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

Revert "[Feature][Quant] Auto-detect quantization format from model f… (#6873)

Browse Source 
This reverts commit 3953dcf784. to keep
the basic functions available

---------

Signed-off-by: wangli <wangli858794774@gmail.com>
This commit is contained in:
Li Wang
2026-03-10 11:27:32 +08:00
committed by GitHub
parent 40f7d93f1a
commit 33234aa0c5
7 changed files with 12 additions and 584 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
vllm_ascend/platform.py
View File

@@ -178,11 +178,6 @@ class NPUPlatform(Platform):
@classmethod
def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
from vllm_ascend.quantization.utils import maybe_auto_detect_quantization
if vllm_config.model_config is not None:
maybe_auto_detect_quantization(vllm_config)
# initialize ascend config from vllm additional_config
cls._fix_incompatible_config(vllm_config)
ascend_config = init_ascend_config(vllm_config)

110
vllm_ascend/quantization/modelslim_config.py
View File

@@ -21,9 +21,6 @@ This module provides the AscendModelSlimConfig class for parsing quantization
configs generated by the ModelSlim tool, along with model-specific mappings.
"""
import glob
import json
import os
from collections.abc import Mapping
from types import MappingProxyType
from typing import Any, Optional
@@ -42,9 +39,6 @@ from vllm_ascend.utils import ASCEND_QUANTIZATION_METHOD
from .methods import get_scheme_class
# The config filename that ModelSlim generates after quantizing a model.
MODELSLIM_CONFIG_FILENAME = "quant_model_description.json"
logger = init_logger(__name__)
# key: model_type
@@ -424,9 +418,9 @@ class AscendModelSlimConfig(QuantizationConfig):
quantized using the ModelSlim tool.
"""
def __init__(self, quant_config: dict[str, Any] | None = None):
def __init__(self, quant_config: dict[str, Any]):
super().__init__()
self.quant_description = quant_config if quant_config is not None else {}
self.quant_description = quant_config
# TODO(whx): remove this adaptation after adding "shared_head"
# to prefix of DeepSeekShareHead in vLLM.
extra_quant_dict = {}
@@ -456,12 +450,7 @@ class AscendModelSlimConfig(QuantizationConfig):
@classmethod
def get_config_filenames(cls) -> list[str]:
# Return empty list so that vllm's get_quant_config() skips the
# file-based lookup (which raises an unfriendly "Cannot find the
# config file for ascend" error when the model is not quantized).
# Instead, the config file is loaded in maybe_update_config(),
# which can provide a user-friendly error message.
return []
return ["quant_model_description.json"]
@classmethod
def from_config(cls, config: dict[str, Any]) -> "AscendModelSlimConfig":
@@ -573,98 +562,5 @@ class AscendModelSlimConfig(QuantizationConfig):
assert is_skipped is not None
return is_skipped
def maybe_update_config(self, model_name: str) -> None:
"""Load the ModelSlim quantization config from model directory.
This method is called by vllm after get_quant_config() returns
successfully. Since we return an empty list from get_config_filenames()
to bypass vllm's built-in file lookup, we do the actual config loading
here and provide user-friendly error messages when the config is missing.
Args:
model_name: Path to the model directory or model name.
"""
# If quant_description is already populated (e.g. from from_config()),
# there is nothing to do.
if self.quant_description:
return
# Try to find and load the ModelSlim config file
if os.path.isdir(model_name):
config_path = os.path.join(model_name, MODELSLIM_CONFIG_FILENAME)
if os.path.isfile(config_path):
with open(config_path) as f:
self.quant_description = json.load(f)
self._apply_extra_quant_adaptations()
return
# Check if there are any json files at all to help diagnose
json_files = glob.glob(os.path.join(model_name, "*.json"))
json_names = [os.path.basename(f) for f in json_files]
else:
json_names = []
# Config file not found - raise a friendly error message
raise ValueError(
"\n"
+ "=" * 80
+ "\n"
+ "ERROR: ModelSlim Quantization Config Not Found\n"
+ "=" * 80
+ "\n"
+ "\n"
+ f"You have enabled '--quantization {ASCEND_QUANTIZATION_METHOD}' "
+ "(ModelSlim quantization),\n"
+ f"but the model at '{model_name}' does not contain the required\n"
+ f"quantization config file ('{MODELSLIM_CONFIG_FILENAME}').\n"
+ "\n"
+ "This usually means the model weights are NOT quantized by "
+ "ModelSlim.\n"
+ "\n"
+ "Please choose one of the following solutions:\n"
+ "\n"
+ " Solution 1: Remove the quantization option "
+ "(for float/unquantized models)\n"
+ " "
+ "-" * 58
+ "\n"
+ f" Remove '--quantization {ASCEND_QUANTIZATION_METHOD}' from "
+ "your command if you want to\n"
+ " run the model with the original (float) weights.\n"
+ "\n"
+ " Example:\n"
+ f" vllm serve {model_name}\n"
+ "\n"
+ " Solution 2: Quantize your model weights with ModelSlim first\n"
+ " "
+ "-" * 58
+ "\n"
+ " Use the ModelSlim tool to quantize your model weights "
+ "before deployment.\n"
+ " After quantization, the model directory should contain "
+ f"'{MODELSLIM_CONFIG_FILENAME}'.\n"
+ " For more information, please refer to:\n"
+ " https://gitee.com/ascend/msit/tree/master/msmodelslim\n"
+ "\n"
+ (f" (Found JSON files in model directory: {json_names})\n" if json_names else "")
+ "=" * 80
)
def _apply_extra_quant_adaptations(self) -> None:
"""Apply extra adaptations to the quant_description dict.
This handles known key transformations such as shared_head and
weight_packed mappings.
"""
extra_quant_dict = {}
for k in self.quant_description:
if "shared_head" in k:
new_k = k.replace(".shared_head.", ".")
extra_quant_dict[new_k] = self.quant_description[k]
if "weight_packed" in k:
new_k = k.replace("weight_packed", "weight")
extra_quant_dict[new_k] = self.quant_description[k]
self.quant_description.update(extra_quant_dict)
def get_scaled_act_names(self) -> list[str]:
return []

