[Kernel/Quant] Remove the original marlin format and qqq (#23204)
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
@@ -1,126 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
import numpy
|
||||
import torch
|
||||
|
||||
from .marlin_utils_test import marlin_permute_weights
|
||||
from .quant_utils import get_pack_factor, qqq_quantize_weights
|
||||
|
||||
|
||||
def marlin_qqq_weights(q_w, size_k, size_n, num_bits, perm, group_size):
|
||||
# Permute
|
||||
q_w = marlin_permute_weights(q_w, size_k, size_n, perm)
|
||||
|
||||
# Pack
|
||||
pack_factor = get_pack_factor(num_bits)
|
||||
orig_device = q_w.device
|
||||
|
||||
q_w = q_w.cpu().numpy().astype(numpy.uint32)
|
||||
|
||||
q_packed = numpy.zeros((q_w.shape[0], q_w.shape[1] // pack_factor),
|
||||
dtype=numpy.uint32)
|
||||
if group_size == size_k:
|
||||
for i in range(pack_factor):
|
||||
q_packed |= (q_w[:, i::pack_factor] & 0xF) << num_bits * i
|
||||
else:
|
||||
for i in range(pack_factor):
|
||||
q_packed |= q_w[:, i::pack_factor] << num_bits * i
|
||||
|
||||
q_packed = torch.from_numpy(q_packed.astype(numpy.int32)).to(orig_device)
|
||||
|
||||
return q_packed
|
||||
|
||||
|
||||
def get_qqq_scale_perms():
|
||||
scale_perm: list[int] = []
|
||||
for i in range(8):
|
||||
scale_perm.extend([i + 8 * j for j in range(8)])
|
||||
scale_perm_single: list[int] = []
|
||||
for i in range(4):
|
||||
scale_perm_single.extend(
|
||||
[2 * i + j for j in [0, 1, 8, 9, 16, 17, 24, 25]])
|
||||
return scale_perm, scale_perm_single
|
||||
|
||||
|
||||
# NOTE(HandH1998): QQQ employs different perms for per-group and per-channel weight quantization. # noqa: E501
|
||||
def get_qqq_weight_perm(num_bits: int, quant_type: str):
|
||||
perm_list: list[int] = []
|
||||
for i in range(32):
|
||||
perm1: list[int] = []
|
||||
col = i // 4
|
||||
for block in [0, 1]:
|
||||
for row in [
|
||||
4 * (i % 4),
|
||||
4 * (i % 4) + 1,
|
||||
4 * (i % 4) + 2,
|
||||
4 * (i % 4) + 3,
|
||||
]:
|
||||
perm1.append(16 * row + col + 8 * block)
|
||||
for j in range(4):
|
||||
perm_list.extend([p + 256 * j for p in perm1])
|
||||
|
||||
perm = numpy.array(perm_list)
|
||||
|
||||
assert quant_type in ["per-channel",
|
||||
"per-group"], "not supported quantization type"
|
||||
if num_bits == 4:
|
||||
if quant_type == "per-channel":
|
||||
interleave = numpy.array([4, 0, 5, 1, 6, 2, 7, 3])
|
||||
else:
|
||||
interleave = numpy.array([0, 2, 4, 6, 1, 3, 5, 7])
|
||||
else:
|
||||
raise Exception("num_bits must be 4, got {}".format(num_bits))
|
||||
|
||||
perm = perm.reshape((-1, len(interleave)))[:, interleave].ravel()
|
||||
perm = torch.from_numpy(perm)
|
||||
return perm
|
||||
|
||||
|
||||
def marlin_qqq_permute_scales(s_group, s_channel, size_k, size_n, group_size):
|
||||
scale_perm, scale_perm_single = get_qqq_scale_perms()
|
||||
if group_size < size_k and group_size != -1:
|
||||
s_group = s_group.reshape((-1, len(scale_perm)))[:, scale_perm]
|
||||
s_channel = s_channel.reshape(
|
||||
(-1, len(scale_perm_single)))[:, scale_perm_single]
|
||||
s_group = s_group.reshape((-1, size_n)).contiguous()
|
||||
else:
|
||||
s_channel = s_channel.reshape(
|
||||
(-1, len(scale_perm_single)))[:, scale_perm_single]
|
||||
s_channel = s_channel.reshape((-1, size_n)).contiguous()
|
||||
|
||||
return s_group, s_channel
|
||||
|
||||
|
||||
def marlin_qqq_quantize(
|
||||
w: torch.Tensor,
|
||||
num_bits: int,
|
||||
group_size: int,
|
||||
):
|
||||
size_k, size_n = w.shape
|
||||
|
||||
# Normalize group_size
|
||||
if group_size == -1:
|
||||
group_size = size_k
|
||||
assert group_size <= size_k
|
||||
quant_type = "per-channel" if group_size == size_k else "per-group"
|
||||
|
||||
# Quantize
|
||||
w_ref, q_w, s_group, s_channel = qqq_quantize_weights(
|
||||
w, num_bits, group_size)
|
||||
|
||||
# Reformat to marlin_qqq
|
||||
weight_perm = get_qqq_weight_perm(num_bits, quant_type)
|
||||
marlin_qqq_q_w = marlin_qqq_weights(q_w, size_k, size_n, num_bits,
|
||||
weight_perm, group_size)
|
||||
marlin_qqq_s_group, marlin_qqq_s_channel = marlin_qqq_permute_scales(
|
||||
s_group, s_channel, size_k, size_n, group_size)
|
||||
|
||||
# Create result
|
||||
res_list = [
|
||||
w_ref, marlin_qqq_q_w, marlin_qqq_s_group, marlin_qqq_s_channel
|
||||
]
|
||||
for i in range(len(res_list)):
|
||||
res_list[i] = res_list[i].to(w.device)
|
||||
|
||||
return res_list
|
||||
@@ -9,8 +9,6 @@ import numpy
|
||||
import torch
|
||||
|
||||
from vllm._custom_ops import cutlass_scaled_mm_supports_fp4
|
||||
from vllm.model_executor.layers.quantization.qqq import (
|
||||
MARLIN_QQQ_SUPPORTED_NUM_BITS)
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.scalar_type import ScalarType, scalar_types
