[Feature][Quantization] MXFP4 support for MOE models (#17888)
Signed-off-by: Felix Marty <felmarty@amd.com> Signed-off-by: Bowen Bao <bowenbao@amd.com> Signed-off-by: Felix Marty <Felix.Marty@amd.com> Co-authored-by: Bowen Bao <bowenbao@amd.com>
This commit is contained in:
@@ -237,12 +237,6 @@ class QuarkConfig(QuantizationConfig):
|
||||
"Quark model is not in MX-FP4 format: not group_size=32")
|
||||
return False
|
||||
|
||||
# Weights need to use static quantization.
|
||||
if weight_quant.get("is_dynamic") is True:
|
||||
logger.debug(
|
||||
"Quark model is not in MX-FP4 format: not weight static")
|
||||
return False
|
||||
|
||||
# Activations need to use dynamic quantization.
|
||||
if input_quant.get("is_dynamic") is False:
|
||||
logger.debug(
|
||||
|
||||
@@ -5,11 +5,12 @@ from typing import Any, Callable, Optional
|
||||
|
||||
import torch
|
||||
|
||||
import vllm.model_executor.layers.fused_moe # noqa
|
||||
from vllm import _custom_ops as ops
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.fused_moe import (FusedMoE, FusedMoEMethodBase,
|
||||
FusedMoeWeightScaleSupported)
|
||||
from vllm.model_executor.layers.quantization.utils.mxfp4_utils import (
|
||||
OCP_MX_BLOCK_SIZE)
|
||||
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
|
||||
all_close_1d, normalize_e4m3fn_to_e4m3fnuz, per_tensor_dequantize)
|
||||
from vllm.model_executor.utils import set_weight_attrs
|
||||
@@ -17,7 +18,9 @@ from vllm.platforms import current_platform
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
__all__ = ["QuarkMoEMethod", "QuarkW8A8Fp8MoEMethod"]
|
||||
__all__ = [
|
||||
"QuarkMoEMethod", "QuarkW8A8Fp8MoEMethod", "QuarkW4A4MXFp4MoEMethod"
|
||||
]
|
||||
|
||||
|
||||
class QuarkMoEMethod(FusedMoEMethodBase):
|
||||
@@ -40,6 +43,8 @@ class QuarkMoEMethod(FusedMoEMethodBase):
|
||||
|
||||
if quant_config._is_fp8_w8a8(weight_config, input_config):
|
||||
return QuarkW8A8Fp8MoEMethod(weight_config, input_config)
|
||||
elif quant_config._is_mx_fp4(weight_config, input_config):
|
||||
return QuarkW4A4MXFp4MoEMethod(weight_config, input_config)
|
||||
else:
|
||||
raise RuntimeError("Unsupported FusedMoe scheme")
|
||||
|
||||
@@ -242,4 +247,163 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
|
||||
w1_scale=layer.w13_weight_scale,
|
||||
w2_scale=layer.w2_weight_scale,
|
||||
a1_scale=layer.w13_input_scale,
|
||||
a2_scale=layer.w2_input_scale)
|
||||
a2_scale=layer.w2_input_scale,
|
||||
activation=activation)
|
||||
|
||||
|
||||
class QuarkW4A4MXFp4MoEMethod(QuarkMoEMethod):
|
||||
|
||||
def __init__(self, weight_config: dict[str, Any], input_config: dict[str,
|
||||
Any]):
|
||||
self.weight_quant = weight_config
|
||||
self.input_quant = input_config
|
||||
|
||||
weight_qscheme = self.weight_quant.get("qscheme")
|
||||
input_qscheme = self.input_quant.get("qscheme")
|
||||
if not (weight_qscheme == "per_group"
|
||||
and input_qscheme == "per_group"):
|
||||
raise ValueError(
|
||||
"For MX(FP4) Fused MoE layers, only per-group scales "
|
||||
"for weights and activations are supported. Found "
|
||||
f"{weight_qscheme}, {input_qscheme}") # noqa E501
|
||||
|
||||
self.static_input_scales = not self.input_quant.get("is_dynamic")
|
||||
|
||||
if self.static_input_scales:
|
||||
raise NotImplementedError(
|
||||
"QuarkW4A4MXFp4MoEMethod with static input scales is currently "
|
||||
"not implemented. Please open an issue.")
|
||||
|
||||
if not current_platform.supports_mx():
|
||||
self.emulate = True
|
||||
logger.warning_once(
|
||||
"The current platform does not support native MXFP4 "
|
||||
"computation. Simulated weight dequantization and activation "
|
||||
"QDQ (quantize and dequantize) will be used, with the linear "
|
||||
"layers computed in high precision.")
|
||||
else:
|
||||
self.emulate = True
|
||||
logger.warning_once(
|
||||
"The current platform supports native MXFP4 "
|
||||
"computation, but kernels are not yet integrated in vLLM. "
|
||||
"Simulated weight dequantization and activation "
|
||||
"QDQ (quantize and dequantize) will be used, with the linear "
|
||||
"layers computed in high precision.")
|
||||
|
||||
def create_weights(self, layer: torch.nn.Module, num_experts: int,
|
||||
hidden_size: int, intermediate_size_per_partition: int,
|
||||
params_dtype: torch.dtype, **extra_weight_attrs):
|
||||
|
||||
# Add the quantization method used (per tensor/grouped/channel)
|
||||
# to ensure the weight scales are loaded in properly
|
||||
extra_weight_attrs.update(
|
||||
{"quant_method": FusedMoeWeightScaleSupported.BLOCK.value})
|
||||
|
||||
params_dtype = torch.uint8
|
||||
|
||||
# WEIGHTS
|
||||
w13_weight = torch.nn.Parameter(torch.empty(
|
||||
num_experts,
|
||||
2 * intermediate_size_per_partition,
|
||||
hidden_size // 2,
|
||||
dtype=params_dtype),
|
||||
requires_grad=False)
|
||||
layer.register_parameter("w13_weight", w13_weight)
|
||||
|
||||
set_weight_attrs(w13_weight, extra_weight_attrs)
|
||||
|
||||
w2_weight = torch.nn.Parameter(torch.empty(
|
||||
num_experts,
|
||||
hidden_size,
|
||||
intermediate_size_per_partition // 2,
|
||||
dtype=params_dtype),
|
||||
requires_grad=False)
|
||||
layer.register_parameter("w2_weight", w2_weight)
|
||||
|
||||
set_weight_attrs(w2_weight, extra_weight_attrs)
|
||||
|
||||
# WEIGHT_SCALES
|
||||
w13_weight_scale = torch.nn.Parameter(
|
||||
torch.ones(
|
||||
num_experts,
|
||||
2 * intermediate_size_per_partition,
|
||||
hidden_size // OCP_MX_BLOCK_SIZE,
|
||||
dtype=params_dtype,
|
||||
),
|
||||
requires_grad=False,
|
||||
)
|
||||
w2_weight_scale = torch.nn.Parameter(
|
||||
torch.ones(
|
||||
num_experts,
|
||||
hidden_size,
|
||||
intermediate_size_per_partition // OCP_MX_BLOCK_SIZE,
|
||||
dtype=params_dtype,
|
||||
),
|
||||
requires_grad=False,
|
||||
)
|
||||
set_weight_attrs(w2_weight_scale, extra_weight_attrs)
|
||||
set_weight_attrs(w13_weight_scale, extra_weight_attrs)
|
||||
|
||||
layer.register_parameter("w13_weight_scale", w13_weight_scale)
|
||||
layer.register_parameter("w2_weight_scale", w2_weight_scale)
|
||||
|
||||
def apply(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
router_logits: torch.Tensor,
|
||||
top_k: int,
|
||||
renormalize: bool,
|
||||
use_grouped_topk: bool = False,
|
||||
topk_group: Optional[int] = None,
|
||||
num_expert_group: Optional[int] = None,
|
||||
global_num_experts: int = -1,
|
||||
expert_map: Optional[torch.Tensor] = None,
|
||||
custom_routing_function: Optional[Callable] = None,
|
||||
scoring_func: str = "softmax",
|
||||
e_score_correction_bias: Optional[torch.Tensor] = None,
|
||||
apply_router_weight_on_input: bool = False,
|
||||
activation: str = "silu",
|
||||
enable_eplb: bool = False,
|
||||
expert_load_view: Optional[torch.Tensor] = None,
|
||||
logical_to_physical_map: Optional[torch.Tensor] = None,
|
||||
logical_replica_count: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
|
||||
if enable_eplb:
|
||||
raise NotImplementedError(
|
||||
"EPLB not supported for `QuarkW4A4MXFp4MoEMethod` yet.")
