[Feature][Humming] Humming feature integration (#56685)
Signed-off-by: jinzhen.ljz <jinzhen.ljz@antgroup.com> Signed-off-by: mgoin <mgoin64@gmail.com> Co-authored-by: Codex <noreply@openai.com> Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
co-authored by
Codex
mgoin
parent
27757dde02
commit
7d99c2c4fd
@@ -101,6 +101,7 @@ from vllm.model_executor.kernels.linear.mxfp6 import (
|
||||
from vllm.model_executor.kernels.linear.mxfp6.emulation import (
|
||||
EmulationMxfp6LinearKernel,
|
||||
)
|
||||
from vllm.model_executor.kernels.linear.mxfp6.humming import HummingMxFp6LinearKernel
|
||||
from vllm.model_executor.kernels.linear.mxfp8 import (
|
||||
Mxfp8LinearKernel,
|
||||
Mxfp8LinearLayerConfig,
|
||||
@@ -295,6 +296,7 @@ _LINEAR_BACKEND_KERNEL_MAP: dict[str, set[type]] = {
|
||||
HummingLinearKernel,
|
||||
HummingMxfp8LinearKernel,
|
||||
HummingMxFp4LinearKernel,
|
||||
HummingMxFp6LinearKernel,
|
||||
HummingNvFp4LinearKernel,
|
||||
},
|
||||
"marlin": {
|
||||
@@ -574,6 +576,7 @@ _POSSIBLE_NVFP4_KERNELS: dict[PlatformEnum, list[type[NvFp4LinearKernel]]] = {
|
||||
|
||||
_POSSIBLE_MXFP6_KERNELS: dict[PlatformEnum, list[type[MxFp6LinearKernel]]] = {
|
||||
PlatformEnum.CUDA: [
|
||||
HummingMxFp6LinearKernel,
|
||||
EmulationMxfp6LinearKernel,
|
||||
],
|
||||
PlatformEnum.ROCM: [
|
||||
@@ -1294,6 +1297,7 @@ __all__ = [
|
||||
"MxFp6LinearLayerConfig",
|
||||
"init_mxfp6_linear_kernel",
|
||||
"EmulationMxfp6LinearKernel",
|
||||
"HummingMxFp6LinearKernel",
|
||||
"AiterMxfp4LinearKernel",
|
||||
"EmulationMxfp4LinearKernel",
|
||||
"FlashInferMxFp4LinearKernel",
|
||||
|
||||
@@ -24,7 +24,7 @@ class HummingLinearKernel(MPLinearKernel):
|
||||
return True, None
|
||||
|
||||
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
||||
from vllm.model_executor.layers.quantization.utils.humming_utils import (
|
||||
from vllm.model_executor.layers.quantization.utils.humming import (
|
||||
convert_linear_layer_to_humming_standard,
|
||||
get_humming_linear_compute_config,
|
||||
prepare_humming_linear_layer_config,
|
||||
@@ -57,13 +57,14 @@ class HummingLinearKernel(MPLinearKernel):
|
||||
x: torch.Tensor,
|
||||
bias: torch.Tensor | None = None,
|
||||
) -> torch.Tensor:
|
||||
from vllm.model_executor.layers.quantization.utils.humming_utils import (
|
||||
from vllm.model_executor.layers.quantization.utils.humming import (
|
||||
apply_humming_linear,
|
||||
)
|
||||
|
||||
return apply_humming_linear(
|
||||
layer,
|
||||
x,
|
||||
skip_bias_add=bias is None,
|
||||
layer_config=self.layer_config,
|
||||
compute_config=self.compute_config,
|
||||
locks=self.locks,
|
||||
|
||||
@@ -3,21 +3,18 @@
|
||||
|
||||
import torch
|
||||
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.quantization.utils.humming_utils import (
|
||||
from vllm.model_executor.layers.quantization.utils.humming import (
|
||||
apply_humming_linear,
|
||||
convert_linear_layer_to_humming_standard,
|
||||
get_humming_linear_compute_config,
|
||||
prepare_humming_linear_layer_config,
|
||||
quant_key_to_input_schema,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import kMxfp4Dynamic
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.utils.import_utils import has_humming
|
||||
|
||||
from .base import MxFp4LinearKernel, MxFp4LinearLayerConfig
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
class HummingMxFp4LinearKernel(MxFp4LinearKernel):
|
||||
"""Humming GEMM Kernel for MXFP4."""
