Skip to content

GitLab

Commit ca4ec0ce authored by lizhigong's avatar lizhigong
Browse files

Merge remote-tracking branch 'origin/v0.7.2-dev' into v0.7.2_zero_overhead

parents 0be169ad ae0ed592
Hide whitespace changes
Inline Side-by-side
Showing with 1230 additions and 5101 deletions
vllm/model_executor/layers/fused_moe/configs/E=64,N=256,device_name=K100_AI_nn.json 0 → 100644
View file @ ca4ec0ce
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 5,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4,
"num_ldmatrixes": 1
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 5,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"num_ldmatrixes": 1
}
}
vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=DCU_K100_AI_nn.json 0 → 100644
View file @ ca4ec0ce
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 4,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 3,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"48": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1536": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"3072": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
}
}
vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=K100_AI_nn.json 0 → 100644
View file @ ca4ec0ce
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 4,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 3,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"48": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1536": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"3072": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
}
}
vllm/model_executor/layers/fused_moe/fused_moe.py
View file @ ca4ec0ce
......@@ -14,11 +14,170 @@ from vllm import _custom_ops as ops
from vllm.logger import init_logger
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)
from vllm.platforms import current_platform
from vllm.utils import direct_register_custom_op
logger = init_logger(__name__)
@triton.jit
def fused_moe_kernel_awq(
# Pointers to matrices
a_ptr, # [4, 7168]
b_ptr, # [256, 512, 3584]
c_ptr, # (8, 8, 512)
b_scale_ptr, # (256, 512, 56)
b_zp_ptr, # (256, 256, 56)
topk_weights_ptr,
sorted_token_ids_ptr, # [0, 1, 2, 3, 4]
expert_ids_ptr,
num_tokens_post_padded_ptr,
# Matrix dimensions
N: tl.constexpr,
K: tl.constexpr,
EM, # pading后的总索引长度
num_valid_tokens, # 有效索引的上限
# 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, #1
stride_bn,
stride_cm,
stride_cn,
stride_bse,
stride_bsk,#1
stride_bsn,
stride_bze,
stride_bzk,
stride_bzn,
block_k_diviable: tl.constexpr,
group_size: tl.constexpr, # 128
# 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,
has_zp: tl.constexpr,
use_int4_w4a16: tl.constexpr,
use_int8_w8a16: tl.constexpr):
pid = tl.program_id(axis=0)
num_pid_m = tl.cdiv(EM, 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 % num_pid_in_group) % group_size_m)
pid_n = (pid % num_pid_in_group) // group_size_m
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
return
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id) # [block_m]
token_mask = offs_token < num_valid_tokens
offs_bn = (pid_n * BLOCK_SIZE_N +
tl.arange(0, BLOCK_SIZE_N)) % N # [block_n]
offs_k = tl.arange(0, BLOCK_SIZE_K) # 0, 1, 2, ...... , 127 # # [block_k]
offs_k2 = tl.arange(0, BLOCK_SIZE_K // 2) # 0, 1, 2, ...... , 127 # # [block_k]
a_ptrs = a_ptr + (offs_token[:, None] // top_k * stride_am +
offs_k[None, :] * stride_ak) # [block_m, block_k]
off_experts = tl.load(expert_ids_ptr + pid_m)
if use_int4_w4a16:
# [0, 1, 2, ...... , 126, 127] --> [0, 0, 1, 1 ...... , 63, 63]
# [128, 129, 130, ...... , 254, 255] --> [64, 64, 65, 65 ...... , 127, 127]
# b_ptrs = b_ptr + off_experts * stride_be + \
# (offs_k[:, None] // 2) * stride_bk + offs_bn[None, :] * stride_bn
b_ptrs = b_ptr + off_experts * stride_be + \
offs_bn[:, None] * stride_bn + (offs_k2[None, :]) * stride_bk
# tl.device_print("stride_bn",stride_bsn)>1
# tl.device_print("stride_bk",stride_bk)=1
b_shifter = (offs_k[:, None] % 2) * 4 # 0, 4
elif use_int8_w8a16:
b_ptrs = b_ptr + off_experts * stride_be + \
offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn
if not has_zp and use_int4_w4a16:
b_zp_num = 8
if not has_zp and use_int8_w8a16:
b_zp_num = 128
elif has_zp and use_int4_w4a16:
b_zp_shifter = (offs_bn[None, :] % 2) * 4 # 0, 4
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
if not block_k_diviable:
k_mask = offs_k[:, None] < K - k * BLOCK_SIZE_K
k_other = 0.0
else:
k_mask = None
k_other = None
a = tl.load(a_ptrs,
mask=token_mask[:, None] &
(offs_k[None, :] < K - k * BLOCK_SIZE_K),
other=0.0)
b = tl.load(b_ptrs)
if use_int4_w4a16:
b = tl.interleave(b, b)
b= b.trans()
b = (b >> b_shifter) & 0xF
b_scale_ptrs = b_scale_ptr + off_experts * stride_bse + \
offs_bn[None, :] * stride_bsk + \
((offs_k[:, None] + BLOCK_SIZE_K * k) // group_size) * stride_bsn
qzeros_scles = tl.load(b_scale_ptrs, mask=k_mask, other=k_other)
scales_int16 = tl.cast(qzeros_scles,tl.uint16)
b_scale = tl.cast(scales_int16,tl.float16,bitcast=True)
# tl.device_print("b_scale dequant",b_scale)
mid = qzeros_scles >> 16
# b_zp = tl.cast(mid,tl.float16,bitcast=False)
b_zp = tl.cast(mid,tl.float16)
