[Misc] Fused MoE Marlin support for GPTQ (#8217)

This commit is contained in:
Dipika Sikka
2024-09-09 23:02:52 -04:00
committed by GitHub
parent c7cb5c3335
commit 6cd5e5b07e
19 changed files with 912 additions and 204 deletions
@@ -1,18 +1,22 @@
from typing import Any, Dict, List, Optional
from typing import Any, Callable, Dict, List, Optional, Union
import torch
from torch.nn import Parameter
from vllm import _custom_ops as ops
from vllm.logger import init_logger
from vllm.model_executor.layers.linear import LinearBase, LinearMethodBase
from vllm.model_executor.layers.fused_moe.layer import (
FusedMoE, FusedMoEMethodBase, FusedMoeWeightScaleSupported)
from vllm.model_executor.layers.linear import (LinearBase, LinearMethodBase,
set_weight_attrs)
from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig)
from vllm.model_executor.layers.quantization.utils.marlin_utils import (
apply_gptq_marlin_linear, check_marlin_supported, marlin_is_k_full,
marlin_make_empty_g_idx, marlin_make_workspace, marlin_permute_scales,
marlin_repeat_scales_on_all_ranks, marlin_sort_g_idx, replace_tensor,
verify_marlin_supported, verify_marlin_supports_shape)
marlin_make_empty_g_idx, marlin_make_workspace, marlin_moe_permute_scales,
marlin_permute_scales, marlin_repeat_scales_on_all_ranks,
marlin_sort_g_idx, replace_tensor, verify_marlin_supported,
verify_marlin_supports_shape)
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
from vllm.model_executor.parameter import (ChannelQuantScaleParameter,
GroupQuantScaleParameter,
@@ -33,8 +37,14 @@ class GPTQMarlinConfig(QuantizationConfig):
(8, True): scalar_types.uint8b128,
}
def __init__(self, weight_bits: int, group_size: int, desc_act: bool,
is_sym: bool, lm_head_quantized: bool) -> None:
def __init__(
self,
weight_bits: int,
group_size: int,
desc_act: bool,
is_sym: bool,
lm_head_quantized: bool,
) -> None:
if desc_act and group_size == -1:
# In this case, act_order == True is the same as act_order == False
# (since we have only one group per output channel)
@@ -105,11 +115,14 @@ class GPTQMarlinConfig(QuantizationConfig):
" faster inference")
return None
def get_quant_method(self, layer: torch.nn.Module,
prefix: str) -> Optional["GPTQMarlinLinearMethod"]:
if (isinstance(layer, LinearBase) or
(isinstance(layer, ParallelLMHead) and self.lm_head_quantized)):
def get_quant_method(
self, layer: torch.nn.Module, prefix: str
) -> Optional[Union["GPTQMarlinLinearMethod", "GPTQMarlinMoEMethod"]]:
if isinstance(layer, LinearBase) or (isinstance(layer, ParallelLMHead)
and self.lm_head_quantized):
return GPTQMarlinLinearMethod(self)
elif isinstance(layer, FusedMoE):
return GPTQMarlinMoEMethod(self)
return None
def get_scaled_act_names(self) -> List[str]:
@@ -179,7 +192,8 @@ class GPTQMarlinLinearMethod(LinearMethodBase):
output_size_per_partition=output_size_per_partition,
input_size_per_partition=input_size_per_partition,
input_size=input_size,
group_size=group_size)
group_size=group_size,
)
# Determine sharding
if marlin_repeat_scales_on_all_ranks(self.quant_config.desc_act,
@@ -299,7 +313,8 @@ class GPTQMarlinLinearMethod(LinearMethodBase):
perm=layer.g_idx_sort_indices,
size_k=layer.input_size_per_partition,
size_n=layer.output_size_per_partition,
num_bits=self.quant_config.quant_type.size_bits)
num_bits=self.quant_config.quant_type.size_bits,
)
replace_tensor(layer, "qweight", marlin_qweight)
# Permute scales from autogptq format to marlin format.
@@ -308,7 +323,8 @@ class GPTQMarlinLinearMethod(LinearMethodBase):
size_k=(layer.input_size if self.quant_config.desc_act else
layer.input_size_per_partition),
size_n=layer.output_size_per_partition,
group_size=self.quant_config.group_size)
group_size=self.quant_config.group_size,
)
replace_tensor(layer, "scales", marlin_scales)
def apply(
@@ -329,4 +345,270 @@ class GPTQMarlinLinearMethod(LinearMethodBase):
output_size_per_partition=layer.output_size_per_partition,
input_size_per_partition=layer.input_size_per_partition,
is_k_full=layer.is_k_full,
bias=bias)
bias=bias,
)
class GPTQMarlinMoEMethod(FusedMoEMethodBase):
"""MoE Marlin method with quantization."""
