forked from Karylab-cklius/vllm
[Refactor] Remove dead code (#42889)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
@@ -39,7 +39,6 @@ from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tenso
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.compressed_tensors.schemes import (
|
||||
WNA16_SUPPORTED_BITS,
|
||||
CompressedTensors24,
|
||||
CompressedTensorsScheme,
|
||||
CompressedTensorsW4A4Fp4,
|
||||
CompressedTensorsW4A4Mxfp4,
|
||||
@@ -760,19 +759,8 @@ class CompressedTensorsConfig(QuantizationConfig):
|
||||
input_quant=input_quant,
|
||||
sparsity_scheme=sparsity_scheme,
|
||||
):
|
||||
# Have a valid sparsity scheme
|
||||
# Validate layer is supported by Cutlass 2:4 Kernel
|
||||
model_compression_config = (
|
||||
None
|
||||
if sparsity_scheme is None or sparsity_scheme.format == "dense"
|
||||
else self.config
|
||||
)
|
||||
|
||||
scheme = CompressedTensors24(
|
||||
quantized=weight_quant is not None or input_quant is not None,
|
||||
weight_quant=weight_quant,
|
||||
input_quant=input_quant,
|
||||
model_compression_config=model_compression_config,
|
||||
raise NotImplementedError(
|
||||
"Sparse24 models are no longer supported by vLLM."
|
||||
)
|
||||
elif weight_quant is None:
|
||||
# Falling back to UnquantizedLinearMethod
|
||||
|
||||
@@ -13,9 +13,6 @@ from .compressed_tensors_w8a8_mxfp8 import CompressedTensorsW8A8Mxfp8
|
||||
from .compressed_tensors_w8a16_fp8 import CompressedTensorsW8A16Fp8
|
||||
from .compressed_tensors_wNa16 import WNA16_SUPPORTED_BITS, CompressedTensorsWNA16
|
||||
|
||||
# This avoids circular import error
|
||||
from .compressed_tensors_24 import CompressedTensors24 # isort: skip
|
||||
|
||||
__all__ = [
|
||||
"CompressedTensorsScheme",
|
||||
"CompressedTensorsWNA16",
|
||||
@@ -23,7 +20,6 @@ __all__ = [
|
||||
"CompressedTensorsW8A8Int8",
|
||||
"CompressedTensorsW8A8Fp8",
|
||||
"WNA16_SUPPORTED_BITS",
|
||||
"CompressedTensors24",
|
||||
"CompressedTensorsW4A16Fp4",
|
||||
"CompressedTensorsW4A4Mxfp4",
|
||||
"CompressedTensorsW4A4Fp4",
|
||||
|
||||
-54
@@ -1,54 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
from collections.abc import Callable
|
||||
from typing import Any
|
||||
|
||||
import torch
|
||||
from compressed_tensors.quantization import (
|
||||
QuantizationArgs,
|
||||
)
|
||||
|
||||
from vllm.model_executor.layers.quantization.compressed_tensors.schemes import (
|
||||
CompressedTensorsScheme,
|
||||
)
|
||||
|
||||
__all__ = ["CompressedTensors24"]
|
||||
|
||||
|
||||
class CompressedTensors24(CompressedTensorsScheme):
|
||||
def __init__(
|
||||
self,
|
||||
quantized: bool = False,
|
||||
weight_quant: QuantizationArgs | None = None,
|
||||
input_quant: QuantizationArgs | None = None,
|
||||
model_compression_config: dict[str, Any] | None = None,
|
||||
):
|
||||
raise NotImplementedError("Sparse24 models are no longer supported by vLLM")
|
||||
|
||||
@classmethod
|
||||
def get_min_capability(cls) -> int:
|
||||
raise NotImplementedError("Sparse24 models are no longer supported by vLLM")
|
||||
|
||||
def create_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
input_size: int,
|
||||
output_partition_sizes: list[int],
|
||||
input_size_per_partition: int,
|
||||
params_dtype: torch.dtype,
|
||||
weight_loader: Callable,
|
||||
**kwargs,
|
||||
):
|
||||
raise NotImplementedError("Sparse24 models are no longer supported by vLLM")
|
||||
|
||||
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
||||
raise NotImplementedError("Sparse24 models are no longer supported by vLLM")
|
||||
|
||||
def apply_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
bias: torch.Tensor | None = None,
|
||||
) -> torch.Tensor:
|
||||
raise NotImplementedError("Sparse24 models are no longer supported by vLLM")
|
||||
@@ -1,90 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
This file contains the Pydantic schemas for various quantization-related
|
||||
parameters. When a relevant quantization technique is specified, these
|
||||
parameters are loaded in the form of a JSON alongside the model weights
|
||||
and augment the model with additional information needed for use of that
|
||||
technique. The format of this JSON should be specified by one or more
|
||||
schemas contained here.
