[Online quantization] Add online MXFP4 quantization support (#49347)
Signed-off-by: Felix Marty <Felix.Marty@amd.com>
This commit is contained in:
@@ -36,6 +36,10 @@ from vllm.model_executor.layers.quantization.online.fp8 import (
|
||||
from vllm.model_executor.layers.quantization.online.int8 import (
|
||||
Int8OnlineMoEMethod,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.online.mxfp4 import (
|
||||
Mxfp4OnlineLinearMethod,
|
||||
Mxfp4OnlineMoEMethod,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.online.mxfp8 import (
|
||||
Mxfp8OnlineLinearMethod,
|
||||
Mxfp8OnlineMoEMethod,
|
||||
@@ -49,6 +53,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
|
||||
kFp8StaticChannelSym,
|
||||
kFp8StaticTensorSym,
|
||||
kInt8StaticChannelSym,
|
||||
kMxfp4Static,
|
||||
kMxfp8Dynamic,
|
||||
kNvfp4Static,
|
||||
)
|
||||
@@ -64,6 +69,7 @@ _ONLINE_LINEAR_METHODS: dict[QuantKey, type] = {
|
||||
kFp8Static128BlockSym: Fp8PerBlockOnlineLinearMethod,
|
||||
kFp8StaticChannelSym: Fp8PtpcOnlineLinearMethod,
|
||||
kMxfp8Dynamic: Mxfp8OnlineLinearMethod,
|
||||
kMxfp4Static: Mxfp4OnlineLinearMethod,
|
||||
}
|
||||
|
||||
_ONLINE_MOE_METHODS: dict[QuantKey, type] = {
|
||||
@@ -71,6 +77,7 @@ _ONLINE_MOE_METHODS: dict[QuantKey, type] = {
|
||||
kFp8Static128BlockSym: Fp8PerBlockOnlineMoEMethod,
|
||||
kFp8StaticChannelSym: Fp8PtpcOnlineMoEMethod,
|
||||
kMxfp8Dynamic: Mxfp8OnlineMoEMethod,
|
||||
kMxfp4Static: Mxfp4OnlineMoEMethod,
|
||||
kInt8StaticChannelSym: Int8OnlineMoEMethod,
|
||||
kNvfp4Static: Nvfp4OnlineMoEMethod,
|
||||
}
|
||||
|
||||
@@ -639,36 +639,6 @@ class Fp8PerBlockOnlineMoEMethod(_Fp8OnlineMoEBase):
|
||||
round_up(intermediate_size_per_partition, block_size),
|
||||
)
|
||||
|
||||
def _zero_padding(self, layer: Module) -> None:
|
||||
hidden_size = layer.moe_config.hidden_dim_unpadded
|
||||
intermediate_size = layer.moe_config.intermediate_size_per_partition_unpadded
|
||||
|
||||
w13_shard = layer.w13_weight.shape[1] // self.moe.w13_num_shards
|
||||
if w13_shard > intermediate_size:
|
||||
for shard in range(self.moe.w13_num_shards):
|
||||
start = shard * w13_shard + intermediate_size
|
||||
layer.w13_weight[:, start : (shard + 1) * w13_shard, :] = 0
|
||||
if layer.w13_weight.shape[2] > hidden_size:
|
||||
layer.w13_weight[:, :, hidden_size:] = 0
|
||||
|
||||
if layer.w2_weight.shape[1] > hidden_size:
|
||||
layer.w2_weight[:, hidden_size:, :] = 0
|
||||
if layer.w2_weight.shape[2] > intermediate_size:
|
||||
layer.w2_weight[:, :, intermediate_size:] = 0
|
||||
|
||||
if getattr(layer, "w13_bias", None) is not None:
|
||||
w13_bias_shard = layer.w13_bias.shape[1] // self.moe.w13_num_shards
|
||||
if w13_bias_shard > intermediate_size:
|
||||
for shard in range(self.moe.w13_num_shards):
|
||||
start = shard * w13_bias_shard + intermediate_size
|
||||
layer.w13_bias[:, start : (shard + 1) * w13_bias_shard] = 0
|
||||
|
||||
if (
|
||||
getattr(layer, "w2_bias", None) is not None
|
||||
and layer.w2_bias.shape[1] > hidden_size
|
||||
):
|
||||
layer.w2_bias[:, hidden_size:] = 0
|
||||
|
||||
def process_weights_after_loading(self, layer: Module) -> None:
|
||||
if getattr(layer, "_already_called_process_weights_after_loading", False):
|
||||
return
|
||||
|
||||
@@ -95,6 +95,36 @@ class OnlineMoEMethodBase(FusedMoEMethodBase):
|
||||
|
||||
initialize_online_processing(layer)
|
||||
|
||||
def _zero_padding(self, layer: torch.nn.Module) -> None:
|
||||
hidden_size = layer.moe_config.hidden_dim_unpadded
|
||||
intermediate_size = layer.moe_config.intermediate_size_per_partition_unpadded
|
||||
|
||||
w13_shard = layer.w13_weight.shape[1] // self.moe.w13_num_shards
|
||||
if w13_shard > intermediate_size:
|
||||
for shard in range(self.moe.w13_num_shards):
|
||||
start = shard * w13_shard + intermediate_size
