[Quant][Feature] Support online MXFP8 quantization for MoE and dense models (#35448)
Signed-off-by: EdalatiAli <aliedalati@cohere.com>
This commit is contained in:
@@ -31,6 +31,7 @@ QuantizationMethods = Literal[
|
||||
"torchao",
|
||||
"inc",
|
||||
"mxfp4",
|
||||
"mxfp8",
|
||||
"petit_nvfp4",
|
||||
"cpu_awq",
|
||||
]
|
||||
@@ -129,6 +130,7 @@ def get_quantization_config(quantization: str) -> type[QuantizationConfig]:
|
||||
)
|
||||
from .moe_wna16 import MoeWNA16Config
|
||||
from .mxfp4 import Mxfp4Config
|
||||
from .mxfp8 import Mxfp8Config
|
||||
from .petit import PetitNvFp4Config
|
||||
from .ptpc_fp8 import PTPCFp8Config
|
||||
from .torchao import TorchAOConfig
|
||||
@@ -156,6 +158,7 @@ def get_quantization_config(quantization: str) -> type[QuantizationConfig]:
|
||||
"auto-round": INCConfig,
|
||||
"inc": INCConfig,
|
||||
"mxfp4": Mxfp4Config,
|
||||
"mxfp8": Mxfp8Config,
|
||||
"petit_nvfp4": PetitNvFp4Config,
|
||||
"cpu_awq": CPUAWQConfig,
|
||||
}
|
||||
|
||||
@@ -25,13 +25,13 @@ from vllm.model_executor.layers.fused_moe.layer import (
|
||||
FusedMoeWeightScaleSupported,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.oracle.fp8 import (
|
||||
Fp8MoeBackend,
|
||||
convert_to_fp8_moe_kernel_format,
|
||||
make_fp8_moe_kernel,
|
||||
make_fp8_moe_quant_config,
|
||||
select_fp8_moe_backend,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.oracle.mxfp8 import (
|
||||
MxFp8MoeBackend,
|
||||
select_mxfp8_moe_backend,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.oracle.nvfp4 import (
|
||||
@@ -1712,8 +1712,7 @@ class ModelOptMxFp8FusedMoE(FusedMoEMethodBase):
|
||||
self.quant_config = quant_config
|
||||
assert self.quant_config.is_checkpoint_mxfp8_serialized
|
||||
|
||||
# Select MXFP8 MoE backend
|
||||
self.mxfp8_backend = select_mxfp8_moe_backend(self.moe)
|
||||
self.mxfp8_backend, _ = select_mxfp8_moe_backend(self.moe)
|
||||
|
||||
def create_weights(
|
||||
self,
|
||||
@@ -1943,7 +1942,7 @@ class ModelOptMxFp8FusedMoE(FusedMoEMethodBase):
|
||||
|
||||
@property
|
||||
def is_monolithic(self) -> bool:
|
||||
return self.mxfp8_backend == MxFp8MoeBackend.FLASHINFER_TRTLLM
|
||||
return self.mxfp8_backend == Fp8MoeBackend.FLASHINFER_TRTLLM
|
||||
|
||||
def apply_monolithic(
|
||||
self,
|
||||
@@ -1956,7 +1955,7 @@ class ModelOptMxFp8FusedMoE(FusedMoEMethodBase):
|
||||
Fp8QuantizationType,
|
||||
)
|
||||
|
||||
assert self.mxfp8_backend == MxFp8MoeBackend.FLASHINFER_TRTLLM
|
||||
assert self.mxfp8_backend == Fp8MoeBackend.FLASHINFER_TRTLLM
|
||||
|
||||
if layer.enable_eplb:
|
||||
raise NotImplementedError(
|
||||
|
||||
@@ -0,0 +1,354 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
"""Online MXFP8 (microscaling FP8, block-32) quantization config and methods."""
