class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
def __init__(
self,
weight_config: dict[str, Any],
input_config: dict[str, Any],
moe: FusedMoEConfig,
):
super().__init__(moe)
self.weight_quant = weight_config
self.input_quant = input_config
self.weight_qscheme = self.weight_quant.get("qscheme")
self.input_qscheme = self.input_quant.get("qscheme")
self.weight_dtype = self.weight_quant.get("dtype", "").replace(
"fp8_e4m3", "fp8"
)
self.input_dtype = self.input_quant.get("dtype", "").replace("fp8_e4m3", "fp8")
per_tensor = (
self.weight_qscheme == "per_tensor" and self.input_qscheme == "per_tensor"
)
per_channel = (
self.weight_qscheme == "per_channel" and self.input_qscheme == "per_channel"
)
self.act_quant_group_shape = (
GroupShape.PER_TOKEN if per_channel else GroupShape.PER_TENSOR
)
if not (per_tensor or per_channel):
raise ValueError(
"For FP8 Fused MoE layers, only per-tensor and per-channel "
"scales for weights and activations are supported. Found "
f"{self.weight_qscheme}, {self.input_qscheme}"
) # noqa E501
self.static_input_scales = not self.input_quant.get("is_dynamic")
if self.static_input_scales and per_channel:
raise ValueError(
"For FP8 Fused MoE layer, we require either per tensor or "
"channelwise, dynamic per token quantization."
)
# For GPUs that lack FP8 hardware support, we can leverage the Marlin
# kernel for fast weight-only FP8 quantization
self.use_marlin = (
not current_platform.has_device_capability(89)
or envs.VLLM_TEST_FORCE_FP8_MARLIN
)
# Disable marlin for rocm
if current_platform.is_rocm():
self.use_marlin = False
self.rocm_aiter_moe_enabled = rocm_aiter_ops.is_fused_moe_enabled()
self.model_type = getattr(
get_current_vllm_config().model_config.hf_config, "model_type", None
)
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,
):
layer.intermediate_size_per_partition = intermediate_size_per_partition
layer.hidden_size = hidden_size
layer.num_experts = num_experts
layer.orig_dtype = params_dtype
layer.weight_block_size = None
params_dtype = torch.float8_e4m3fn
# WEIGHTS
w13_weight = torch.nn.Parameter(
torch.empty(
num_experts,
2 * intermediate_size_per_partition,
hidden_size,
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,
dtype=params_dtype,
),
requires_grad=False,
)
layer.register_parameter("w2_weight", w2_weight)
set_weight_attrs(w2_weight, extra_weight_attrs)
# WEIGHT_SCALES
if self.weight_qscheme == "per_tensor":
# Allocate 2 scales for w1 and w3 respectively.
# They are combined to a single scale after weight loading.
if self.model_type != "gpt_oss":
w13_weight_scale = torch.nn.Parameter(
torch.ones(num_experts, 2, dtype=torch.float32), requires_grad=False
)
else:
# For gpt_oss, the w1(gate) & w3(up) are fused as one.
# Therefore, only one weight scale for each expert.
w13_weight_scale = torch.nn.Parameter(
torch.ones(num_experts, 1, dtype=torch.float32), requires_grad=False
)
layer.register_parameter("w13_weight_scale", w13_weight_scale)
w2_weight_scale = torch.nn.Parameter(
torch.ones(num_experts, dtype=torch.float32), requires_grad=False
)
layer.register_parameter("w2_weight_scale", w2_weight_scale)
# Add PER-TENSOR quantization for FusedMoE.weight_loader.
extra_weight_attrs.update(
{"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
)
set_weight_attrs(w13_weight_scale, extra_weight_attrs)
set_weight_attrs(w2_weight_scale, extra_weight_attrs)
elif self.weight_qscheme == "per_channel":
# quark's scale is 1 dim.
w13_weight_scale = torch.nn.Parameter(
torch.ones(
num_experts,
2 * intermediate_size_per_partition,
dtype=torch.float32,
),
requires_grad=False,
)
layer.register_parameter("w13_weight_scale", w13_weight_scale)
w2_weight_scale = torch.nn.Parameter(
torch.ones(num_experts, hidden_size, dtype=torch.float32),
requires_grad=False,
)
layer.register_parameter("w2_weight_scale", w2_weight_scale)
