modules():
if not getattr(layer, "b12x_block_fp8_linear", False):
continue
weight = layer.weight
weight_scale = getattr(layer, "weight_scale_inv", None)
if weight_scale is None:
weight_scale = layer.weight_scale
n, k = map(int, weight.shape)
signature = (
weight.device,
n,
k,
weight.dtype,
weight_scale.dtype,
output_dtype,
)
layer_map.setdefault(signature, layer)
if not layer_map:
return 0
blockscaled = _import_b12x_blockscaled()
if blockscaled is None:
return 0
token_counts = b12x_warmup_token_counts(
max_tokens=max_tokens,
cudagraph_capture_sizes=cudagraph_capture_sizes,
)
warmed = 0
with torch.inference_mode():
for signature, layer in layer_map.items():
weight = layer.weight
weight_scale = getattr(layer, "weight_scale_inv", None)
if weight_scale is None:
weight_scale = layer.weight_scale
k = signature[2]
for tokens in token_counts:
a = torch.empty(
(tokens, k),
dtype=weight.dtype,
device=weight.device,
)
a_scale = torch.empty(
(tokens, k // 128),
dtype=torch.float32,
device=weight.device,
)
_run_b12x_fp8_block_scaled_mm(
a,
weight,
a_scale,
weight_scale,
output_dtype,
)
warmed += 1
return warmed
class B12xFp8BlockScaledMMKernel(Fp8BlockScaledMMLinearKernel):
"""K128 block-FP8 linear through the native B12X SM120 dense GEMM."""
@classmethod
def is_supported(
cls,
compute_capability: int | None = None,
) -> tuple[bool, str | None]:
del compute_capability
if not current_platform.is_cuda():
return False, "B12X FP8 kernels are only available on CUDA"
if not current_platform.is_device_capability_family(120):
return False, "B12X FP8 kernels require a Blackwell 12x device"
blockscaled = _import_b12x_blockscaled()
if blockscaled is None:
return False, "Install the B12X backend with `pip install vllm[b12x]`"
if not blockscaled.is_supported():
return False, "B12X regular block-FP8 GEMM is not supported"
return True, None
@classmethod
def can_implement(
cls,
config: FP8ScaledMMLinearLayerConfig,
) -> tuple[bool, str | None]:
can_implement_base, reason = super().can_implement(config)
if not can_implement_base:
return can_implement_base, reason
if config.input_dtype not in (torch.bfloat16, torch.float16):
return False, "Supports only bf16/fp16 input dtype"
if config.input_dtype != config.out_dtype:
return False, "Input and output dtype must match"
act_group_shape = config.activation_quant_key.scale.group_shape
if act_group_shape != GroupShape(1, 128):
return (
False,
"Supports only dynamic per-token group activation quantization "
"with group_shape=(1,128)",
)
weight_group_shape = config.weight_quant_key.scale.group_shape
if weight_group_shape != GroupShape(128, 128):
return False, "Supports only 128x128 block-scaled FP8 weights"
out_features, in_features = config.weight_shape
if in_features <= 0 or in_features % 128 != 0:
return False, "Input features must be a positive multiple of 128"
if out_features <= 0 or out_features % 128 != 0:
return False, "Output features must be a positive multiple of 128"
return True, None
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
super().process_weights_after_loading(layer)
params = self._get_layer_params(layer)
if params.weight_scale_inv is not None:
weight_scale = params.weight_scale_inv
scale_attr = params.WEIGHT_SCALE_INV
else:
weight_scale = params.weight_scale
scale_attr = params.WEIGHT_SCALE
if weight_scale is not None and weight_scale.dtype in (
torch.float8_e8m0fnu,
torch.uint8,
):