# b_zp = tl.cast(zeros_int16,tl.float16,bitcast=False)
# tl.device_print("bzp",b_zp)
# We accumulate along the K dimension.
b = ((b - b_zp) * b_scale).to(tl.float16)
accumulator = tl.dot(a, b, acc=accumulator)
# Advance the ptrs to the next K block.
a_ptrs += BLOCK_SIZE_K * stride_ak
if use_int4_w4a16:
b_ptrs += (BLOCK_SIZE_K // 2) * stride_bk
else:
b_ptrs += BLOCK_SIZE_K * stride_bk
if MUL_ROUTED_WEIGHT:
moe_weight = tl.load(topk_weights_ptr + offs_token,
mask=token_mask,
other=0)
accumulator = accumulator * moe_weight[:, None]
accumulator = accumulator.to(compute_type)
# -----------------------------------------------------------
# Write back the block of the output
offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[
None, :]
c_mask = token_mask[:, None] & (offs_cn[None, :] < N)
tl.store(c_ptrs, accumulator, mask=c_mask)
@triton.jit
def fused_moe_kernel_gptq_awq(
......@@ -265,6 +424,7 @@ def fused_moe_kernel(
top_k: tl.constexpr,
compute_type: tl.constexpr,
use_fp8_w8a8: tl.constexpr,
use_int8_w8a8: tl.constexpr,
use_int8_w8a16: tl.constexpr):
"""
Implements the fused computation for a Mixture of Experts (MOE) using
......@@ -346,7 +506,7 @@ 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:
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
......@@ -376,7 +536,7 @@ 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
......@@ -402,7 +562,7 @@ 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:
......@@ -558,7 +718,7 @@ def moe_align_block_size_triton(
def moe_align_block_size(
topk_ids: torch.Tensor, block_size: int,
num_experts: int) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
num_experts: int, num_token: Optional[int] = None) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Aligns the token distribution across experts to be compatible with block
size for matrix multiplication.
......@@ -596,11 +756,18 @@ def moe_align_block_size(
- The padding ensures that the total number of tokens is now divisible
by block_size for proper block matrix operations.
"""
max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1)
sorted_ids = torch.empty((max_num_tokens_padded, ),
dtype=torch.int32,
device=topk_ids.device)
sorted_ids.fill_(topk_ids.numel())
if num_token:
if num_token < block_size:
max_num_tokens_padded = min(topk_ids.numel() * block_size, topk_ids.numel() + num_experts * (block_size - 1))
else:
max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1)
sorted_ids = torch.full((max_num_tokens_padded,), fill_value=topk_ids.numel(), dtype=torch.int32, device=topk_ids.device)
else:
max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1)
sorted_ids = torch.empty((max_num_tokens_padded, ),
dtype=torch.int32,
device=topk_ids.device)
sorted_ids.fill_(topk_ids.numel())
max_num_m_blocks = triton.cdiv(max_num_tokens_padded, block_size)
expert_ids = torch.empty((max_num_m_blocks, ),
dtype=torch.int32,
......@@ -709,6 +876,7 @@ 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,
block_shape: Optional[List[int]] = None,
......@@ -727,6 +895,19 @@ def invoke_fused_moe_kernel(A: torch.Tensor,
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:
A, A_scale = ops.scaled_int8_quant(A, A_scale)
else:
assert len(block_shape) == 2
block_n, 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
......@@ -749,44 +930,82 @@ def invoke_fused_moe_kernel(A: torch.Tensor,
block_shape is not None and block_shape[1] > 0:
assert B_scale is not None and B_scale.ndim == 3
assert B_zp is None or B_zp.ndim == 3
fused_moe_kernel_gptq_awq[grid](
A,
B,
C,
B_scale,
B_zp,
topk_weights,
sorted_token_ids,
expert_ids,
num_tokens_post_padded,
B.shape[1],
A.shape[1],
EM,
topk_ids.numel(),
A.stride(0),
A.stride(1),
B.stride(0),
B.stride(2),
B.stride(1),
C.stride(1),
C.stride(2),
B_scale.stride(0),
B_scale.stride(2),
B_scale.stride(1),
B_zp.stride(0) if B_zp is not None else 0,
B_zp.stride(2) if B_zp is not None else 0,
B_zp.stride(1) if B_zp is not None else 0,
