Skip to content

GitLab

Commit 5b7f2c7d authored by SAC_fanth's avatar SAC_fanth
Browse files

fused_moe_fp8接入lmslim

parent c2e7f4ad
Hide whitespace changes
Inline Side-by-side
Showing with 21 additions and 417 deletions
vllm/model_executor/layers/fused_moe/fused_moe.py
View file @ 5b7f2c7d
......@@ -1791,7 +1791,7 @@ def fused_experts_impl(
device=hidden_states.device,
dtype=hidden_states.dtype)
if use_int8_w8a8 is True:
if use_int8_w8a8 or use_fp8_w8a8:
return fused_experts_impl_int8(hidden_states=hidden_states,
w1=w1,
w2=w2,
......@@ -1801,8 +1801,8 @@ def fused_experts_impl(
inplace=inplace,
activation=activation,
apply_router_weight_on_input=apply_router_weight_on_input,
use_fp8_w8a8=False,
use_int8_w8a8=True,
use_fp8_w8a8=use_fp8_w8a8,
use_int8_w8a8=use_int8_w8a8,
use_int8_w8a16=False,
use_int4_w4a16=False,
per_channel_quant=per_channel_quant,
......
vllm/model_executor/layers/quantization/utils/fp8_utils.py
View file @ 5b7f2c7d
......@@ -21,6 +21,10 @@ from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
from vllm.platforms import current_platform
from vllm.triton_utils import tl, triton
from vllm.utils import cdiv, direct_register_custom_op, has_deep_gemm
try:
from lmslim.layers.gemm.fp8_utils import per_token_group_quant_fp8,w8a8_block_fp8_matmul
except Exception:
print("INFO: Please updata lmslim if you want to use fp8_utils.\n")
logger = init_logger(__name__)
......@@ -256,330 +260,6 @@ def block_quant_to_tensor_quant(
return x_q_tensor, scale
@triton.jit
def _per_token_group_quant_fp8(
# Pointers to inputs and output
y_ptr,
y_q_ptr,
y_s_ptr,
group_size,
# Num columns of y
y_num_columns,
y_row_stride,
# Avoid to divide zero
eps,
# Information for float8
fp8_min,
fp8_max,
# Meta-parameters
BLOCK: tl.constexpr,
):
"""A Triton-accelerated function to perform per-token-group
quantization on a tensor.
This function converts the tensor values into float8 values.
"""
groups_per_row = y_num_columns // group_size
# Map the program id to the row of X and Y it should compute.
g_id = tl.program_id(0)
row = g_id // groups_per_row
row_g_id = g_id % groups_per_row
# Ensure offset calculations use int64 to prevent overflow
y_ptr_offset = (row.to(tl.int64) * y_row_stride) + (row_g_id.to(tl.int64) *
group_size)
y_ptr += y_ptr_offset
y_q_ptr_offset = g_id.to(tl.int64) * group_size
y_q_ptr += y_q_ptr_offset
y_s_ptr += g_id
cols = tl.arange(0, BLOCK) # N <= BLOCK
mask = cols < group_size
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
# Quant
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
y_s = _absmax / fp8_max
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
tl.store(y_q_ptr + cols, y_q, mask=mask)
tl.store(y_s_ptr, y_s)
@triton.jit
def _per_token_group_quant_fp8_colmajor(
# Pointers to inputs and output
y_ptr,
y_q_ptr,
y_s_ptr,
group_size,
# Num columns of y
y_num_columns,
y_row_stride,
# Stride from one column to the next of y_s
y_s_col_stride,
# Avoid to divide zero
eps,
# Information for float8
fp8_min,
fp8_max,
# Meta-parameters
BLOCK: tl.constexpr,
):
"""A Triton-accelerated function to perform per-token-group
quantization on a tensor.
This function converts the tensor values into float8 values.
"""