# TODO: Remove once B12X supports 128x128 UE8M0 block scales.
replace_parameter(
layer,
scale_attr,
_upcast_e8m0_to_fp32(weight_scale).contiguous(),
)
layer.b12x_block_fp8_linear = True
def apply_block_scaled_mm(
self,
A: torch.Tensor,
B: torch.Tensor,
As: torch.Tensor,
Bs: torch.Tensor,
) -> torch.Tensor:
return _run_b12x_fp8_block_scaled_mm(
A,
B,
As,
Bs,
self.config.out_dtype,
)
__all__ = [
"B12xFp8BlockScaledMMKernel",
"warmup_b12x_block_fp8_linear",
]
@@ -0,0 +1,238 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
from collections.abc import Iterable
from typing import Any
import torch
from vllm.model_executor.utils import replace_parameter
from vllm.platforms import current_platform
from vllm.utils.b12x import (
b12x_warmup_token_counts,
reuse_packed_weight_storage,
)
from vllm.utils.b12x import (
get_b12x_tensor_fp8_linear as _import_b12x_tensor_fp8,
)
from vllm.utils.torch_utils import current_stream
from .ScaledMMLinearKernel import (
FP8ScaledMMLinearKernel,
FP8ScaledMMLinearLayerConfig,
)
def _apply_b12x_tensor_fp8_packed_linear(
layer: torch.nn.Module,
x_q: torch.Tensor,
bias: torch.Tensor | None,
out_dtype: torch.dtype,
) -> torch.Tensor:
packed_weight = getattr(layer, "b12x_tensor_fp8_packed_weight", None)
if packed_weight is None:
raise RuntimeError(
"b12x tensor FP8 packed weights are missing; "
"process_weights_after_loading did not run for this layer"
)
tensor_fp8 = _import_b12x_tensor_fp8()
assert tensor_fp8 is not None
input_2d = x_q.reshape(-1, x_q.shape[-1]).contiguous()
output_shape = [*x_q.shape[:-1], int(packed_weight.out_features)]
output = tensor_fp8.mm(
input_2d,
packed_weight,
bias=bias,
out_dtype=out_dtype,
expected_m=max(1, int(input_2d.shape[0])),
)
return output.view(*output_shape)
def warmup_b12x_tensor_fp8_linear(
model: torch.nn.Module,
*,
max_tokens: int,
cudagraph_capture_sizes: Iterable[int] = (),
output_dtype: torch.dtype = torch.bfloat16,
) -> int:
if not current_platform.is_cuda():
return 0
if not current_platform.is_device_capability_family(120):
return 0
if output_dtype not in (torch.bfloat16, torch.float16):
output_dtype = torch.bfloat16
layer_map: dict[tuple[Any, ...], Any] = {}
for layer in model.modules():
packed_weight = getattr(
layer,
"b12x_tensor_fp8_packed_weight",
None,
)
if packed_weight is None:
continue
device = torch.device(packed_weight.values.device)
signature = (
device,
int(packed_weight.in_features),
int(packed_weight.padded_in_features),
int(packed_weight.out_features),
output_dtype,
)
layer_map.setdefault(signature, packed_weight)
if not layer_map:
return 0
tensor_fp8 = _import_b12x_tensor_fp8()
if tensor_fp8 is None:
return 0
token_counts = b12x_warmup_token_counts(
max_tokens=max_tokens,
cudagraph_capture_sizes=cudagraph_capture_sizes,
)
warmed = 0
last_device: torch.device | None = None
with torch.inference_mode():
for signature, packed_weight in layer_map.items():
last_device = signature[0]
warmed += int(
tensor_fp8.prewarm(
packed_weight,
token_counts,
out_dtype=output_dtype,
stream=current_stream().cuda_stream,
)
)
if warmed > 0 and last_device is not None and last_device.type == "cuda":
torch.accelerator.synchronize(last_device)
return warmed
class B12xTensorFP8ScaledMMLinearKernel(FP8ScaledMMLinearKernel):
"""Static per-tensor FP8 linear through the B12X SM12x dense GEMM."""
@classmethod
def is_supported(
cls,
compute_capability: int | None = None,
) -> tuple[bool, str | None]:
del compute_capability
if not current_platform.is_cuda():
return False, "b12x tensor FP8 kernels are only available on CUDA"
if not current_platform.is_device_capability_family(120):
return False, "b12x tensor FP8 kernels require a Blackwell 12x device"
tensor_fp8 = _import_b12x_tensor_fp8()
if tensor_fp8 is None:
return False, "Install the B12X backend with `pip install vllm[b12x]`"
if not tensor_fp8.is_supported():
return False, "b12x.gemm.tensor_fp8_linear is not supported"
return True, None
@classmethod
def can_implement(
cls,
config: FP8ScaledMMLinearLayerConfig,
) -> tuple[bool, str | None]:
activation_scale = config.activation_quant_key.scale
weight_scale = config.weight_quant_key.scale
if (
not activation_scale.static
or not activation_scale.group_shape.is_per_tensor()
):
return False, "requires static per-tensor activation scales"
if not weight_scale.static or not weight_scale.group_shape.is_per_tensor():
return False, "requires static per-tensor weight scales"
if config.input_dtype not in (torch.bfloat16, torch.float16):
return False, "supports only bf16/fp16 input dtype"
if config.out_dtype not in (torch.bfloat16, torch.float16):
return False, "supports only bf16/fp16 output dtype"