block_k_diviable=A.shape[1] % config["BLOCK_SIZE_K"] == 0,
group_size=block_shape[1],
MUL_ROUTED_WEIGHT=mul_routed_weight,
top_k=top_k,
compute_type=compute_type,
has_zp=B_zp is not None,
use_int4_w4a16=use_int4_w4a16,
use_int8_w8a16=use_int8_w8a16,
**config,
)
if os.environ.get('AWQ_MOE_SZ') == '1':
fused_moe_kernel_awq[grid](
A,
B,
C,
B_scale,
B_zp,
topk_weights,
sorted_token_ids,
expert_ids,
num_tokens_post_padded,
B.shape[1],
A.shape[1],
EM,
topk_ids.numel(),
A.stride(0),
A.stride(1),
B.stride(0),
B.stride(2),
B.stride(1),
C.stride(1),
C.stride(2),
B_scale.stride(0),
B_scale.stride(2),
B_scale.stride(1),
B_zp.stride(0) if B_zp is not None else 0,
B_zp.stride(2) if B_zp is not None else 0,
B_zp.stride(1) if B_zp is not None else 0,
block_k_diviable=A.shape[1] % config["BLOCK_SIZE_K"] == 0,
group_size=block_shape[1],
MUL_ROUTED_WEIGHT=mul_routed_weight,
top_k=top_k,
compute_type=compute_type,
has_zp=B_zp is not None,
use_int4_w4a16=use_int4_w4a16,
use_int8_w8a16=use_int8_w8a16,
**config,
)
else:
fused_moe_kernel_gptq_awq[grid](
A,
B,
C,
B_scale,
B_zp,
topk_weights,
sorted_token_ids,
expert_ids,
num_tokens_post_padded,
B.shape[1],
A.shape[1],
EM,
topk_ids.numel(),
A.stride(0),
A.stride(1),
B.stride(0),
B.stride(2),
B.stride(1),
C.stride(1),
C.stride(2),
B_scale.stride(0),
B_scale.stride(2),
B_scale.stride(1),
B_zp.stride(0) if B_zp is not None else 0,
B_zp.stride(2) if B_zp is not None else 0,
B_zp.stride(1) if B_zp is not None else 0,
block_k_diviable=A.shape[1] % config["BLOCK_SIZE_K"] == 0,
group_size=block_shape[1],
MUL_ROUTED_WEIGHT=mul_routed_weight,
top_k=top_k,
compute_type=compute_type,
has_zp=B_zp is not None,
use_int4_w4a16=use_int4_w4a16,
use_int8_w8a16=use_int8_w8a16,
**config,
)
else:
fused_moe_kernel[grid](
......@@ -826,6 +1045,7 @@ 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,
**config,
)
......@@ -872,6 +1092,15 @@ def get_moe_configs(
config_file_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name)
if torch.cuda.get_device_properties(torch.cuda.current_device()).multi_processor_count == 120:
config_file_path_120 = config_file_path.replace(".json","_120.json")
if os.path.exists(config_file_path_120):
with open(config_file_path_120) as f:
logger.info("Using configuration from %s for MoE layer.",
config_file_path_120)
# If a configuration has been found, return it
return {int(key): val for key, val in json.load(f).items()}
if os.path.exists(config_file_path):
with open(config_file_path) as f:
logger.info("Using configuration from %s for MoE layer.",
......@@ -1060,9 +1289,12 @@ def grouped_topk(hidden_states: torch.Tensor,
def get_config_dtype_str(dtype: torch.dtype,
use_int4_w4a16: Optional[bool] = False,
use_int8_w8a16: Optional[bool] = False,
use_fp8_w8a8: Optional[bool] = False):
use_fp8_w8a8: Optional[bool] = False,
use_int8_w8a8: Optional[bool] = False):
if use_fp8_w8a8:
return "fp8_w8a8"
elif use_int8_w8a8:
return "int8_w8a8"
elif use_int8_w8a16:
return "int8_w8a16"
elif use_int4_w4a16:
......@@ -1080,6 +1312,7 @@ def inplace_fused_experts(hidden_states: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
use_int4_w4a16: bool = False,
w1_scale: Optional[torch.Tensor] = None,
......@@ -1094,7 +1327,7 @@ def inplace_fused_experts(hidden_states: torch.Tensor,
start_expert: Optional[int] = -1,
end_expert: Optional[int] = -1) -> None:
fused_experts_impl(hidden_states, w1, w2, topk_weights, topk_ids, True,
use_fp8_w8a8, use_int8_w8a16, use_int4_w4a16, w1_scale,
use_fp8_w8a8,use_int8_w8a8, use_int8_w8a16, use_int4_w4a16, w1_scale,
w2_scale, w1_zp, w2_zp, a1_scale, a2_scale, block_shape,
use_nn_moe, moe_ep_size=moe_ep_size,
start_expert=start_expert, end_expert=end_expert)
......@@ -1107,6 +1340,7 @@ def inplace_fused_experts_fake(
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
use_int4_w4a16: bool = False,
w1_scale: Optional[torch.Tensor] = None,
......@@ -1138,6 +1372,7 @@ def outplace_fused_experts(