|
||||
|
||||
from vllm.model_executor.layers.fused_moe import fused_experts
|
||||
|
||||
topk_weights, topk_ids = FusedMoE.select_experts(
|
||||
hidden_states=x,
|
||||
router_logits=router_logits,
|
||||
use_grouped_topk=use_grouped_topk,
|
||||
top_k=top_k,
|
||||
renormalize=renormalize,
|
||||
topk_group=topk_group,
|
||||
num_expert_group=num_expert_group,
|
||||
custom_routing_function=custom_routing_function,
|
||||
scoring_func=scoring_func,
|
||||
e_score_correction_bias=e_score_correction_bias)
|
||||
|
||||
out = fused_experts(
|
||||
x,
|
||||
layer.w13_weight,
|
||||
layer.w2_weight,
|
||||
topk_weights=topk_weights,
|
||||
topk_ids=topk_ids,
|
||||
inplace=True,
|
||||
use_mxfp4_w4a4=True,
|
||||
global_num_experts=global_num_experts,
|
||||
apply_router_weight_on_input=apply_router_weight_on_input,
|
||||
expert_map=expert_map,
|
||||
w1_scale=layer.w13_weight_scale,
|
||||
w2_scale=layer.w2_weight_scale,
|
||||
a1_scale=None,
|
||||
a2_scale=None,
|
||||
block_shape=None,
|
||||
activation=activation,
|
||||
)
|
||||
return out
|
||||
|
||||
@@ -6,14 +6,16 @@ from typing import Any, Callable, Optional
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
|
||||
import vllm.envs as envs
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.quantization.quark.schemes import QuarkScheme
|
||||
from vllm.model_executor.layers.quantization.utils.mxfp4_utils import (
|
||||
OCP_MX_BLOCK_SIZE, per_token_group_quant_mxfp4)
|
||||
OCP_MX_BLOCK_SIZE, dequant_mxfp4, quant_dequant_mxfp4)
|
||||
from vllm.model_executor.parameter import (GroupQuantScaleParameter,
|
||||
PackedvLLMParameter)
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
__all__ = ["QuarkW4A4MXFP4"]
|
||||
|
||||
|
||||
@@ -25,7 +27,29 @@ class QuarkW4A4MXFP4(QuarkScheme):
|
||||
self.qscheme = "per_group"
|
||||
self.weight_quant_spec = weight_quant_spec
|
||||
self.input_quant_spec = input_quant_spec
|
||||
self.emulate = not current_platform.supports_mx()
|
||||
|
||||
self.static_input_scales = not input_quant_spec.get("is_dynamic")
|
||||
|
||||
if self.static_input_scales:
|
||||
raise NotImplementedError(
|
||||
"QuarkW4A4MXFP4 with static input scales is currently not "
|
||||
"implemented. Please open an issue.")
|
||||
|
||||
if not current_platform.supports_mx():
|
||||
self.emulate = True
|
||||
logger.warning_once(
|
||||
"The current platform does not support native MXFP4 "
|
||||
"computation. Simulated weight dequantization and activation "
|
||||
"QDQ (quantize and dequantize) will be used, with the linear "
|
||||
"layers computed in high precision.")