|
||||
@@ -39,14 +36,10 @@ class HummingMxFp4LinearKernel(MxFp4LinearKernel):
|
||||
|
||||
@classmethod
|
||||
def can_implement(cls, config: MxFp4LinearLayerConfig) -> tuple[bool, str | None]:
|
||||
if config.activation_quant_key not in (None, kMxfp4Dynamic):
|
||||
return False, "only supports MXFP4 dynamic or unquantized activations"
|
||||
if config.activation_quant_key is not None:
|
||||
logger.warning_once(
|
||||
"HummingMxFp4LinearKernel is a weight-only (A16) kernel; "
|
||||
"the requested activation quantization (%s) is ignored.",
|
||||
config.activation_quant_key,
|
||||
)
|
||||
try:
|
||||
quant_key_to_input_schema(config.activation_quant_key)
|
||||
except ValueError as error:
|
||||
return False, str(error)
|
||||
return True, None
|
||||
|
||||
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
||||
@@ -62,7 +55,10 @@ class HummingMxFp4LinearKernel(MxFp4LinearKernel):
|
||||
}
|
||||
|
||||
convert_linear_layer_to_humming_standard(layer=layer, name_map=name_map)
|
||||
self.layer_config = prepare_humming_linear_layer_config(layer, quant_config)
|
||||
input_schema = quant_key_to_input_schema(self.config.activation_quant_key)
|
||||
self.layer_config = prepare_humming_linear_layer_config(
|
||||
layer, quant_config, input_schema=input_schema
|
||||
)
|
||||
self.compute_config = get_humming_linear_compute_config()
|
||||
self.locks = torch.zeros(1024, dtype=torch.int32, device=layer.weight.device)
|
||||
|
||||
@@ -75,6 +71,7 @@ class HummingMxFp4LinearKernel(MxFp4LinearKernel):
|
||||
return apply_humming_linear(
|
||||
layer,
|
||||
x,
|
||||
skip_bias_add=bias is None,
|
||||
layer_config=self.layer_config,
|
||||
compute_config=self.compute_config,
|
||||
locks=self.locks,
|
||||
|
||||
@@ -13,8 +13,8 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import QuantKey
|
||||
class MxFp6LinearLayerConfig:
|
||||
"""Configuration for an MXFP6 linear layer.
|
||||
|
||||
All MXFP6 layers share the same structure: packed uint8 weights (2 FP4 values per
|
||||
byte) and per-block weight scales (group size 32).
|
||||
All MXFP6 layers share the same structure: packed uint8 weights (4 FP6 values
|
||||
per 3 bytes) and per-block weight scales (group size 32).
|
||||
|
||||
Attributes:
|
||||
weight_quant_key: Identifies the weight quantization format. Can be
|
||||
|
||||
@@ -0,0 +1,84 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
import torch
|
||||
|
||||
from vllm.model_executor.layers.quantization.utils.humming import (
|
||||
apply_humming_linear,
|
||||
convert_linear_layer_to_humming_standard,
|
||||
get_humming_linear_compute_config,
|
||||
prepare_humming_linear_layer_config,
|
||||
quant_key_to_input_schema,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import kMxfp6E2M3Static
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.utils.import_utils import has_humming
|
||||
|
||||
from .base import MxFp6LinearKernel, MxFp6LinearLayerConfig
|
||||
|
||||
|
||||
class HummingMxFp6LinearKernel(MxFp6LinearKernel):
|
||||
"""Humming GEMM for packed MXFP6 E2M3 and E3M2 weights."""