|
||||
layer.w13_weight[:, start : (shard + 1) * w13_shard, :] = 0
|
||||
if layer.w13_weight.shape[2] > hidden_size:
|
||||
layer.w13_weight[:, :, hidden_size:] = 0
|
||||
|
||||
if layer.w2_weight.shape[1] > hidden_size:
|
||||
layer.w2_weight[:, hidden_size:, :] = 0
|
||||
if layer.w2_weight.shape[2] > intermediate_size:
|
||||
layer.w2_weight[:, :, intermediate_size:] = 0
|
||||
|
||||
if getattr(layer, "w13_bias", None) is not None:
|
||||
w13_bias_shard = layer.w13_bias.shape[1] // self.moe.w13_num_shards
|
||||
if w13_bias_shard > intermediate_size:
|
||||
for shard in range(self.moe.w13_num_shards):
|
||||
start = shard * w13_bias_shard + intermediate_size
|
||||
layer.w13_bias[:, start : (shard + 1) * w13_bias_shard] = 0
|
||||
|
||||
if (
|
||||
getattr(layer, "w2_bias", None) is not None
|
||||
and layer.w2_bias.shape[1] > hidden_size
|
||||
):
|
||||
layer.w2_bias[:, hidden_size:] = 0
|
||||
|
||||
@abstractmethod
|
||||
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
||||
pass
|
||||
|
||||
@@ -0,0 +1,317 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
"""Online MXFP4 (microscaling FP4, block-32) quantization methods."""
|
||||
|
||||
from typing import TYPE_CHECKING
|
||||
|
||||
import torch
|
||||
from torch.nn import Module
|
||||
|
||||
if TYPE_CHECKING:
|
||||
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
FusedMoEParallelConfig,
|
||||
FusedMoEQuantConfig,
|
||||
RoutedExperts,
|
||||
)
|
||||
|
||||
from vllm.model_executor.kernels.linear import init_mxfp4_linear_kernel
|
||||
from vllm.model_executor.layers.fused_moe.oracle.mxfp4 import (
|
||||
TRITON_BACKENDS,
|
||||
Mxfp4MoeBackend,
|
||||
convert_weight_to_mxfp4_moe_kernel_format,
|
||||
make_mxfp4_moe_kernel,
|
||||
make_mxfp4_moe_quant_config,
|
||||
mxfp4_round_up_hidden_size_and_intermediate_size,
|
||||
select_mxfp4_moe_backend,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.online.fp8 import (
|
||||
_Fp8OnlineLinearBase,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.online.moe_base import (
|
||||
OnlineMoEMethodBase,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.utils.mxfp4_utils import (
|
||||
mxfp4_quantize,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import kMxfp4Dynamic
|
||||
from vllm.model_executor.utils import replace_parameter
|
||||
|
||||
# MXFP4 block size (elements per shared e8m0 scale), same as the checkpoint
|
||||
# and MoE oracle paths.
|
||||
MXFP4_BLOCK_SIZE = 32
|
||||
|
||||
|
||||
def _quantize_mxfp4_moe_weight(
|
||||
weight: torch.Tensor,
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
"""Batch quantization: bf16/fp16 stacked expert weights -> MXFP4.
|
||||
|
||||
Returns packed FP4 weights (uint8, two values per byte) and per-block
|
||||
(group-32) e8m0 scales (uint8), one pair per expert.
|
||||
"""
|
||||
num_experts = weight.size(0)
|
||||
first_quant, first_scale = mxfp4_quantize(weight[0])
|
||||
# Pre-allocate the output tensors rather than stacking.
|
||||
# This is important for consistent memory layout.
|
||||
w_quant = torch.empty(
|
||||
(num_experts, *first_quant.shape),
|
||||
dtype=first_quant.dtype,
|
||||
device=weight.device,
|
||||
)
|
||||
w_scales = torch.empty(
|
||||
(num_experts, *first_scale.shape),
|
||||
dtype=first_scale.dtype,
|
||||
device=weight.device,
|
||||
)
|
||||
w_quant[0] = first_quant
|
||||
w_scales[0] = first_scale
|
||||
for i in range(1, num_experts):
|
||||
w_quant[i], w_scales[i] = mxfp4_quantize(weight[i])
|
||||
|
||||
return w_quant, w_scales
|
||||
|
||||
|
||||
class Mxfp4OnlineLinearMethod(_Fp8OnlineLinearBase):
|
||||
"""Online MXFP4 linear method.
|
||||
Loads bf16/fp16 checkpoints and quantizes weights to MXFP4 (microscaling
|
||||
FP4 with block-32 scales) during weight loading.