groups_per_row = y_num_columns // group_size
# Map the program id to the row of X and Y it should compute.
g_id = tl.program_id(0)
row = g_id // groups_per_row
row_g_id = g_id % groups_per_row
# Ensure offset calculations use int64 to prevent overflow
y_ptr_offset = (row.to(tl.int64) * y_row_stride) + (row_g_id.to(tl.int64) *
group_size)
y_ptr += y_ptr_offset
y_q_ptr_offset = g_id.to(tl.int64) * group_size
y_q_ptr += y_q_ptr_offset
# Convert g_id the flattened block coordinate to 2D so we can index
# into the output y_scales matrix
blocks_per_row = y_num_columns // group_size
scale_col = g_id % blocks_per_row
scale_row = g_id // blocks_per_row
# Ensure offset calculation uses int64 for y_s_ptr
y_s_ptr_offset = (scale_col.to(tl.int64) * y_s_col_stride) + scale_row.to(
tl.int64)
y_s_ptr += y_s_ptr_offset
cols = tl.arange(0, BLOCK) # group_size <= BLOCK
mask = cols < group_size
y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
# Quant
_absmax = tl.maximum(tl.max(tl.abs(y)), eps)
y_s = _absmax / fp8_max
y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
tl.store(y_q_ptr + cols, y_q, mask=mask)
tl.store(y_s_ptr, y_s)
def per_token_group_quant_fp8(
x: torch.Tensor,
group_size: int,
eps: float = 1e-10,
dtype: Optional[torch.dtype] = None,
column_major_scales: bool = False,
out_q: Optional[torch.Tensor] = None,
) -> tuple[torch.Tensor, torch.Tensor]:
"""Function to perform per-token-group quantization on an input tensor `x`.
It converts the tensor values into signed float8 values and returns the
quantized tensor along with the scaling factor used for quantization.
Args:
x: The input tensor with ndim >= 2.
group_size: The group size used for quantization.
eps: The minimum to avoid dividing zero.
dtype: The dype of output tensor. Note that only `torch.float8_e4m3fn`
is supported for now.
column_major_scales: Outputs scales in column major.
out_q: Optional output tensor. If not provided, function will create.
Returns:
tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the
scaling factor for quantization.
"""
dtype = current_platform.fp8_dtype() if dtype is None else dtype
assert (x.shape[-1] % group_size == 0), (
f"the last dimension of `x` {x.shape[-1]} must be divisible "
f"by `group_size` {group_size}")
assert x.stride(-1) == 1, "`x` groups must be contiguous"
finfo = torch.finfo(dtype)
fp8_min = finfo.min
fp8_max = finfo.max
assert out_q is None or out_q.shape == x.shape
x_q = out_q
if x_q is None:
x_q = torch.empty_like(x, device=x.device, dtype=dtype)
M = x.numel() // group_size
N = group_size
if column_major_scales:
shape = (x.shape[-1] // group_size, ) + x.shape[:-1]
x_s = torch.empty(shape, device=x.device,
dtype=torch.float32).permute(-1, -2)
else:
shape = x.shape[:-1] + (x.shape[-1] // group_size, )
x_s = torch.empty(shape, device=x.device, dtype=torch.float32)
BLOCK = triton.next_power_of_2(N)
# heuristics for number of warps
num_warps = min(max(BLOCK // 256, 1), 8)
num_stages = 1
if column_major_scales:
_per_token_group_quant_fp8_colmajor[(M, )](
x,
x_q,
x_s,
group_size,
x.shape[1],
x.stride(0),
x_s.stride(1),
eps,
fp8_min=fp8_min,
fp8_max=fp8_max,
BLOCK=BLOCK,
num_warps=num_warps,
num_stages=num_stages,
)
else:
_per_token_group_quant_fp8[(M, )](
x,
x_q,
x_s,
group_size,
x.shape[1],
x.stride(0),
eps,
fp8_min=fp8_min,
fp8_max=fp8_max,
BLOCK=BLOCK,
num_warps=num_warps,
num_stages=num_stages,
)
return x_q, x_s
@triton.jit
def _w8a8_block_fp8_matmul(
# Pointers to inputs and output
A,
B,
C,
As,
Bs,
# Shape for matmul
M,
N,
K,
# Block size for block-wise quantization
group_n,
group_k,
# Stride for inputs and output
stride_am,
stride_ak,
stride_bk,
stride_bn,
stride_cm,
stride_cn,
stride_As_m,
stride_As_k,
stride_Bs_k,
stride_Bs_n,
# Meta-parameters
BLOCK_SIZE_M: tl.constexpr,
BLOCK_SIZE_N: tl.constexpr,
BLOCK_SIZE_K: tl.constexpr,
GROUP_SIZE_M: tl.constexpr,
):
"""Triton-accelerated function used to perform linear operations (dot
product) on input tensors `A` and `B` with block-wise quantization, and
store the result in output tensor `C`.