out_features, in_features = config.weight_shape
if out_features <= 0 or in_features <= 0 or in_features % 32 != 0:
return False, "weight dimensions must be positive with K divisible by 32"
return True, None
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
weight, weight_scale, input_scale, _ = self._get_layer_params(layer)
if weight.dtype != torch.float8_e4m3fn:
raise ValueError(
f"b12x tensor FP8 requires float8_e4m3fn weight, got {weight.dtype}"
)
if weight_scale.numel() != 1 or input_scale is None or input_scale.numel() != 1:
raise ValueError(
"b12x tensor FP8 requires scalar weight and activation scales"
)
out_features, in_features = map(int, self.config.weight_shape)
if tuple(weight.shape) != (in_features, out_features):
raise ValueError(
"b12x tensor FP8 expects the processed weight in [K,N] layout, "
f"got {tuple(weight.shape)} for N={out_features}, K={in_features}"
)
tensor_fp8 = _import_b12x_tensor_fp8()
assert tensor_fp8 is not None
output_scale = (
input_scale.detach().to(torch.float32).reshape(1)
* weight_scale.detach().to(torch.float32).reshape(1)
).contiguous()
packed_weight = tensor_fp8.pack_weight(
weight.detach().T.contiguous(),
output_scale,
)
layer.b12x_tensor_fp8_packed_weight = reuse_packed_weight_storage(
getattr(layer, "b12x_tensor_fp8_packed_weight", None),
packed_weight,
)
weight_name, weight_scale_name, _, _ = self.layer_param_names
replace_parameter(layer, weight_name, weight.new_empty((0,)))
replace_parameter(layer, weight_scale_name, weight_scale.new_empty((0,)))
def apply_weights(
self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
if not isinstance(x, torch.Tensor):
raise TypeError("b12x tensor FP8 linear requires a Tensor input")
_, _, input_scale, input_scale_ub = self._get_layer_params(layer)
input_2d = x.reshape(-1, x.shape[-1])
x_q, _ = self.quant_fp8(input_2d, input_scale, input_scale_ub)
out_dtype = self.config.out_dtype
output = _apply_b12x_tensor_fp8_packed_linear(
layer,
x_q,
bias,
out_dtype,
)
return output.view(*x.shape[:-1], output.shape[-1])
def apply_scaled_mm(
self,
*,
A: torch.Tensor,
B: torch.Tensor,
out_dtype: torch.dtype,
As: torch.Tensor,
Bs: torch.Tensor,
bias: torch.Tensor | None,
output_shape: list,
) -> torch.Tensor:
del A, B, out_dtype, As, Bs, bias, output_shape
raise NotImplementedError("b12x tensor FP8 linear overrides apply_weights")
__all__ = [
"B12xTensorFP8ScaledMMLinearKernel",
"warmup_b12x_tensor_fp8_linear",
]
+60
View File
@@ -0,0 +1,60 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Warm B12X JIT kernels used by a loaded model."""
from typing import TYPE_CHECKING
import torch
from vllm.logger import init_logger
from vllm.model_executor.kernels.linear.mxfp4.b12x import (
warmup_b12x_mxfp4_linear,
)
from vllm.model_executor.kernels.linear.mxfp8.b12x import (
warmup_b12x_mxfp8_linear,
)
from vllm.model_executor.kernels.linear.nvfp4.b12x import (
warmup_b12x_nvfp4_linear,
)
from vllm.model_executor.kernels.linear.scaled_mm.b12x_block import (
warmup_b12x_block_fp8_linear,
)
from vllm.model_executor.kernels.linear.scaled_mm.b12x_tensor import (
warmup_b12x_tensor_fp8_linear,
)
if TYPE_CHECKING:
from vllm.v1.worker.gpu_worker import Worker
logger = init_logger(__name__)
def b12x_warmup(worker: "Worker", cudagraph_capture_sizes: list[int]) -> None:
model = worker.get_model()
max_tokens = worker.scheduler_config.max_num_batched_tokens
output_dtype = getattr(
getattr(worker, "model_config", None),
"dtype",
torch.bfloat16,
)
warmup_kwargs = {
"max_tokens": max_tokens,
"cudagraph_capture_sizes": cudagraph_capture_sizes,
"output_dtype": output_dtype,
}
providers = (
("block-FP8", warmup_b12x_block_fp8_linear),
("MXFP8", warmup_b12x_mxfp8_linear),
("tensor FP8", warmup_b12x_tensor_fp8_linear),
("MXFP4", warmup_b12x_mxfp4_linear),
("NVFP4", warmup_b12x_nvfp4_linear),
)
for name, warmup in providers:
warmed = warmup(model, **warmup_kwargs)
if warmed:
logger.info_once(
"Warmed up %d B12X %s linear GEMM signatures.",
warmed,
name,
)
+7 -3
View File
@@ -12,6 +12,7 @@ import torch
import vllm.envs as envs
from vllm.logger import init_logger
from vllm.model_executor.warmup.b12x_warmup import b12x_warmup
from vllm.model_executor.warmup.cutedsl_warmup import cutedsl_warmup
from vllm.model_executor.warmup.deep_gemm_warmup import deep_gemm_warmup
from vllm.model_executor.warmup.deepseek_v4_mhc_warmup import (
@@ -113,15 +114,16 @@ def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):
qwen_triton_warmup(worker.model_runner, worker.vllm_config.model_config)
compilation_config = worker.vllm_config.compilation_config
cudagraph_capture_sizes = list(compilation_config.cudagraph_capture_sizes or [])