|
||||
"""
|
||||
|
||||
activation_quant_key = kMxfp4Dynamic
|
||||
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self.kernel = init_mxfp4_linear_kernel(
|
||||
activation_quant_key=self.activation_quant_key
|
||||
)
|
||||
|
||||
def create_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
input_size_per_partition: int,
|
||||
output_partition_sizes: list[int],
|
||||
input_size: int,
|
||||
output_size: int,
|
||||
params_dtype: torch.dtype,
|
||||
**extra_weight_attrs,
|
||||
):
|
||||
if input_size_per_partition % MXFP4_BLOCK_SIZE != 0:
|
||||
raise ValueError(
|
||||
f"MXFP4 requires input_size_per_partition "
|
||||
f"({input_size_per_partition}) to be divisible by "
|
||||
f"{MXFP4_BLOCK_SIZE}."
|
||||
)
|
||||
|
||||
super().create_weights(
|
||||
layer,
|
||||
input_size_per_partition,
|
||||
output_partition_sizes,
|
||||
input_size,
|
||||
output_size,
|
||||
params_dtype,
|
||||
**extra_weight_attrs,
|
||||
)
|
||||
|
||||
def process_weights_after_loading(self, layer: Module) -> None:
|
||||
if getattr(layer, "_already_called_process_weights_after_loading", False):
|
||||
return
|
||||
|
||||
weight_fp4, weight_scale = mxfp4_quantize(layer.weight.contiguous())
|
||||
|
||||
layer.input_scale = None
|
||||
replace_parameter(layer, "weight", weight_fp4.data)
|
||||
replace_parameter(layer, "weight_scale", weight_scale.data)
|
||||
|
||||
self.kernel.process_weights_after_loading(layer)
|
||||
|
||||
layer._already_called_process_weights_after_loading = True
|
||||
|
||||
def apply(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
bias: torch.Tensor | None = None,
|
||||
) -> torch.Tensor:
|
||||
return self.kernel.apply_weights(layer, x, bias)
|
||||
|
||||
|
||||
class Mxfp4OnlineMoEMethod(OnlineMoEMethodBase):
|
||||
"""MoE method for online MXFP4 (block) quantization."""
|
||||
|
||||
mxfp4_backend: Mxfp4MoeBackend
|
||||
experts_cls: "type[mk.FusedMoEExperts] | None"
|
||||
activation_quant_key = kMxfp4Dynamic
|
||||
|
||||
def __init__(self, *, layer: torch.nn.Module):
|
||||
super().__init__(layer.moe_config)
|
||||
self.weight_block_size: list[int] = [1, MXFP4_BLOCK_SIZE]
|
||||
self.weight_scale_name = "weight_scale"
|
||||
|
||||
self.mxfp4_backend, self.experts_cls = select_mxfp4_moe_backend(
|
||||
config=self.moe, activation_key=self.activation_quant_key
|
||||
)
|
||||
|
||||
def maybe_roundup_sizes(
|
||||
self,
|
||||
hidden_size: int,
|
||||
intermediate_size_per_partition: int,
|
||||
act_dtype: torch.dtype,
|
||||
moe_parallel_config: "FusedMoEParallelConfig",
|
||||
) -> tuple[int, int]:
|
||||
hidden_size, intermediate_size_per_partition = super().maybe_roundup_sizes(
|
||||
hidden_size=hidden_size,
|
||||
intermediate_size_per_partition=intermediate_size_per_partition,
|
||||
act_dtype=act_dtype,
|
||||
moe_parallel_config=moe_parallel_config,
|
||||
)
|
||||
return mxfp4_round_up_hidden_size_and_intermediate_size(
|
||||
self.mxfp4_backend, hidden_size, intermediate_size_per_partition
|
||||
)
|
||||
|
||||
def create_weights(
|
||||
self,
|
||||
layer: Module,
|
||||
num_experts: int,
|
||||
hidden_size: int,
|
||||
intermediate_size_per_partition: int,
|
||||
params_dtype: torch.dtype,
|
||||
**extra_weight_attrs,
|
||||
):
|
||||
if (
|
||||
hidden_size % MXFP4_BLOCK_SIZE != 0
|
||||
or intermediate_size_per_partition % MXFP4_BLOCK_SIZE != 0
|
||||
):
|
||||
raise ValueError(
|
||||
"Online MXFP4 MoE requires hidden/intermediate sizes divisible "
|
||||
f"by {MXFP4_BLOCK_SIZE}."
|
||||
)
|
||||
|
||||
super().create_weights(
|
||||
layer=layer,
|
||||
num_experts=num_experts,
|
||||
hidden_size=hidden_size,
|
||||
intermediate_size_per_partition=intermediate_size_per_partition,
|
||||
params_dtype=params_dtype,
|
||||
**extra_weight_attrs,
|
||||
)
|
||||
|
||||
layer.weight_block_size = [1, MXFP4_BLOCK_SIZE]
|
||||
|
||||
def _setup_kernel(
|
||||
self,
|
||||
layer: "RoutedExperts",
|
||||
w13: torch.Tensor,
|
||||
w2: torch.Tensor,
|
||||
w13_scale: torch.Tensor,
|
||||
w2_scale: torch.Tensor,
|
||||
w13_bias: torch.Tensor | None,
|
||||
w2_bias: torch.Tensor | None,
|
||||
) -> None:
|
||||
w13, w2, w13_scale, w2_scale, w13_bias, w2_bias = (
|
||||
convert_weight_to_mxfp4_moe_kernel_format(
|
||||
mxfp4_backend=self.mxfp4_backend,
|
||||
layer=layer,
|
||||
w13_weight=w13,
|
||||
w2_weight=w2,
|
||||
w13_weight_scale=w13_scale,
|
||||
w2_weight_scale=w2_scale,
|
||||
w13_bias=w13_bias,
|
||||
w2_bias=w2_bias,
|
||||
)
|
||||
)