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
use_int4_w4a16: bool = False,
w1_scale: Optional[torch.Tensor] = None,
......@@ -1152,7 +1387,7 @@ def outplace_fused_experts(
start_expert: Optional[int] = -1,
end_expert: Optional[int] = -1) -> torch.Tensor:
return fused_experts_impl(hidden_states, w1, w2, topk_weights, topk_ids,
False, use_fp8_w8a8, use_int8_w8a16,
False, use_fp8_w8a8,use_int8_w8a8,use_int8_w8a16,
use_int4_w4a16, w1_scale, w2_scale, w1_zp, w2_zp,
a1_scale, a2_scale, block_shape,
use_nn_moe, moe_ep_size=moe_ep_size,
......@@ -1166,6 +1401,7 @@ def outplace_fused_experts_fake(
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
use_int4_w4a16: bool = False,
w1_scale: Optional[torch.Tensor] = None,
......@@ -1197,6 +1433,7 @@ def fused_experts(hidden_states: torch.Tensor,
topk_ids: torch.Tensor,
inplace: bool = False,
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
use_int4_w4a16: bool = False,
w1_scale: Optional[torch.Tensor] = None,
......@@ -1213,7 +1450,7 @@ def fused_experts(hidden_states: torch.Tensor,
if inplace:
torch.ops.vllm.inplace_fused_experts(hidden_states, w1, w2,
topk_weights, topk_ids,
use_fp8_w8a8, use_int8_w8a16,
use_fp8_w8a8,use_int8_w8a8,use_int8_w8a16,
use_int4_w4a16, w1_scale,
w2_scale, w1_zp, w2_zp, a1_scale,
a2_scale, block_shape,
......@@ -1224,7 +1461,7 @@ def fused_experts(hidden_states: torch.Tensor,
return hidden_states
else:
return torch.ops.vllm.outplace_fused_experts(
hidden_states, w1, w2, topk_weights, topk_ids, use_fp8_w8a8,
hidden_states, w1, w2, topk_weights, topk_ids, use_fp8_w8a8,use_int8_w8a8,
use_int8_w8a16, use_int4_w4a16, w1_scale, w2_scale, w1_zp, w2_zp,
a1_scale, a2_scale, block_shape,
use_nn_moe, moe_ep_size=moe_ep_size,
......@@ -1239,6 +1476,7 @@ def fused_experts_impl(hidden_states: torch.Tensor,
topk_ids: torch.Tensor,
inplace: bool = False,
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
use_int4_w4a16: bool = False,
w1_scale: Optional[torch.Tensor] = None,
......@@ -1279,6 +1517,7 @@ def fused_experts_impl(hidden_states: torch.Tensor,
CHUNK_SIZE = envs.VLLM_FUSED_MOE_CHUNK_SIZE
M = min(num_tokens, CHUNK_SIZE)
config_dtype = get_config_dtype_str(use_fp8_w8a8=use_fp8_w8a8,
use_int8_w8a8=use_int8_w8a8,
use_int8_w8a16=use_int8_w8a16,
use_int4_w4a16=use_int4_w4a16,
dtype=hidden_states.dtype)
......@@ -1346,8 +1585,12 @@ def fused_experts_impl(hidden_states: torch.Tensor,
curr_topk_weights = topk_weights[begin_chunk_idx:end_chunk_idx]
if moe_ep_size == 1:
sorted_token_ids, expert_ids, num_tokens_post_padded = (
moe_align_block_size(curr_topk_ids, config['BLOCK_SIZE_M'], E))
if use_int4_w4a16:
sorted_token_ids, expert_ids, num_tokens_post_padded = (
moe_align_block_size(curr_topk_ids, config['BLOCK_SIZE_M'], E, curr_hidden_states.shape[0]))
else:
sorted_token_ids, expert_ids, num_tokens_post_padded = (
moe_align_block_size(curr_topk_ids, config['BLOCK_SIZE_M'], E))
else:
sorted_token_ids, expert_ids, num_tokens_post_padded = (
moe_ep_align_block_size(curr_topk_ids, config['BLOCK_SIZE_M'], E,
......@@ -1369,6 +1612,7 @@ 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,
block_shape=block_shape,
......@@ -1393,6 +1637,7 @@ 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,
block_shape=block_shape,
......@@ -1416,6 +1661,7 @@ 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,
w1_scale: Optional[torch.Tensor] = None,
......@@ -1426,7 +1672,7 @@ def fused_moe(
a2_scale: Optional[torch.Tensor] = None,
block_shape: Optional[List[int]] = None,
use_nn_moe: Optional[bool] = False,
moe_ep_size: Optional[int] = None,
moe_ep_size: Optional[int] = 1,
start_expert: Optional[int] = None,
end_expert: Optional[int] = None,
) -> torch.Tensor:
......@@ -1492,6 +1738,7 @@ def fused_moe(
topk_ids,
inplace=inplace,
use_fp8_w8a8=use_fp8_w8a8,
use_int8_w8a8=use_int8_w8a8,
use_int8_w8a16=use_int8_w8a16,
use_int4_w4a16=use_int4_w4a16,
w1_scale=w1_scale,
......