|
||||
|
||||
from typing import Any
|
||||
|
||||
import torch
|
||||
from torch.nn import Module
|
||||
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
FusedMoE,
|
||||
FusedMoEMethodBase,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.layer import UnquantizedFusedMoEMethod
|
||||
from vllm.model_executor.layers.fused_moe.oracle.mxfp8 import (
|
||||
select_mxfp8_moe_backend,
|
||||
)
|
||||
from vllm.model_executor.layers.linear import (
|
||||
LinearBase,
|
||||
UnquantizedLinearMethod,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization import QuantizationMethods
|
||||
from vllm.model_executor.layers.quantization.base_config import (
|
||||
QuantizeMethodBase,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.fp8 import (
|
||||
Fp8Config,
|
||||
Fp8KVCacheMethod,
|
||||
Fp8OnlineLinearMethod,
|
||||
Fp8OnlineMoEMethod,
|
||||
_copy_missing_attrs,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.utils.mxfp8_utils import (
|
||||
MXFP8_BLOCK_SIZE,
|
||||
Mxfp8LinearBackend,
|
||||
Mxfp8LinearOp,
|
||||
mxfp8_e4m3_quantize,
|
||||
swizzle_mxfp8_scale,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import (
|
||||
is_layer_skipped,
|
||||
)
|
||||
from vllm.model_executor.model_loader.weight_utils import (
|
||||
initialize_single_dummy_weight,
|
||||
)
|
||||
from vllm.model_executor.parameter import ModelWeightParameter
|
||||
from vllm.model_executor.utils import replace_parameter, set_weight_attrs
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
class Mxfp8Config(Fp8Config):
|
||||
"""Config class for online MXFP8 MoE quantization."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
activation_scheme: str = "dynamic",
|
||||
ignored_layers: list[str] | None = None,
|
||||
) -> None:
|
||||
if activation_scheme != "dynamic":
|
||||
raise ValueError("mxfp8 only supports dynamic activation scheme.")
|
||||
super().__init__(
|
||||
is_checkpoint_fp8_serialized=False,
|
||||
activation_scheme=activation_scheme,
|
||||
ignored_layers=ignored_layers,
|
||||
weight_block_size=None,
|
||||
)
|
||||
|
||||
@classmethod
|
||||
def get_name(cls) -> QuantizationMethods:
|
||||
return "mxfp8"
|
||||
|
||||
@classmethod
|
||||
def get_min_capability(cls) -> int:
|
||||
return 100
|
||||
|
||||
@classmethod
|
||||
def from_config(cls, config: dict[str, Any]) -> "Mxfp8Config":
|
||||
activation_scheme = cls.get_from_keys_or(
|
||||
config, ["activation_scheme"], "dynamic"
|
||||
)
|
||||
ignored_layers = cls.get_from_keys_or(config, ["ignored_layers"], None)
|
||||
if not ignored_layers:
|
||||
ignored_layers = cls.get_from_keys_or(
|
||||
config, ["modules_to_not_convert"], None
|
||||
)
|
||||
return cls(
|
||||
activation_scheme=activation_scheme,
|
||||
ignored_layers=ignored_layers,
|
||||
)
|
||||
|
||||
def get_quant_method(
|
||||
self, layer: torch.nn.Module, prefix: str
|
||||
) -> "QuantizeMethodBase | None":
|
||||
if isinstance(layer, LinearBase):
|
||||
if is_layer_skipped(
|
||||
prefix=prefix,
|
||||
ignored_layers=self.ignored_layers,
|
||||
fused_mapping=self.packed_modules_mapping,
|
||||
skip_with_substr=True,
|
||||
):
|
||||
return UnquantizedLinearMethod()
|
||||
return Mxfp8OnlineLinearMethod(self)
|
||||
elif isinstance(layer, FusedMoE):
|
||||
if is_layer_skipped(
|
||||
prefix=prefix,
|
||||
ignored_layers=self.ignored_layers,
|
||||
fused_mapping=self.packed_modules_mapping,
|
||||
skip_with_substr=True,
|
||||
):
|
||||
return UnquantizedFusedMoEMethod(layer.moe_config)
|
||||
return Mxfp8OnlineMoEMethod(self, layer)
|
||||
elif isinstance(layer, Attention):
|
||||
return Fp8KVCacheMethod(self)
|
||||
return None
|
||||
|
||||
|
||||
class Mxfp8OnlineLinearMethod(Fp8OnlineLinearMethod):
|
||||
"""Online MXFP8 linear method.
|
||||
Loads bf16/fp16 checkpoints and quantizes weights to MXFP8 (microscaling
|
||||
FP8 with block-32 scales) during weight loading.