|
||||
|
||||
For example, when the KV cache is quantized to FP8-E4M3 (currently only
|
||||
possible on ROCm), the model can be optionally augmented with KV cache
|
||||
scaling factors.
|
||||
"""
|
||||
|
||||
from pydantic import BaseModel, ConfigDict, ValidationInfo, model_validator
|
||||
|
||||
|
||||
class KVCacheQuantSchema(BaseModel):
|
||||
dtype: str
|
||||
# Each key is a TP rank. Each value is a dictionary mapping a TP rank's
|
||||
# layer indices to their per-tensor KV cache scaling factor.
|
||||
# TODO: Consider pulling this and its validation methods out into its
|
||||
# own schema class (tricky as its members are variable)
|
||||
scaling_factor: dict[int, dict[int, float]]
|
||||
|
||||
@model_validator(mode="after")
|
||||
def check_is_fp8(self) -> "KVCacheQuantSchema":
|
||||
assert self.dtype == "float8_e4m3fn", (
|
||||
"Loaded scaling factors intended for KV cache dtype = "
|
||||
f"{self.dtype} rather than float8_e4m3fn!"
|
||||
)
|
||||
return self
|
||||
|
||||
@model_validator(mode="after")
|
||||
def check_tp_ranks(self, info: ValidationInfo) -> "KVCacheQuantSchema":
|
||||
context = info.context
|
||||
if context:
|
||||
tp_size = context["tp_size"]
|
||||
num_hidden_layers = context["num_hidden_layers"]
|
||||
assert len(self.scaling_factor) == tp_size, (
|
||||
f"Loaded dictionary has TP size {len(self.scaling_factor)} "
|
||||
f"but LLM engine is currently running with TP size {tp_size}."
|
||||
)
|
||||
for tp_rank, layer_maps in self.scaling_factor.items():
|
||||
assert len(layer_maps) == num_hidden_layers, (
|
||||
f"KV cache scales map for TP rank {tp_rank} is malformed. "
|
||||
f"Expected {num_hidden_layers} layers, got "
|
||||
f"{len(layer_maps)}."
|
||||
)
|
||||
for i in range(tp_size):
|
||||
assert i in self.scaling_factor, (
|
||||
f"KV cache scales map for TP rank {i} not found."
|
||||
)
|
||||
return self
|
||||
|
||||
@model_validator(mode="after")
|
||||
def check_current_rank(self, info: ValidationInfo) -> "KVCacheQuantSchema":
|
||||
context = info.context
|
||||
if context:
|
||||
tp_rank = context["tp_rank"]
|
||||
num_hidden_layers = context["num_hidden_layers"]
|
||||
layer_scales_map = self.scaling_factor[tp_rank]
|
||||
for i in range(num_hidden_layers):
|
||||
assert i in layer_scales_map, (
|
||||
f"Could not find KV cache scales for layer {i} in "
|
||||
f"TP rank {tp_rank}."
|
||||
)
|
||||
return self
|
||||
|
||||
|
||||
class QuantParamSchema(BaseModel):
|
||||
# TODO: Generalize and extend with more fields
|
||||
# (e.g. weights/activations params) once functionality is enabled
|
||||
model_config = ConfigDict(protected_namespaces=())
|
||||
model_type: str | None
|
||||
kv_cache: KVCacheQuantSchema
|
||||
|
||||
@model_validator(mode="after")
|
||||
def check_model_type(self, info: ValidationInfo) -> "QuantParamSchema":
|
||||
context = info.context
|
||||
if context:
|
||||
model_type = context.get("model_type", None)
|
||||
if model_type is not None:
|
||||
assert model_type == self.model_type, (
|
||||
f"Model type is {model_type} but loaded "
|
||||
f"scaling factors belonging to different "
|
||||
f"model type {self.model_type}!"
|
||||
)
|
||||
return self
|
||||
@@ -1,6 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""TurboQuant quantizer utilities.
|
||||
|
||||
Triton kernels handle all quantization, packing, and dequantization on GPU.
|
||||
"""
|
||||
Reference in New Issue
Block a user