[Kernel]Support W4A8 Grouped GEMM on Hopper (#29691)

Signed-off-by: czhu-cohere <conway.zhu@cohere.com>
This commit is contained in:
czhu-cohere
2025-12-08 19:29:06 -08:00
committed by GitHub
parent ea657f2078
commit f6227c22ab
22 changed files with 2045 additions and 101 deletions
@@ -63,8 +63,10 @@ if HAS_TRITON:
from vllm.model_executor.layers.fused_moe.cutlass_moe import (
CutlassBatchedExpertsFp8,
CutlassExpertsFp8,
CutlassExpertsW4A8Fp8,
cutlass_moe_fp4,
cutlass_moe_fp8,
cutlass_moe_w4a8_fp8,
)
from vllm.model_executor.layers.fused_moe.deep_gemm_moe import DeepGemmExperts
from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
@@ -88,8 +90,10 @@ if HAS_TRITON:
"grouped_topk",
"cutlass_moe_fp8",
"cutlass_moe_fp4",
"cutlass_moe_w4a8_fp8",
"CutlassExpertsFp8",
"CutlassBatchedExpertsFp8",
"CutlassExpertsW4A8Fp8",
"TritonExperts",
"BatchedTritonExperts",
"DeepGemmExperts",
@@ -143,6 +143,7 @@ class FusedMoEQuantDesc:
scale: Union[torch.Tensor, "PrecisionConfig", None] = None
# Quantization alphas or gscales, used for nvfp4 types.
# W4A8 FP8: used for per-channel scales
# TODO(bnell): put some of these in subclasses
alpha_or_gscale: torch.Tensor | None = None
@@ -442,7 +443,9 @@ class FusedMoEQuantConfig:
- a1_scale: Optional scale to be used for a1.
- a2_scale: Optional scale to be used for a2.
- g1_alphas: Optional global quantization scales for w1 (for nvfp4).
per-channel scales for w1 (for W4A8 FP8).
- g2_alphas: Optional global quantization scales for w2 (for nvfp4).
per-channel scales for w2 (for W4A8 FP8).
- a1_gscale: Optional global quantization scales for a1 (for nvfp4).
- a2_gscale: Optional global quantization scales for a2 (for nvfp4).
- w1_bias: Optional biases for w1 (GPT OSS Triton).
@@ -461,6 +464,7 @@ class FusedMoEQuantConfig:
"mxfp4",
"mxfp6_e3m2",
"mxfp6_e2m3",
"int4",
}
if weight_dtype is None:
@@ -671,6 +675,31 @@ def int8_w8a16_moe_quant_config(
)
def int4_w4afp8_moe_quant_config(
w1_scale: torch.Tensor,
w2_scale: torch.Tensor,
g1_alphas: torch.Tensor,
g2_alphas: torch.Tensor,
per_act_token_quant: bool = False,
per_out_ch_quant: bool = False,
block_shape: list[int] | None = None,
) -> FusedMoEQuantConfig:
"""
Construct a quant config for fp8 activations and int4 weights.
"""
return FusedMoEQuantConfig.make(
torch.float8_e4m3fn, # quant dtype for activations
w1_scale=w1_scale,
w2_scale=w2_scale,
g1_alphas=g1_alphas,
g2_alphas=g2_alphas,
per_act_token_quant=per_act_token_quant,
per_out_ch_quant=per_out_ch_quant,
block_shape=block_shape,
weight_dtype="int4", # weight dtype for weights
)
def biased_moe_quant_config(
w1_bias: torch.Tensor | None,
w2_bias: torch.Tensor | None,
@@ -1052,3 +1052,404 @@ def run_cutlass_block_scaled_fused_experts(
return (
c2[c_map].view(m, topk, k) * topk_weights.view(m, topk, 1).to(out_dtype)
).sum(dim=1)
# W4A8
def run_cutlass_moe_w4a8_fp8(
output: torch.Tensor,
hidden_states: torch.Tensor,
w1: torch.Tensor,
w2: torch.Tensor,
topk_ids: torch.Tensor,
activation_callable: Callable,
global_num_experts: int,
expert_map: torch.Tensor | None,
w1_scale: torch.Tensor | None,
w2_scale: torch.Tensor | None,
a1q_scale: torch.Tensor | None,
a2_scale: torch.Tensor | None,
w1_chan_scale: torch.Tensor,
w2_chan_scale: torch.Tensor,
a_strides1: torch.Tensor,
a_strides2: torch.Tensor,
b_strides1: torch.Tensor,