# DSv4 mHC TileLang kernels (hc_pre/hc_post/hc_head_op) run every decoder
# layer per token; warm them across token sizes first so the first real
# request doesn't pay JIT cost. No-op for non-DSv4 models (gated inside).
deepseek_v4_mhc_warmup(
worker.get_model(),
max_tokens=worker.scheduler_config.max_num_batched_tokens,
cudagraph_capture_sizes=(
worker.vllm_config.compilation_config.cudagraph_capture_sizes or []
),
cudagraph_capture_sizes=cudagraph_capture_sizes,
)
# Run next so input-prep kernels JIT against pristine runner state.
@@ -156,6 +158,8 @@ def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):
max_tokens = worker.scheduler_config.max_num_batched_tokens
deep_gemm_warmup(model, max_tokens)
b12x_warmup(worker, cudagraph_capture_sizes)
minimax_m3_msa_warmup(worker)
enable_flashinfer_autotune = (
+101
View File
@@ -0,0 +1,101 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Lazy accessors for the optional ``b12x`` package."""
import functools
import importlib
import importlib.util
from collections.abc import Iterable
from dataclasses import fields, is_dataclass
from types import ModuleType
from typing import Any
import torch
@functools.cache
def has_b12x() -> bool:
"""Return whether the B12X package is installed."""
return importlib.util.find_spec("b12x") is not None
@functools.cache
def _get_submodule(module_name: str) -> ModuleType | None:
if not has_b12x():
return None
try:
return importlib.import_module(module_name)
except (ImportError, ModuleNotFoundError):
return None
def get_b12x_blockscaled() -> ModuleType | None:
return _get_submodule("b12x.gemm.blockscaled")
def get_b12x_intrinsics() -> ModuleType | None:
return _get_submodule("b12x._lib.intrinsics")
def get_b12x_mxfp8_linear() -> ModuleType | None:
return _get_submodule("b12x.gemm.mxfp8_linear")
def get_b12x_tensor_fp8_linear() -> ModuleType | None:
return _get_submodule("b12x.gemm.tensor_fp8_linear")
def b12x_warmup_token_counts(
*,
max_tokens: int,
cudagraph_capture_sizes: Iterable[int] = (),
) -> tuple[int, ...]:
# B12X deduplicates shapes that select the same internal kernel policy.
# Keep the complete serving shape set here rather than duplicating its
# policy-selection heuristics in vLLM.
counts = {1}
counts.update(int(size) for size in cudagraph_capture_sizes if int(size) > 0)
if int(max_tokens) > 0:
counts.add(int(max_tokens))
return tuple(sorted(counts))
def _same_packed_layout(current: Any, replacement: Any) -> bool:
if type(current) is not type(replacement):
return False
if isinstance(current, torch.Tensor):
return (
current.shape == replacement.shape
and current.stride() == replacement.stride()
and current.dtype == replacement.dtype
and current.device == replacement.device
)
if is_dataclass(current):
return all(
_same_packed_layout(
getattr(current, field.name),
getattr(replacement, field.name),
)
for field in fields(current)
)
return bool(current == replacement)
def _copy_packed_tensors(current: Any, replacement: Any) -> None:
if isinstance(current, torch.Tensor):
current.copy_(replacement)
elif is_dataclass(current):
for field in fields(current):
_copy_packed_tensors(
getattr(current, field.name),
getattr(replacement, field.name),
)
@torch.no_grad()
def reuse_packed_weight_storage(current: Any, replacement: Any) -> Any:
"""Reuse packed tensor addresses when a compatible weight is reloaded."""
if current is None or not _same_packed_layout(current, replacement):
return replacement
_copy_packed_tensors(current, replacement)
return current