vllm/model_executor/layers/fused_moe/layer.py
View file @ ca4ec0ce
......@@ -363,6 +363,9 @@ class FusedMoE(torch.nn.Module):
if (self.quant_method.__class__.__name__ ==
"CompressedTensorsWNA16MoEMethod"):
moe_quant_params["intermediate_size_full"] = intermediate_size
if (self.quant_method.__class__.__name__ in ("BlockInt8MoEMethod")):
moe_quant_params["intermediate_size"] = self.intermediate_size_per_partition
self.quant_method.create_weights(layer=self, **moe_quant_params)
......
vllm/model_executor/layers/linear.py
View file @ ca4ec0ce
......@@ -37,7 +37,7 @@ WEIGHT_LOADER_V2_SUPPORTED = [
"MarlinLinearMethod", "QQQLinearMethod", "GPTQMarlin24LinearMethod",
"TPUInt8LinearMethod", "GPTQLinearMethod", "FBGEMMFp8LinearMethod",
"ModelOptFp8LinearMethod", "IPEXAWQLinearMethod", "IPEXGPTQLinearMethod",
"HQQMarlinMethod", "QuarkLinearMethod"
"HQQMarlinMethod", "QuarkLinearMethod", "BlockInt8LinearMethod",
]
......@@ -664,9 +664,12 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
if isinstance(param, BlockQuantScaleParameter):
from vllm.model_executor.layers.quantization.fp8 import (
Fp8LinearMethod, Fp8MoEMethod)
from vllm.model_executor.layers.quantization.blockwise_int8 import (
BlockInt8LinearMethod, BlockInt8MoEMethod)
assert self.quant_method is not None
assert isinstance(self.quant_method,
(Fp8LinearMethod, Fp8MoEMethod))
(Fp8LinearMethod, Fp8MoEMethod, BlockInt8LinearMethod, BlockInt8MoEMethod))
weight_block_size = self.quant_method.quant_config.weight_block_size
assert weight_block_size is not None
block_n, _ = weight_block_size[0], weight_block_size[1]
......
vllm/model_executor/layers/quantization/__init__.py
View file @ ca4ec0ce
......@@ -29,7 +29,8 @@ QUANTIZATION_METHODS: List[str] = [
"neuron_quant",
"ipex",
"quark",
"moe_wna16"
"moe_wna16",
"blockwise_int8"
]
# The customized quantization methods which will be added to this dict.
......@@ -101,6 +102,7 @@ def get_quantization_config(quantization: str) -> Type[QuantizationConfig]:
from .neuron_quant import NeuronQuantConfig
from .qqq import QQQConfig
from .tpu_int8 import Int8TpuConfig
from .blockwise_int8 import BlockInt8Config
method_to_config: Dict[str, Type[QuantizationConfig]] = {
"aqlm": AQLMConfig,
......@@ -127,6 +129,7 @@ def get_quantization_config(quantization: str) -> Type[QuantizationConfig]:
"ipex": IPEXConfig,
"quark": QuarkConfig,
"moe_wna16": MoeWNA16Config,
"blockwise_int8": BlockInt8Config,
}
# Update the `method_to_config` with customized quantization methods.
method_to_config.update(_CUSTOMIZED_METHOD_TO_QUANT_CONFIG)
......
vllm/model_executor/layers/quantization/blockwise_int8.py 0 → 100755
View file @ ca4ec0ce
# SPDX-License-Identifier: Apache-2.0
# Adapted from https://github.com/sgl-project/sglang/pull/3730
import logging
from typing import Any, Callable, Dict, List, Optional
import torch
from torch.nn import Module
from vllm.model_executor.layers.quantization.utils.quant_utils import (
is_layer_skipped)
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.model_executor.layers.linear import (LinearBase, LinearMethodBase,
UnquantizedLinearMethod)
from vllm.model_executor.layers.fused_moe import (FusedMoE, FusedMoEMethodBase,
FusedMoeWeightScaleSupported)
from vllm.model_executor.parameter import (BlockQuantScaleParameter,
ModelWeightParameter,
PerTensorScaleParameter)
from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig, QuantizeMethodBase)
from vllm.model_executor.layers.quantization.utils.int8_utils import (
apply_w8a8_block_int8_linear)
from vllm.model_executor.utils import set_weight_attrs
ACTIVATION_SCHEMES = ["static", "dynamic"]
logger = logging.getLogger(__name__)
class BlockInt8Config(QuantizationConfig):
"""Config class for INT8."""
def __init__(
self,
is_checkpoint_int8_serialized: bool = False,
activation_scheme: str = "dynamic",
ignored_layers: Optional[List[str]] = None,
weight_block_size: Optional[List[int]] = None,
) -> None:
self.is_checkpoint_int8_serialized = is_checkpoint_int8_serialized
if is_checkpoint_int8_serialized:
logger.warning(
"Detected int8 checkpoint. Please note that the "
"format is experimental and subject to change."
