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

[Lint]Style: Convert vllm-ascend/ to ruff format(Batch #7) (#6023)

Browse Source 
### What this PR does / why we need it?
**Scope of Changes**:
| File Path |
| :--- |
|` vllm_ascend/quantization/compressed_tensors/compressed_tensors.py`|
|` vllm_ascend/quantization/quant_config.py`|
|` vllm_ascend/quantization/utils.py`|
|` vllm_ascend/quantization/w4a16.py`|
|` vllm_ascend/quantization/w4a4_flatquant_dynamic.py`|
|` vllm_ascend/quantization/w4a8_dynamic.py`|
|` vllm_ascend/quantization/w8a16.py`|
|` vllm_ascend/quantization/w8a8.py`|
|` vllm_ascend/quantization/w8a8_dynamic.py`|
|` vllm_ascend/quantization/w8a8_pdmix.py`|
|` vllm_ascend/quantization/w8a8mxfp8.py`|
|` vllm_ascend/sample/rejection_sampler.py`|
|` vllm_ascend/sample/sampler.py`|
|` vllm_ascend/worker/block_table.py`|

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.13.0
- vLLM main:
2c24bc6996

Signed-off-by: MrZ20 <2609716663@qq.com>
This commit is contained in:
SILONG ZENG
2026-02-06 14:56:53 +08:00
committed by GitHub
parent d0bc16859c
commit 99aedaff63
20 changed files with 997 additions and 1307 deletions
Download Patch File Download Diff File Expand all files Collapse all files

253
vllm_ascend/quantization/modelslim_config.py
View File

@@ -21,20 +21,18 @@ This module provides the AscendModelSlimConfig class for parsing quantization
configs generated by the ModelSlim tool, along with model-specific mappings.
"""
from collections.abc import Mapping
from types import MappingProxyType
from typing import Any, Dict, List, Mapping, Optional
from typing import Any, Optional
import torch
from vllm.config import get_current_vllm_config
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.linear import LinearBase
from vllm.model_executor.layers.quantization import \
register_quantization_config
from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig, QuantizeMethodBase)
from vllm.model_executor.layers.vocab_parallel_embedding import (
UnquantizedEmbeddingMethod, VocabParallelEmbedding)
from vllm.model_executor.layers.quantization import register_quantization_config
from vllm.model_executor.layers.quantization.base_config import QuantizationConfig, QuantizeMethodBase
from vllm.model_executor.layers.vocab_parallel_embedding import UnquantizedEmbeddingMethod, VocabParallelEmbedding
from vllm.model_executor.models.utils import WeightsMapper
from vllm_ascend.utils import ASCEND_QUANTIZATION_METHOD
@@ -45,7 +43,7 @@ logger = init_logger(__name__)
# key: model_type
# value: orig_to_new_prefix
QUANT_MODEL_PREFIX_MAPPINGS: Dict[str, Dict[str, str]] = {
QUANT_MODEL_PREFIX_MAPPINGS: dict[str, dict[str, str]] = {
"qwen3_vl_moe": {
"visual.": "model.visual.",
"language_model.lm_head.": "lm_head.",
@@ -60,7 +58,7 @@ QUANT_MODEL_PREFIX_MAPPINGS: Dict[str, Dict[str, str]] = {
# key: model_type
# value: dict of fused module name -> list of original module names
packed_modules_model_mapping: Dict[str, Dict[str, List[str]]] = {
packed_modules_model_mapping: dict[str, dict[str, list[str]]] = {
"qwen3_moe": {
"qkv_proj": [
"q_proj",
@@ -71,52 +69,44 @@ packed_modules_model_mapping: Dict[str, Dict[str, List[str]]] = {
"gate_proj",
"up_proj",
],
"experts":
["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"experts": ["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
},
"deepseek_v2": {
"gate_up_proj": ["gate_proj", "up_proj"],
"experts":
["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"]
"experts": ["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"],
},
"deepseek_v3": {
"gate_up_proj": ["gate_proj", "up_proj"],
"experts":
["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"]
"experts": ["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"],
},
"pangu_ultra_moe": {
"gate_up_proj": ["gate_proj", "up_proj"],
"experts":
["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"]
"experts": ["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"],
},
"kimi_k2": {
"gate_up_proj": ["gate_proj", "up_proj"],
"experts":
["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"]
"experts": ["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"],
},
"deepseek_v32": {
"gate_up_proj": ["gate_proj", "up_proj"],
"experts":
["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"]