b_strides2: torch.Tensor,
c_strides1: torch.Tensor,
c_strides2: torch.Tensor,
s_strides1: torch.Tensor,
s_strides2: torch.Tensor,
workspace13: torch.Tensor,
workspace2: torch.Tensor,
expert_num_tokens: torch.Tensor | None,
out_dtype: torch.dtype,
per_act_token: bool,
per_out_ch: bool,
use_batched_format: bool,
topk_weights: torch.Tensor | None,
group_size: int,
):
a1q = hidden_states
M = a1q.size(0)
local_E = w1.size(0)
device = a1q.device
_, K, N_packed = w2.shape
N = N_packed * 8 # logical N, pack 8 int4 into 1 int32
assert per_act_token, "W4A8 must use per-token scales"
assert per_out_ch, "W4A8 must use per-channel scales"
assert w1_scale is not None
assert w2_scale is not None
assert w1_scale.dtype == torch.float8_e4m3fn
assert w2_scale.dtype == torch.float8_e4m3fn
assert w1.dtype == torch.int32
assert w2.dtype == torch.int32
assert w1_chan_scale.dtype == torch.float32
assert w2_chan_scale.dtype == torch.float32
assert w1.size(0) == w2.size(0), "Weights expert number mismatch"
assert a1q_scale is not None
assert a2_scale is None
assert out_dtype in [torch.bfloat16], f"Invalid output dtype: {out_dtype}"
if expert_map is not None:
assert expert_num_tokens is None
assert not use_batched_format, "batched format not supported yet"
assert group_size == 128, f"Only group size 128 supported but got {group_size=}"
assert global_num_experts != -1
assert w1.size(2) * 8 == K, (
f"w1 hidden size mismatch: got {w1.size(2) * 8}, expected {K=}"
)
# Translate info from expert_map to topk_ids
if expert_map is not None:
local_topk_ids = torch.where(
expert_map[topk_ids] != -1, expert_map[topk_ids], -1
)
else:
local_topk_ids = topk_ids
topk = local_topk_ids.size(1)
a1q_perm = _resize_cache(workspace2.view(dtype=torch.float8_e4m3fn), (M * topk, K))
mm1_out = _resize_cache(workspace13, (M * topk, N * 2))
act_out = _resize_cache(workspace2, (M * topk, N))
# original workspace are based on input hidden_states dtype (bf16)
quant_out = _resize_cache(
workspace13.view(dtype=torch.float8_e4m3fn), (M * topk, N)
)
mm2_out = _resize_cache(workspace2, (M * topk, K))
problem_sizes1 = torch.empty(
(global_num_experts, 3), dtype=torch.int32, device=device
)
problem_sizes2 = torch.empty(
(global_num_experts, 3), dtype=torch.int32, device=device
)
num_expert = global_num_experts if expert_map is None else expert_map.size(0)
# permuted a1q reuses workspace2
a1q, a1q_scale, expert_offsets, inv_perm, _ = moe_permute(
a1q,
a1q_scale,
topk_ids,
num_expert,
local_E,
expert_map,
permuted_hidden_states=a1q_perm,
)
expert_offsets = expert_offsets[:-1]
# For RS gemm SwapAB is always enabled (swap logical M, N in the problem shape)
ops.get_cutlass_moe_mm_problem_sizes(
local_topk_ids,
problem_sizes1,
problem_sizes2,
global_num_experts,
N,
K,
force_swap_ab=True,
)
ops.cutlass_w4a8_moe_mm(
mm1_out,
a1q,
w1,
a1q_scale,
w1_chan_scale,
w1_scale,
group_size,
expert_offsets,
problem_sizes1,
a_strides1,
b_strides1,
c_strides1,
s_strides1,
)
activation_callable(act_out, mm1_out)
a2q, a2q_scale = ops.scaled_fp8_quant(
act_out, a2_scale, use_per_token_if_dynamic=per_act_token, output=quant_out
)
if expert_map is not None:
mm2_out.fill_(0)
ops.cutlass_w4a8_moe_mm(
mm2_out,
a2q,
w2,
a2q_scale,
w2_chan_scale,
w2_scale,
group_size,
expert_offsets,
problem_sizes2,
a_strides2,
b_strides2,
c_strides2,
s_strides2,
)