)
if activation_scheme not in ACTIVATION_SCHEMES:
raise ValueError("Unsupported activation scheme"
f" {activation_scheme}")
self.activation_scheme = activation_scheme
self.ignored_layers = ignored_layers or []
if weight_block_size is not None:
if not is_checkpoint_int8_serialized:
raise ValueError(
f"The block-wise quantization only supports "
"int8-serialized checkpoint for now."
)
if len(weight_block_size) != 2:
raise ValueError(
f"The quantization block size of weight must have 2 "
"dimensions, but got {len(weight_block_size)} dimensions."
)
if activation_scheme != "dynamic":
raise ValueError(
f"The block-wise quantization only supports dynamic "
"activation scheme for now, but got "
"{activation_scheme} activation scheme."
)
self.weight_block_size = weight_block_size
@classmethod
def get_name(cls) -> str:
return "blockwise_int8"
@classmethod
def get_supported_act_dtypes(cls) -> List[torch.dtype]:
return [torch.bfloat16, torch.half]
@classmethod
def get_min_capability(cls) -> int:
return 80
@classmethod
def get_config_filenames(cls) -> List[str]:
return []
@classmethod
def from_config(cls, config: Dict[str, Any]) -> "BlockInt8Config":
quant_method = cls.get_from_keys(config, ["quant_method"])
is_checkpoint_int8_serialized = "int8" in quant_method
activation_scheme = cls.get_from_keys(config, ["activation_scheme"])
ignored_layers = cls.get_from_keys_or(config, ["ignored_layers"], None)
weight_block_size = cls.get_from_keys_or(config,
["weight_block_size"], None)
return cls(
is_checkpoint_int8_serialized=is_checkpoint_int8_serialized,
activation_scheme=activation_scheme,
ignored_layers=ignored_layers,
weight_block_size=weight_block_size,
)
def get_quant_method(
self, layer: torch.nn.Module, prefix: str
) -> Optional["QuantizeMethodBase"]:
if isinstance(layer, LinearBase):
if is_layer_skipped(prefix, self.ignored_layers):
return UnquantizedLinearMethod()
return BlockInt8LinearMethod(self)
elif isinstance(layer, FusedMoE):
return BlockInt8MoEMethod(self)
return None
def get_scaled_act_names(self) -> List[str]:
return []
class BlockInt8LinearMethod(LinearMethodBase):
"""Linear method for INT8.
Supports loading INT8 checkpoints with static weight scale and
dynamic activation scale.
Limitations:
Only support block-wise int8 quantization and int8 checkpoint
Args:
quant_config: The quantization config.
"""
def __init__(self, quant_config: BlockInt8Config):
self.quant_config = quant_config
assert self.quant_config.weight_block_size is not None
assert self.quant_config.is_checkpoint_int8_serialized
def create_weights(
self,
layer: torch.nn.Module,
input_size_per_partition: int,
output_partition_sizes: Optional[List[int]],
input_size: int,
output_size: int,
params_dtype: torch.dtype,
**extra_weight_attrs,
):
# assert output_partition_sizes is not None, (
# "output_partition_sizes must be provided for quantization")
output_size_per_partition = sum(output_partition_sizes)
weight_loader = extra_weight_attrs.get("weight_loader")
tp_size = get_tensor_model_parallel_world_size()
block_n, block_k = (
self.quant_config.weight_block_size[0],
self.quant_config.weight_block_size[1],
)
# Required by row parallel
if tp_size > 1 and input_size // input_size_per_partition == tp_size:
if input_size_per_partition % block_k != 0:
raise ValueError(
f"Weight input_size_per_partition = "
f"{input_size_per_partition} is not divisible by "
f"weight quantization block_k = {block_k}."
)
# Required by collum parallel or enabling merged weights
if (tp_size > 1 and output_size // output_size_per_partition == tp_size) or len(
output_partition_sizes
) > 1:
for output_partition_size in output_partition_sizes:
if output_partition_size % block_n != 0:
raise ValueError(
f"Weight output_partition_size = "
f"{output_partition_size} is not divisible by "
f"weight quantization block_n = {block_n}."
