EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[AMD] Support a new flag to disable quant on parallelLinear layer if required (#11811)

Browse Source
This commit is contained in:
jacky.cheng
2025-10-23 10:16:15 +08:00
committed by GitHub
parent 6ade6a02d4
commit 99c92ff24b
2 changed files with 9 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
python/sglang/srt/environ.py
View File

@@ -158,6 +158,7 @@ class Envs:
# AMD & ROCm # AMD & ROCm
SGLANG_USE_AITER = EnvBool(False) SGLANG_USE_AITER = EnvBool(False)
SGLANG_ROCM_FUSED_DECODE_MLA = EnvBool(False) SGLANG_ROCM_FUSED_DECODE_MLA = EnvBool(False)
SGLANG_ROCM_DISABLE_LINEARQUANT = EnvBool(False)
# Quantization # Quantization
SGLANG_INT4_WEIGHT = EnvBool(False) SGLANG_INT4_WEIGHT = EnvBool(False)

9
python/sglang/srt/layers/linear.py
View File

@@ -32,7 +32,7 @@ from sglang.srt.layers.parameter import (
) )
from sglang.srt.layers.quantization.unquant import UnquantizedLinearMethod from sglang.srt.layers.quantization.unquant import UnquantizedLinearMethod
from sglang.srt.layers.utils import pad_or_narrow_weight from sglang.srt.layers.utils import pad_or_narrow_weight
from sglang.srt.utils import is_cpu, is_npu, set_weight_attrs from sglang.srt.utils import get_bool_env_var, is_cpu, is_hip, is_npu, set_weight_attrs
if TYPE_CHECKING: if TYPE_CHECKING:
from sglang.srt.layers.quantization.base_config import ( from sglang.srt.layers.quantization.base_config import (
@@ -40,6 +40,11 @@ if TYPE_CHECKING:
QuantizeMethodBase, QuantizeMethodBase,
) )
_is_hip = is_hip()
_disable_hip_linear_quant = _is_hip and get_bool_env_var(
"SGLANG_ROCM_DISABLE_LINEARQUANT"
)
logger = logging.getLogger(__name__) logger = logging.getLogger(__name__)
WEIGHT_LOADER_V2_SUPPORTED = [ WEIGHT_LOADER_V2_SUPPORTED = [
@@ -824,6 +829,7 @@ class QKVParallelLinear(ColumnParallelLinear):
self.num_kv_heads * self.head_size * tp_size, # v_proj self.num_kv_heads * self.head_size * tp_size, # v_proj
] ]
self.use_presharded_weights = load_presharded_attn self.use_presharded_weights = load_presharded_attn
quant_config = None if _disable_hip_linear_quant else quant_config
super().__init__( super().__init__(
input_size=input_size, input_size=input_size,
@@ -1225,6 +1231,7 @@ class RowParallelLinear(LinearBase):
tp_size: Optional[int] = None, tp_size: Optional[int] = None,
use_presharded_weights: bool = False, use_presharded_weights: bool = False,
): ):
quant_config = None if _disable_hip_linear_quant else quant_config
super().__init__( super().__init__(
input_size, output_size, skip_bias_add, params_dtype, quant_config, prefix input_size, output_size, skip_bias_add, params_dtype, quant_config, prefix
) )