[Kernel] Zero point support in fused MarlinMoE kernel + AWQ Fused MoE (#8973)

Co-authored-by: Dipika <dipikasikka1@gmail.com>
Co-authored-by: Dipika Sikka <ds3822@columbia.edu>
This commit is contained in:
ElizaWszola
2024-10-04 12:34:44 -06:00
committed by GitHub
co-authored by Dipika Dipika Sikka
parent 0dcc8cbe5a
commit 05d686432f
23 changed files with 969 additions and 223 deletions
@@ -10,15 +10,24 @@ from vllm.model_executor.layers.fused_moe.fused_moe import (
from vllm.scalar_type import scalar_types
def get_scalar_type(num_bits: int, has_zp: bool):
if has_zp:
assert num_bits == 4
return scalar_types.uint4
else:
return scalar_types.uint4b8 if num_bits == 4 else scalar_types.uint8b128
def single_marlin_moe(
hidden_states: torch.Tensor,
w: torch.Tensor,
scales: torch.Tensor,
gating_output: torch.Tensor,
g_idx: torch.Tensor,
perm: torch.Tensor,
topk: int,
renormalize: bool,
g_idx: Optional[torch.Tensor] = None,
sort_indices: Optional[torch.Tensor] = None,
w_zeros: Optional[torch.Tensor] = None,
override_config: Optional[Dict[str, Any]] = None,
num_bits: int = 8,
is_k_full: bool = True,
@@ -34,10 +43,12 @@ def single_marlin_moe(
- scales (torch.Tensor): The quantization scales.
- gating_output (torch.Tensor): The output of the gating operation
(before softmax).
- g_idx (torch.Tensor): The act_order indices.
- perm (torch.Tensor): The act_order input permutation.
- g_idx (Optional[torch.Tensor]): Optional act_order indices.
- sort_indices (Optional[torch.Tensor]): Optional act_order input
permutation.
- topk (int): The number of top-k experts to select.
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
- override_config (Optional[Dict[str, Any]]): Optional override
for the kernel configuration.
- num_bits (bool): The number of bits in expert weights quantization.
@@ -79,16 +90,34 @@ def single_marlin_moe(
max_workspace_size = (N // 64) * 16
workspace = torch.zeros(max_workspace_size,
dtype=torch.int,
device="cuda",
device=hidden_states.device,
requires_grad=False)
scalar_type = (scalar_types.uint4b8
if num_bits == 4 else scalar_types.uint8b128)
has_zero_point = w_zeros is not None
if w_zeros is None:
w_zeros = torch.empty((0, 0),
dtype=hidden_states.dtype,
device=hidden_states.device,
requires_grad=False)
if g_idx is None:
g_idx = torch.empty((0, 0),
dtype=torch.int32,
device=hidden_states.device,
requires_grad=False)
if sort_indices is None:
sort_indices = torch.empty((0),
dtype=torch.int32,
device=hidden_states.device,
requires_grad=False)
scalar_type = get_scalar_type(num_bits, has_zero_point)
intermediate_cache = torch.ops._moe_C.marlin_gemm_moe(
hidden_states, w, sorted_token_ids, topk_weights, topk_ids, scales,
g_idx, perm, workspace, scalar_type, M, N, K, is_k_full, E, topk,
block_size_m, True, False)
w_zeros, g_idx, sort_indices, workspace, scalar_type, M, N, K,
is_k_full, E, topk, block_size_m, True, False)
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
@@ -97,16 +126,18 @@ def fused_marlin_moe(
hidden_states: torch.Tensor,
w1: torch.Tensor,
w2: torch.Tensor,
w1_scale: torch.Tensor,
w2_scale: torch.Tensor,
gating_output: torch.Tensor,
g_idx1: torch.Tensor,
g_idx2: torch.Tensor,
perm1: torch.Tensor,
perm2: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
g_idx1: Optional[torch.Tensor] = None,
g_idx2: Optional[torch.Tensor] = None,
sort_indices1: Optional[torch.Tensor] = None,
sort_indices2: Optional[torch.Tensor] = None,
w1_zeros: Optional[torch.Tensor] = None,
w2_zeros: Optional[torch.Tensor] = None,
override_config: Optional[Dict[str, Any]] = None,
w1_scale: Optional[torch.Tensor] = None,
w2_scale: Optional[torch.Tensor] = None,
num_bits: int = 8,
is_k_full: bool = True,
) -> torch.Tensor:
@@ -118,21 +149,22 @@ def fused_marlin_moe(
- hidden_states (torch.Tensor): The input tensor to the MoE layer.
- w1 (torch.Tensor): The first set of expert weights.
- w2 (torch.Tensor): The second set of expert weights.