)
layer.logical_widths = output_partition_sizes
layer.input_size_per_partition = input_size_per_partition
layer.output_size_per_partition = output_size_per_partition
layer.orig_dtype = params_dtype
# WEIGHT
weight_dtype = (
torch.int8
if self.quant_config.is_checkpoint_int8_serialized
else params_dtype
)
weight = ModelWeightParameter(
data=torch.empty(
output_size_per_partition, input_size_per_partition, dtype=weight_dtype
),
input_dim=1,
output_dim=0,
weight_loader=weight_loader,
)
layer.register_parameter("weight", weight)
# WEIGHT SCALE
scale = BlockQuantScaleParameter(
data=torch.empty(
(output_size_per_partition + block_n - 1) // block_n,
(input_size_per_partition + block_k - 1) // block_k,
dtype=torch.float32,
),
input_dim=1,
output_dim=0,
weight_loader=weight_loader,
)
scale[:] = torch.finfo(torch.float32).min
layer.register_parameter("weight_scale_inv", scale)
# INPUT ACTIVATION SCALE
assert self.quant_config.activation_scheme == "dynamic"
layer.register_parameter("input_scale", None)
def process_weights_after_loading(self, layer: Module) -> None:
# Block quant doesn't need to process weights after loading
# Use torch Parameter to avoid cuda graph capturing issue
layer.weight = torch.nn.Parameter(layer.weight.data, requires_grad=False)
layer.weight_scale_inv = torch.nn.Parameter(
layer.weight_scale_inv.data, requires_grad=False
)
def apply(
self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: Optional[torch.Tensor] = None,
) -> torch.Tensor:
return apply_w8a8_block_int8_linear(
input=x,
weight=layer.weight,
block_size=self.quant_config.weight_block_size,
weight_scale=layer.weight_scale_inv,
input_scale=None,
bias=bias,
)
class BlockInt8MoEMethod:
"""MoE method for INT8.
Supports loading INT8 checkpoints with static weight scale and
dynamic activation scale.
Limitations:
Only support block-wise int8 quantization and int8 checkpoint
Args:
quant_config: The quantization config.
"""
def __new__(cls, *args, **kwargs):
from vllm.model_executor.layers.fused_moe import FusedMoE, FusedMoEMethodBase
if not hasattr(cls, "_initialized"):
original_init = cls.__init__
new_cls = type(
cls.__name__,
(FusedMoEMethodBase,),
{
"__init__": original_init,
**{k: v for k, v in cls.__dict__.items() if k != "__dict__"},
},
)
obj = super(new_cls, new_cls).__new__(new_cls)
obj.__init__(*args, **kwargs)
return obj
return super().__new__(cls)
def __init__(self, quant_config):
self.quant_config = quant_config
assert self.quant_config.weight_block_size is not None
assert self.quant_config.is_checkpoint_int8_serialized
def create_weights(
self,
layer: Module,
num_experts: int,
hidden_size: int,
intermediate_size: int,
params_dtype: torch.dtype,
**extra_weight_attrs,
):
from vllm.model_executor.layers.fused_moe import FusedMoeWeightScaleSupported
if self.quant_config.is_checkpoint_int8_serialized:
params_dtype = torch.int8
tp_size = get_tensor_model_parallel_world_size()
block_n, block_k = (
self.quant_config.weight_block_size[0],
self.quant_config.weight_block_size[1],
)
# NOTE(HandH1998): To ensure proper alignment of the block-wise quantization scales, the output_size of the weights for both the gate and up layers must be divisible by block_n.
# Required by collum parallel or enabling merged weights
if intermediate_size % block_n != 0:
raise ValueError(
f"The output_size of gate's and up's weight = "
f"{intermediate_size} is not divisible by "
f"weight quantization block_n = {block_n}."
)
if tp_size > 1:
# Required by row parallel
if intermediate_size % block_k != 0:
raise ValueError(
f"The input_size of down's weight = "
f"{intermediate_size} is not divisible by "
f"weight quantization block_k = {block_k}."
)
# WEIGHTS
w13_weight = torch.nn.Parameter(
torch.empty(
num_experts, 2 * intermediate_size, hidden_size, dtype=params_dtype
),
requires_grad=False,
)
layer.register_parameter("w13_weight", w13_weight)
set_weight_attrs(w13_weight, extra_weight_attrs)
w2_weight = torch.nn.Parameter(
torch.empty(
num_experts, hidden_size, intermediate_size, dtype=params_dtype
),
requires_grad=False,
)
layer.register_parameter("w2_weight", w2_weight)
set_weight_attrs(w2_weight, extra_weight_attrs)
# WEIGHT_SCALES
w13_weight_scale = torch.nn.Parameter(
torch.ones(
num_experts,
2 * ((intermediate_size + block_n - 1) // block_n),
(hidden_size + block_k - 1) // block_k,
dtype=torch.float32,
),
requires_grad=False,
)
w2_weight_scale = torch.nn.Parameter(
torch.ones(
num_experts,
(hidden_size + block_n - 1) // block_n,
(intermediate_size + block_k - 1) // block_k,
dtype=torch.float32,
),
requires_grad=False,
)
layer.register_parameter("w13_weight_scale_inv", w13_weight_scale)
layer.register_parameter("w2_weight_scale_inv", w2_weight_scale)
extra_weight_attrs.update(
{"quant_method": FusedMoeWeightScaleSupported.BLOCK.value}
)
set_weight_attrs(w13_weight_scale, extra_weight_attrs)
set_weight_attrs(w2_weight_scale, extra_weight_attrs)
# INPUT_SCALES
assert self.quant_config.activation_scheme == "dynamic"
layer.w13_input_scale = None
layer.w2_input_scale = None
def process_weights_after_loading(self, layer: Module) -> None:
# Block quant doesn't need to process weights after loading
return
def apply(
self,
layer: torch.nn.Module,
x: torch.Tensor,
router_logits: torch.Tensor,
top_k: int,
renormalize: bool,
use_grouped_topk: bool,
topk_group: Optional[int] = None,
use_nn_moe: Optional[bool] = False,
num_expert_group: Optional[int] = None,
custom_routing_function: Optional[Callable] = None,
scoring_func: str = "softmax",
e_score_correction_bias: Optional[torch.Tensor] = None,
moe_ep_size: Optional[int] = 1,
start_expert: Optional[int] = -1,
end_expert: Optional[int] = -1
) -> torch.Tensor:
from vllm.model_executor.layers.fused_moe import fused_experts
#print("===========fused_experts========================")
# Expert selection
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,
scoring_func=scoring_func,
e_score_correction_bias=e_score_correction_bias
)
# Expert fusion with INT8 quantization
return fused_experts(
x,
layer.w13_weight,
layer.w2_weight,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=True,
use_int8_w8a8=True,
w1_scale=(layer.w13_weight_scale_inv),
w2_scale=(layer.w2_weight_scale_inv),
a1_scale=layer.w13_input_scale,
a2_scale=layer.w2_input_scale,
block_shape=self.quant_config.weight_block_size,
use_nn_moe=use_nn_moe,
moe_ep_size=moe_ep_size,
start_expert=start_expert,
end_expert=end_expert
)