|
||||
|
||||
Args:
|
||||
quant_config: The MXFP8 quantization config.
|
||||
"""
|
||||
|
||||
uses_meta_device: bool = True
|
||||
|
||||
def __init__(self, quant_config: "Mxfp8Config"):
|
||||
self.quant_config = quant_config
|
||||
self.out_dtype = torch.get_default_dtype()
|
||||
self.mxfp8_linear = Mxfp8LinearOp(self._select_backend())
|
||||
logger.info_once(
|
||||
"Using %s backend for MXFP8 GEMM", self.mxfp8_linear.backend.value
|
||||
)
|
||||
|
||||
@staticmethod
|
||||
def _select_backend() -> Mxfp8LinearBackend:
|
||||
try:
|
||||
from vllm.utils import flashinfer as fi
|
||||
|
||||
_ = fi.mm_mxfp8
|
||||
return Mxfp8LinearBackend.FLASHINFER_CUTLASS
|
||||
except Exception:
|
||||
logger.warning(
|
||||
"FlashInfer mm_mxfp8 not available, "
|
||||
"falling back to MXFP8 emulation backend."
|
||||
)
|
||||
return Mxfp8LinearBackend.EMULATION
|
||||
|
||||
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 % MXFP8_BLOCK_SIZE != 0:
|
||||
raise ValueError(
|
||||
f"MXFP8 requires input_size_per_partition "
|
||||
f"({input_size_per_partition}) to be divisible by "
|
||||
f"{MXFP8_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
|
||||
|
||||
if layer.weight.device == torch.device("meta"):
|
||||
weight = ModelWeightParameter(
|
||||
data=torch.empty_like(layer.weight, device=layer._load_device),
|
||||
input_dim=1,
|
||||
output_dim=0,
|
||||
weight_loader=layer.weight.weight_loader,
|
||||
)
|
||||
_copy_missing_attrs(layer.weight, weight)
|
||||
layer.register_parameter("weight", weight)
|
||||
initialize_single_dummy_weight(layer.weight)
|
||||
|
||||
weight_fp8, weight_scale = mxfp8_e4m3_quantize(layer.weight.contiguous())
|
||||
|
||||
if self.mxfp8_linear.backend == Mxfp8LinearBackend.FLASHINFER_CUTLASS:
|
||||
N, K = layer.weight.shape[0], layer.weight.shape[1]
|
||||
weight_scale = swizzle_mxfp8_scale(weight_scale, N, K)
|
||||
|
||||
layer.input_scale = None
|
||||
replace_parameter(layer, "weight", weight_fp8.data)
|
||||
replace_parameter(layer, "weight_scale", weight_scale.data)
|
||||
|
||||
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.mxfp8_linear.apply(
|
||||
input=x,
|
||||
weight=layer.weight,
|
||||
weight_scale=layer.weight_scale,
|
||||
out_dtype=self.out_dtype,
|
||||
bias=bias,
|
||||
)
|
||||
|
||||
|
||||
class Mxfp8OnlineMoEMethod(Fp8OnlineMoEMethod):
|
||||
"""MoE method for online MXFP8 (block) quantization."""