|
||||
|
||||
# Handle weight/scale assignment based on backend type
|
||||
# Copied from `quark_moe.py`.
|
||||
# TODO: This should not be here, replace_parameter should handle
|
||||
# triton_kernels.tensor.Tensor.
|
||||
if self.mxfp4_backend in TRITON_BACKENDS or self.mxfp4_backend in (
|
||||
Mxfp4MoeBackend.AITER_MXFP4_FP8,
|
||||
):
|
||||
# Triton-based backends: w13/w2 are triton_kernels.tensor.Tensor
|
||||
# Store on layer for apply(), scales are PrecisionConfig
|
||||
layer.w13_weight = w13
|
||||
layer.w2_weight = w2
|
||||
self.w13_precision_config = w13_scale
|
||||
self.w2_precision_config = w2_scale
|
||||
else:
|
||||
# Standard backends: replace parameters
|
||||
replace_parameter(layer, "w13_weight", w13)
|
||||
replace_parameter(layer, "w2_weight", w2)
|
||||
replace_parameter(layer, "w13_weight_scale", w13_scale)
|
||||
replace_parameter(layer, "w2_weight_scale", w2_scale)
|
||||
|
||||
if w13_bias is not None and w2_bias is not None:
|
||||
replace_parameter(layer, "w13_bias", w13_bias)
|
||||
replace_parameter(layer, "w2_bias", w2_bias)
|
||||
|
||||
self.moe_quant_config = self.get_fused_moe_quant_config(layer)
|
||||
if self.moe_quant_config is not None and self.experts_cls is not None:
|
||||
self.moe_kernel = make_mxfp4_moe_kernel(
|
||||
moe_quant_config=self.moe_quant_config,
|
||||
moe_config=self.moe,
|
||||
mxfp4_backend=self.mxfp4_backend,
|
||||
experts_cls=self.experts_cls,
|
||||
routing_tables=layer._expert_routing_tables(),
|
||||
layer=layer,
|
||||
)
|
||||
|
||||
def get_fused_moe_quant_config(
|
||||
self, layer: torch.nn.Module
|
||||
) -> "FusedMoEQuantConfig | None":