|
||||
else:
|
||||
self.emulate = True
|
||||
logger.warning_once(
|
||||
"The current platform supports native MXFP4 "
|
||||
"computation, but kernels are not yet integrated in vLLM. "
|
||||
"Simulated weight dequantization and activation "
|
||||
"QDQ (quantize and dequantize) will be used, with the linear "
|
||||
"layers computed in high precision.")
|
||||
|
||||
@classmethod
|
||||
def get_min_capability(cls) -> int:
|
||||
@@ -37,43 +61,6 @@ class QuarkW4A4MXFP4(QuarkScheme):
|
||||
layer.weight_scale = torch.nn.Parameter(layer.weight_scale.data,
|
||||
requires_grad=False)
|
||||
|
||||
if self.emulate:
|
||||
try:
|
||||
from quark.torch.export.nn.modules import realquantizer
|
||||
from quark.torch.quantization.config.config import (
|
||||
QuantizationSpec)
|
||||
except ImportError as err:
|
||||
raise ImportError(
|
||||
"The package `amd-quark` is required to use AMD Quark "
|
||||
"MX-FP4 models. Please install it with `pip install "
|
||||
"amd-quark`.") from err
|
||||
|
||||
weight_quant_spec = QuantizationSpec.from_dict(
|
||||
self.weight_quant_spec)
|
||||
|
||||
weight_quantizer = realquantizer.get_real_quantizer(
|
||||
qspec=weight_quant_spec,
|
||||
quantizer=None,
|
||||
real_quantized=True,
|
||||
reorder=False,
|
||||
float_dtype=self.out_dtype,
|
||||
scale_shape=layer.weight_scale.shape,
|
||||
zero_point_shape=None,
|
||||
)
|
||||
weight_quantizer.scale.data = layer.weight_scale.data
|
||||
|
||||
if not envs.VLLM_QUARK_EMU_MEM_OPT:
|
||||
layer.weight = torch.nn.Parameter(
|
||||
weight_quantizer(layer.weight.data).to(self.out_dtype),
|
||||
requires_grad=False,
|
||||
)
|
||||
else:
|
||||
self.weight_quantizer = weight_quantizer
|
||||
layer.weight_scale = None
|
||||
|
||||
# This call is necessary to release the scales memory.
|
||||
torch.cuda.empty_cache()
|
||||
|
||||
def create_weights(self, layer: torch.nn.Module,
|
||||
output_partition_sizes: list[int],
|
||||
input_size_per_partition: int,
|
||||
@@ -116,11 +103,10 @@ class QuarkW4A4MXFP4(QuarkScheme):
|
||||
bias: Optional[torch.Tensor] = None) -> torch.Tensor:
|
||||
|
||||
if self.emulate:
|
||||
if envs.VLLM_QUARK_EMU_MEM_OPT:
|
||||
dq_w = self.weight_quantizer(layer.weight).to(self.out_dtype)
|
||||
else:
|
||||
dq_w = layer.weight
|
||||
qdq_x, _ = per_token_group_quant_mxfp4(x, OCP_MX_BLOCK_SIZE)
|
||||
return F.linear(qdq_x, dq_w, bias)
|
||||
dq_w = dequant_mxfp4(layer.weight, layer.weight_scale, x.dtype)
|
||||
|
||||
x = quant_dequant_mxfp4(x)
|
||||
|
||||
return F.linear(x, dq_w, bias)
|
||||
else:
|
||||
raise NotImplementedError()
|
||||
|
||||
@@ -1,45 +1,67 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
import torch
|
||||
|
||||
from vllm.utils import direct_register_custom_op
|
||||
|
||||
OCP_MX_BLOCK_SIZE = 32
|
||||
|
||||
|
||||
def per_token_group_quant_mxfp4(x: torch.Tensor,
|
||||
block_k: int,
|
||||
scale_calculation_mode: str = "even"