# for non-chunking mode the output is resized from workspace13
# so we need to make sure mm2_out uses workspace2.
moe_unpermute(
out=output,
permuted_hidden_states=mm2_out,
topk_weights=topk_weights,
inv_permuted_idx=inv_perm,
)
class CutlassExpertsW4A8Fp8(mk.FusedMoEPermuteExpertsUnpermute):
def __init__(
self,
out_dtype: torch.dtype | None,
a_strides1: torch.Tensor,
a_strides2: torch.Tensor,
b_strides1: torch.Tensor,
b_strides2: torch.Tensor,
c_strides1: torch.Tensor,
c_strides2: torch.Tensor,
s_strides1: torch.Tensor,
s_strides2: torch.Tensor,
quant_config: FusedMoEQuantConfig,
group_size: int,
):
super().__init__(quant_config)
self.out_dtype = out_dtype
self.a_strides1 = a_strides1
self.a_strides2 = a_strides2
self.b_strides1 = b_strides1
self.b_strides2 = b_strides2
self.c_strides1 = c_strides1
self.c_strides2 = c_strides2
self.s_strides1 = s_strides1
self.s_strides2 = s_strides2
self.group_size = group_size
@property
def activation_formats(
self,
) -> tuple[mk.FusedMoEActivationFormat, mk.FusedMoEActivationFormat]:
return (
mk.FusedMoEActivationFormat.Standard,
mk.FusedMoEActivationFormat.Standard,
)
def supports_chunking(self) -> bool:
return True
def supports_expert_map(self) -> bool:
return True
def finalize_weight_and_reduce_impl(self) -> mk.TopKWeightAndReduce:
# topk weights and reduction are fused in moe_unpermute cuda kernel
return TopKWeightAndReduceNoOP()
def workspace_dtype(self, act_dtype: torch.dtype) -> torch.dtype:
return self.out_dtype if self.out_dtype is not None else act_dtype
def workspace_shapes(
self,
M: int,
N: int,
K: int,
topk: int,
global_num_experts: int,
local_num_experts: int,
expert_tokens_meta: mk.ExpertTokensMetadata | None,
) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
workspace1 = (M * topk, max(N, K))
workspace2 = (M * topk, max(N // 2, K))
output = (M, K)
return (workspace1, workspace2, output)
def apply(
self,
output: torch.Tensor,
hidden_states: torch.Tensor,
w1: torch.Tensor,
w2: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
activation: str,
global_num_experts: int,
expert_map: torch.Tensor | None,
a1q_scale: torch.Tensor | None,
a2_scale: torch.Tensor | None,
workspace13: torch.Tensor | None,
workspace2: torch.Tensor | None,
expert_tokens_meta: mk.ExpertTokensMetadata | None,
apply_router_weight_on_input: bool,
):
assert self.w1_zp is None, "w1_zp is not supported in CUTLASS MoE"
assert self.w2_zp is None, "w2_zp is not supported in CUTLASS MoE"
expert_num_tokens = None
activation_callable = lambda o, i: self.activation(activation, o, i)
use_batched_format = (
self.activation_formats[0] == mk.FusedMoEActivationFormat.BatchedExperts
)
assert not use_batched_format, "batched format not supported"
in_dtype = hidden_states.dtype
run_cutlass_moe_w4a8_fp8(
output,
hidden_states,
w1,
w2,
topk_ids,
activation_callable,
global_num_experts,
expert_map,
self.w1_scale,
self.w2_scale,
a1q_scale,
a2_scale,
self.g1_alphas, # per-channel scales
self.g2_alphas, # per-channel scales
self.a_strides1,
self.a_strides2,
self.b_strides1,
self.b_strides2,
self.c_strides1,
self.c_strides2,
self.s_strides1,
self.s_strides2,
workspace13,
workspace2,
expert_num_tokens,
self.out_dtype if self.out_dtype is not None else in_dtype,
self.per_act_token_quant,
self.per_out_ch_quant,
use_batched_format,
topk_weights,
self.group_size,
)
def cutlass_moe_w4a8_fp8(
a: torch.Tensor,
w1_q: torch.Tensor,
w2_q: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
a_strides1: torch.Tensor,
a_strides2: torch.Tensor,
b_strides1: torch.Tensor,
b_strides2: torch.Tensor,
c_strides1: torch.Tensor,
c_strides2: torch.Tensor,
s_strides1: torch.Tensor,
s_strides2: torch.Tensor,
quant_config: FusedMoEQuantConfig,
activation: str = "silu",
expert_map: torch.Tensor | None = None,
apply_router_weight_on_input: bool = False,
global_num_experts: int = -1,
group_size: int = 128,
) -> torch.Tensor:
"""
This function computes a w4a8-quantized Mixture of Experts (MoE) layer
using two sets of quantized weights, w1_q and w2_q, and top-k gating
mechanism. The matrix multiplications are implemented with CUTLASS
mixed-dtype grouped gemm.
