[Kernel] Support W8A8 channel-wise weights and per-token activations in triton fused_moe_kernel (#16366)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2025-04-11 17:54:08 +00:00
committed by GitHub
parent 4d022cbc75
commit f41647ee6b
7 changed files with 1229 additions and 158 deletions
+181 -85
View File
@@ -16,7 +16,10 @@ from vllm.model_executor.layers.fused_moe.deep_gemm_moe import (
_valid_deep_gemm, deep_gemm_moe_fp8)
from vllm.model_executor.layers.fused_moe.moe_align_block_size import (
moe_align_block_size)
from vllm.model_executor.layers.fused_moe.utils import _fp8_quantize
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
per_token_group_quant_fp8)
from vllm.model_executor.layers.quantization.utils.int8_utils import (
per_token_group_quant_int8, per_token_quant_int8)
from vllm.platforms import current_platform
from vllm.utils import direct_register_custom_op
@@ -251,50 +254,53 @@ def fused_moe_kernel_gptq_awq(
@triton.jit
def fused_moe_kernel(
# Pointers to matrices
a_ptr,
b_ptr,
c_ptr,
a_scale_ptr,
b_scale_ptr,
topk_weights_ptr,
sorted_token_ids_ptr,
expert_ids_ptr,
num_tokens_post_padded_ptr,
# Matrix dimensions
N,
K,
EM,
num_valid_tokens,
# The stride variables represent how much to increase the ptr by when
# moving by 1 element in a particular dimension. E.g. `stride_am` is
# how much to increase `a_ptr` by to get the element one row down
# (A has M rows).
stride_am,
stride_ak,
stride_be,
stride_bk,
stride_bn,
stride_cm,
stride_cn,
stride_asm,
stride_ask,
stride_bse,
stride_bsk,
stride_bsn,
# Block size for block-wise quantization
group_n: tl.constexpr,
group_k: tl.constexpr,
# Meta-parameters
BLOCK_SIZE_M: tl.constexpr,
BLOCK_SIZE_N: tl.constexpr,
BLOCK_SIZE_K: tl.constexpr,
GROUP_SIZE_M: tl.constexpr,
MUL_ROUTED_WEIGHT: tl.constexpr,
top_k: tl.constexpr,
compute_type: tl.constexpr,
use_fp8_w8a8: tl.constexpr,
use_int8_w8a16: tl.constexpr):
# Pointers to matrices
a_ptr,
b_ptr,
c_ptr,
a_scale_ptr,
b_scale_ptr,
topk_weights_ptr,
sorted_token_ids_ptr,
expert_ids_ptr,
num_tokens_post_padded_ptr,
# Matrix dimensions
N,
K,
EM,
num_valid_tokens,
# The stride variables represent how much to increase the ptr by when
# moving by 1 element in a particular dimension. E.g. `stride_am` is
# how much to increase `a_ptr` by to get the element one row down
# (A has M rows).
stride_am,
stride_ak,
stride_be,
stride_bk,
stride_bn,
stride_cm,
stride_cn,
stride_asm,
stride_ask,
stride_bse,
stride_bsk,
stride_bsn,
# Block size for block-wise quantization
group_n: tl.constexpr,
group_k: tl.constexpr,
# Meta-parameters
BLOCK_SIZE_M: tl.constexpr,
BLOCK_SIZE_N: tl.constexpr,
BLOCK_SIZE_K: tl.constexpr,
GROUP_SIZE_M: tl.constexpr,
MUL_ROUTED_WEIGHT: tl.constexpr,
top_k: tl.constexpr,
compute_type: tl.constexpr,
use_fp8_w8a8: tl.constexpr,
use_int8_w8a8: tl.constexpr,
use_int8_w8a16: tl.constexpr,
per_channel_quant: tl.constexpr,
):
"""
Implements the fused computation for a Mixture of Experts (MOE) using
token and expert matrices.