"""
pid = tl.program_id(axis=0)
num_pid_m = tl.cdiv(M, BLOCK_SIZE_M)
num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
num_pid_in_group = GROUP_SIZE_M * num_pid_n
group_id = pid // num_pid_in_group
first_pid_m = group_id * GROUP_SIZE_M
group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
pid_m = first_pid_m + (pid % group_size_m)
pid_n = (pid % num_pid_in_group) // group_size_m
offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
offs_k = tl.arange(0, BLOCK_SIZE_K)
a_ptrs = A + (offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak)
b_ptrs = B + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn)
As_ptrs = As + offs_am * stride_As_m
offs_bsn = offs_bn // group_n
Bs_ptrs = Bs + offs_bsn * stride_Bs_n
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
a = tl.load(a_ptrs,
mask=offs_k[None, :] < K - k * BLOCK_SIZE_K,
other=0.0)
b = tl.load(b_ptrs,
mask=offs_k[:, None] < K - k * BLOCK_SIZE_K,
other=0.0)
k_start = k * BLOCK_SIZE_K
offs_ks = k_start // group_k
a_s = tl.load(As_ptrs + offs_ks * stride_As_k)
b_s = tl.load(Bs_ptrs + offs_ks * stride_Bs_k)
accumulator += tl.dot(a, b) * a_s[:, None] * b_s[None, :]
a_ptrs += BLOCK_SIZE_K * stride_ak
b_ptrs += BLOCK_SIZE_K * stride_bk
if C.dtype.element_ty == tl.bfloat16:
c = accumulator.to(tl.bfloat16)
elif C.dtype.element_ty == tl.float16:
c = accumulator.to(tl.float16)
else:
c = accumulator.to(tl.float32)
offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
c_ptrs = C + stride_cm * offs_cm[:, None] + stride_cn * offs_cn[None, :]
c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
tl.store(c_ptrs, c, mask=c_mask)
@functools.lru_cache
def get_w8a8_block_fp8_configs(N: int, K: int, block_n: int,
block_k: int) -> Optional[dict[int, Any]]:
"""
Return optimized configurations for the w8a8 block fp8 kernel.
The return value will be a dictionary that maps an irregular grid of
batch sizes to configurations of the w8a8 block fp8 kernel. To evaluate the
kernel on a given batch size bs, the closest batch size in the grid should
be picked and the associated configuration chosen to invoke the kernel.
"""
# First look up if an optimized configuration is available in the configs
# directory
device_name = current_platform.get_device_name().replace(" ", "_")
json_file_name = f"N={N},K={K},device_name={device_name},dtype=fp8_w8a8,block_shape=[{block_n},{block_k}].json" # noqa: E501
config_file_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name)
if os.path.exists(config_file_path):
with open(config_file_path) as f:
logger.info(
"Using configuration from %s for W8A8 Block FP8 kernel.",
config_file_path,
)
# If a configuration has been found, return it
return {int(key): val for key, val in json.load(f).items()}
# If no optimized configuration is available, we will use the default
# configuration
logger.warning(
"Using default W8A8 Block FP8 kernel config. Performance might "
"be sub-optimal! Config file not found at %s",
config_file_path,
)
return None
def hipblaslt_w8a8_block_fp8_matmul(
......@@ -605,88 +285,4 @@ def hipblaslt_w8a8_block_fp8_matmul(
enum_block_size, None)
return d
def w8a8_block_fp8_matmul(
A: torch.Tensor,
B: torch.Tensor,
As: torch.Tensor,
Bs: torch.Tensor,
block_size: list[int],
output_dtype: torch.dtype = torch.float16,
) -> torch.Tensor:
"""This function performs matrix multiplication with block-wise
quantization.
It takes two input tensors `A` and `B` with scales `As` and `Bs`.
The output is returned in the specified `output_dtype`.
Args:
A: The input tensor, e.g., activation.
B: The input tensor, e.g., weight.
As: The per-token-group quantization scale for `A`.
Bs: The per-block quantization scale for `B`.
block_size: The block size for per-block quantization. It should
be 2-dim, e.g., [128, 128].
output_dytpe: The dtype of the returned tensor.
Returns:
torch.Tensor: The result of matmul.