Parameters:
- a (torch.Tensor): The input tensor to the MoE layer.
Shape: [M, K]
- w1_q (torch.Tensor): The first set of fp8-quantized expert weights.
Shape: [num_experts, 2*N, K // packed_factor]
- w2_q (torch.Tensor): The second set of fp8-quantized expert weights.
Shape: [num_experts, K, N // packed_factor]
- topk_weights (torch.Tensor): The weights of each token->expert mapping.
- topk_ids (torch.Tensor): The token->expert mappings.
- a_strides1 (torch.Tensor): The input strides for the first gemm.
Shape: [num_experts]
- a_strides2 (torch.Tensor): The input strides for the second gemm.
Shape: [num_experts]
- b_strides1 (torch.Tensor): The packed layout for the first gemm weights.
Shape: [num_experts, 3]
dtype: torch.int32
- b_strides2 (torch.Tensor): The packed layout for the second gemm weights.
Shape: [num_experts, 3]
dtype: torch.int32
- c_strides1 (torch.Tensor): The output strides for the first gemm.
Shape: [num_experts]
- c_strides2 (torch.Tensor): The output strides for the second gemm.
Shape: [num_experts]
- s_strides1 (torch.Tensor): strides for the group-wise scales for the first gemm.
Shape: [num_experts, 2]
dtype: torch.int64
- s_strides2 (torch.Tensor): strides for the group-wise scales for the second gemm.
Shape: [num_experts, 2]
dtype: torch.int64
- per_act_token (Optional[bool]): Whether the scale is per-token or
per-tensor.
- activation (str): The activation function to use.
- expert_map (Optional[torch.Tensor]): In the case of Expert parallel,
every Rank is responsible for a subset of experts. expert_map is a
mapping from global expert-id to local expert-id. When expert_map[i]
is -1, it means that this Rank is not responsible for global
expert-id i.
- apply_router_weight_on_input (bool): When true, the topk weights are
applied directly on the inputs. This is only applicable when topk is 1.
- global_num_experts (int): The total number of experts.
- group_size (int): The number of weights per scale factor
Returns:
- torch.Tensor: The bf16 output tensor after applying the MoE layer.
"""
assert quant_config is not None
num_experts = global_num_experts if global_num_experts != -1 else w1_q.size(0)
fn = mk.FusedMoEModularKernel(
MoEPrepareAndFinalizeNoEP(),
CutlassExpertsW4A8Fp8(
out_dtype=a.dtype,
a_strides1=a_strides1,
a_strides2=a_strides2,
b_strides1=b_strides1,
b_strides2=b_strides2,
c_strides1=c_strides1,
c_strides2=c_strides2,
s_strides1=s_strides1,
s_strides2=s_strides2,
quant_config=quant_config,
group_size=group_size,
),
)
return fn(
a,
w1_q,
w2_q,
topk_weights,
topk_ids,
activation=activation,
global_num_experts=num_experts,
expert_map=expert_map,
apply_router_weight_on_input=apply_router_weight_on_input,
)
@@ -367,7 +367,7 @@ class FusedMoEPrepareAndFinalize(ABC):
class FusedMoEPermuteExpertsUnpermute(ABC):
"""
An abstract base class for the [Permute-Experts-Unpermute] step described
above.
above.