@@ -371,12 +377,23 @@ def fused_moe_kernel(
None, :] * stride_bsn
b_scale = tl.load(b_scale_ptrs)
if use_fp8_w8a8:
if use_fp8_w8a8 or use_int8_w8a8:
# block-wise
if group_k > 0 and group_n > 0:
a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm
offs_bsn = offs_bn // group_n
b_scale_ptrs = (b_scale_ptr + off_experts * stride_bse +
offs_bsn * stride_bsn)
# channel-wise
elif per_channel_quant:
b_scale_ptrs = b_scale_ptr + off_experts * stride_bse + offs_bn[
None, :] * stride_bsn
b_scale = tl.load(b_scale_ptrs)
# Load per-token scale for activations
a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm
a_scale = tl.load(a_scale_ptrs, mask=token_mask, other=0.0)[:,
None]
# tensor-wise
else:
a_scale = tl.load(a_scale_ptr)
b_scale = tl.load(b_scale_ptr + off_experts)
@@ -400,7 +417,7 @@ def fused_moe_kernel(
# We accumulate along the K dimension.
if use_int8_w8a16:
accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)
elif use_fp8_w8a8:
elif use_fp8_w8a8 or use_int8_w8a8:
if group_k > 0 and group_n > 0:
k_start = k * BLOCK_SIZE_K
offs_ks = k_start // group_k
@@ -412,7 +429,11 @@ def fused_moe_kernel(
accumulator += tl.dot(a, b) * a_scale[:,
None] * b_scale[None, :]
else:
accumulator = tl.dot(a, b, acc=accumulator)
if use_fp8_w8a8:
# acc used to enable fp8_fast_accum
accumulator = tl.dot(a, b, acc=accumulator)
else:
accumulator += tl.dot(a, b)
else:
accumulator += tl.dot(a, b)
# Advance the ptrs to the next K block.
@@ -426,7 +447,7 @@ def fused_moe_kernel(
accumulator = accumulator * moe_weight[:, None]
if use_int8_w8a16:
accumulator = (accumulator * b_scale).to(compute_type)
elif use_fp8_w8a8:
elif use_fp8_w8a8 or use_int8_w8a8:
if group_k > 0 and group_n > 0:
accumulator = accumulator.to(compute_type)
else:
@@ -457,27 +478,15 @@ def invoke_fused_moe_kernel(A: torch.Tensor,
config: Dict[str, Any],
compute_type: tl.dtype,
use_fp8_w8a8: bool,
use_int8_w8a8: bool,
use_int8_w8a16: bool,
use_int4_w4a16: bool,
per_channel_quant: bool,
block_shape: Optional[List[int]] = None) -> None:
assert topk_weights is not None or not mul_routed_weight
assert topk_weights is None or topk_weights.stride(1) == 1
assert sorted_token_ids.stride(0) == 1
if use_fp8_w8a8:
assert B_scale is not None
assert (block_shape is None or triton.cdiv(B.shape[-2], block_shape[0])
== B_scale.shape[-2])
assert (block_shape is None or triton.cdiv(B.shape[-1], block_shape[1])
== B_scale.shape[-1])
elif use_int8_w8a16 or use_int4_w4a16:
assert B_scale is not None
assert block_shape is None or block_shape[0] == 0
else:
assert A_scale is None
assert B_scale is None
M = A.shape[0]
num_tokens = M * top_k
@@ -604,7 +613,9 @@ def invoke_fused_moe_kernel(A: torch.Tensor,
top_k=top_k,
compute_type=compute_type,
use_fp8_w8a8=use_fp8_w8a8,
use_int8_w8a8=use_int8_w8a8,
use_int8_w8a16=use_int8_w8a16,
per_channel_quant=per_channel_quant,
BLOCK_SIZE_K=BLOCK_SIZE_K,
**config,
)
@@ -956,8 +967,10 @@ def inplace_fused_experts(hidden_states: torch.Tensor,
activation: str = "silu",
apply_router_weight_on_input: bool = False,
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
use_int4_w4a16: bool = False,
per_channel_quant: bool = False,
global_num_experts: int = -1,
expert_map: Optional[torch.Tensor] = None,
w1_scale: Optional[torch.Tensor] = None,
@@ -969,9 +982,10 @@ def inplace_fused_experts(hidden_states: torch.Tensor,
block_shape: Optional[List[int]] = None) -> None:
fused_experts_impl(hidden_states, w1, w2, topk_weights, topk_ids, True,
activation, apply_router_weight_on_input, use_fp8_w8a8,
use_int8_w8a16, use_int4_w4a16, global_num_experts,
expert_map, w1_scale, w2_scale, w1_zp, w2_zp, a1_scale,
a2_scale, block_shape)
use_int8_w8a8, use_int8_w8a16, use_int4_w4a16,
per_channel_quant, global_num_experts, expert_map,
w1_scale, w2_scale, w1_zp, w2_zp, a1_scale, a2_scale,
block_shape)
def inplace_fused_experts_fake(
