EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[Quantization] Support compressed tensors w8a8 static and w8a8 dynamic weight (#4036)

Browse Source 
### What this PR does / why we need it?

While using the LLM Compressor quantization tool from the VLLM community
to generate quantized weights, the VLLM Ascend engine needs to be
adapted to support the compressed tensors quantization format.

1. Add AscendCompressedTensorsConfig to replace CompressedTensorsConfig
in vllm.
2. Support CompressedTensorsW8A8 static weight.
- weight: per-channel, int8, symmetric; activation: per-tensor, int8,
symmetric.
4. Support CompressedTensorsW8A8Dynamic weight.
- weight: per-channel, int8, symmetric; activation: per-token, int8,
symmetric, dynamic.
5. Modify the override_quantization_method in AscendQuantConfig.

Co-authored-by: taoqun110 taoqun@huawei.com
Co-authored-by: chenxi-hh chen464822955@163.com

- vLLM version: v0.11.2

---------

Signed-off-by: LHXuuu <scut_xlh@163.com>
Signed-off-by: chenxi-hh <chen464822955@163.com>
Signed-off-by: chenxi-hh <32731611+chenxi-hh@users.noreply.github.com>
Co-authored-by: chenxi-hh <chen464822955@163.com>
Co-authored-by: chenxi-hh <32731611+chenxi-hh@users.noreply.github.com>
This commit is contained in:
LHXuuu
2025-11-28 14:09:39 +08:00
committed by GitHub
parent ab37a7d5ae
commit bdc66972db
18 changed files with 707 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
vllm_ascend/platform.py
View File

@@ -30,12 +30,13 @@ from vllm_ascend.ascend_config import (check_ascend_config, get_ascend_config,
init_ascend_config)
from vllm_ascend.torchair.utils import (check_torchair_cache_exist,
delete_torchair_cache_file)
from vllm_ascend.utils import (ASCEND_QUANTIZATION_METHOD, AscendDeviceType,
enable_sp, get_ascend_device_type, is_vl_model,
prefill_context_parallel_enable,
update_aclgraph_sizes,
update_cudagraph_capture_sizes,
update_default_aclgraph_sizes)
# isort: off
from vllm_ascend.utils import (
ASCEND_QUANTIZATION_METHOD, COMPRESSED_TENSORS_METHOD, AscendDeviceType,
enable_sp, get_ascend_device_type, is_vl_model,
prefill_context_parallel_enable, update_aclgraph_sizes,
update_cudagraph_capture_sizes, update_default_aclgraph_sizes)
if TYPE_CHECKING:
from vllm.config import ModelConfig, VllmConfig
@@ -56,7 +57,9 @@ class NPUPlatform(Platform):
device_control_env_var: str = "ASCEND_RT_VISIBLE_DEVICES"
dispatch_key: str = "PrivateUse1"
supported_quantization: list[str] = [ASCEND_QUANTIZATION_METHOD]
supported_quantization: list[str] = [
ASCEND_QUANTIZATION_METHOD, COMPRESSED_TENSORS_METHOD
]
def is_sleep_mode_available(self) -> bool:
return True
@@ -79,6 +82,8 @@ class NPUPlatform(Platform):
if ASCEND_QUANTIZATION_METHOD not in quant_action.choices:
quant_action.choices.append(ASCEND_QUANTIZATION_METHOD)
from vllm_ascend.quantization.compressed_tensors.compressed_tensors import \
AscendCompressedTensorsConfig # noqa: F401
from vllm_ascend.quantization.quant_config import \
AscendQuantConfig # noqa: F401