|
||||
|
||||
@@ -386,89 +384,6 @@ def gptq_quantize_weights(w: torch.Tensor,
|
||||
return w_ref, w_q, w_s, g_idx, rand_perm
|
||||
|
||||
|
||||
# QQQ employs different quant schemes for per-group and
|
||||
# per-channel quantization.
|
||||
def qqq_quantize_weights(w: torch.Tensor, num_bits: int, group_size: int):
|
||||
orig_device = w.device
|
||||
size_k, size_n = w.shape
|
||||
|
||||
assert w.is_floating_point(), "w must be float"
|
||||
assert num_bits in MARLIN_QQQ_SUPPORTED_NUM_BITS, \
|
||||
f"Unsupported num_bits = {num_bits}"
|
||||
assert group_size in SUPPORTED_GROUP_SIZES + [
|
||||
size_k
|
||||
], f"Unsupported groupsize = {group_size}"
|
||||
|
||||
if group_size == -1:
|
||||
group_size = size_k
|
||||
assert group_size <= size_k
|
||||
|
||||
if group_size < size_k:
|
||||
# Reshape to [groupsize, -1]
|
||||
w = w.reshape((-1, group_size, size_n))
|
||||
w = w.permute(1, 0, 2)
|
||||
w = w.reshape((group_size, -1))
|
||||
|
||||
max_q_val = 2**num_bits - 1
|
||||
half_q_val = (max_q_val + 1) // 2
|
||||
|
||||
# Compute scale for each group
|
||||
s_group = torch.max(torch.abs(w), 0, keepdim=True)[0]
|
||||
s_group *= 2 / max_q_val # 2 => symmetric
|
||||
|
||||
# Quantize
|
||||
q_w = torch.round(w / s_group).int()
|
||||
q_w += half_q_val
|
||||
q_w = torch.clamp(q_w, 0, max_q_val)
|
||||
# Compute ref (dequantized)
|
||||
w_ref = (q_w - half_q_val).half() * s_group
|
||||
|
||||
# Restore original shapes
|
||||
def reshape_w(w):
|
||||
w = w.reshape((group_size, -1, size_n))
|
||||
w = w.permute(1, 0, 2)
|
||||
w = w.reshape((size_k, size_n)).contiguous()
|
||||
return w
|
||||
|
||||
q_w = reshape_w(q_w)
|
||||
w_ref = reshape_w(w_ref)
|
||||
|
||||
# Compute int8 quantization scale for each channel
|
||||
s_channel = torch.max(torch.abs(w_ref), 0, keepdim=True)[0]
|
||||
s_channel /= 127.0
|
||||
t_int8 = (w_ref / s_channel).round().clamp(-128, 127).to(torch.int8)
|
||||
w_ref = t_int8.half() * s_channel
|
||||
s_channel = s_channel.reshape(1, -1).to(dtype=torch.float)
|
||||
|
||||
# Fuse scales
|
||||
s_group = (s_group.reshape(-1, size_n).contiguous() /
|
||||
s_channel).to(dtype=torch.half)
|
||||
else:
|
||||
max_q_val = 2**(num_bits - 1) - 1
|
||||
|
||||
# Compute scale for each channel
|
||||
s_channel = torch.max(torch.abs(w), 0, keepdim=True)[0]
|
||||
s_channel /= max_q_val
|
||||
|
||||
# Quantize
|
||||
q_w = torch.round(w / s_channel).int()
|
||||
q_w = torch.clamp(q_w, -max_q_val, max_q_val)
|
||||
# Compute ref (dequantized)
|
||||
w_ref = q_w.half() * s_channel
|
||||
|
||||
s_group = torch.tensor([], dtype=torch.half)
|
||||
# div 2 ** (8 - self.bits)) to offset right shift in unpacking
|
||||
s_channel /= (2**(8 - num_bits))
|
||||
s_channel = s_channel.reshape(-1, size_n).contiguous().to(torch.float)
|
||||
|
||||
return (
|
||||
w_ref.to(device=orig_device),
|
||||
q_w.to(device=orig_device),
|
||||
s_group.to(device=orig_device),
|
||||
s_channel.to(device=orig_device),
|
||||
)
|
||||
|
||||
|
||||
def sort_weights(q_w: torch.Tensor, g_idx: torch.Tensor):
|
||||
orig_device = q_w.device
|
||||
|
||||
|
||||
Reference in New Issue
Block a user