"""
def __init__(
@@ -256,7 +256,7 @@ class CompressedTensorsConfig(QuantizationConfig):
if format is not None
else is_activation_quantization_format(quant_format)
)
# TODO(czhu): w4a8fp8 is in packed-quantized format
# w4a8fp8 is in packed-quantized format
# but needs input activation quantization
input_activations = quant_config.get("input_activations")
if act_quant_format or input_activations:
@@ -33,6 +33,7 @@ from vllm.model_executor.layers.fused_moe.config import (
FusedMoEQuantConfig,
fp8_w8a8_moe_quant_config,
int4_w4a16_moe_quant_config,
int4_w4afp8_moe_quant_config,
int8_w8a8_moe_quant_config,
int8_w8a16_moe_quant_config,
nvfp4_moe_quant_config,
@@ -79,7 +80,11 @@ from vllm.model_executor.layers.quantization.utils.marlin_utils_fp4 import (
from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
prepare_moe_fp8_layer_for_marlin,
)
from vllm.model_executor.layers.quantization.utils.quant_utils import swizzle_blockscale
from vllm.model_executor.layers.quantization.utils.quant_utils import (
convert_bf16_scales_to_fp8,
convert_packed_uint4b8_to_signed_int4_inplace,
swizzle_blockscale,
)
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
all_close_1d,
normalize_e4m3fn_to_e4m3fnuz,
@@ -204,6 +209,11 @@ class CompressedTensorsMoEMethod(FusedMoEMethodBase):
return CompressedTensorsW8A8Int8MoEMethod(
weight_quant, input_quant, layer.moe_config
)
elif quant_config._is_fp8_w4a8_sm90(weight_quant, input_quant):
logger.info_once("Using CompressedTensorsW4A8Fp8MoEMethod")
return CompressedTensorsW4A8Fp8MoEMethod(
weight_quant, input_quant, layer.moe_config
)
elif quant_config._is_dynamic_token_w4a8_int(weight_quant, input_quant):
return CompressedTensorsW4A8Int8MoEMethod(
weight_quant, input_quant, layer.moe_config
@@ -2428,3 +2438,331 @@ class CompressedTensorsW4A8Int8MoEMethod(CompressedTensorsMoEMethod):
apply_router_weight_on_input,
int(_act_kind(activation)),
)
class CompressedTensorsW4A8Fp8MoEMethod(CompressedTensorsMoEMethod):
def __init__(
self,
weight_quant: QuantizationArgs,
input_quant: QuantizationArgs,
moe: FusedMoEConfig,
layer_name: str | None = None,
):
super().__init__(moe)
self.weight_quant = weight_quant
self.input_quant = input_quant
self.group_size = self.weight_quant.group_size
self.num_bits = self.weight_quant.num_bits
self.packed_factor = 32 // self.num_bits
assert self.weight_quant.symmetric, (
"Only symmetric quantization is supported for W4A8 MoE"
)
assert self.weight_quant.actorder != "group"
assert self.group_size == 128, "Only group size 128 supported for W4A8 MoE"
self.disable_expert_map = False
self.layer_name = layer_name
from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
from vllm.model_executor.layers.quantization.utils.quant_utils import (
GroupShape,
)
self.quant_fp8 = QuantFP8(static=False, group_shape=GroupShape.PER_TOKEN)
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
# requirement for CUTLASS reorder_tensor
assert hidden_size % 256 == 0, f"{hidden_size=} must be divisible by 256"
assert intermediate_size_per_partition % 256 == 0, (
f"{intermediate_size_per_partition=} must be divisible by 256"
)
# storage type, pack 8xint4 into int32
params_dtype = torch.int32
# WEIGHTS
w13_weight_packed = torch.nn.Parameter(
torch.empty(
num_experts,
2 * intermediate_size_per_partition,
hidden_size // self.packed_factor,
dtype=params_dtype,
),
requires_grad=False,
)
layer.register_parameter("w13_weight_packed", w13_weight_packed)
set_weight_attrs(w13_weight_packed, extra_weight_attrs)
w2_weight_packed = torch.nn.Parameter(
torch.empty(
num_experts,
hidden_size,
intermediate_size_per_partition // self.packed_factor,
dtype=params_dtype,
),
requires_grad=False,
)
layer.register_parameter("w2_weight_packed", w2_weight_packed)
set_weight_attrs(w2_weight_packed, extra_weight_attrs)
# SCALES
# weight_scale refers to the group-wise scales
# they are initially loaded as bf16, we will convert to fp8
# after loading
w13_weight_scale = torch.nn.Parameter(
torch.ones(
num_experts,
2 * intermediate_size_per_partition,
hidden_size // self.group_size,
dtype=layer.orig_dtype,
),
requires_grad=False,
)
layer.register_parameter("w13_weight_scale", w13_weight_scale)
w2_weight_scale = torch.nn.Parameter(
torch.ones(
num_experts,
hidden_size,
intermediate_size_per_partition // self.group_size,
dtype=layer.orig_dtype,
),
requires_grad=False,
)
layer.register_parameter("w2_weight_scale", w2_weight_scale)
# Add PER-GROUP quantization for FusedMoE.weight_loader.
extra_weight_attrs.update(
{"quant_method": FusedMoeWeightScaleSupported.GROUP.value}
)
set_weight_attrs(w13_weight_scale, extra_weight_attrs)
set_weight_attrs(w2_weight_scale, extra_weight_attrs)
# weight shapes
w2_weight_shape = torch.nn.Parameter(
torch.empty(num_experts, 2), requires_grad=False
)
layer.register_parameter("w2_weight_shape", w2_weight_shape)
set_weight_attrs(w2_weight_shape, extra_weight_attrs)
w13_weight_shape = torch.nn.Parameter(
torch.empty(num_experts, 2), requires_grad=False
)
layer.register_parameter("w13_weight_shape", w13_weight_shape)
set_weight_attrs(w13_weight_shape, extra_weight_attrs)
# don't use input scales
layer.w13_input_scale = None
layer.w2_input_scale = None
def process_weights_after_loading(self, layer):
device = layer.w13_weight_packed.device
# STRIDES
# A, C
self.a_strides1_c_strides2 = torch.full(
(layer.local_num_experts,),
layer.hidden_size,
device=device,
dtype=torch.int64,
)
self.a_strides2 = torch.full(
(layer.local_num_experts,),
layer.intermediate_size_per_partition,
device=device,
dtype=torch.int64,
)
self.c_strides1 = torch.full(
(layer.local_num_experts,),
2 * layer.intermediate_size_per_partition,
device=device,
dtype=torch.int64,
)
# S (group-wise scales)
# sizeof(StrideS) = 16 bytes, so we need to use 2xint64 to encode it
self.s_strides1 = torch.zeros(
(layer.local_num_experts, 2), device=device, dtype=torch.int64
)
self.s_strides1[:, 0] = 2 * layer.intermediate_size_per_partition
self.s_strides2 = torch.zeros(
(layer.local_num_experts, 2), device=device, dtype=torch.int64
)
self.s_strides2[:, 0] = layer.hidden_size
# encode and reorder weight tensors, and get the layout to pass to
# the grouped gemm kernel. `b_strides1/2` specifies the entire layout
convert_packed_uint4b8_to_signed_int4_inplace(layer.w13_weight_packed)
w13_weight_shuffled, self.b_strides1 = (
ops.cutlass_encode_and_reorder_int4b_grouped(layer.w13_weight_packed)
)
replace_parameter(layer, "w13_weight_packed", w13_weight_shuffled)
convert_packed_uint4b8_to_signed_int4_inplace(layer.w2_weight_packed)
w2_weight_shuffled, self.b_strides2 = (
ops.cutlass_encode_and_reorder_int4b_grouped(layer.w2_weight_packed)
)
replace_parameter(layer, "w2_weight_packed", w2_weight_shuffled)
# convert bf16 scales to (fp8_scales, channel_scales)
w13_weight_scale, w13_weight_chan_scale = convert_bf16_scales_to_fp8(
self.quant_fp8, layer.w13_weight_scale
)
w2_weight_scale, w2_weight_chan_scale = convert_bf16_scales_to_fp8(
self.quant_fp8, layer.w2_weight_scale
)
# register channel scales
layer.register_parameter(
"w13_weight_chan_scale",
torch.nn.Parameter(w13_weight_chan_scale, requires_grad=False),
)
layer.register_parameter(
"w2_weight_chan_scale",
torch.nn.Parameter(w2_weight_chan_scale, requires_grad=False),
)
# The scales are stored as (E, N, K // 128) but the kernel expects
# (E, K // 128, N) in row-major format, so we need to permute the last 2 dims
# and make it contiguous
w13_weight_scale_packed = ops.cutlass_pack_scale_fp8(
w13_weight_scale.permute(0, 2, 1).contiguous()
)
replace_parameter(layer, "w13_weight_scale", w13_weight_scale_packed)
w2_weight_scale_packed = ops.cutlass_pack_scale_fp8(
w2_weight_scale.permute(0, 2, 1).contiguous()
)
replace_parameter(layer, "w2_weight_scale", w2_weight_scale_packed)
def maybe_make_prepare_finalize(
self,
routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None,
) -> mk.FusedMoEPrepareAndFinalize | None:
return super().maybe_make_prepare_finalize(routing_tables)
def get_fused_moe_quant_config(
self, layer: torch.nn.Module
) -> FusedMoEQuantConfig | None:
# Store quantization scales; both per-group and per-channel
# Note we haven't specified the group size here because
# the quant config logic assumes group-wise scaling
# and channel-wise scaling are exclusive.
return int4_w4afp8_moe_quant_config(
w1_scale=layer.w13_weight_scale, # group scale
w2_scale=layer.w2_weight_scale, # group scale
g1_alphas=layer.w13_weight_chan_scale,
g2_alphas=layer.w2_weight_chan_scale,
per_act_token_quant=True, # always use dynamc per-token
per_out_ch_quant=True, # always use per-channel
)
def select_gemm_impl(
self,
prepare_finalize: mk.FusedMoEPrepareAndFinalize,
layer: torch.nn.Module,
) -> mk.FusedMoEPermuteExpertsUnpermute:
assert self.moe_quant_config is not None
assert (
prepare_finalize.activation_format == FusedMoEActivationFormat.Standard
), "BatchedExperts not supported"
from vllm.model_executor.layers.fused_moe import CutlassExpertsW4A8Fp8
experts: FusedMoEPermuteExpertsUnpermute
logger.debug("CutlassExpertsW4A8Fp8(%s)", self.__class__.__name__)
experts = CutlassExpertsW4A8Fp8(
out_dtype=self.moe.in_dtype,
a_strides1=self.a_strides1_c_strides2,
a_strides2=self.a_strides2,
b_strides1=self.b_strides1,
b_strides2=self.b_strides2,
c_strides1=self.c_strides1,
c_strides2=self.a_strides1_c_strides2,
s_strides1=self.s_strides1,
s_strides2=self.s_strides2,
quant_config=self.moe_quant_config,
group_size=self.group_size,
)
num_dispatchers = prepare_finalize.num_dispatchers()
self.disable_expert_map = (
num_dispatchers > 1 or not experts.supports_expert_map()
)
return experts
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: int | None = None,
num_expert_group: int | None = None,
global_num_experts: int = -1,
expert_map: torch.Tensor | None = None,
custom_routing_function: Callable | None = None,
scoring_func: str = "softmax",
routed_scaling_factor: float = 1.0,
e_score_correction_bias: torch.Tensor | None = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
enable_eplb: bool = False,
expert_load_view: torch.Tensor | None = None,
logical_to_physical_map: torch.Tensor | None = None,
logical_replica_count: torch.Tensor | None = None,
):
if enable_eplb:
raise NotImplementedError(
"EPLB not supported for `CompressedTensorsW4A8Fp8MoEMethod` yet."
)
assert self.moe_quant_config is not None
topk_weights, topk_ids, _ = layer.select_experts(
hidden_states=x,
router_logits=router_logits,
)
from vllm.model_executor.layers.fused_moe.cutlass_moe import (
cutlass_moe_w4a8_fp8,
)
return cutlass_moe_w4a8_fp8(
x,
layer.w13_weight_packed,
layer.w2_weight_packed,
topk_weights,
topk_ids,
quant_config=self.moe_quant_config,
activation=activation,
global_num_experts=global_num_experts,
expert_map=None if self.disable_expert_map else expert_map,
a_strides1=self.a_strides1_c_strides2,
a_strides2=self.a_strides2,
b_strides1=self.b_strides1,
b_strides2=self.b_strides2,
c_strides1=self.c_strides1,
c_strides2=self.a_strides1_c_strides2,
s_strides1=self.s_strides1,
s_strides2=self.s_strides2,
group_size=self.group_size,
)
@property
def supports_eplb(self) -> bool:
return False
@@ -128,14 +128,15 @@ class CompressedTensorsW4A8Fp8(CompressedTensorsScheme):
),
)