|
||||
|
||||
uses_meta_device: bool = True
|
||||
|
||||
def __init__(self, quant_config: Fp8Config, layer: torch.nn.Module):
|
||||
FusedMoEMethodBase.__init__(self, layer.moe_config)
|
||||
self.quant_config = quant_config
|
||||
assert not quant_config.is_checkpoint_fp8_serialized
|
||||
assert quant_config.activation_scheme == "dynamic"
|
||||
|
||||
self.weight_block_size = [1, MXFP8_BLOCK_SIZE]
|
||||
self.block_quant = True
|
||||
self.weight_scale_name = "weight_scale"
|
||||
|
||||
self.fp8_backend, self.experts_cls = select_mxfp8_moe_backend(config=self.moe)
|
||||
|
||||
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 % MXFP8_BLOCK_SIZE != 0
|
||||
or intermediate_size_per_partition % MXFP8_BLOCK_SIZE != 0
|
||||
):
|
||||
raise ValueError(
|
||||
"Online MXFP8 MoE requires hidden/intermediate sizes divisible "
|
||||
f"by {MXFP8_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,
|
||||
)
|
||||
|
||||
w13_weight_scale = torch.nn.Parameter(
|
||||
torch.zeros(
|
||||
num_experts,
|
||||
2 * intermediate_size_per_partition,
|
||||
hidden_size // MXFP8_BLOCK_SIZE,
|
||||
dtype=torch.uint8,
|
||||
),
|
||||
requires_grad=False,
|
||||
)
|
||||
w2_weight_scale = torch.nn.Parameter(
|
||||
torch.zeros(
|
||||
num_experts,
|
||||
hidden_size,
|
||||
intermediate_size_per_partition // MXFP8_BLOCK_SIZE,
|
||||
dtype=torch.uint8,
|
||||
),
|
||||
requires_grad=False,
|
||||
)
|
||||
layer.register_parameter("w13_weight_scale", w13_weight_scale)
|
||||
layer.register_parameter("w2_weight_scale", w2_weight_scale)
|
||||
set_weight_attrs(w13_weight_scale, extra_weight_attrs)
|
||||
set_weight_attrs(w2_weight_scale, extra_weight_attrs)
|
||||
layer.weight_block_size = [1, MXFP8_BLOCK_SIZE]
|
||||
|
||||
def _quantize_mxfp8_moe_weight(
|
||||
self, weight: torch.Tensor
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
"""Batch quantization: bf16/fp16 weights -> MXFP8 (fp8 + uint8 scales)."""
|
||||
num_batches = weight.size(0)
|
||||
w_quant = []
|
||||
w_scales = []
|
||||
for i in range(num_batches):
|
||||
mx_fp8_quant, mx_fp8_scale = mxfp8_e4m3_quantize(
|
||||
weight[i], is_sf_swizzled_layout=False
|
||||
)
|
||||
w_quant.append(mx_fp8_quant)
|
||||
w_scales.append(mx_fp8_scale)
|
||||
|
||||
return torch.stack(w_quant), torch.stack(w_scales)
|
||||
|
||||
def process_weights_after_loading(self, layer: Module) -> None:
|
||||
if getattr(layer, "_already_called_process_weights_after_loading", False):
|
||||
return
|
||||
|
||||
if layer.w13_weight.device == torch.device("meta"):
|
||||
w13_weight = torch.nn.Parameter(
|
||||
torch.empty_like(layer.w13_weight, device=layer._load_device),
|
||||
requires_grad=False,
|
||||
)
|
||||
set_weight_attrs(
|
||||
w13_weight, {"weight_loader": layer.w13_weight.weight_loader}
|
||||
)
|
||||
_copy_missing_attrs(layer.w13_weight, w13_weight)
|
||||
layer.register_parameter("w13_weight", w13_weight)
|
||||
initialize_single_dummy_weight(layer.w13_weight)
|
||||
if layer.w2_weight.device == torch.device("meta"):
|
||||
w2_weight = torch.nn.Parameter(
|
||||
torch.empty_like(layer.w2_weight, device=layer._load_device),
|
||||
requires_grad=False,
|
||||
)
|
||||
set_weight_attrs(
|
||||
w2_weight, {"weight_loader": layer.w2_weight.weight_loader}
|
||||
)
|
||||
_copy_missing_attrs(layer.w2_weight, w2_weight)
|
||||
layer.register_parameter("w2_weight", w2_weight)
|
||||
initialize_single_dummy_weight(layer.w2_weight)
|
||||
|
||||
fp8_dtype = current_platform.fp8_dtype()
|
||||
w13 = torch.empty_like(layer.w13_weight, dtype=fp8_dtype)
|
||||
w2 = torch.empty_like(layer.w2_weight, dtype=fp8_dtype)
|
||||
w13_scale = layer.w13_weight_scale
|
||||
w2_scale = layer.w2_weight_scale
|
||||
|
||||
w13, w13_scale = self._quantize_mxfp8_moe_weight(layer.w13_weight)
|
||||
w2, w2_scale = self._quantize_mxfp8_moe_weight(layer.w2_weight)
|
||||
|
||||
self._setup_kernel(
|
||||
layer,
|
||||
w13,
|
||||
w2,
|
||||
w13_scale,
|
||||
w2_scale,
|
||||
layer.w13_input_scale,
|
||||
layer.w2_input_scale,
|
||||
)
|
||||
|
||||
layer._already_called_process_weights_after_loading = True
|
||||
@@ -305,6 +305,81 @@ def align_fp8_moe_weights_for_fi(
|
||||
return padded_w13, padded_w2, padded_intermediate
|
||||
|
||||
|
||||
def _shuffle_mxfp8_moe_weights(
|
||||
w13: torch.Tensor,
|
||||
w2: torch.Tensor,
|
||||
w13_scale: torch.Tensor,
|
||||
w2_scale: torch.Tensor,
|
||||
is_gated: bool,
|
||||
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
|
||||
"""Preprocess MXFP8 weights and scales for the FlashInfer TRT-LLM kernel.