- w1_scale (torch.Tensor): Scale to be used for w1.
- w2_scale (torch.Tensor): Scale to be used for w2.
- gating_output (torch.Tensor): The output of the gating operation
(before softmax).
- g_idx1 (torch.Tensor): The first set of act_order indices.
- g_idx2 (torch.Tensor): The second set of act_order indices.
- perm1 (torch.Tensor): The first act_order input permutation.
- perm2 (torch.Tensor): The second act_order input permutation.
- g_idx1 (Optional[torch.Tensor]): The first set of act_order indices.
- g_idx2 (Optional[torch.Tensor]): The second set of act_order indices.
- sort_indices1 (Optional[torch.Tensor]): The first act_order input
permutation.
- sort_indices2 (Optional[torch.Tensor]): The second act_order input
permutation.
- topk_weights (torch.Tensor): Top-k weights.
- topk_ids (torch.Tensor): Indices of topk-k elements.
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
- override_config (Optional[Dict[str, Any]]): Optional override
for the kernel configuration.
- w1_scale (Optional[torch.Tensor]): Optional scale to be used for
w1.
- w2_scale (Optional[torch.Tensor]): Optional scale to be used for
w2.
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
- num_bits (bool): The number of bits in expert weights quantization.
Returns:
@@ -152,6 +184,20 @@ def fused_marlin_moe(
assert hidden_states.dtype == torch.float16
assert num_bits in [4, 8]
has_no_act_order = (g_idx1 is None and g_idx2 is None
and sort_indices1 is None and sort_indices2 is None)
has_all_act_order = (g_idx1 is not None and g_idx2 is not None
and sort_indices1 is not None
and sort_indices2 is not None)
assert has_no_act_order or has_all_act_order, (
"g_idx and sorted_indices "
"must be all not None or must be all None")
has_no_zp = w1_zeros is None and w2_zeros is None
has_all_zp = w1_zeros is not None and w2_zeros is not None
assert has_no_zp or has_all_zp, ("zero points must be both not None or "
"must be both None")
M, K = hidden_states.shape
E = w1.shape[0]
N = w2.shape[1] * 16
@@ -172,14 +218,42 @@ def fused_marlin_moe(
sorted_token_ids, _, _ = moe_align_block_size(topk_ids, block_size_m, E)
max_workspace_size = ((M + 255) // 256) * (max(2 * N, K) // 64) * 16
max_workspace_size = (max(2 * N, K) // 64) * 16
workspace = torch.zeros(max_workspace_size,
dtype=torch.int,
device="cuda",
requires_grad=False)
scalar_type = (scalar_types.uint4b8
if num_bits == 4 else scalar_types.uint8b128)
if has_no_zp:
w1_zeros = torch.empty((0, 0),
dtype=hidden_states.dtype,
device=hidden_states.device,
requires_grad=False)
w2_zeros = torch.empty((0, 0),
dtype=hidden_states.dtype,
device=hidden_states.device,
requires_grad=False)
if has_no_act_order:
g_idx1 = torch.empty((0, 0),
dtype=torch.int32,
device=hidden_states.device,
requires_grad=False)
g_idx2 = torch.empty((0, 0),
dtype=torch.int32,
device=hidden_states.device,
requires_grad=False)
sort_indices1 = torch.empty((0),
dtype=torch.int32,
device=hidden_states.device,
requires_grad=False)
sort_indices2 = torch.empty((0, 0),
dtype=torch.int32,
device=hidden_states.device,
requires_grad=False)
scalar_type1 = get_scalar_type(num_bits, has_all_zp)
scalar_type2 = get_scalar_type(num_bits, has_all_zp)
intermediate_cache2 = torch.empty(
(M * topk_ids.shape[1], N),
@@ -194,10 +268,11 @@ def fused_marlin_moe(
topk_weights,
topk_ids,
w1_scale,
w1_zeros,
g_idx1,
perm1,
sort_indices1,
workspace,
scalar_type,
scalar_type1,
M,
2 * N,
K,
@@ -218,10 +293,11 @@ def fused_marlin_moe(
topk_weights,
topk_ids,
w2_scale,
w2_zeros,
g_idx2,
perm2,
sort_indices2,
workspace,
scalar_type,
scalar_type2,
M,
K,
N,
@@ -1,16 +1,21 @@
from typing import Any, Dict, List, Optional
from typing import Any, Callable, Dict, List, Optional
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)
QuantizationConfig, QuantizeMethodBase)
from vllm.model_executor.layers.quantization.utils import replace_parameter
from vllm.model_executor.layers.quantization.utils.marlin_utils import (
apply_awq_marlin_linear, awq_to_marlin_zero_points, check_marlin_supported,
marlin_make_empty_g_idx, marlin_make_workspace, marlin_permute_scales,
marlin_make_empty_g_idx, marlin_make_workspace, marlin_moe_permute_scales,
marlin_permute_scales, moe_awq_to_marlin_zero_points,
verify_marlin_supported, verify_marlin_supports_shape)