"experts": ["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"],
},
# NOTE 1.The quantized MTP layer of deepseek on the NPU is not quantized;
# NOTE 2.The description file generated by the current msmodelslim tool does not have
# MTP layer info. Please manually add it and set the value to FLOAT.
"deepseek_mtp": {
"gate_up_proj": ["gate_proj", "up_proj"],
"experts":
["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"]
"experts": ["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
},
"pangu_ultra_moe_mtp": {
"gate_up_proj": ["gate_proj", "up_proj"],
"experts":
["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"]
"experts": ["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"],
},
"qwen3_next": {
"qkv_proj": [
@@ -126,8 +116,7 @@ packed_modules_model_mapping: Dict[str, Dict[str, List[str]]] = {
],
"gate_up_proj": ["gate_proj", "up_proj"],
"in_proj": ["in_proj_qkvz", "in_proj_ba"],
"experts":
["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"]
"experts": ["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
},
"qwen2_5_vl": {
"qkv_proj": [
@@ -150,8 +139,7 @@ packed_modules_model_mapping: Dict[str, Dict[str, List[str]]] = {
"gate_proj",
"up_proj",
],
"experts":
["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"experts": ["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
},
"glm4_moe": {
"qkv_proj": [
@@ -163,20 +151,17 @@ packed_modules_model_mapping: Dict[str, Dict[str, List[str]]] = {
"gate_proj",
"up_proj",
],
"experts":
["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"]
"experts": ["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
},
"glm4_moe_lite": {
"glm4_moe_lite": {
"gate_up_proj": ["gate_proj", "up_proj"],
"experts":
["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"]
"experts": ["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"],
},
"longcat_flash": {
"gate_up_proj": ["gate_proj", "up_proj"],
"experts":
["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"]
"experts": ["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"],
"fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"],
},
"minimax_m2": {
"qkv_proj": [
@@ -184,17 +169,17 @@ packed_modules_model_mapping: Dict[str, Dict[str, List[str]]] = {
"k_proj",
"v_proj",
],
"experts": ["experts.0.w1", "experts.0.w2", "experts.0.w3"]
}
"experts": ["experts.0.w1", "experts.0.w2", "experts.0.w3"],
},
}
def get_packed_modules_mapping(model_type: str) -> Dict[str, List[str]]:
def get_packed_modules_mapping(model_type: str) -> dict[str, list[str]]:
"""Get packed modules mapping for a model type.
Args:
model_type: The model type string (e.g., "deepseek_v3").
Returns:
Dictionary mapping fused module names to their component module names.
Returns empty dict if model_type is not found.
@@ -202,12 +187,12 @@ def get_packed_modules_mapping(model_type: str) -> Dict[str, List[str]]:
return packed_modules_model_mapping.get(model_type, {})
def get_prefix_mapping(model_type: str) -> Dict[str, str]:
def get_prefix_mapping(model_type: str) -> dict[str, str]:
"""Get prefix mapping for a model type.
Args:
model_type: The model type string (e.g., "qwen3_vl_moe").
Returns:
Dictionary mapping original prefixes to new prefixes.
Returns empty dict if model_type is not found.
@@ -216,15 +201,15 @@ def get_prefix_mapping(model_type: str) -> Dict[str, str]:
def get_linear_quant_type(
quant_description: Dict[str, Any], prefix: str,
packed_modules_mapping: Dict[str, Any]) -> Optional[str]:
quant_description: dict[str, Any], prefix: str, packed_modules_mapping: dict[str, Any]
) -> str | None:
"""Determine the quantization type for a linear layer.
Args:
quant_description: The quantization description dictionary.
prefix: The layer prefix.
packed_modules_mapping: Mapping for packed/fused modules.
Returns:
The quantization type string (e.g., "W8A8_DYNAMIC").