|
||||
|
||||
Following flashinfer/tests/moe/test_trtllm_gen_fused_moe.py:
|
||||
1. reorder_rows_for_gated_act_gemm (interleave gate/up rows)
|
||||
2. shuffle_matrix_a (weight data layout shuffle)
|
||||
3. shuffle_matrix_sf_a (scale factor layout shuffle)
|
||||
"""
|
||||
from flashinfer import (
|
||||
reorder_rows_for_gated_act_gemm,
|
||||
shuffle_matrix_a,
|
||||
shuffle_matrix_sf_a,
|
||||
)
|
||||
|
||||
epilogue_tile_m = 128
|
||||
num_experts = w13.shape[0]
|
||||
intermediate_size = w13.shape[1] // 2
|
||||
hidden_size = w13.shape[2]
|
||||
|
||||
w13_interleaved: list[torch.Tensor] = []
|
||||
w13_scale_interleaved: list[torch.Tensor] = []
|
||||
for i in range(num_experts):
|
||||
if is_gated:
|
||||
w13_interleaved.append(
|
||||
reorder_rows_for_gated_act_gemm(
|
||||
w13[i].reshape(2 * intermediate_size, -1)
|
||||
)
|
||||
)
|
||||
w13_scale_interleaved.append(
|
||||
reorder_rows_for_gated_act_gemm(
|
||||
w13_scale[i].reshape(2 * intermediate_size, -1)
|
||||
)
|
||||
)
|
||||
else:
|
||||
w13_interleaved.append(w13[i])
|
||||
w13_scale_interleaved.append(w13_scale[i])
|
||||
|
||||
w13_shuffled: list[torch.Tensor] = []
|
||||
w2_shuffled: list[torch.Tensor] = []
|
||||
w13_scale_shuffled: list[torch.Tensor] = []
|
||||
w2_scale_shuffled: list[torch.Tensor] = []
|
||||
for i in range(num_experts):
|
||||
w13_shuffled.append(
|
||||
shuffle_matrix_a(w13_interleaved[i].view(torch.uint8), epilogue_tile_m)
|
||||
)
|
||||
w2_shuffled.append(shuffle_matrix_a(w2[i].view(torch.uint8), epilogue_tile_m))
|
||||
w13_scale_shuffled.append(
|
||||
shuffle_matrix_sf_a(
|
||||
w13_scale_interleaved[i]
|
||||
.view(torch.uint8)
|
||||
.reshape(2 * intermediate_size, -1),
|
||||
epilogue_tile_m,
|
||||
)
|
||||
)
|
||||
w2_scale_shuffled.append(
|
||||
shuffle_matrix_sf_a(
|
||||
w2_scale[i].view(torch.uint8).reshape(hidden_size, -1),
|
||||
epilogue_tile_m,
|
||||
)
|
||||
)
|
||||
|
||||
w13_out = torch.stack(w13_shuffled).view(torch.float8_e4m3fn)
|
||||
w2_out = torch.stack(w2_shuffled).view(torch.float8_e4m3fn)
|
||||
w13_scale_out = torch.stack(w13_scale_shuffled).reshape(w13_scale.shape)
|
||||
w2_scale_out = torch.stack(w2_scale_shuffled).reshape(w2_scale.shape)
|
||||
|
||||
return w13_out, w2_out, w13_scale_out, w2_scale_out
|
||||
|
||||
|
||||
def prepare_fp8_moe_layer_for_fi(
|
||||
layer: torch.nn.Module,
|
||||
w13: torch.Tensor,
|
||||
@@ -314,7 +389,7 @@ def prepare_fp8_moe_layer_for_fi(
|
||||
w2_scale: torch.Tensor,
|
||||
w2_input_scale: torch.Tensor | None,
|
||||
is_trtllm: bool = False,
|
||||
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
||||
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
|
||||
"""
|
||||
Convert Fp8 MoE weights to flashinfer kernel format
|
||||
|
||||
@@ -329,10 +404,33 @@ def prepare_fp8_moe_layer_for_fi(
|
||||
block_quant = (
|
||||
hasattr(layer, "weight_block_size") and layer.weight_block_size is not None
|
||||
)