@@ -983,8 +997,10 @@ def inplace_fused_experts_fake(
activation: str = "silu",
apply_router_weight_on_input: bool = False,
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
use_int4_w4a16: bool = False,
per_channel_quant: bool = False,
global_num_experts: int = -1,
expert_map: Optional[torch.Tensor] = None,
w1_scale: Optional[torch.Tensor] = None,
@@ -1015,8 +1031,10 @@ def outplace_fused_experts(
activation: str = "silu",
apply_router_weight_on_input: bool = False,
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
use_int4_w4a16: bool = False,
per_channel_quant: bool = False,
global_num_experts: int = -1,
expert_map: Optional[torch.Tensor] = None,
w1_scale: Optional[torch.Tensor] = None,
@@ -1028,7 +1046,8 @@ def outplace_fused_experts(
block_shape: Optional[List[int]] = None) -> torch.Tensor:
return fused_experts_impl(hidden_states, w1, w2, topk_weights, topk_ids,
False, activation, apply_router_weight_on_input,
use_fp8_w8a8, use_int8_w8a16, use_int4_w4a16,
use_fp8_w8a8, use_int8_w8a8, use_int8_w8a16,
use_int4_w4a16, per_channel_quant,
global_num_experts, expert_map, w1_scale,
w2_scale, w1_zp, w2_zp, a1_scale, a2_scale,
block_shape)
@@ -1042,8 +1061,10 @@ def outplace_fused_experts_fake(
topk_ids: torch.Tensor,
activation: str = "silu",
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
use_int4_w4a16: bool = False,
per_channel_quant: bool = False,
global_num_experts: int = -1,
expert_map: Optional[torch.Tensor] = None,
w1_scale: Optional[torch.Tensor] = None,
@@ -1092,8 +1113,10 @@ def fused_experts(hidden_states: torch.Tensor,
activation: str = "silu",
apply_router_weight_on_input: bool = False,
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
use_int4_w4a16: bool = False,
per_channel_quant: bool = False,
global_num_experts: int = -1,
expert_map: Optional[torch.Tensor] = None,
w1_scale: Optional[torch.Tensor] = None,
@@ -1132,8 +1155,10 @@ def fused_experts(hidden_states: torch.Tensor,
activation=activation,
apply_router_weight_on_input=apply_router_weight_on_input,
use_fp8_w8a8=use_fp8_w8a8,
use_int8_w8a8=use_int8_w8a8,
use_int8_w8a16=use_int8_w8a16,
use_int4_w4a16=use_int4_w4a16,
per_channel_quant=per_channel_quant,
global_num_experts=global_num_experts,
expert_map=expert_map,
w1_scale=w1_scale,
@@ -1145,6 +1170,59 @@ def fused_experts(hidden_states: torch.Tensor,
block_shape=block_shape)
def moe_kernel_prepare_input(
A: torch.Tensor,
B: torch.Tensor,
A_scale: Optional[torch.Tensor],
B_scale: Optional[torch.Tensor],
use_fp8_w8a8: bool,
use_int8_w8a8: bool,
use_int8_w8a16: bool,
use_int4_w4a16: bool,
per_channel_quant: bool,
block_shape: Optional[List[int]] = None,
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
if use_fp8_w8a8:
assert B_scale is not None
if block_shape is None:
# If weights are per-channel (per_channel_quant=True), then
# activations apply per-token quantization. Otherwise, assume
# activation tensor-wise fp8 quantization, dynamic or static
A, A_scale = ops.scaled_fp8_quant(
A, A_scale, use_per_token_if_dynamic=per_channel_quant)
else:
# activation block-wise fp8 quantization
assert len(block_shape) == 2
_, block_k = block_shape[0], block_shape[1]
A, A_scale = per_token_group_quant_fp8(A, block_k)
assert triton.cdiv(A.shape[-1], block_k) == A_scale.shape[-1]
# assert triton.cdiv(B.shape[-2], block_n) == B_scale.shape[-2]
# assert triton.cdiv(B.shape[-1], block_k) == B_scale.shape[-1]
elif use_int8_w8a8:
assert B_scale is not None
if block_shape is None:
# activation channel-wise int8 quantization
assert (per_channel_quant
), "int8 quantization only supports block or channel-wise"
A, A_scale = per_token_quant_int8(A)
else:
# activation block-wise int8 quantization
assert len(block_shape) == 2
_, block_k = block_shape[0], block_shape[1]
A, A_scale = per_token_group_quant_int8(A, block_k)
assert triton.cdiv(A.shape[-1], block_k) == A_scale.shape[-1]