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
from vllm.model_executor.parameter import (GroupQuantScaleParameter,
@@ -35,12 +40,13 @@ class AWQMarlinConfig(QuantizationConfig):
self.group_size = group_size
self.has_zp = has_zp
self.lm_head_quantized = lm_head_quantized
self.weight_bits = weight_bits
if weight_bits not in self.TYPE_MAP:
raise ValueError(f"Unsupported num_bits = {weight_bits}. "
if self.weight_bits not in self.TYPE_MAP:
raise ValueError(f"Unsupported num_bits = {self.weight_bits}. "
f"Supported num_bits = {self.TYPE_MAP.keys()}")
self.quant_type = self.TYPE_MAP[weight_bits]
self.quant_type = self.TYPE_MAP[self.weight_bits]
verify_marlin_supported(self.quant_type,
group_size=self.group_size,
@@ -98,10 +104,12 @@ class AWQMarlinConfig(QuantizationConfig):
return None
def get_quant_method(self, layer: torch.nn.Module,
prefix: str) -> Optional["AWQMarlinLinearMethod"]:
prefix: str) -> Optional["QuantizeMethodBase"]:
if (isinstance(layer, LinearBase) or
(isinstance(layer, ParallelLMHead) and self.lm_head_quantized)):
return AWQMarlinLinearMethod(self)
elif isinstance(layer, FusedMoE):
return AWQMoEMethod(self)
return None
def get_scaled_act_names(self) -> List[str]:
@@ -271,4 +279,182 @@ class AWQMarlinLinearMethod(LinearMethodBase):
quant_type=self.quant_config.quant_type,
output_size_per_partition=layer.output_size_per_partition,
input_size_per_partition=layer.input_size_per_partition,
bias=bias)
bias=bias)
class AWQMoEMethod(FusedMoEMethodBase):
def __init__(self, quant_config: AWQMarlinConfig):
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):
extra_weight_attrs.update({
"is_transposed":
True,
"quant_method":
FusedMoeWeightScaleSupported.GROUP.value,
})
w13_qweight = Parameter(torch.empty(num_experts,
hidden_size,
2 * intermediate_size //
self.quant_config.pack_factor,
dtype=torch.int32),
requires_grad=False)
layer.register_parameter("w13_qweight", w13_qweight)
set_weight_attrs(w13_qweight, extra_weight_attrs)
w2_qweight = Parameter(torch.empty(num_experts,
intermediate_size,
hidden_size //
self.quant_config.pack_factor,
dtype=torch.int32),
requires_grad=False)
layer.register_parameter("w2_qweight", w2_qweight)
set_weight_attrs(w2_qweight, extra_weight_attrs)
num_groups_w13 = hidden_size // self.quant_config.group_size
num_groups_w2 = intermediate_size // self.quant_config.group_size
# WEIGHT_SCALES
# Allocate 2 scales for w1 and w3 respectively.
w13_scales = Parameter(torch.empty(num_experts,
num_groups_w13,
intermediate_size * 2,
dtype=params_dtype),
requires_grad=False)
layer.register_parameter("w13_scales", w13_scales)
set_weight_attrs(w13_scales, extra_weight_attrs)
w2_scales = Parameter(torch.empty(num_experts,
num_groups_w2,
hidden_size,
dtype=params_dtype),
requires_grad=False)
layer.register_parameter("w2_scales", w2_scales)
set_weight_attrs(w2_scales, extra_weight_attrs)
# WEIGHT_ZERO_POINT
# Allocate 2 zero points for w1 and w3 respectively.
w13_qzeros = Parameter(torch.empty(num_experts,
num_groups_w13,
2 * intermediate_size //
self.quant_config.pack_factor,
dtype=torch.int32),
requires_grad=False)
layer.register_parameter("w13_qzeros", w13_qzeros)
set_weight_attrs(w13_qzeros, extra_weight_attrs)
w2_qzeros = Parameter(torch.empty(num_experts,
num_groups_w2,
hidden_size //
self.quant_config.pack_factor,
dtype=torch.int32),
requires_grad=False)
layer.register_parameter("w2_qzeros", w2_qzeros)
set_weight_attrs(w2_qzeros, extra_weight_attrs)
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
num_experts = layer.w13_qweight.shape[0]
device = layer.w13_qweight.device
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,
)
marlin_w13_qweight = ops.awq_marlin_moe_repack(
layer.w13_qweight,
layer.w13_g_idx_sort_indices,
size_k=layer.w13_qweight.shape[1],
size_n=layer.w13_qweight.shape[2] * self.quant_config.pack_factor,
num_bits=self.quant_config.weight_bits,
)
replace_parameter(layer, "w13_qweight", marlin_w13_qweight)
marlin_w2_qweight = ops.awq_marlin_moe_repack(
layer.w2_qweight,
layer.w2_g_idx_sort_indices,
size_k=layer.w2_qweight.shape[1],
size_n=layer.w2_qweight.shape[2] * self.quant_config.pack_factor,
num_bits=self.quant_config.weight_bits,
)
replace_parameter(layer, "w2_qweight", marlin_w2_qweight)
# Why does this take the intermediate size for size_k?
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_parameter(layer, "w13_scales", marlin_w13_scales)
marlin_w2_scales = marlin_moe_permute_scales(
s=layer.w2_scales,
size_k=layer.intermediate_size_per_partition,
size_n=layer.w2_scales.shape[2],
group_size=self.quant_config.group_size,
)
replace_parameter(layer, "w2_scales", marlin_w2_scales)
marlin_w13_zp = moe_awq_to_marlin_zero_points(
layer.w13_qzeros,
size_k=layer.w13_qzeros.shape[1],
size_n=layer.w13_qzeros.shape[2] * self.quant_config.pack_factor,
num_bits=self.quant_config.weight_bits)
replace_parameter(layer, "w13_qzeros", marlin_w13_zp)
marlin_w2_zp = moe_awq_to_marlin_zero_points(
layer.w2_qzeros,
size_k=layer.w2_qzeros.shape[1],
size_n=layer.w2_qzeros.shape[2] * self.quant_config.pack_factor,
num_bits=self.quant_config.weight_bits)
replace_parameter(layer, "w2_qzeros", marlin_w2_zp)
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)
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)
return fused_marlin_moe(
x,
layer.w13_qweight,
layer.w2_qweight,
layer.w13_scales,
layer.w2_scales,
router_logits,
topk_weights,
topk_ids,
w1_zeros=layer.w13_qzeros,
w2_zeros=layer.w2_qzeros,
num_bits=self.quant_config.weight_bits,
)
@@ -498,14 +498,14 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
x,
layer.w13_weight_packed,
layer.w2_weight_packed,
layer.w13_weight_scale,
layer.w2_weight_scale,
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_weight_scale,
w2_scale=layer.w2_weight_scale,
g_idx1=layer.w13_g_idx,
g_idx2=layer.w2_g_idx,
sort_indices1=layer.w13_g_idx_sort_indices,
sort_indices2=layer.w2_g_idx_sort_indices,
num_bits=self.num_bits,
)
@@ -557,14 +557,14 @@ class GPTQMarlinMoEMethod(FusedMoEMethodBase):
x,
layer.w13_qweight,
layer.w2_qweight,
layer.w13_scales,
layer.w2_scales,
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,
g_idx1=layer.w13_g_idx,
g_idx2=layer.w2_g_idx,
sort_indices1=layer.w13_g_idx_sort_indices,
sort_indices2=layer.w2_g_idx_sort_indices,
num_bits=self.quant_config.quant_type.size_bits,
).to(orig_dtype)
@@ -208,6 +208,7 @@ def marlin_moe_permute_scales(
device=s.device,
dtype=s.dtype,
)
for e in range(num_experts):
output[e] = marlin_permute_scales(s[e], size_k, size_n, group_size)
return output
@@ -258,6 +259,20 @@ def awq_to_marlin_zero_points(q_zp_packed: torch.Tensor, size_k: int,
return marlin_zp
def moe_awq_to_marlin_zero_points(q_zp_packed: torch.Tensor, size_k: int,
size_n: int, num_bits: int):
num_experts = q_zp_packed.shape[0]
output = torch.empty(
(num_experts, q_zp_packed.shape[1], q_zp_packed.shape[2]),
device=q_zp_packed.device,
dtype=q_zp_packed.dtype,
)
for e in range(num_experts):
output[e] = awq_to_marlin_zero_points(q_zp_packed[e], size_k, size_n,
num_bits)
return output
def apply_gptq_marlin_linear(
input: torch.Tensor,
weight: torch.Tensor,
+3 -1
View File
@@ -23,7 +23,9 @@ def get_model_architecture(
architectures = getattr(model_config.hf_config, "architectures", [])
# Special handling for quantized Mixtral.
# FIXME(woosuk): This is a temporary hack.
mixtral_supported = ["fp8", "compressed-tensors", "gptq_marlin"]
mixtral_supported = [
"fp8", "compressed-tensors", "gptq_marlin", "awq_marlin"
]
if (model_config.quantization is not None
and model_config.quantization not in mixtral_supported