# Add PER-CHANNEL quantization for FusedMoE.weight_loader.
extra_weight_attrs.update(
{"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value}
)
set_weight_attrs(w13_weight_scale, extra_weight_attrs)
set_weight_attrs(w2_weight_scale, extra_weight_attrs)
# INPUT_SCALES
if self.static_input_scales:
w13_input_scale = torch.nn.Parameter(
torch.ones(num_experts, dtype=torch.float32), requires_grad=False
)
layer.register_parameter("w13_input_scale", w13_input_scale)
set_weight_attrs(w13_input_scale, extra_weight_attrs)
w2_input_scale = torch.nn.Parameter(
torch.ones(num_experts, dtype=torch.float32), requires_grad=False
)
layer.register_parameter("w2_input_scale", w2_input_scale)
set_weight_attrs(w2_input_scale, extra_weight_attrs)
else:
layer.w13_input_scale = None
layer.w2_input_scale = None
if self.has_bias:
w13_bias = torch.nn.Parameter(
torch.zeros(
num_experts,
2 * intermediate_size_per_partition,
dtype=torch.float32,
),
requires_grad=False,
)
layer.register_parameter("w13_bias", w13_bias)
set_weight_attrs(w13_bias, extra_weight_attrs)
w2_bias = torch.nn.Parameter(
torch.zeros(num_experts, hidden_size, dtype=torch.float32),
requires_grad=False,
)
layer.register_parameter("w2_bias", w2_bias)
set_weight_attrs(w2_bias, extra_weight_attrs)
else:
layer.w13_bias, layer.w2_bias = None, None
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
# Fp8 moe kernels require a single activation scale.
# We take the max of all the scales in case they differ.
if self.static_input_scales:
if layer.w13_input_scale is None or layer.w2_input_scale is None:
raise ValueError(
"QuantConfig has static quantization, but found "
"activation scales are None."
)
if not all_close_1d(layer.w13_input_scale) or not all_close_1d(
layer.w2_input_scale
):
logger.warning_once(
"Found input_scales that are not equal for "
"fp8 MoE layer. Using the maximum across experts "
"for each layer. "
)
layer.w13_input_scale = torch.nn.Parameter(
layer.w13_input_scale.max(), requires_grad=False
)
layer.w2_input_scale = torch.nn.Parameter(
layer.w2_input_scale.max(), requires_grad=False
)
if current_platform.is_fp8_fnuz():
# Normalize the weights and scales
w13_weight, w13_weight_scale, w13_input_scale = (
normalize_e4m3fn_to_e4m3fnuz(
layer.w13_weight, layer.w13_weight_scale, layer.w13_input_scale
)
)
w2_weight, w2_weight_scale, w2_input_scale = normalize_e4m3fn_to_e4m3fnuz(
layer.w2_weight, layer.w2_weight_scale, layer.w2_input_scale
)
# Reset the parameter
layer.w13_weight = torch.nn.Parameter(w13_weight, requires_grad=False)
layer.w13_weight_scale = torch.nn.Parameter(
w13_weight_scale, requires_grad=False
)
if w13_input_scale is not None:
layer.w13_input_scale = torch.nn.Parameter(
w13_input_scale, requires_grad=False
)
layer.w2_weight = torch.nn.Parameter(w2_weight, requires_grad=False)
layer.w2_weight_scale = torch.nn.Parameter(
w2_weight_scale, requires_grad=False
)
if w2_input_scale is not None:
layer.w2_input_scale = torch.nn.Parameter(
w2_input_scale, requires_grad=False
)
# For per-tensor case, Fp8 moe kernel needs single weight scale
# for w13 per expert. Use max then dequant and requant each expert.
if self.weight_qscheme == "per_tensor":