# TODO(czhu): allocate the packed fp8 scales memory here?
# the scales will be expanded by 8x via `cutlass_pack_scale_fp8`
# After loading, we will transform bf16 -> fp8 ->
# expand by 8x via `cutlass_pack_scale_fp8`
# and construct per-channel fp32 scales.
weight_scale_args = {
"weight_loader": weight_loader,
"data": torch.empty(
output_size_per_partition,
scales_and_zp_size,
dtype=torch.float8_e4m3fn,
dtype=params_dtype,
),
}
@@ -152,17 +153,9 @@ class CompressedTensorsW4A8Fp8(CompressedTensorsScheme):
data=torch.empty(2, dtype=torch.int64), weight_loader=weight_loader
)
# per-channel scales
weight_chan_scale = ChannelQuantScaleParameter(
data=torch.empty((output_size_per_partition, 1), dtype=torch.float32),
output_dim=0,
weight_loader=weight_loader,
)
layer.register_parameter("weight_packed", weight)
layer.register_parameter("weight_scale", weight_scale)
layer.register_parameter("weight_shape", weight_shape)
layer.register_parameter("weight_chan_scale", weight_chan_scale)
self.kernel = kernel_type(
mp_linear_kernel_config,
@@ -6,7 +6,11 @@ import torch
from vllm import _custom_ops as ops
from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
from vllm.model_executor.layers.quantization.utils.quant_utils import (
GroupShape,
convert_bf16_scales_to_fp8,
convert_packed_uint4b8_to_signed_int4_inplace,
)
from vllm.model_executor.parameter import BasevLLMParameter, permute_param_layout_
from vllm.platforms import current_platform
from vllm.scalar_type import scalar_types
@@ -48,7 +52,6 @@ class CutlassW4A8LinearKernel(MPLinearKernel):
"CUTLASS W4A8, only supported int4",
)
# TODO(czhu): support -1 (column-wise)
if c.group_size != 128:
return False, "Only group_size 128 is supported"
@@ -71,9 +74,9 @@ class CutlassW4A8LinearKernel(MPLinearKernel):
# `weight_packed` is: {input_dim = 0, output_dim = 1, packed_dim = 0}
# `weight_scale` is: {input_dim = 0, output_dim = 1}
def process_weights_after_loading(self, layer: torch.nn.Module):
# TODO(czhu): optimize speed/mem usage
def transform_w_q(x):
assert isinstance(x, BasevLLMParameter)
convert_packed_uint4b8_to_signed_int4_inplace(x.data)
permute_param_layout_(x, input_dim=0, output_dim=1, packed_dim=0)
x.data = ops.cutlass_encode_and_reorder_int4b(x.data.t().contiguous().t())
return x
@@ -85,10 +88,18 @@ class CutlassW4A8LinearKernel(MPLinearKernel):
x.data = ops.cutlass_pack_scale_fp8(x.data)
return x
w_s = getattr(layer, self.w_s_name)
fp8_scales, chan_scales = convert_bf16_scales_to_fp8(self.quant_fp8, w_s.data)
w_s.data = fp8_scales
# register per-channel scales
layer.register_parameter(
"weight_chan_scale", torch.nn.Parameter(chan_scales, requires_grad=False)
)
# Encode/reorder weights and pack scales
self._transform_param(layer, self.w_q_name, transform_w_q)
self._transform_param(layer, self.w_s_name, transform_w_s)
self._transform_param(layer, "weight_chan_scale", lambda x: x)
def apply_weights(
self,
@@ -2,7 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""This file is used for /tests and /benchmarks"""
from collections.abc import Mapping
from collections.abc import Callable, Mapping
from dataclasses import dataclass
from types import MappingProxyType
from typing import ClassVar, NamedTuple
@@ -691,3 +691,51 @@ def cutlass_fp4_supported() -> bool:
capability_tuple = current_platform.get_device_capability()
capability = -1 if capability_tuple is None else capability_tuple.to_int()
return cutlass_scaled_mm_supports_fp4(capability)
def convert_bf16_scales_to_fp8(
quant_fp8: Callable, scales: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
"""
Convert a BF16 scale tensor into the pair of (fp8_scales, channel_scales)
expected by W4A8 GEMM kernels.
"""
assert scales.is_contiguous(), (
f"scale tensor must be contiguous, got {scales.stride()=}"
)
assert scales.is_cuda, "scales must be on gpu"
orig_shape = scales.shape
k_groups = orig_shape[-1]
flat_scales = scales.view(-1, k_groups)
fp8_scales, chan_scales = quant_fp8(flat_scales)
fp8_scales = (fp8_scales.float() / 8.0).to(torch.float8_e4m3fn)
chan_scales *= 8.0
# restore original shape
fp8_scales = fp8_scales.view(orig_shape)
chan_scales = chan_scales.view(orig_shape[:-1], -1)
return fp8_scales, chan_scales
def convert_packed_uint4b8_to_signed_int4_inplace(t: torch.Tensor) -> torch.Tensor:
"""
Convert int4b8 (packed to int32) to signed int4
"""
assert t.is_cuda, "tensor must be on gpu"
assert t.dtype == torch.int32, f"expected int32 packed weights but got {t.dtype}"
# loop through the 8 4-bit nibbles in each int32 entry
for i in range(8):
shift = 4 * i
# extract the i-th 4-bit nibble
nib = (t >> shift) & 0xF
# clear the original nibble by masking out
t &= ~(0xF << shift)
# convert int4b8 [0..15] to signed int4 [-8..7] by subtracting 8
# and update in-place
t |= ((nib - 8) & 0xF) << shift
return t