"""
assert len(block_size) == 2
block_n, block_k = block_size[0], block_size[1]
assert A.shape[-1] == B.shape[-1]
assert A.shape[:-1] == As.shape[:-1] and A.is_contiguous()
assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1]
M = A.numel() // A.shape[-1]
assert B.ndim == 2 and Bs.ndim == 2
N, K = B.shape
assert triton.cdiv(N, block_n) == Bs.shape[0]
assert triton.cdiv(K, block_k) == Bs.shape[1]
C_shape = A.shape[:-1] + (N, )
C = A.new_empty(C_shape, dtype=output_dtype)
configs = get_w8a8_block_fp8_configs(N, K, block_size[0], block_size[1])
if configs:
# Get the optimal config if there is one
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
else:
# Default config
# Block-wise quant: BLOCK_SIZE_N must be divisible by block_size[0]
# BLOCK_SIZE_K must be divisible by block_size[1]
config = {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": block_size[0],
"BLOCK_SIZE_K": block_size[1],
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2,
}
def grid(META):
return (triton.cdiv(M, META["BLOCK_SIZE_M"]) *
triton.cdiv(N, META["BLOCK_SIZE_N"]), )
_w8a8_block_fp8_matmul[grid](
A,
B,
C,
As,
Bs,
M,
N,
K,
block_n,
block_k,
A.stride(-2),
A.stride(-1),
B.stride(1),
B.stride(0),
C.stride(-2),
C.stride(-1),
As.stride(-2),
As.stride(-1),
Bs.stride(1),
Bs.stride(0),
**config,
)
return C
vllm/utils/__init__.py
View file @ 5b7f2c7d
......@@ -1958,7 +1958,7 @@ class W8a8GetCacheJSON:
self.moe_weight_shapes=[]
arch_name = torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0]
arch_cu = torch.cuda.get_device_properties(torch.cuda.current_device()).multi_processor_count
self.cache_json_data = {}
device_name =arch_name+'_'+str(arch_cu)+'cu'
self.device_name=device_name
self.topk=1
......@@ -2060,19 +2060,27 @@ class W8a8GetCacheJSON:
return self.triton_json_dir+f"/linear_{n}_{k}_block[{block_n},{block_k}]_{self.device_name}.json"
def get_moeint8json_name(self,E,N1,N2,K,TOPK,
block_size:Optional[list]=None,use_int4_w4a8:Optional[bool]=False):
block_size:Optional[list]=None,use_int4_w4a8:Optional[bool]=False,use_int8_w8a8:Optional[bool]=False):
if use_int4_w4a8:
if block_size is not None:
return self.triton_json_dir+f"/MOE_W4A8INT8[{block_size[0]},{block_size[1]}]_E={E}_N1={N1}_N2={N2}_K={K}_TOPK{TOPK}_{self.device_name}.json"
else:
return self.triton_json_dir+f"/MOE_W4A8INT8_E={E}_N1={N1}_N2={N2}_K={K}_TOPK{TOPK}_{self.device_name}.json"
return self.triton_json_dir+f"/MOE_W4A8INT8_E={E}_N1={N1}_N2={N2}_K={K}_TOPK{TOPK}_{self.device_name}.json"
elif use_int8_w8a8:
if block_size is not None:
return self.triton_json_dir + f"/MOE_BLOCKINT8[{block_size[0]},{block_size[1]}]_E={E}_N1={N1}_N2={N2}_K={K}_TOPK{TOPK}_{self.device_name}.json"
else:
return self.triton_json_dir + f"/MOE_W8A8INT8_E={E}_N1={N1}_N2={N2}_K={K}_TOPK{TOPK}_{self.device_name}.json"
else:
if block_size is not None:
return self.triton_json_dir+f"/MOE_BLOCKINT8[{block_size[0]},{block_size[1]}]_E={E}_N1={N1}_N2={N2}_K={K}_TOPK{TOPK}_{self.device_name}.json"
return self.triton_json_dir + f"/MOE_BLOCKFP8[{block_size[0]},{block_size[1]}]_E={E}_N1={N1}_N2={N2}_K={K}_TOPK{TOPK}_{self.device_name}.json"
else:
return self.triton_json_dir+f"/MOE_W8A8INT8_E={E}_N1={N1}_N2={N2}_K={K}_TOPK{TOPK}_{self.device_name}.json"
return self.triton_json_dir + f"/MOE_W8A8FP8_E={E}_N1={N1}_N2={N2}_K={K}_TOPK{TOPK}_{self.device_name}.json"
def get_moeint8_triton_cache(self,file_path,E,N1,N2,K,TOPK):
if file_path in self.cache_json_data:
# 直接返回缓存数据，避免重复读取
return self.cache_json_data[file_path]
cache_json_file=file_path
if os.path.exists(file_path):
......@@ -2088,7 +2096,7 @@ class W8a8GetCacheJSON:
for sub_key, sub_value in value.items():
configs_key= f"{sub_key}_{key}"
configs_dict[configs_key]=sub_value
self.cache_json_data[file_path] = configs_dict
return configs_dict
# Adapted from: https://stackoverflow.com/a/47212782/5082708
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment