Skip to content

GitLab

Commit 7ef78993 authored by gaoqiong's avatar gaoqiong
Browse files

增加w8a8 triton warmup代码

parent c56b26cd
Show whitespace changes
Inline Side-by-side
Showing with 1475 additions and 5 deletions
setup.py
View file @ 7ef78993
...@@ -543,7 +543,7 @@ if _build_custom_ops(): ...@@ -543,7 +543,7 @@ if _build_custom_ops():
ext_modules.append(CMakeExtension(name="vllm._C")) ext_modules.append(CMakeExtension(name="vllm._C"))
package_data = { package_data = {
"vllm": ["py.typed", "model_executor/layers/fused_moe/configs/*.json", "benchmarks/*.py"] "vllm": ["py.typed", "model_executor/layers/fused_moe/configs/*.json", "benchmarks/*.py","model_executor/layers/quantization/configs/*.json"]
} }
if envs.VLLM_USE_PRECOMPILED: if envs.VLLM_USE_PRECOMPILED:
ext_modules = [] ext_modules = []
......
vllm/_custom_ops.py
View file @ 7ef78993
import contextlib import contextlib
import functools import functools
from typing import List, Optional, Tuple, Union from typing import List, Optional, Tuple, Union,Type
import torch import torch
...@@ -11,6 +11,7 @@ from vllm.platforms import current_platform ...@@ -11,6 +11,7 @@ from vllm.platforms import current_platform
try: try:
from lmslim import quant_ops from lmslim import quant_ops
from lmslim import quant_tools
except Exception: except Exception:
print("INFO: Please install lmslim if you want to infer gptq or awq or w8a8 model.\n") print("INFO: Please install lmslim if you want to infer gptq or awq or w8a8 model.\n")
...@@ -909,6 +910,18 @@ def triton_scaled_mm(a: torch.Tensor, ...@@ -909,6 +910,18 @@ def triton_scaled_mm(a: torch.Tensor,
return quant_ops.triton_scaled_mm(a, b,scale_a,scale_b,out_dtype,bias) return quant_ops.triton_scaled_mm(a, b,scale_a,scale_b,out_dtype,bias)
def triton_int8_gemm_helper(m: int,
n: int,
k: int,
per_token_act_quant: bool,
per_out_channel_weight_quant: bool,
use_bias: bool,
out_dtype: Type[torch.dtype] = torch.float16,
device: str = "cuda",
best_config:Optional[list] = None):
return quant_tools.triton_int8_gemm_helper(m,n,k,per_token_act_quant,per_out_channel_weight_quant,use_bias,out_dtype,device,est_config)
def cutlass_scaled_mm_azp(a: torch.Tensor, def cutlass_scaled_mm_azp(a: torch.Tensor,
b: torch.Tensor, b: torch.Tensor,
scale_a: torch.Tensor, scale_a: torch.Tensor,
......
vllm/model_executor/layers/quantization/configs/w8a8/W8A8_12288_4096_K100_AI.json 0 → 100755
View file @ 7ef78993
{
"12288_4096": {
"1": {
"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
},
"2": {
"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
},
"3": {
"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
},
"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": 4,
"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": 4,
"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": 8,
"SPLIT_K": 1,
"num_stages": 1,
"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": 1,
"num_warps": 8
},
"9": {
"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
},
"10": {
"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
},
"11": {
"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
},
"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": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"14": {
"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
},
"15": {
"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
},
"16": {
"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
},
"17": {
"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
},
"18": {
"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
},
"19": {
"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
},
"20": {
"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
},
"21": {
"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
},
"22": {
"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
},
"23": {
"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": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"25": {
"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
},
"26": {
"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
},
"27": {
"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
},
"28": {
"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
},
"29": {
"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
},
"30": {
"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
},
"31": {
"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
},
"32": {
"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
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"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
},
"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": 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_22016_4096_K100_AI.json 0 → 100755
View file @ 7ef78993
{
"22016_4096": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 2
},
"2": {
"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
},
"3": {
"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
},
"4": {
"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
},
"5": {
"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
},
"6": {
"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
},
"7": {
"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
},
"8": {
"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
},
"9": {
"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
},
"10": {
"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
},
"11": {
"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
},
"12": {
"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
},
"13": {
"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
},
"14": {
"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
},
"15": {
"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
},
"16": {
"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
},
"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": 8
},
"18": {
"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
},
"19": {
"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
},
"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
},
"21": {
"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
},
"22": {
"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
},
"23": {
"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": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"25": {
"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
},
"26": {
"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
},
"27": {
"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
},
"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
},
"29": {
"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
},
"30": {
"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
},
"31": {
"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": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"64": {
"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
},
"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": 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": 8,
"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_4096_11008_K100_AI.json 0 → 100755
View file @ 7ef78993
{
"4096_11008": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"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": 256,
"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": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"17": {
"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
},
"18": {
"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
},
"19": {
"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
},
"20": {
"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
},
"21": {
"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
},
"22": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"23": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"25": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"26": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"27": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"29": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"30": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"31": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"256": {
"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
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"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": 64,
"GROUP_SIZE_M": 4,
"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": 4,
"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": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
vllm/model_executor/layers/quantization/configs/w8a8/W8A8_4096_4096_K100_AI.json 0 → 100755
View file @ 7ef78993
{
"4096_4096": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"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": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"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": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"17": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"18": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"19": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"21": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"22": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"23": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"25": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"26": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"27": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"29": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"30": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"31": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"64": {
"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": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"256": {
"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
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"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": 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": 4,
"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": 4,
"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": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
vllm/model_executor/models/llama.py
View file @ 7ef78993
...@@ -691,7 +691,7 @@ class LlamaForCausalLM(nn.Module, SupportsLoRA): ...@@ -691,7 +691,7 @@ class LlamaForCausalLM(nn.Module, SupportsLoRA):
m=int(key.split('_')[0]) m=int(key.split('_')[0])
n=int(key.split('_')[1]) n=int(key.split('_')[1])
k=int(key.split('_')[2]) k=int(key.split('_')[2])
ops._int8_gemm_helper(m=m,n=n,k=k,per_token_act_quant=True,per_out_channel_weight_quant=True,use_bias=False,best_config=value) ops.triton_int8_gemm_helper(m=m,n=n,k=k,per_token_act_quant=True,per_out_channel_weight_quant=True,use_bias=False,best_config=value)
# If this function is called, it should always initialize KV cache scale # If this function is called, it should always initialize KV cache scale
# factors (or else raise an exception). Thus, handled exceptions should # factors (or else raise an exception). Thus, handled exceptions should
......
vllm/utils.py
View file @ 7ef78993
...@@ -1345,7 +1345,8 @@ class W8a8GetCacheJSON: ...@@ -1345,7 +1345,8 @@ class W8a8GetCacheJSON:
return cls._instance return cls._instance
def _initialize(self): def _initialize(self):
self.triton_json_dir=(os.getenv('TRITON_JSON_DIR', './cache')) current_folder_path = os.path.dirname(os.path.abspath(__file__))
self.triton_json_dir=(os.getenv('TRITON_JSON_DIR', current_folder_path+'/model_executor/layers/quantization/configs/w8a8'))
self.triton_json_dict=[] self.triton_json_dict=[]
def getspec_config(self,configs_dict,M,N,K): def getspec_config(self,configs_dict,M,N,K):
...@@ -1418,5 +1419,5 @@ class W8a8GetCacheJSON: ...@@ -1418,5 +1419,5 @@ class W8a8GetCacheJSON:
def get_w8a8json_name(self,n,k): def get_w8a8json_name(self,n,k):
device_name = current_platform.get_device_name().replace(" ", "_") device_name = current_platform.get_device_name().replace(" ", "_")
return self.triton_json_dir+f"/W8A8_{n}_{k}_HCU{device_name}.json" return self.triton_json_dir+f"/W8A8_{n}_{k}_{device_name}.json"
\ 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