147
vllm_ascend/quantization/utils.py
View File

@@ -1,147 +0,0 @@
#
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# This file is a part of the vllm-ascend project.
#
import json
import os
from vllm.logger import init_logger
from vllm_ascend.quantization.modelslim_config import MODELSLIM_CONFIG_FILENAME
from vllm_ascend.utils import ASCEND_QUANTIZATION_METHOD, COMPRESSED_TENSORS_METHOD
logger = init_logger(__name__)
def detect_quantization_method(model_path: str) -> str | None:
"""Auto-detect the quantization method from model directory files.
This function performs a lightweight check (JSON files and file existence
only — no .safetensors or .bin inspection) to determine which quantization
method was used to produce the weights in *model_path*.
Detection priority:
1. **ModelSlim (Ascend)** – ``quant_model_description.json`` exists
in the model directory.
2. **LLM-Compressor (compressed-tensors)** – ``config.json`` contains
a ``quantization_config`` section with
``"quant_method": "compressed-tensors"``.
3. **None** – neither condition is met; the caller should fall back to
the default (float) behaviour.
Args:
model_path: Path to the local model directory.
Returns:
``"ascend"`` for ModelSlim models,
``"compressed-tensors"`` for LLM-Compressor models,
or ``None`` if no quantization signature is found.
"""
if not os.path.isdir(model_path):
return None
# Case 1: ModelSlim — look for quant_model_description.json
modelslim_config_path = os.path.join(model_path, MODELSLIM_CONFIG_FILENAME)
if os.path.isfile(modelslim_config_path):
return ASCEND_QUANTIZATION_METHOD
# Case 2: LLM-Compressor — look for compressed-tensors in config.json
config_json_path = os.path.join(model_path, "config.json")
if os.path.isfile(config_json_path):
try:
with open(config_json_path) as f:
config = json.load(f)
quant_cfg = config.get("quantization_config")
if isinstance(quant_cfg, dict):
quant_method = quant_cfg.get("quant_method", "")
if quant_method == COMPRESSED_TENSORS_METHOD:
return COMPRESSED_TENSORS_METHOD
except (json.JSONDecodeError, OSError):
# Malformed or unreadable config.json — skip silently.
pass
# Case 3: No quantization signature found.
return None
def maybe_auto_detect_quantization(vllm_config) -> None:
"""Auto-detect and apply the quantization method on *vllm_config*.
This should be called during engine initialisation (from
``NPUPlatform.check_and_update_config``) **after** ``VllmConfig`` has been
created but **before** heavy weights are loaded.
Because ``check_and_update_config`` runs *after*
``VllmConfig.__post_init__`` has already evaluated
``_get_quantization_config`` (which returned ``None`` when
``model_config.quantization`` was not set), we must:
1. Set ``model_config.quantization`` to the detected value.
2. Recreate ``vllm_config.quant_config`` so that the quantization
pipeline (``get_quant_config`` → ``QuantizationConfig`` →
``get_quant_method`` for every layer) is properly initialised.
Rules:
* If the user explicitly set ``--quantization``, that value is
respected. A warning is emitted when the detected method differs.
* If no ``--quantization`` was given, the detected method (if any) is
applied automatically.
Args:
vllm_config: A ``vllm.config.VllmConfig`` instance (mutable).
"""
model_config = vllm_config.model_config
model_path = model_config.model
user_quant = model_config.quantization
detected = detect_quantization_method(model_path)
if detected is None:
# No quantization signature found — nothing to do.
return
if user_quant is not None:
# User explicitly specified a quantization method.
if user_quant != detected:
logger.warning(
"Auto-detected quantization method '%s' from model "
"files at '%s', but user explicitly specified "
"'--quantization %s'. Respecting the user-specified "
"value. If you encounter errors during model loading, "
"consider using '--quantization %s' instead.",
detected,
model_path,
user_quant,
detected,
)
return
# No user-specified quantization — apply auto-detected value.
model_config.quantization = detected
logger.info(
"Auto-detected quantization method '%s' from model files "
"at '%s'. To override, pass '--quantization <method>' explicitly.",
detected,
model_path,
)
# Recreate quant_config on VllmConfig. The original __post_init__
# already ran _get_quantization_config(), but at that point
# model_config.quantization was None so it returned None. Now that
# we've set it, we need to build the actual QuantizationConfig so the
# downstream model-loading code can use it.
from vllm.config import VllmConfig as _VllmConfig
vllm_config.quant_config = _VllmConfig._get_quantization_config(model_config, vllm_config.load_config)