# assert triton.cdiv(B.shape[-2], block_n) == B_scale.shape[-2]
# assert triton.cdiv(B.shape[-1], block_k) == B_scale.shape[-1]
elif use_int8_w8a16 or use_int4_w4a16:
assert B_scale is not None
assert block_shape is None or block_shape[0] == 0
else:
assert A_scale is None
assert B_scale is None
return A, A_scale
def fused_experts_impl(hidden_states: torch.Tensor,
w1: torch.Tensor,
w2: torch.Tensor,
@@ -1154,8 +1232,10 @@ def fused_experts_impl(hidden_states: torch.Tensor,
activation: str = "silu",
apply_router_weight_on_input: bool = False,
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
use_int4_w4a16: bool = False,
per_channel_quant: bool = False,
global_num_experts: int = -1,
expert_map: Optional[torch.Tensor] = None,
w1_scale: Optional[torch.Tensor] = None,
@@ -1257,14 +1337,17 @@ def fused_experts_impl(hidden_states: torch.Tensor,
curr_topk_ids = topk_ids[begin_chunk_idx:end_chunk_idx]
curr_topk_weights = topk_weights[begin_chunk_idx:end_chunk_idx]
a1q_scale: Optional[torch.Tensor] = None
if use_fp8_w8a8:
qcurr_hidden_states, a1q_scale = _fp8_quantize(
curr_hidden_states, a1_scale, block_shape)
else:
qcurr_hidden_states = curr_hidden_states
a1q_scale = a1_scale
qcurr_hidden_states, qa1_scale = moe_kernel_prepare_input(
A=curr_hidden_states,
B=w1,
A_scale=a1_scale,
B_scale=w1_scale,
use_fp8_w8a8=use_fp8_w8a8,
use_int8_w8a8=use_int8_w8a8,
use_int8_w8a16=use_int8_w8a16,
use_int4_w4a16=use_int4_w4a16,
per_channel_quant=per_channel_quant,
block_shape=block_shape)
sorted_token_ids, expert_ids, num_tokens_post_padded = (
moe_align_block_size(curr_topk_ids, config['BLOCK_SIZE_M'],
@@ -1273,7 +1356,7 @@ def fused_experts_impl(hidden_states: torch.Tensor,
invoke_fused_moe_kernel(qcurr_hidden_states,
w1,
intermediate_cache1,
a1q_scale,
qa1_scale,
w1_scale,
w1_zp,
curr_topk_weights,
@@ -1285,8 +1368,10 @@ def fused_experts_impl(hidden_states: torch.Tensor,
config,
compute_type=compute_type,
use_fp8_w8a8=use_fp8_w8a8,
use_int8_w8a8=use_int8_w8a8,
use_int8_w8a16=use_int8_w8a16,
use_int4_w4a16=use_int4_w4a16,
per_channel_quant=per_channel_quant,
block_shape=block_shape)
if activation == "silu":
@@ -1298,19 +1383,22 @@ def fused_experts_impl(hidden_states: torch.Tensor,
else:
raise ValueError(f"Unsupported FusedMoe activation: {activation}")
a2q_scale: Optional[torch.Tensor] = None
if use_fp8_w8a8:
qintermediate_cache2, a2q_scale = _fp8_quantize(
intermediate_cache2, a2_scale, block_shape)
else:
qintermediate_cache2 = intermediate_cache2
a2q_scale = a2_scale
qintermediate_cache2, qa2_scale = moe_kernel_prepare_input(
A=intermediate_cache2,
B=w2,
A_scale=a2_scale,
B_scale=w2_scale,
use_fp8_w8a8=use_fp8_w8a8,
use_int8_w8a8=use_int8_w8a8,
use_int8_w8a16=use_int8_w8a16,
use_int4_w4a16=use_int4_w4a16,
per_channel_quant=per_channel_quant,
block_shape=block_shape)
invoke_fused_moe_kernel(qintermediate_cache2,
w2,
intermediate_cache3,
a2q_scale,
qa2_scale,
w2_scale,
w2_zp,
curr_topk_weights,
@@ -1322,8 +1410,10 @@ def fused_experts_impl(hidden_states: torch.Tensor,
config,
compute_type=compute_type,
use_fp8_w8a8=use_fp8_w8a8,
use_int8_w8a8=use_int8_w8a8,
use_int8_w8a16=use_int8_w8a16,
use_int4_w4a16=use_int4_w4a16,
per_channel_quant=per_channel_quant,
block_shape=block_shape)
ops.moe_sum(intermediate_cache3.view(*intermediate_cache3.shape),
@@ -1346,8 +1436,10 @@ def fused_moe(
topk_group: Optional[int] = None,
custom_routing_function: Optional[Callable] = None,
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
use_int4_w4a16: bool = False,
per_channel_quant: bool = False,
global_num_experts: int = -1,
expert_map: Optional[torch.Tensor] = None,
w1_scale: Optional[torch.Tensor] = None,
@@ -1380,6 +1472,8 @@ def fused_moe(
note: Deepseekv2 model uses grouped_topk
- use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
products for w1 and w2. Defaults to False.