|
||||
|
||||
@classmethod
|
||||
def is_supported(
|
||||
cls, compute_capability: int | None = None
|
||||
) -> tuple[bool, str | None]:
|
||||
if not current_platform.is_cuda():
|
||||
return False, "Humming only supported on CUDA"
|
||||
|
||||
if not has_humming():
|
||||
return False, "Humming is not installed"
|
||||
|
||||
if not current_platform.has_device_capability(75):
|
||||
return False, "Humming only supported on SM75+"
|
||||
|
||||
return True, None
|
||||
|
||||
@classmethod
|
||||
def can_implement(cls, config: MxFp6LinearLayerConfig) -> tuple[bool, str | None]:
|
||||
try:
|
||||
quant_key_to_input_schema(config.activation_quant_key)
|
||||
except ValueError as error:
|
||||
return False, str(error)
|
||||
return True, None
|
||||
|
||||
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
||||
weight_dtype = "float6e3m2"
|
||||
if self.config.weight_quant_key == kMxfp6E2M3Static:
|
||||
weight_dtype = "float6e2m3"
|
||||
|
||||
quant_config = {
|
||||
"quant_method": "humming",
|
||||
"dtype": weight_dtype,
|
||||
"scale_dtype": "float8e8m0",
|
||||
"group_size": 32,
|
||||
"weight_scale_type": "group",
|
||||
}
|
||||
layer.weight_scale.data = layer.weight_scale.data.view(torch.float8_e8m0fnu)
|
||||
|
||||
convert_linear_layer_to_humming_standard(
|
||||
layer=layer,
|
||||
name_map={"weight": "weight", "weight_scale": "weight_scale"},
|
||||
)
|
||||
input_schema = quant_key_to_input_schema(self.config.activation_quant_key)
|
||||
self.layer_config = prepare_humming_linear_layer_config(
|
||||
layer, quant_config, input_schema=input_schema
|
||||
)
|
||||
self.compute_config = get_humming_linear_compute_config()
|
||||
self.locks = torch.zeros(1024, dtype=torch.int32, device=layer.weight.device)
|
||||
|
||||
def apply_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
bias: torch.Tensor | None = None,
|
||||
) -> torch.Tensor:
|
||||
return apply_humming_linear(
|
||||
layer,
|
||||
x,
|
||||
skip_bias_add=bias is None,
|
||||
layer_config=self.layer_config,
|
||||
compute_config=self.compute_config,
|
||||
locks=self.locks,
|
||||
)
|
||||
@@ -3,12 +3,14 @@
|
||||
|
||||
import torch
|
||||
|
||||
from vllm.model_executor.layers.quantization.utils.humming_utils import (
|
||||
from vllm.model_executor.layers.quantization.utils.humming import (
|
||||
apply_humming_linear,
|
||||
convert_linear_layer_to_humming_standard,
|
||||
get_humming_linear_compute_config,
|
||||
prepare_humming_linear_layer_config,
|
||||
quant_key_to_input_schema,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import kMxfp8Dynamic
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.utils.import_utils import has_humming
|
||||
|
||||
@@ -50,7 +52,10 @@ class HummingMxfp8LinearKernel(Mxfp8LinearKernel):
|
||||
}
|
||||
|
||||
convert_linear_layer_to_humming_standard(layer=layer, name_map=name_map)
|
||||
self.layer_config = prepare_humming_linear_layer_config(layer, quant_config)
|
||||
input_schema = quant_key_to_input_schema(kMxfp8Dynamic)
|
||||
self.layer_config = prepare_humming_linear_layer_config(
|
||||
layer, quant_config, input_schema=input_schema
|
||||
)
|
||||
self.compute_config = get_humming_linear_compute_config()
|
||||
self.locks = torch.zeros(1024, dtype=torch.int32, device=layer.weight.device)
|
||||
|
||||
@@ -63,6 +68,7 @@ class HummingMxfp8LinearKernel(Mxfp8LinearKernel):
|
||||
return apply_humming_linear(
|
||||
layer,
|
||||
x,
|
||||
skip_bias_add=bias is None,
|
||||
layer_config=self.layer_config,
|
||||
compute_config=self.compute_config,
|
||||
locks=self.locks,
|
||||
|
||||
@@ -4,11 +4,13 @@
|
||||
import torch
|
||||
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.quantization.utils.humming_utils import (
|
||||
from vllm.model_executor.layers.quantization.utils.humming import (
|
||||
apply_humming_linear,
|
||||
get_humming_linear_compute_config,
|
||||
prepare_humming_linear_layer_config,
|
||||
quant_key_to_input_schema,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import kNvfp4Dynamic
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.utils.import_utils import has_humming
|
||||
|
||||
@@ -40,9 +42,7 @@ class HummingNvFp4LinearKernel(NvFp4LinearKernel):