|
||||
# NOTE: unlike `QuarkOCP_MX_MoEMethod.get_fused_moe_quant_config`, this
|
||||
# does not branch on `self.mxfp4_backend in TRITON_BACKENDS or
|
||||
# AITER_MXFP4_FP8` to read `self.w13_precision_config`/
|
||||
# `self.w2_precision_config` instead of `layer.w13_weight_scale`/
|
||||
# `layer.w2_weight_scale`. For those backends,
|
||||
# `convert_weight_to_mxfp4_moe_kernel_format` deletes the latter, so
|
||||
# `getattr` below would raise `AttributeError`. This is currently
|
||||
# unreachable because `activation_quant_key` is hardcoded to
|
||||
# `kMxfp4Dynamic`
|
||||
# TODO: When supporting online activation quant key override,
|
||||
# fix this convert_weight_to_mxfp4_moe_kernel_format issue.
|
||||
w1_scale = getattr(layer, f"w13_{self.weight_scale_name}")
|
||||
w2_scale = getattr(layer, f"w2_{self.weight_scale_name}")
|
||||
|
||||
return make_mxfp4_moe_quant_config(
|
||||
mxfp4_backend=self.mxfp4_backend,
|
||||
w1_scale=w1_scale,
|
||||
w2_scale=w2_scale,
|
||||
w1_bias=getattr(layer, "w13_bias", None),
|
||||
w2_bias=getattr(layer, "w2_bias", None),
|
||||
a1_scale=layer.w13_input_scale,
|
||||
a2_scale=layer.w2_input_scale,
|
||||
swiglu_limit=getattr(layer, "swiglu_limit", None),
|
||||
gemm1_alpha=getattr(layer, "swiglu_alpha", None),
|
||||
gemm1_beta=getattr(layer, "swiglu_beta", None),
|
||||
layer=layer,
|
||||
)
|
||||
|
||||
def process_weights_after_loading(self, layer: Module) -> None:
|
||||
if getattr(layer, "_already_called_process_weights_after_loading", False):
|
||||
return
|
||||
|
||||
self._zero_padding(layer)
|
||||
|
||||
if self.mxfp4_backend == Mxfp4MoeBackend.NONE:
|
||||
layer._already_called_process_weights_after_loading = True
|
||||
return
|
||||
|
||||
layer.w13_input_scale = None
|
||||
layer.w2_input_scale = None
|
||||
|
||||
w13, w13_scale = _quantize_mxfp4_moe_weight(layer.w13_weight)
|
||||
w2, w2_scale = _quantize_mxfp4_moe_weight(layer.w2_weight)
|
||||
|
||||
self._setup_kernel(
|
||||
layer,
|
||||
w13,
|
||||
w2,
|
||||
w13_scale,
|
||||
w2_scale,
|
||||
getattr(layer, "w13_bias", None),
|
||||
getattr(layer, "w2_bias", None),
|
||||
)
|
||||
|
||||
layer._already_called_process_weights_after_loading = True
|
||||
@@ -6,6 +6,7 @@ import torch
|
||||
|
||||
from vllm.logger import init_logger
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.triton_utils import tl, triton
|
||||
from vllm.utils.import_utils import has_triton_kernels
|
||||
from vllm.utils.torch_utils import direct_register_custom_op, is_torch_equal_or_newer
|
||||
|
||||
@@ -184,3 +185,319 @@ except AttributeError as error:
|
||||
|
||||
def xpu_mxfp4_quantize(x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
return torch.ops.vllm.xpu_mxfp4_quantize(x)
|
||||
|
||||
|
||||
def permute_shape(shape: tuple[int, ...], permute_order: list[int]) -> tuple[int, ...]:
|
||||
return tuple(shape[i] for i in permute_order)
|
||||
|
||||
|
||||
@triton.jit
|
||||
def _compute_quant_and_scale(src_tensor, valid_src_mask):
|
||||
BLOCK_SIZE_OUT_DIM: tl.constexpr = src_tensor.shape[0]
|
||||
BLOCK_SIZE_QUANT_DIM: tl.constexpr = src_tensor.shape[1]
|
||||
BLOCK_SIZE_QUANT_MX_SCALE: tl.constexpr = src_tensor.shape[1] // 32
|
||||
|
||||
# Explicit cast to fp32 since most ops are not supported on bfloat16. We
|
||||
# avoid needless conversions to and from bf16
|
||||
f32_tensor = src_tensor.to(tl.float32)
|
||||
abs_tensor = tl.abs(f32_tensor)
|
||||
abs_tensor = tl.where(
|
||||
valid_src_mask, abs_tensor, -1.0
|
||||
) # Don't consider padding tensors in scale computation
|
||||
abs_tensor = tl.reshape(
|
||||
abs_tensor, [BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_MX_SCALE, 32]
|
||||
)
|
||||
max_val = tl.max(abs_tensor, axis=2, keep_dims=True)
|
||||
# Round-to-nearest-even block scale (OCP MX spec), matching hardware/aiter
|
||||
# MXFP4 quantizers: compute 2 ** (floor(log2(rounding(max_abs(v)))) - max_exp).
|
||||
max_val = max_val.to(tl.int32, bitcast=True)
|
||||
max_val = (max_val + 0x200000).to(tl.uint32, bitcast=True) & 0x7F800000
|
||||
max_val = max_val.to(tl.float32, bitcast=True)
|
||||
# Add epsilon to prevent log2(0) = -inf when max_val is zero (all-zero input block)
|
||||
eps = tl.where(max_val == 0.0, 2 ** (-126), 0.0)
|
||||
scale_e8m0_unbiased = tl.log2(max_val + eps).floor() - 2
|
||||
scale_e8m0_unbiased = tl.clamp(scale_e8m0_unbiased, min=-127, max=127)
|
||||
dequant_scale_rounded = tl.exp2(scale_e8m0_unbiased)
|
||||
dequant_scale_exponent = dequant_scale_rounded.to(tl.uint32, bitcast=True)
|
||||
|
||||
dequant_scale_rounded = dequant_scale_exponent.to(tl.float32, bitcast=True)
|
||||
quant_scale = tl.where(dequant_scale_rounded == 0, 0, 1.0 / dequant_scale_rounded)
|
||||
|
||||
f32_tensor = tl.reshape(
|
||||
f32_tensor, [BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_MX_SCALE, 32]
|
||||
)
|
||||
quant_tensor = f32_tensor * quant_scale
|
||||
|
||||
# Reshape the tensors after scaling
|
||||
quant_tensor = quant_tensor.reshape([BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_DIM])
|
||||
# Set the invalid portions of the tensor to 0. This will ensure that any
|
||||
# padding tensors are 0 in the mx format.
|
||||
quant_tensor = tl.where(valid_src_mask, quant_tensor, 0)
|
||||
dequant_scale_exponent = dequant_scale_exponent.reshape(
|
||||
[BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_MX_SCALE]
|
||||
)
|
||||
|
||||
# First, we simply extract the exponent part of the scales and store the result
|
||||
dequant_scale_exponent = (dequant_scale_exponent >> 23).to(tl.uint8)
|
||||
# Now we must convert the tensor to packed e2m1 (mxfp4).
|
||||
quant_tensor = quant_tensor.to(tl.uint32, bitcast=True)
|
||||
signs = quant_tensor & 0x80000000
|
||||
exponents = (quant_tensor >> 23) & 0xFF
|
||||
mantissas = quant_tensor & 0x7FFFFF
|
||||
|
||||
# 0.25 <= x < 0.75 maps to 0.5, a denormal number
|
||||
E8_BIAS = 127
|
||||
E2_BIAS = 1
|
||||
# Move implicit bit 1 at the beginning to mantissa for denormals. Bits
|
||||
# shifted out are OR-reduced into the LSB as a sticky bit, so the
|
||||
# round-to-even logic below can tell an exact tie from a value that is
|
||||
# merely close to one.
|
||||
adjusted_exponents = tl.core.sub(E8_BIAS, exponents + 1, sanitize_overflow=False)
|
||||
folded_mantissas = 0x400000 | (mantissas >> 1)
|
||||
denormal_mantissas = folded_mantissas >> adjusted_exponents
|
||||
denormal_mantissas |= (
|
||||
(denormal_mantissas << adjusted_exponents) != folded_mantissas
|
||||
).to(tl.uint32)
|
||||
mantissas = tl.where(exponents < E8_BIAS, denormal_mantissas, mantissas)
|
||||
|
||||
# For normal numbers, we change the bias from 127 to 1, and for subnormals,
|
||||
# we keep exponent as 0.
|
||||
exponents = tl.maximum(exponents, E8_BIAS - E2_BIAS) - (E8_BIAS - E2_BIAS)
|
||||
|
||||
# Combine sign, exponent, and mantissa, rounding to nearest with ties
|
||||
# broken to even (OCP MX spec, matching hardware/aiter MXFP4
|
||||
# quantizers) via the standard "add a rounding bias, then truncate"
|
||||
# trick: the bias's LSB is the kept mantissa bit, so an exact tie
|
||||
# rounds to whichever candidate is even, while any lower (sticky) bit
|
||||
# being set always rounds up.
|
||||
rounding_bias = ((mantissas >> 22) & 1) + 0x1FFFFF
|
||||
rounded_mantissa = (mantissas + rounding_bias) >> 22 # in {0, 1, 2}
|
||||
e2m1_tmp = tl.minimum((exponents << 1) + rounded_mantissa, 0x7)
|
||||
e2m1_value = ((signs >> 28) | e2m1_tmp).to(tl.uint8)
|
||||
|
||||
e2m1_value = tl.reshape(
|
||||
e2m1_value, [BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_DIM // 2, 2]
|
||||
)
|
||||
evens, odds = tl.split(e2m1_value)
|
||||
out_tensor = evens | (odds << 4)
|
||||
|
||||
return out_tensor, dequant_scale_exponent
|
||||
|
||||
|
||||
@triton.jit
|
||||
def _downcast_to_mxfp(
|
||||
mx_tensor_ptr,
|
||||
stride_mxt_outer,
|
||||
stride_mxt_quant: tl.constexpr,
|
||||
mx_scale_ptr,
|
||||
stride_mx_scale_outer,
|
||||
stride_mx_scale_quant,
|
||||
src_ptr,
|
||||
stride_src_outer,
|
||||
stride_src_quant,
|
||||
outer_dim,
|
||||
quant_dim,
|
||||
BLOCK_SIZE_OUT_DIM: tl.constexpr,
|
||||
BLOCK_SIZE_QUANT_DIM: tl.constexpr,
|
||||
):
|
||||
tl.static_assert(
|
||||
stride_mxt_quant == 1, f"Output stride, {stride_mxt_quant=} must be 1."