- use_int8_w8a8 (bool): If True, use int8 arithmetic to compute the inner
products for w1 and w2. Defaults to False.
- use_int8_w8a16 (bool): If True, use matmul of int8 weight and bf16/fp16
activation to compute the inner products for w1 and w2.
Defaults to False.
@@ -1426,8 +1520,10 @@ def fused_moe(
inplace=inplace,
activation=activation,
use_fp8_w8a8=use_fp8_w8a8,
use_int8_w8a8=use_int8_w8a8,
use_int8_w8a16=use_int8_w8a16,
use_int4_w4a16=use_int4_w4a16,
per_channel_quant=per_channel_quant,
global_num_experts=global_num_experts,
expert_map=expert_map,
w1_scale=w1_scale,
@@ -0,0 +1,459 @@
# SPDX-License-Identifier: Apache-2.0
# Adapted from https://github.com/sgl-project/sglang/blob/4cb53ecd0cffceb6dee5c011a58f65997a86f151/python/sglang/srt/layers/quantization/int8_kernel.py
import functools
import json
import logging
import os
from typing import Any, Dict, List, Optional, Tuple
import torch
import triton
import triton.language as tl
from vllm.platforms import current_platform
logger = logging.getLogger(__name__)
def apply_w8a8_block_int8_linear(
input: torch.Tensor,
weight: torch.Tensor,
block_size: List[int],
weight_scale: torch.Tensor,
input_scale: Optional[torch.Tensor] = None,
bias: Optional[torch.Tensor] = None,
) -> torch.Tensor:
assert input_scale is None
# View input as 2D matrix for fp8 methods
input_2d = input.view(-1, input.shape[-1])
output_shape = [*input.shape[:-1], weight.shape[0]]
q_input, x_scale = per_token_group_quant_int8(input_2d, block_size[1])
output = w8a8_block_int8_matmul(q_input,
weight,
x_scale,
weight_scale,
block_size,
output_dtype=input.dtype)
if bias is not None:
output = output + bias
return output.to(dtype=input.dtype).view(*output_shape)
def input_to_int8(
x: torch.Tensor,
dtype: torch.dtype = torch.int8) -> Tuple[torch.Tensor, torch.Tensor]:
"""This function quantizes input values to int8 values with
tensor-wise quantization."""
iinfo = torch.iinfo(dtype)
min_val, max_val = x.aminmax()
amax = torch.maximum(min_val.abs(), max_val.abs()).clamp(min=1e-12)
int8_min, int8_max = iinfo.min, iinfo.max
scale = int8_max / amax
x_scl_sat = (x * scale).clamp(min=int8_min, max=int8_max)
return x_scl_sat.to(dtype).contiguous(), scale.float().reciprocal()
def block_dequant(
x_q_block: torch.Tensor,
x_s: torch.Tensor,
block_size: List[int],
) -> torch.Tensor:
"""This function conducts block-wise dequantization.
The inputs are block-wise quantization tensor `x_q_block`,
block-wise quantization scale and the block size.
The outputs are dequantized tensor.