"""
@@ -232,11 +217,10 @@ def get_linear_quant_type(
if proj_name in packed_modules_mapping:
quant_type = None
shard_prefixes = [
prefix.replace(proj_name, shard_proj_name)
for shard_proj_name in packed_modules_mapping[proj_name]
prefix.replace(proj_name, shard_proj_name) for shard_proj_name in packed_modules_mapping[proj_name]
]
for shard_prefix in shard_prefixes:
shard_quant_type = quant_description[shard_prefix + '.weight']
shard_quant_type = quant_description[shard_prefix + ".weight"]
if quant_type is None:
quant_type = shard_quant_type
@@ -244,72 +228,68 @@ def get_linear_quant_type(
raise ValueError(
f"Not all shards of {prefix} are quantized with same quant type."
f"Shard {proj_name} uses {shard_quant_type}, but another shard"
f"use {quant_type}. Please check quantization config.")
f"use {quant_type}. Please check quantization config."
)
else:
quant_type = quant_description[prefix + '.weight']
quant_type = quant_description[prefix + ".weight"]
return quant_type
def get_quant_type_for_layer(
quant_description: Dict[str, Any],
prefix: str,
layer_type: str,
packed_modules_mapping: Optional[Dict[str,
Any]] = None) -> Optional[str]:
quant_description: dict[str, Any],
prefix: str,
layer_type: str,
packed_modules_mapping: dict[str, Any] | None = None,
) -> str | None:
"""Determine the quantization type for a layer.
Args:
quant_description: The quantization description dictionary.
prefix: The layer prefix.
layer_type: The type of layer ("linear", "moe", "attention").
packed_modules_mapping: Mapping for packed/fused modules.
Returns:
The quantization type string (e.g., "W8A8_DYNAMIC").
"""
if packed_modules_mapping is None:
packed_modules_mapping = dict()
# Attention
if layer_type == "attention" and 'fa_quant_type' in quant_description.keys(
):
return quant_description['fa_quant_type']
if layer_type == "attention" and "fa_quant_type" in quant_description:
return quant_description["fa_quant_type"]
# Linear / MoE
return get_linear_quant_type(quant_description, prefix,
packed_modules_mapping)
return get_linear_quant_type(quant_description, prefix, packed_modules_mapping)
def create_scheme_for_layer(
quant_description: Dict[str, Any],
prefix: str,
layer_type: str,
packed_modules_mapping: Optional[Dict[str, Any]] = None):
quant_description: dict[str, Any],
prefix: str,
layer_type: str,
packed_modules_mapping: dict[str, Any] | None = None,
):
"""Create a quantization scheme instance for a layer.
Args:
quant_description: The quantization description dictionary.
prefix: The layer prefix.
layer_type: The type of layer ("linear", "moe", "attention").
packed_modules_mapping: Mapping for packed/fused modules.
Returns:
An instance of the appropriate quantization scheme class.
"""
logger.info_once("Using the vLLM Ascend modelslim Quantization now!")
quant_type = get_quant_type_for_layer(quant_description, prefix,
layer_type, packed_modules_mapping)
quant_type = get_quant_type_for_layer(quant_description, prefix, layer_type, packed_modules_mapping)
if quant_type is None:
raise ValueError(
f"Could not determine quantization type for layer {prefix}.")
raise ValueError(f"Could not determine quantization type for layer {prefix}.")
# Use registry to get scheme class
scheme_cls = get_scheme_class(quant_type, layer_type)
if scheme_cls is not None:
return scheme_cls()
raise NotImplementedError(
f"Currently, vLLM Ascend doesn't support {quant_type} for {layer_type}."
)
raise NotImplementedError(f"Currently, vLLM Ascend doesn't support {quant_type} for {layer_type}.")
@register_quantization_config(ASCEND_QUANTIZATION_METHOD)
@@ -321,13 +301,13 @@ class AscendModelSlimConfig(QuantizationConfig):
quantized using the ModelSlim tool.
"""
def __init__(self, quant_config: Dict[str, Any]):
def __init__(self, quant_config: dict[str, Any]):
super().__init__()
self.quant_description = quant_config
# TODO(whx): remove this adaptation after adding "shared_head"
# to prefix of DeepSeekShareHead in vLLM.
extra_quant_dict = {}
for k in self.quant_description.keys():
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]
@@ -344,25 +324,23 @@ class AscendModelSlimConfig(QuantizationConfig):
return ASCEND_QUANTIZATION_METHOD
@classmethod
def get_supported_act_dtypes(cls) -> List[torch.dtype]:
def get_supported_act_dtypes(cls) -> list[torch.dtype]:
return [torch.int8, torch.float16, torch.bfloat16]
@classmethod
def get_min_capability(cls) -> int:
raise NotImplementedError(
"Ascend hardware dose not support \"get_min_capability\" feature.")
raise NotImplementedError('Ascend hardware dose not support "get_min_capability" feature.')
@classmethod
def get_config_filenames(cls) -> List[str]:
def get_config_filenames(cls) -> list[str]:
return ["quant_model_description.json"]
@classmethod
def from_config(cls, config: Dict[str, Any]) -> "AscendModelSlimConfig":
def from_config(cls, config: dict[str, Any]) -> "AscendModelSlimConfig":
return cls(config)
@classmethod
def override_quantization_method(cls, hf_quant_cfg,
user_quant) -> Optional[str]:
def override_quantization_method(cls, hf_quant_cfg, user_quant) -> str | None:
if hf_quant_cfg is not None:
quant_method = hf_quant_cfg.get("quant_method", None)
if not quant_method and torch.npu.is_available():
@@ -373,15 +351,17 @@ class AscendModelSlimConfig(QuantizationConfig):
# TODO (Levi-JQ): will be removed when QuantizationConfig.apply_vllm_mapper is implemented
prefix_mapping = QUANT_MODEL_PREFIX_MAPPINGS.get(model_type)
if prefix_mapping:
hf_to_vllm_mapper = WeightsMapper(
orig_to_new_prefix=prefix_mapping)
hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix=prefix_mapping)
return hf_to_vllm_mapper._map_name(prefix)
return prefix
def get_quant_method(self, layer: torch.nn.Module,
prefix: str) -> Optional["QuantizeMethodBase"]:
from .method_adapters import (AscendEmbeddingMethod, AscendFusedMoEMethod,
AscendKVCacheMethod, AscendLinearMethod)
def get_quant_method(self, layer: torch.nn.Module, prefix: str) -> Optional["QuantizeMethodBase"]:
from .method_adapters import (
AscendEmbeddingMethod,
AscendFusedMoEMethod,
AscendKVCacheMethod,
AscendLinearMethod,
)
vllm_config = get_current_vllm_config()
model_type = vllm_config.model_config.hf_config.model_type
@@ -390,81 +370,67 @@ class AscendModelSlimConfig(QuantizationConfig):
# Adapt to Minimax architecture: update layer names to MoE convention
prefix = prefix.replace("mlp", "block_sparse_moe")
# Normalize the prefix by stripping specific expert indices (e.g., 'experts.0' -> 'experts')
parts = prefix.split('.')
parts = prefix.split(".")