|
||||
)
|
||||
tl.static_assert(
|
||||
BLOCK_SIZE_QUANT_DIM % 32 == 0,
|
||||
f"{BLOCK_SIZE_QUANT_DIM=} must be a multiple of 32",
|
||||
)
|
||||
|
||||
# uint8 signifies two fp4 e2m1 values packed into a single byte
|
||||
tl.static_assert(
|
||||
mx_tensor_ptr.dtype.element_ty == tl.uint8,
|
||||
f"Invalid {mx_tensor_ptr.dtype.element_ty=}. Must be uint8.",
|
||||
)
|
||||
|
||||
src_dtype: tl.constexpr = src_ptr.dtype.element_ty
|
||||
tl.static_assert(
|
||||
mx_scale_ptr.dtype.element_ty == tl.uint8,
|
||||
f"{mx_scale_ptr.dtype.element_ty=} must be uint8",
|
||||
)
|
||||
tl.static_assert(
|
||||
(src_dtype == tl.bfloat16) or (src_dtype == tl.float16),
|
||||
f"{src_dtype=} must be bfloat16 or float16",
|
||||
)
|
||||
|
||||
outer_block = tl.program_id(0).to(tl.int64)
|
||||
quant_block = tl.program_id(1).to(tl.int64)
|
||||
|
||||
BLOCK_SIZE_QUANT_MX_SCALE: tl.constexpr = BLOCK_SIZE_QUANT_DIM // 32
|
||||
BLOCK_SIZE_QUANT_MX_TENSOR: tl.constexpr = BLOCK_SIZE_QUANT_DIM // 2
|
||||
|
||||
start_src_quant = quant_block * BLOCK_SIZE_QUANT_DIM
|
||||
start_mx_scale_quant = quant_block * BLOCK_SIZE_QUANT_MX_SCALE
|
||||
start_mx_quant = quant_block * BLOCK_SIZE_QUANT_MX_TENSOR
|
||||
start_out = outer_block * BLOCK_SIZE_OUT_DIM
|
||||
|
||||
src_ptr += start_src_quant * stride_src_quant + start_out * stride_src_outer
|
||||
mx_scale_ptr += (
|
||||
start_mx_scale_quant * stride_mx_scale_quant + start_out * stride_mx_scale_outer
|
||||
)
|
||||
mx_tensor_ptr += start_mx_quant * stride_mxt_quant + start_out * stride_mxt_outer
|
||||
|
||||
offs_src_quant = tl.arange(0, BLOCK_SIZE_QUANT_DIM)[None, :].to(tl.int64)
|
||||
offs_mxt_quant = tl.arange(0, BLOCK_SIZE_QUANT_MX_TENSOR)[None, :].to(tl.int64)
|
||||
offs_scale_quant = tl.arange(0, BLOCK_SIZE_QUANT_MX_SCALE)[None, :].to(tl.int64)
|
||||
offs_outer = tl.arange(0, BLOCK_SIZE_OUT_DIM)[:, None].to(tl.int64)
|
||||
|
||||
mask_src_quant = start_src_quant + offs_src_quant < quant_dim
|
||||
mask_n = start_out + offs_outer < outer_dim
|
||||
full_mask_src = mask_src_quant and mask_n
|
||||
|
||||
mask_mxt_quant = start_mx_quant + offs_mxt_quant < tl.cdiv(quant_dim, 2)
|
||||
full_mask_mxt = mask_mxt_quant and mask_n
|
||||
|
||||
scale_mask_k = start_mx_scale_quant + offs_scale_quant < tl.cdiv(quant_dim, 32)
|
||||
full_scale_mask = scale_mask_k and mask_n
|
||||
|
||||
src_tensor_offsets = (
|
||||
offs_src_quant * stride_src_quant + offs_outer * stride_src_outer
|
||||
)
|
||||
mx_scale_offsets = (
|
||||
offs_scale_quant * stride_mx_scale_quant + offs_outer * stride_mx_scale_outer
|
||||
)
|
||||
mx_tensor_offsets = (
|
||||
offs_mxt_quant * stride_mxt_quant + offs_outer * stride_mxt_outer
|
||||
)
|
||||
src_tensor = tl.load(src_ptr + src_tensor_offsets, mask=full_mask_src)
|
||||
|
||||
out_tensor, scale_tensor = _compute_quant_and_scale(src_tensor, full_mask_src)
|
||||
|
||||
tl.store(mx_scale_ptr + mx_scale_offsets, scale_tensor, mask=full_scale_mask)
|
||||
tl.store(mx_tensor_ptr + mx_tensor_offsets, out_tensor, mask=full_mask_mxt)
|
||||
|
||||
|
||||
def downcast_to_mxfp(
|
||||
src_tensor: torch.Tensor,
|
||||
axis: int,
|
||||
out_quant_tensor: torch.Tensor | None = None,
|
||||
out_scale: torch.Tensor | None = None,
|
||||
BLOCK_OUT_DIM: int = 128,
|
||||
BLOCK_QUANT_DIM: int = 32,
|
||||
):
|
||||
"""
|
||||
Convert the src weights to MXFP4. The src weight is quantized along the
|
||||
axis dimension into packed e2m1 values (torch.uint8, two values per
|
||||
byte), so the size of that dimension in the output is half of the
|
||||
logical (unpacked) size.
|
||||
"""
|
||||
out_quant_type = torch.uint8
|
||||
ndim = src_tensor.ndim
|
||||
assert -ndim <= axis < ndim, f"Invalid axis {axis=}"
|
||||
axis = axis if axis >= 0 else axis + ndim
|
||||
|
||||
L = src_tensor.shape[axis]
|
||||
# We make this assertion since we can't track if the "real" shape was odd,
|
||||
# and we padded it to be even.
|
||||
# We want to maintain the property dequant(quant(x)).shape == x.shape
|
||||
assert L % 2 == 0, f"axis dim must be divisible by 2 for e2m1. Got {L}"
|
||||
|
||||
device = src_tensor.device
|
||||
|
||||
packed_quant_dim = triton.cdiv(L, 2)
|
||||
out_scale_dim = triton.cdiv(L, 32)