def __init__(self, quant_config: GPTQMarlinConfig) -> None:
self.quant_config = quant_config
def create_weights(
self,
layer: torch.nn.Module,
num_experts: int,
hidden_size: int,
intermediate_size: int,
params_dtype: torch.dtype,
**extra_weight_attrs,
):
# Currently assuming is_k_full is always True
# (input size per partition is the same as full input size)
# Supports only sym for now (no zp)
if self.quant_config.group_size != -1:
scales_size13 = hidden_size // self.quant_config.group_size
scales_size2 = intermediate_size // self.quant_config.group_size
strategy = FusedMoeWeightScaleSupported.GROUP.value
else:
scales_size13 = 1
scales_size2 = 1
strategy = FusedMoeWeightScaleSupported.CHANNEL.value
extra_weight_attrs.update({
"quant_method": strategy,
"is_transposed": True
})
# Fused gate_up_proj (column parallel)
w13_qweight = torch.nn.Parameter(
torch.empty(
num_experts,
hidden_size // self.quant_config.pack_factor,
2 * intermediate_size,
dtype=torch.int32,
),
requires_grad=False,
)
layer.register_parameter("w13_qweight", w13_qweight)
set_weight_attrs(w13_qweight, extra_weight_attrs)
# down_proj (row parallel)
w2_qweight = torch.nn.Parameter(
torch.empty(
num_experts,
intermediate_size // self.quant_config.pack_factor,
hidden_size,
dtype=torch.int32,
),
requires_grad=False,
)
layer.register_parameter("w2_qweight", w2_qweight)
set_weight_attrs(w2_qweight, extra_weight_attrs)
# up_proj scales
w13_scales = torch.nn.Parameter(
torch.empty(num_experts,
scales_size13,
2 * intermediate_size,
dtype=torch.half),
requires_grad=False,
)
layer.register_parameter("w13_scales", w13_scales)
set_weight_attrs(w13_scales, extra_weight_attrs)
# down_proj scales
w2_scales = torch.nn.Parameter(
torch.empty(num_experts,
scales_size2,
hidden_size,
dtype=torch.half),
requires_grad=False,
)
layer.register_parameter("w2_scales", w2_scales)
set_weight_attrs(w2_scales, extra_weight_attrs)
# up_proj scales
w13_qzeros = torch.nn.Parameter(
torch.empty(num_experts,
scales_size13,
2 * intermediate_size // self.quant_config.pack_factor,
dtype=params_dtype),
requires_grad=False,
)
layer.register_parameter("w13_qzeros", w13_qzeros)
set_weight_attrs(w13_qzeros, extra_weight_attrs)
# down_proj scales
w2_qzeros = torch.nn.Parameter(
torch.empty(num_experts,
scales_size2,
hidden_size // self.quant_config.pack_factor,
dtype=params_dtype),
requires_grad=False,
)
layer.register_parameter("w2_qzeros", w2_qzeros)
set_weight_attrs(w2_qzeros, extra_weight_attrs)
w13_g_idx = torch.nn.Parameter(
torch.empty(
num_experts,
hidden_size,
dtype=torch.int32,
),
requires_grad=False,
)
layer.register_parameter("w13_g_idx", w13_g_idx)
set_weight_attrs(w13_g_idx, extra_weight_attrs)
w2_g_idx = torch.nn.Parameter(
torch.empty(
num_experts,
intermediate_size,
dtype=torch.int32,
),
requires_grad=False,
)
layer.register_parameter("w2_g_idx", w2_g_idx)
set_weight_attrs(w2_g_idx, extra_weight_attrs)
w13_g_idx_sort_indices = torch.nn.Parameter(
torch.empty(
num_experts,
hidden_size,
dtype=torch.int32,
),
requires_grad=False,
)
layer.register_parameter("w13_g_idx_sort_indices",
w13_g_idx_sort_indices)
set_weight_attrs(w13_g_idx_sort_indices, extra_weight_attrs)
w2_g_idx_sort_indices = torch.nn.Parameter(
torch.empty(
num_experts,
intermediate_size,
dtype=torch.int32,
),
requires_grad=False,
)
layer.register_parameter("w2_g_idx_sort_indices",
w2_g_idx_sort_indices)
set_weight_attrs(w2_g_idx_sort_indices, extra_weight_attrs)
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
