Skip to content

GitLab

Commit 66b809cc authored by zhuwenwen's avatar zhuwenwen
Browse files

Merge tag 'v0.7.2' into v0.7.2-dev

parents 37b63c24 0408efc6
Hide whitespace changes
Inline Side-by-side
Showing with 641 additions and 24 deletions
vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json 0 → 100644
View file @ 66b809cc
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 5
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 5
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 5
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 5
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 5
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 3
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 5
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 5
},
"96": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 3
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 3
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
}
}
vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json 0 → 100644
View file @ 66b809cc
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 5
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 5
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 5
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
},
"96": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 5
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 3
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 2
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 2
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
}
}
vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json 0 → 100644
View file @ 66b809cc
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 5
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 5
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 3
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 5
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 5
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 2
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 2
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
}
}
vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json 0 → 100644
View file @ 66b809cc
{
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 5
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 2
},
"8": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 5
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2
},
"32": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 5
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 3
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"96": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 5
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 3
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 3
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
}
}
vllm/model_executor/layers/quantization/utils/fp8_utils.py
View file @ 66b809cc
# SPDX-License-Identifier: Apache-2.0
# Adapted from https://github.com/sgl-project/sglang/pull/2575 # Adapted from https://github.com/sgl-project/sglang/pull/2575
import functools import functools
import json import json
...@@ -13,7 +15,7 @@ from vllm.logger import init_logger ...@@ -13,7 +15,7 @@ from vllm.logger import init_logger
from vllm.model_executor.layers.quantization.utils.quant_utils import ( from vllm.model_executor.layers.quantization.utils.quant_utils import (
_normalize_quant_group_shape, scaled_dequantize) _normalize_quant_group_shape, scaled_dequantize)
from vllm.model_executor.layers.quantization.utils.w8a8_utils import ( from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
apply_fp8_linear) CUTLASS_BLOCK_FP8_SUPPORTED, CUTLASS_FP8_SUPPORTED, apply_fp8_linear)
from vllm.platforms import current_platform from vllm.platforms import current_platform
logger = init_logger(__name__) logger = init_logger(__name__)
...@@ -36,7 +38,7 @@ def apply_w8a8_block_fp8_linear( ...@@ -36,7 +38,7 @@ def apply_w8a8_block_fp8_linear(
weight_scale: torch.Tensor, weight_scale: torch.Tensor,
input_scale: Optional[torch.Tensor] = None, input_scale: Optional[torch.Tensor] = None,
bias: Optional[torch.Tensor] = None, bias: Optional[torch.Tensor] = None,
cutlass_block_fp8_supported: bool = True, cutlass_block_fp8_supported: bool = CUTLASS_BLOCK_FP8_SUPPORTED,
) -> torch.Tensor: ) -> torch.Tensor:
assert input_scale is None assert input_scale is None
# View input as 2D matrix for fp8 methods # View input as 2D matrix for fp8 methods
...@@ -45,6 +47,16 @@ def apply_w8a8_block_fp8_linear( ...@@ -45,6 +47,16 @@ def apply_w8a8_block_fp8_linear(
shape_supported_by_cutlass = (weight.shape[0] % 128 == 0 shape_supported_by_cutlass = (weight.shape[0] % 128 == 0
and weight.shape[1] % 128 == 0) and weight.shape[1] % 128 == 0)
if current_platform.is_rocm():
scale_a_shape = ((input_2d.shape[-1] // block_size[1], ) +
input_2d.shape[:-1])[::-1]
scale_b_shape = (weight_scale.view(-1, 1)
if weight_scale.dim() <= 1 else weight_scale.T).shape
ar, ac = scale_a_shape
br, bc = scale_b_shape
if (ac > 1 or bc > 1 or ar not in (1, input_2d.shape[0])
or br not in (1, weight.shape[0])):
shape_supported_by_cutlass = False
if cutlass_block_fp8_supported and shape_supported_by_cutlass: if cutlass_block_fp8_supported and shape_supported_by_cutlass:
q_input, x_scale = per_token_group_quant_fp8(input_2d, q_input, x_scale = per_token_group_quant_fp8(input_2d,
block_size[1], block_size[1],
...@@ -73,12 +85,14 @@ def apply_w8a8_block_fp8_linear( ...@@ -73,12 +85,14 @@ def apply_w8a8_block_fp8_linear(
# `apply_fp8_linear` # `apply_fp8_linear`
# NOTE(lucas): this is quite messy, we should think through this more formally # NOTE(lucas): this is quite messy, we should think through this more formally
def apply_fp8_linear_generic( def apply_fp8_linear_generic(
input: torch.Tensor, input: torch.Tensor,
weight: torch.Tensor, weight: torch.Tensor,
weight_scale: torch.Tensor, weight_scale: torch.Tensor,
input_group_shape: Tuple[int, int], input_group_shape: Tuple[int, int],
weight_group_shape: Tuple[int, int], weight_group_shape: Tuple[int, int],
input_scale: Optional[torch.Tensor] = None, # static scale if one input_scale: Optional[torch.Tensor] = None, # static scale if one
cutlass_fp8_supported: bool = CUTLASS_FP8_SUPPORTED,
cutlass_block_fp8_supported: bool = CUTLASS_BLOCK_FP8_SUPPORTED,
) -> torch.Tensor: ) -> torch.Tensor:
# View input as 2D matrix for fp8 methods # View input as 2D matrix for fp8 methods
input = input.view(-1, input.shape[-1]) input = input.view(-1, input.shape[-1])
...@@ -93,14 +107,18 @@ def apply_fp8_linear_generic( ...@@ -93,14 +107,18 @@ def apply_fp8_linear_generic(
if is_dim_blocked(0, weight.shape, weight_group_shape[0])\ if is_dim_blocked(0, weight.shape, weight_group_shape[0])\
and is_dim_blocked(1, weight.shape, weight_group_shape[1]) and\ and is_dim_blocked(1, weight.shape, weight_group_shape[1]) and\
input_group_shape == (1, weight_group_shape[1]): input_group_shape == (1, weight_group_shape[1]):
return apply_w8a8_block_fp8_linear(input, weight, return apply_w8a8_block_fp8_linear(
list(weight_group_shape), input,
weight_scale) weight,
list(weight_group_shape),
weight_scale,
cutlass_block_fp8_supported=cutlass_block_fp8_supported)
else: else:
# Despite having linear in the it doesn't conform to # Despite having linear in the it doesn't conform to
# `torch.nn.functional.linear` which is defined as `input @ weight.T` # `torch.nn.functional.linear` which is defined as `input @ weight.T`
# so we explicitly transpose the weight matrix here # so we explicitly transpose the weight matrix here
return apply_fp8_linear(input, weight.T, weight_scale.T, return apply_fp8_linear(input, weight.T, weight_scale.T,
cutlass_fp8_supported=cutlass_fp8_supported,
use_per_token_if_dynamic=\ use_per_token_if_dynamic=\
(input_group_shape == (1, input.shape[1]))) (input_group_shape == (1, input.shape[1])))
...@@ -405,7 +423,7 @@ def get_w8a8_block_fp8_configs(N: int, K: int, block_n: int, ...@@ -405,7 +423,7 @@ def get_w8a8_block_fp8_configs(N: int, K: int, block_n: int,
# First look up if an optimized configuration is available in the configs # First look up if an optimized configuration is available in the configs
# directory # directory
device_name = current_platform.get_device_name().replace(" ", "_") device_name = current_platform.get_device_name().replace(" ", "_")
json_file_name = f"N={N},K={K},device_name={device_name},dtype=fp8_w8a8,block_shape=[{block_n}, {block_k}].json" # noqa: E501 json_file_name = f"N={N},K={K},device_name={device_name},dtype=fp8_w8a8,block_shape=[{block_n},{block_k}].json" # noqa: E501
config_file_path = os.path.join( config_file_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name) os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name)
......
vllm/model_executor/layers/quantization/utils/layer_utils.py
View file @ 66b809cc
# SPDX-License-Identifier: Apache-2.0
from typing import Union from typing import Union
import torch import torch
......
vllm/model_executor/layers/quantization/utils/machete_utils.py
View file @ 66b809cc
# SPDX-License-Identifier: Apache-2.0
from typing import List, Optional, Tuple from typing import List, Optional, Tuple
import torch import torch
......
vllm/model_executor/layers/quantization/utils/marlin_utils.py
View file @ 66b809cc
# SPDX-License-Identifier: Apache-2.0
from typing import List, Optional, Tuple from typing import List, Optional, Tuple
import numpy import numpy
......
vllm/model_executor/layers/quantization/utils/marlin_utils_fp8.py
View file @ 66b809cc
# SPDX-License-Identifier: Apache-2.0
from typing import Optional from typing import Optional
import torch import torch
......
vllm/model_executor/layers/quantization/utils/marlin_utils_test.py
View file @ 66b809cc
# SPDX-License-Identifier: Apache-2.0
"""Utility functions used for tests and benchmarks""" """Utility functions used for tests and benchmarks"""
from typing import List, Optional from typing import List, Optional
......
vllm/model_executor/layers/quantization/utils/marlin_utils_test_24.py
View file @ 66b809cc
# SPDX-License-Identifier: Apache-2.0
"""Utility functions used for tests and benchmarks""" """Utility functions used for tests and benchmarks"""
import random import random
......
vllm/model_executor/layers/quantization/utils/marlin_utils_test_qqq.py
View file @ 66b809cc
# SPDX-License-Identifier: Apache-2.0
from typing import List from typing import List
import numpy import numpy
......
vllm/model_executor/layers/quantization/utils/quant_utils.py
View file @ 66b809cc
# SPDX-License-Identifier: Apache-2.0
"""This file is used for /tests and /benchmarks""" """This file is used for /tests and /benchmarks"""
from typing import List, Optional, Tuple from types import MappingProxyType
from typing import List, Mapping, Optional, Tuple
import numpy import numpy
import torch import torch
...@@ -11,14 +13,6 @@ from vllm.scalar_type import ScalarType, scalar_types ...@@ -11,14 +13,6 @@ from vllm.scalar_type import ScalarType, scalar_types
SUPPORTED_GPTQ_QUANT_TYPES = [scalar_types.uint4b8, scalar_types.uint8b128] SUPPORTED_GPTQ_QUANT_TYPES = [scalar_types.uint4b8, scalar_types.uint8b128]
SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128] SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128]
# Note: this is a hack. We should update each model to register the
# stacked params and get it from there instead in a future PR.
# fused_name: List[shard_name]
FUSED_LAYER_NAME_MAPPING = {
"qkv_proj": ["q_proj", "k_proj", "v_proj"],
"gate_up_proj": ["gate_proj", "up_proj"]
}
# Normalize the group_shape to the full extent for any dims that are -1 # Normalize the group_shape to the full extent for any dims that are -1
def _normalize_quant_group_shape(x: torch.Tensor, group_shape: Tuple[int, def _normalize_quant_group_shape(x: torch.Tensor, group_shape: Tuple[int,
...@@ -177,14 +171,23 @@ def unpack_quantized_values_into_int32(w_q: torch.Tensor, ...@@ -177,14 +171,23 @@ def unpack_quantized_values_into_int32(w_q: torch.Tensor,
return res.permute(inv_perm) return res.permute(inv_perm)
def is_layer_skipped(prefix: str, ignored_layers: List[str]) -> bool: def is_layer_skipped(
prefix: str,
ignored_layers: List[str],
fused_mapping: Mapping[str, List[str]] = MappingProxyType({})
) -> bool:
# prefix: model.layers.0.self_attn.q_proj # prefix: model.layers.0.self_attn.q_proj
# proj_name: q_proj # proj_name: q_proj
proj_name = prefix.split(".")[-1] proj_name = prefix.split(".")[-1]
if proj_name in FUSED_LAYER_NAME_MAPPING:
# Fused layers like gate_up_proj or qkv_proj will not be fused
# in the safetensors checkpoint. So, we convert the name
# from the fused version to unfused + check to make sure that
# each shard of the fused layer has the same scheme.
if proj_name in fused_mapping:
shard_prefixes = [ shard_prefixes = [
prefix.replace(proj_name, shard_proj_name) prefix.replace(proj_name, shard_proj_name)
for shard_proj_name in FUSED_LAYER_NAME_MAPPING[proj_name] for shard_proj_name in fused_mapping[proj_name]
] ]
is_skipped = None is_skipped = None
......
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