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
def _dequant_mxfp4(x: torch.Tensor, scale: torch.Tensor,
|
||||
float_dtype: torch.dtype) -> torch.Tensor:
|
||||
try:
|
||||
from quark.torch.kernel.hw_emulation.hw_emulation_interface import (
|
||||
fake_quantize_fp4_fp6_per_group_with_scale)
|
||||
from quark.torch.quantization.utils import (even_round,
|
||||
reshape_to_blocks)
|
||||
from quark.torch.kernel import mx
|
||||
except ImportError as err:
|
||||
raise ImportError("The package `amd-quark` is required to use "
|
||||
"MX-FP4 models. Please install it with `pip install "
|
||||
"amd-quark`.") from err
|
||||
|
||||
axis = -1
|
||||
block_x = reshape_to_blocks(x, block_k, axis)
|
||||
amax, _ = torch.max(torch.abs(block_x), dim=-1, keepdim=True)
|
||||
amax = amax.squeeze(-1)
|
||||
return mx.dq_mxfp4(x, scale, float_dtype)
|
||||
|
||||
# TODO: there are other rounding strategies supported in quark and in the
|
||||
# config.json that we do not check for here!
|
||||
if scale_calculation_mode != "even":
|
||||
raise NotImplementedError(
|
||||
f"Scale calculation mode {scale_calculation_mode} is not yet "
|
||||
"supported in MX-FP4 quantization")
|
||||
scale = even_round(amax, "fp4")
|
||||
|
||||
# Apply dequantize(quantize(x)).
|
||||
x = fake_quantize_fp4_fp6_per_group_with_scale(
|
||||
x,
|
||||
scale.to(x.device),
|
||||
axis=axis,
|
||||
group_size=block_k,
|
||||
quant_dtype="fp4",
|
||||
def _dequant_mxfp4_fake(x: torch.Tensor, scale: torch.Tensor,
|
||||
float_dtype: torch.dtype) -> torch.Tensor:
|
||||
return torch.empty((*x.shape[:-1], x.shape[-1] * 2),
|
||||
dtype=float_dtype,
|
||||
device=x.device)
|
||||
|
||||
|
||||
def _quant_dequant_mxfp4(x: torch.Tensor,
|
||||
scale_calculation_mode: str = "even") -> torch.Tensor:
|
||||
try:
|
||||
from quark.torch.kernel import mx
|
||||
except ImportError as err:
|
||||
raise ImportError("The package `amd-quark` is required to use "
|
||||
"MX-FP4 models. Please install it with `pip install "
|
||||
"amd-quark`.") from err
|
||||
|
||||
return mx.qdq_mxfp4(x, scale_calculation_mode)
|
||||
|
||||
|
||||
def _quant_dequant_mxfp4_fake(x: torch.Tensor,
|
||||
scale_calculation_mode: str = "even"
|
||||
) -> torch.Tensor:
|
||||
return torch.empty_like(x)
|
||||
|
||||
|
||||
try:
|
||||
direct_register_custom_op(
|
||||
op_name="dequant_mxfp4",
|
||||
op_func=_dequant_mxfp4,
|
||||
mutates_args=[],
|
||||
fake_impl=_dequant_mxfp4_fake,
|
||||
)
|
||||
dequant_mxfp4 = torch.ops.vllm.dequant_mxfp4
|
||||
except AttributeError as error:
|
||||
raise error
|
||||
|
||||
return x, scale
|
||||
try:
|
||||
direct_register_custom_op(
|
||||
op_name="quant_dequant_mxfp4",
|
||||
op_func=_quant_dequant_mxfp4,
|
||||
mutates_args=[],
|
||||
fake_impl=_quant_dequant_mxfp4_fake,
|
||||
)
|
||||
quant_dequant_mxfp4 = torch.ops.vllm.quant_dequant_mxfp4
|
||||
except AttributeError as error:
|
||||
raise error
|
||||
|
||||
Reference in New Issue
Block a user