|
||||
|
||||
# Move the quantization axis to the end for the kernel, then permute back.
|
||||
permute_order = list(range(ndim))
|
||||
permute_order[axis], permute_order[-1] = permute_order[-1], permute_order[axis]
|
||||
|
||||
prmted_quant_tensor_shape = permute_shape(src_tensor.shape, permute_order)[:-1] + (
|
||||
packed_quant_dim,
|
||||
)
|
||||
prmted_scale_shape = permute_shape(src_tensor.shape, permute_order)[:-1] + (
|
||||
out_scale_dim,
|
||||
)
|
||||
prmted_src_tensor = src_tensor.permute(permute_order)
|
||||
|
||||
if out_quant_tensor is None:
|
||||
out_quant_tensor = torch.empty(
|
||||
prmted_quant_tensor_shape, dtype=out_quant_type, device=device
|
||||
)
|
||||
else:
|
||||
expected_shape = (
|
||||
src_tensor.shape[:axis] + (packed_quant_dim,) + src_tensor.shape[axis + 1 :]
|
||||
)
|
||||
assert out_quant_tensor.shape == expected_shape, (
|
||||
f"{out_quant_tensor.shape=} != {expected_shape=}"
|
||||
)
|
||||
assert out_quant_tensor.dtype == out_quant_type, (
|
||||
f"{out_quant_tensor.dtype=} != {out_quant_type=}"
|
||||
)
|
||||
assert out_quant_tensor.stride(axis) == 1, (
|
||||
f"{out_quant_tensor.stride(axis)=} != 1"
|
||||
)
|
||||
# We expect the axis dimension to be last, so permute the tensor
|
||||
out_quant_tensor = out_quant_tensor.permute(permute_order)
|
||||
|
||||
if out_scale is None:
|
||||
out_scale = torch.empty(prmted_scale_shape, dtype=torch.uint8, device=device)
|
||||
else:
|
||||
expected_scale_shape = permute_shape(prmted_scale_shape, permute_order)
|
||||
assert out_scale.shape == expected_scale_shape, (
|
||||
f"{out_scale.shape=} {expected_scale_shape=}"
|
||||
)
|
||||
assert out_scale.dtype == torch.uint8, f"{out_scale.dtype=} != torch.uint8"
|
||||
out_scale = out_scale.permute(permute_order)
|
||||
|
||||
# Flatten input tensor for kernel. This will typically make a copy
|
||||
reshaped_src_tensor = prmted_src_tensor.reshape(-1, L)
|
||||
blocks_quant_dim = triton.cdiv(reshaped_src_tensor.shape[-1], BLOCK_QUANT_DIM)
|
||||
blocks_out_dim = triton.cdiv(reshaped_src_tensor.shape[0], BLOCK_OUT_DIM)
|
||||
|
||||
# Flatten the output tensors for the kernel, this should be a view always
|
||||
kernel_quant_tensor = out_quant_tensor.reshape(-1, packed_quant_dim)
|
||||
kernel_scale = out_scale.reshape(-1, out_scale_dim)
|
||||
assert kernel_quant_tensor.data_ptr() == out_quant_tensor.data_ptr()
|
||||
assert kernel_scale.data_ptr() == out_scale.data_ptr()
|
||||
|
||||
_downcast_to_mxfp[(blocks_out_dim, blocks_quant_dim)](
|
||||
kernel_quant_tensor,
|
||||
*kernel_quant_tensor.stride(),
|
||||
kernel_scale,
|
||||
*kernel_scale.stride(),
|
||||
reshaped_src_tensor,
|
||||
*reshaped_src_tensor.stride(),
|
||||
*reshaped_src_tensor.shape,
|
||||
BLOCK_OUT_DIM,
|
||||
BLOCK_QUANT_DIM,
|
||||
num_warps=8,
|
||||
)
|
||||
|
||||
out_quant_tensor = out_quant_tensor.permute(permute_order)
|
||||
out_scale = out_scale.permute(permute_order).contiguous()
|
||||
return out_quant_tensor, out_scale, permute_shape(prmted_scale_shape, permute_order)
|
||||
|
||||
|
||||
def mxfp4_quantize(x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
"""Quantize a bf16/fp16 tensor to MXFP4 along its last dimension.
|
||||
|
||||
Dispatches to the fastest backend available on the current platform:
|
||||
|
||||
- the native XPU custom op on XPU,
|
||||
- aiter on ROCm when `aiter` is installed,
|
||||
- and the portable Triton kernel (``downcast_to_mxfp``) otherwise.
|
||||
|
||||
Returns packed FP4 values (uint8, two values per byte) and per-block
|
||||
(group-32) e8m0 scales (uint8).
|
||||
"""
|
||||
if current_platform.is_xpu():
|
||||
return xpu_mxfp4_quantize(x)
|
||||
|
||||
from vllm._aiter_ops import is_aiter_found_and_supported
|
||||
|
||||
if is_aiter_found_and_supported() and x.dtype == torch.bfloat16:
|
||||
from vllm.model_executor.layers.quantization.quark.utils import (
|
||||
quark_quantize_weight_to_mxfp4,
|
||||
)
|
||||
|
||||
return quark_quantize_weight_to_mxfp4(x)
|
||||
|
||||
quant_tensor, scale, _ = downcast_to_mxfp(x, axis=-1)
|
||||
|
||||
return quant_tensor, scale
|
||||
|
||||
Reference in New Issue
Block a user