if "experts" in parts and len(parts) > 2:
exp_idx = parts.index("experts")
if exp_idx + 1 < len(parts) and parts[exp_idx + 1].isdigit():
parts = parts[:exp_idx + 1]
parts = parts[: exp_idx + 1]
prefix = ".".join(parts)
if model_type in packed_modules_model_mapping:
self.packed_modules_mapping = packed_modules_model_mapping[
model_type]
self.packed_modules_mapping = packed_modules_model_mapping[model_type]
prefix = self.quant_prefix_mapper(model_type, prefix)
from vllm_ascend.utils import vllm_version_is
if vllm_version_is("v0.15.0"):
from vllm.attention.layer import Attention # type: ignore
from vllm.attention.layer import Attention # type: ignore
else:
from vllm.model_executor.layers.attention import Attention
if prefix.startswith("language_model"):
prefix = prefix.split('.', 1)[-1]
prefix = prefix.split(".", 1)[-1]
if isinstance(layer, LinearBase):
if self.is_layer_skipped_ascend(prefix,
self.packed_modules_mapping):
if self.is_layer_skipped_ascend(prefix, self.packed_modules_mapping):
# Delayed import to avoid circular import
from vllm_ascend.ops.linear import \
AscendUnquantizedLinearMethod
from vllm_ascend.ops.linear import AscendUnquantizedLinearMethod
return AscendUnquantizedLinearMethod()
scheme = create_scheme_for_layer(self.quant_description, prefix,
"linear",
self.packed_modules_mapping)
scheme = create_scheme_for_layer(self.quant_description, prefix, "linear", self.packed_modules_mapping)
return AscendLinearMethod(scheme)
elif isinstance(layer, Attention) and \
'fa_quant_type' in self.quant_description.keys() and \
self.quant_description['fa_quant_type'] is not None:
scheme = create_scheme_for_layer(self.quant_description, prefix,
"attention",
self.packed_modules_mapping)
elif (
isinstance(layer, Attention)
and "fa_quant_type" in self.quant_description
and self.quant_description["fa_quant_type"] is not None
):
scheme = create_scheme_for_layer(self.quant_description, prefix, "attention", self.packed_modules_mapping)
return AscendKVCacheMethod(scheme)
elif isinstance(layer, FusedMoE):
if self.is_layer_skipped_ascend(prefix,
self.packed_modules_mapping):
if self.is_layer_skipped_ascend(prefix, self.packed_modules_mapping):
# Delayed import to avoid circular import
from vllm_ascend.ops.fused_moe.fused_moe import \
AscendUnquantizedFusedMoEMethod
from vllm_ascend.ops.fused_moe.fused_moe import AscendUnquantizedFusedMoEMethod
return AscendUnquantizedFusedMoEMethod(layer.moe_config)
scheme = create_scheme_for_layer(self.quant_description, prefix,
"moe",
self.packed_modules_mapping)
scheme = create_scheme_for_layer(self.quant_description, prefix, "moe", self.packed_modules_mapping)
return AscendFusedMoEMethod(scheme, layer.moe_config)
elif isinstance(layer, VocabParallelEmbedding):
if self.is_layer_skipped_ascend(prefix,
self.packed_modules_mapping):
if self.is_layer_skipped_ascend(prefix, self.packed_modules_mapping):
return UnquantizedEmbeddingMethod()
scheme = create_scheme_for_layer(self.quant_description, prefix,
"linear",
self.packed_modules_mapping)
scheme = create_scheme_for_layer(self.quant_description, prefix, "linear", self.packed_modules_mapping)
return AscendEmbeddingMethod(scheme)
return None
def is_layer_skipped_ascend(
self,
prefix: str,
fused_mapping: Mapping[str, List[str]] = MappingProxyType({})):
def is_layer_skipped_ascend(self, prefix: str, fused_mapping: Mapping[str, list[str]] = MappingProxyType({})):
# adapted from vllm.model_executor.layers.quantization.utils.quant_utils.is_layer_skipped
proj_name = prefix.split(".")[-1]
if proj_name in fused_mapping:
shard_prefixes = [
prefix.replace(proj_name, shard_proj_name)
for shard_proj_name in fused_mapping[proj_name]
prefix.replace(proj_name, shard_proj_name) for shard_proj_name in fused_mapping[proj_name]
]
is_skipped = None
for shard_prefix in shard_prefixes:
is_shard_skipped = self.quant_description[shard_prefix +
'.weight'] == "FLOAT"
is_shard_skipped = self.quant_description[shard_prefix + ".weight"] == "FLOAT"
if is_skipped is None:
is_skipped = is_shard_skipped
@@ -472,12 +438,13 @@ class AscendModelSlimConfig(QuantizationConfig):
raise ValueError(
f"Detected some but not all shards of {prefix} "
"are quantized. All shards of fused layers "
"to have the same precision.")
"to have the same precision."
)
else:
is_skipped = self.quant_description[prefix + '.weight'] == "FLOAT"
is_skipped = self.quant_description[prefix + ".weight"] == "FLOAT"
assert is_skipped is not None
return is_skipped
def get_scaled_act_names(self) -> List[str]:
def get_scaled_act_names(self) -> list[str]:
return []