|
||||
return True, None
|
||||
|
||||
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
||||
# Route through humming's compressed-tensors nvfp4 loader (same path as
|
||||
# the MoE oracle); the native group_tensor schema mishandles a scalar
|
||||
# global scale.
|
||||
# Reuse the CT loader for packed FP4 weights and tensor scales.
|
||||
quant_config = {
|
||||
"quant_method": "compressed-tensors",
|
||||
"format": "nvfp4-pack-quantized",
|
||||
@@ -60,7 +60,16 @@ class HummingNvFp4LinearKernel(NvFp4LinearKernel):
|
||||
layer.weight_global_scale = torch.nn.Parameter(
|
||||
1.0 / layer.weight_global_scale, requires_grad=False
|
||||
)
|
||||
self.layer_config = prepare_humming_linear_layer_config(layer, quant_config)
|
||||
input_quant_key = None
|
||||
if hasattr(layer, "input_global_scale"):
|
||||
layer.input_scale_2 = torch.nn.Parameter(
|
||||
layer.input_global_scale.reshape(-1), requires_grad=False
|
||||
)
|
||||
input_quant_key = kNvfp4Dynamic
|
||||
input_schema = quant_key_to_input_schema(input_quant_key)
|
||||
self.layer_config = prepare_humming_linear_layer_config(
|
||||
layer, quant_config, input_schema=input_schema
|
||||
)
|
||||
self.compute_config = get_humming_linear_compute_config()
|
||||
self.locks = torch.zeros(1024, dtype=torch.int32, device=layer.weight.device)
|
||||
|
||||
@@ -73,6 +82,7 @@ class HummingNvFp4LinearKernel(NvFp4LinearKernel):
|
||||
return apply_humming_linear(
|
||||
layer,
|
||||
x,
|
||||
skip_bias_add=bias is None,
|
||||
layer_config=self.layer_config,
|
||||
compute_config=self.compute_config,
|
||||
locks=self.locks,
|
||||
|
||||
@@ -4,7 +4,7 @@
|
||||
import torch
|
||||
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.quantization.utils.humming_utils import (
|
||||
from vllm.model_executor.layers.quantization.utils.humming import (
|
||||
apply_humming_linear,
|
||||
convert_linear_layer_to_humming_standard,
|
||||
get_humming_linear_compute_config,
|
||||
@@ -92,6 +92,7 @@ class HummingFP8ScaledMMLinearKernel(FP8ScaledMMLinearKernel):
|
||||
return apply_humming_linear(
|
||||
layer,
|
||||
x,
|
||||
skip_bias_add=bias is None,
|
||||
layer_config=self.layer_config,
|
||||
compute_config=self.compute_config,
|
||||
locks=self.locks,
|
||||
@@ -158,6 +159,7 @@ class HummingInt8ScaledMMLinearKernel(Int8ScaledMMLinearKernel):
|
||||
return apply_humming_linear(
|
||||
layer,
|
||||
x,
|
||||
skip_bias_add=bias is None,
|
||||
layer_config=self.layer_config,
|
||||
compute_config=self.compute_config,
|
||||
locks=self.locks,
|
||||
|
||||
Reference in New Issue
Block a user