assert layer.w13_weight_scale is not None
shard_size = layer.intermediate_size_per_partition
max_w13_scales = layer.w13_weight_scale.max(dim=1).values
# For gpt_oss, w1 and w3 are fused into a single combined
# gate_up_proj tensor with size 2*intermediate_size_per_partition
# and only one scale per expert.
# Process the entire weight tensor as one shard.
if self.model_type == "gpt_oss":
for expert_id in range(layer.local_num_experts):
# Process all 2*intermediate_size_per_partition rows at once
dq_weight = per_tensor_dequantize(
layer.w13_weight[expert_id],
layer.w13_weight_scale[expert_id][0],
)
layer.w13_weight[expert_id], _ = ops.scaled_fp8_quant(
dq_weight, max_w13_scales[expert_id]
)
else:
# For non-gpt_oss, process w1 and w3 shards separately
for expert_id in range(layer.local_num_experts):
start = 0
for shard_id in range(2):
dq_weight = per_tensor_dequantize(
layer.w13_weight[expert_id][start : start + shard_size, :],
layer.w13_weight_scale[expert_id][shard_id],
)
(
layer.w13_weight[expert_id][start : start + shard_size, :],
_,
) = ops.scaled_fp8_quant(dq_weight, max_w13_scales[expert_id])
start += shard_size
layer.w13_weight_scale = torch.nn.Parameter(
max_w13_scales, requires_grad=False
)
# quark's scale is 1 dim.
elif self.weight_qscheme == "per_channel":
if self.act_quant_group_shape == GroupShape.PER_TOKEN:
w13_weight_scale = layer.w13_weight_scale.unsqueeze(-1)
layer.w13_weight_scale = torch.nn.Parameter(
w13_weight_scale, requires_grad=False
)
w2_weight_scale = layer.w2_weight_scale.unsqueeze(-1)
layer.w2_weight_scale = torch.nn.Parameter(
w2_weight_scale, requires_grad=False
)
# Property to determine if AITER is used
if self.rocm_aiter_moe_enabled:
# reshaping weights is required for aiter moe kernel.
shuffled_w13, shuffled_w2 = rocm_aiter_ops.shuffle_weights(
layer.w13_weight.data, layer.w2_weight.data
)
layer.w13_weight = torch.nn.Parameter(shuffled_w13, requires_grad=False)
layer.w2_weight = torch.nn.Parameter(shuffled_w2, requires_grad=False)
elif self.use_marlin:
w13_weight, w2_weight, w13_weight_scale, w2_weight_scale = (
prepare_fp8_moe_layer_for_marlin(
layer,
layer.w13_weight,
layer.w2_weight,
layer.w13_weight_scale,
layer.w2_weight_scale,
)
)
# TODO(rob): once we apply refactor to Quark, switch to using
# replace_parameter for compatibility with reloading in RL.
layer.w13_weight = torch.nn.Parameter(w13_weight, requires_grad=False)
layer.w2_weight = torch.nn.Parameter(w2_weight, requires_grad=False)
layer.w13_weight_scale = torch.nn.Parameter(
w13_weight_scale, requires_grad=False
)
layer.w2_weight_scale = torch.nn.Parameter(
w2_weight_scale, requires_grad=False
)
def get_fused_moe_quant_config(
self, layer: torch.nn.Module
) -> FusedMoEQuantConfig | None:
return fp8_w8a8_moe_quant_config(
w1_scale=layer.w13_weight_scale,
w2_scale=layer.w2_weight_scale,
a1_scale=layer.w13_input_scale,
a2_scale=layer.w2_input_scale,
w1_bias=layer.w13_bias,
w2_bias=layer.w2_bias,
per_act_token_quant=self.input_qscheme == "per_channel",
per_out_ch_quant=self.weight_qscheme == "per_channel",
)
def apply(
self,
layer: FusedMoE,
x: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
if self.rocm_aiter_moe_enabled:
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
rocm_aiter_fused_experts,
)
return rocm_aiter_fused_experts(
hidden_states=x,
w1=layer.w13_weight,
w2=layer.w2_weight,
topk_weights=topk_weights,
topk_ids=topk_ids,
activation=layer.activation,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
quant_config=self.moe_quant_config,
expert_map=layer.expert_map,
)
elif self.use_marlin:
assert layer.activation == "silu", (
f"{layer.activation} not supported for Marlin MoE."
)
return fused_marlin_moe(
x,
layer.w13_weight,
layer.w2_weight,
None,
None,
layer.w13_weight_scale,
layer.w2_weight_scale,
topk_weights,
topk_ids,
quant_type_id=scalar_types.float8_e4m3fn.id,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
global_num_experts=layer.global_num_experts,
expert_map=layer.expert_map,
inplace=not self.moe.disable_inplace,
)
else:
from vllm.model_executor.layers.fused_moe import fused_experts
return fused_experts(
hidden_states=x,
w1=layer.w13_weight,
w2=layer.w2_weight,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=not self.moe.disable_inplace,
activation=layer.activation,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
global_num_experts=layer.global_num_experts,
expert_map=layer.expert_map,
quant_config=self.moe_quant_config,
)