|
||||
is_mxfp8 = block_quant and w13_scale.dtype == torch.uint8
|
||||
is_gated = layer.activation.is_gated
|
||||
|
||||
# MXFP8 TRT-LLM requires W31 swap + reorder + shuffle.
|
||||
if is_mxfp8 and is_trtllm:
|
||||
# FlashInfer TRT-LLM SwiGLU expects [up; gate] but vLLM stores
|
||||
# [gate; up]. Swap both weights and scales before interleaving.
|
||||
if layer.moe_config.is_act_and_mul:
|
||||
w13 = swap_w13_to_w31(w13)
|
||||
# Scales may be 2D [E, flat] from _quantize_mxfp8_moe_weight;
|
||||
# reshape to 3D so swap_w13_to_w31 can flip the two halves,
|
||||
# then flatten back.
|
||||
if w13_scale.ndim == 2:
|
||||
num_rows = w13.shape[1] # 2 * intermediate_size
|
||||
w13_scale = w13_scale.reshape(w13_scale.shape[0], num_rows, -1)
|
||||
w13_scale = swap_w13_to_w31(w13_scale)
|
||||
w13_scale = w13_scale.reshape(w13_scale.shape[0], -1)
|
||||
else:
|
||||
w13_scale = swap_w13_to_w31(w13_scale)
|
||||
|
||||
w13, w2, w13_scale, w2_scale = _shuffle_mxfp8_moe_weights(
|
||||
w13, w2, w13_scale, w2_scale, is_gated
|
||||
)
|
||||
return w13, w2, w13_scale, w2_scale
|
||||
|
||||
# Some FI MoE kernels require internal alignment of 16
|
||||
# for the gate-up proj. Pad the weights to respect this.
|
||||
is_gated = layer.activation.is_gated
|
||||
if not block_quant:
|
||||
min_alignment = 16 if is_gated else 128
|
||||
w13, w2, new_intermediate = align_fp8_moe_weights_for_fi(
|
||||
@@ -369,4 +467,4 @@ def prepare_fp8_moe_layer_for_fi(
|
||||
w13_scale.clamp_(min=_FI_CUTLASS_MIN_BLOCK_SCALE)
|
||||
w2_scale.clamp_(min=_FI_CUTLASS_MIN_BLOCK_SCALE)
|
||||
|
||||
return w13, w2, w13_scale
|
||||
return w13, w2, w13_scale, w2_scale
|
||||
|
||||
@@ -149,6 +149,12 @@ kFp8Dynamic128Sym = QuantKey(FP8_DTYPE, kDynamic128Scale, symmetric=True)
|
||||
kStatic128BlockScale = ScaleDesc(torch.float32, True, GroupShape(128, 128))
|
||||
kFp8Static128BlockSym = QuantKey(FP8_DTYPE, kStatic128BlockScale, symmetric=True)
|
||||
|
||||
kMxfp8StaticScale = ScaleDesc(torch.uint8, True, GroupShape(1, 32))
|
||||
kMxfp8Static = QuantKey(FP8_DTYPE, kMxfp8StaticScale, symmetric=True)
|
||||
|
||||
kMxfp8DynamicScale = ScaleDesc(torch.uint8, False, GroupShape(1, 32))
|
||||
kMxfp8Dynamic = QuantKey(FP8_DTYPE, kMxfp8DynamicScale, symmetric=True)
|
||||
|
||||
kDynamic64Scale = ScaleDesc(torch.float32, False, GroupShape(1, 64))
|
||||
kFp8Dynamic64Sym = QuantKey(FP8_DTYPE, kDynamic64Scale, symmetric=True)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user