vllm/model_executor/layers/quantization/configs/w8a8/W8A8_12288_4096_K100_AI.json deleted 100755 → 0
View file @ 0be169ad
{
"12288_4096": {
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"36": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"44": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
vllm/model_executor/layers/quantization/configs/w8a8/W8A8_1280_8192_K100_AI.json deleted 100644 → 0
View file @ 0be169ad
{
"1280_8192": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"44": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"72": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 4
},
"80": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 4
},
"88": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 4
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 2,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"120": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 2
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 2,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 2,
"num_warps": 8
},
"512": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 2,
"num_warps": 8
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 2,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 8
}
}
}
\ No newline at end of file
vllm/model_executor/layers/quantization/configs/w8a8/W8A8_13824_5120_K100_AI.json deleted 100644 → 0
View file @ 0be169ad
{
"13824_5120": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"5": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"6": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"44": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
}
}
}
\ No newline at end of file
vllm/model_executor/layers/quantization/configs/w8a8/W8A8_14336_8192_K100_AI.json deleted 100644 → 0
View file @ 0be169ad
{
"14336_8192": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"3": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"5": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"6": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"7": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"9": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"10": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"11": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"12": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"13": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"17": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
vllm/model_executor/layers/quantization/configs/w8a8/W8A8_15360_5120_K100_AI.json deleted 100644 → 0
View file @ 0be169ad
{
"15360_5120": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
}
}
}
\ No newline at end of file
vllm/model_executor/layers/quantization/configs/w8a8/W8A8_22016_4096_K100_AI.json deleted 100755 → 0
View file @ 0be169ad
{
"22016_4096": {
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"36": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"60": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"5": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"6": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
vllm/model_executor/layers/quantization/configs/w8a8/W8A8_2560_8192_K100_AI.json deleted 100644 → 0
View file @ 0be169ad
{
"2560_8192": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"17": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"144": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"152": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"160": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"16384": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
vllm/model_executor/layers/quantization/configs/w8a8/W8A8_27648_5120_K100_AI.json deleted 100644 → 0
View file @ 0be169ad
{
"27648_5120": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
vllm/model_executor/layers/quantization/configs/w8a8/W8A8_28672_4096_K100_AI.json deleted 100644 → 0
View file @ 0be169ad
{
"28672_4096": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"3": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"5": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"6": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"7": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"9": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"10": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"11": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"44": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
}
}
}
\ No newline at end of file
vllm/model_executor/layers/quantization/configs/w8a8/W8A8_28672_8192_K100_AI.json deleted 100644 → 0
View file @ 0be169ad
{
"28672_8192": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"3": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"5": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"6": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"7": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"9": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"10": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"11": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"12": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"13": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"44": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"52": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"56": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"60": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
vllm/model_executor/layers/quantization/configs/w8a8/W8A8_32000_4096_K100_AI.json deleted 100644 → 0
View file @ 0be169ad
{
"32000_4096": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"3": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"5": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"6": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"7": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"9": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"11": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"44": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
}
}
}
\ No newline at end of file
vllm/model_executor/layers/quantization/configs/w8a8/W8A8_3584_18944_K100_AI.json deleted 100644 → 0
View file @ 0be169ad
{
"3584_18944": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 2,
"num_warps": 2
},
"6": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"56": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"60": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
}
}
}
\ No newline at end of file
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