# Process act_order
if self.quant_config.desc_act:
# Get sorting based on g_idx
num_experts = layer.w13_g_idx.shape[0]
w13_g_idx_sort_indices = torch.empty_like(layer.w13_g_idx)
w2_g_idx_sort_indices = torch.empty_like(layer.w2_g_idx)
w13_sorted_g_idx = torch.empty_like(layer.w13_g_idx)
w2_sorted_g_idx = torch.empty_like(layer.w2_g_idx)
for e in range(num_experts):
w13_g_idx_sort_indices[e] = torch.argsort(
layer.w13_g_idx[e]).to(torch.int32)
w2_g_idx_sort_indices[e] = torch.argsort(layer.w2_g_idx[e]).to(
torch.int32)
w13_sorted_g_idx[e] = layer.w13_g_idx[e][
w13_g_idx_sort_indices[e]]
w2_sorted_g_idx[e] = layer.w2_g_idx[e][
w2_g_idx_sort_indices[e]]
replace_tensor(layer, "w13_g_idx", w13_sorted_g_idx)
replace_tensor(layer, "w2_g_idx", w2_sorted_g_idx)
replace_tensor(layer, "w13_g_idx_sort_indices",
w13_g_idx_sort_indices)
replace_tensor(layer, "w2_g_idx_sort_indices",
w2_g_idx_sort_indices)
else:
# Reset g_idx related tensors
num_experts = layer.w13_g_idx.shape[0]
device = layer.w13_g_idx.device
layer.w13_g_idx = torch.nn.Parameter(
torch.empty((num_experts, 0), dtype=torch.int32,
device=device),
requires_grad=False,
)
layer.w2_g_idx = torch.nn.Parameter(
torch.empty((num_experts, 0), dtype=torch.int32,
device=device),
requires_grad=False,
)
layer.w13_g_idx_sort_indices = torch.nn.Parameter(
torch.empty((num_experts, 0), dtype=torch.int32,
device=device),
requires_grad=False,
)
layer.w2_g_idx_sort_indices = torch.nn.Parameter(
torch.empty((num_experts, 0), dtype=torch.int32,
device=device),
requires_grad=False,
)
# Repack weights
marlin_w13_qweight = ops.gptq_marlin_moe_repack(
layer.w13_qweight,
layer.w13_g_idx_sort_indices,
layer.w13_qweight.shape[1] * self.quant_config.pack_factor,
layer.w13_qweight.shape[2],
self.quant_config.quant_type.size_bits,
)
replace_tensor(layer, "w13_qweight", marlin_w13_qweight)
marlin_w2_qweight = ops.gptq_marlin_moe_repack(
layer.w2_qweight,
layer.w2_g_idx_sort_indices,
layer.w2_qweight.shape[1] * self.quant_config.pack_factor,
layer.w2_qweight.shape[2],
self.quant_config.quant_type.size_bits,
)
replace_tensor(layer, "w2_qweight", marlin_w2_qweight)
# Repack scales
marlin_w13_scales = marlin_moe_permute_scales(
s=layer.w13_scales,
size_k=layer.intermediate_size_per_partition,
size_n=layer.w13_scales.shape[2],
group_size=self.quant_config.group_size,
)
replace_tensor(layer, "w13_scales", marlin_w13_scales)
marlin_w2_scales = marlin_moe_permute_scales(
s=layer.w2_scales,
size_k=layer.w2_scales.shape[1] * self.quant_config.pack_factor,
size_n=layer.w2_scales.shape[2],
group_size=self.quant_config.group_size,
)
replace_tensor(layer, "w2_scales", marlin_w2_scales)
def apply(
self,
layer: torch.nn.Module,
x: torch.Tensor,
router_logits: torch.Tensor,
top_k: int,
renormalize: bool = True,
use_grouped_topk: bool = False,
num_expert_group: Optional[int] = None,
topk_group: Optional[int] = None,
custom_routing_function: Optional[Callable] = None,
) -> torch.Tensor:
from vllm.model_executor.layers.fused_moe.fused_marlin_moe import (
fused_marlin_moe)
# The input must currently be float16
orig_dtype = x.dtype
x = x.half()
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=None)
return fused_marlin_moe(
x,
layer.w13_qweight,
layer.w2_qweight,
router_logits,
layer.w13_g_idx,
layer.w2_g_idx,
layer.w13_g_idx_sort_indices,
layer.w2_g_idx_sort_indices,
topk_weights,
topk_ids,
w1_scale=layer.w13_scales,
w2_scale=layer.w2_scales,
).to(orig_dtype)