"""
block_n, block_k = block_size[0], block_size[1]
n, k = x_q_block.shape
n_tiles = (n + block_n - 1) // block_n
k_tiles = (k + block_k - 1) // block_k
assert n_tiles == x_s.shape[0]
assert k_tiles == x_s.shape[1]
x_dq_block = x_q_block.to(torch.float32)
for i in range(k_tiles):
for j in range(n_tiles):
x_dq_block[
j * block_n:min((j + 1) * block_n, n),
i * block_k:min((i + 1) * block_k, k),
] *= x_s[j][i]
return x_dq_block
@triton.jit
def _per_token_quant_int8(
x_ptr,
xq_ptr,
scale_ptr,
stride_x,
stride_xq,
N,
BLOCK: tl.constexpr,
):
# Adapted from https://github.com/InternLM/lmdeploy/blob/086481ed84b59bee3b8e4274e5fc69620040c048/lmdeploy/pytorch/kernels/cuda/w8a8_triton_kernels.py#L282
row_id = tl.program_id(0)
cols = tl.arange(0, BLOCK)
mask = cols < N
x = tl.load(x_ptr + row_id * stride_x + cols, mask=mask,
other=0.0).to(tl.float32)
absmax = tl.maximum(tl.max(tl.abs(x)), 1e-10)
scale_x = absmax / 127
x_q = x * (127 / absmax)
x_q = tl.extra.cuda.libdevice.round(x_q).to(tl.int8)
tl.store(xq_ptr + row_id * stride_xq + cols, x_q, mask=mask)
tl.store(scale_ptr + row_id, scale_x)
def per_token_quant_int8(x):
M = x.numel() // x.shape[-1]
N = x.shape[-1]
x_q = torch.empty_like(x, device=x.device, dtype=torch.int8)
scales = torch.empty(x.shape[:-1] + (1, ),
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)
assert x.is_contiguous()
_per_token_quant_int8[(M, )](
x,
x_q,
scales,
stride_x=x.stride(-2),
stride_xq=x_q.stride(-2),
N=N,
BLOCK=BLOCK,
num_warps=num_warps,
num_stages=1,
)
return x_q, scales
@triton.jit
def _per_token_group_quant_int8(
# Pointers to inputs and output
y_ptr,
y_q_ptr,
y_s_ptr,
# Stride of input
y_stride,
# Columns of input
N,
# Avoid to divide zero
eps,
# Information for int8
int8_min,
int8_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 int8 values.
"""
# Map the program id to the row of X and Y it should compute.
g_id = tl.program_id(0)
y_ptr += g_id * y_stride
y_q_ptr += g_id * y_stride
y_s_ptr += g_id
cols = tl.arange(0, BLOCK) # N <= BLOCK
mask = cols < N
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 / int8_max
y_q = tl.clamp(y / y_s, int8_min, int8_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_int8(
x: torch.Tensor,
group_size: int,
eps: float = 1e-10,
dtype: torch.dtype = torch.int8,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""Function to perform per-token-group quantization on an input tensor `x`.
It converts the tensor values into signed int8 values and returns the
quantized tensor along with the scaling factor used for quantization.
Args:
x: The input tenosr 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.int8`
is supported for now.
Returns:
Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the
scaling factor for quantization.
"""
assert (x.shape[-1] % group_size == 0
), "the last dimension of `x` cannot be divisible by `group_size`"
assert x.is_contiguous(), "`x` is not contiguous"
iinfo = torch.iinfo(dtype)
int8_max = iinfo.max
int8_min = iinfo.min
x_q = torch.empty_like(x, device=x.device, dtype=dtype)
M = x.numel() // group_size
N = group_size
x_s = torch.empty(
x.shape[:-1] + (x.shape[-1] // group_size, ),
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
_per_token_group_quant_int8[(M, )](
x,
x_q,
x_s,
group_size,
N,
eps,
int8_min=int8_min,
int8_max=int8_max,
BLOCK=BLOCK,
num_warps=num_warps,
num_stages=num_stages,
)
return x_q, x_s
@triton.jit
def _w8a8_block_int8_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).to(tl.float32) * 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_int8_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=int8_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 INT8 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 INT8 kernel config. Performance might "
"be sub-optimal! Config file not found at %s"),
config_file_path,
)
return None
def w8a8_block_int8_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 B.is_contiguous() 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_int8_configs(N, K, block_size[0], block_size[1])
if configs:
# If an optimal configuration map has been found, look up the
# optimal config
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
else:
# Default config
# Block-wise quant: 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": 3,
}
def grid(META):
return (triton.cdiv(M, META["BLOCK_SIZE_M"]) *
triton.cdiv(N, META["BLOCK_SIZE_N"]), )
_w8a8_block_int8_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