Integration of TurboMind AWQ (#2828)

Co-authored-by: root <bjmsong@126.com>

python/pyproject.toml

@@ -28,7 +28,7 @@ runtime_common = [
 srt = [
     "sglang[runtime_common]", "cuda-python",
     "sgl-kernel>=0.0.2.post11", "torch", "vllm>=0.6.3.post1,<=0.6.4.post1",
-    "flashinfer==0.1.6"
+    "flashinfer==0.1.6", "turbomind"
 ]
 
 # HIP (Heterogeneous-computing Interface for Portability) for AMD

python/sglang/srt/configs/model_config.py

@@ -14,6 +14,7 @@
 
 import json
 import logging
+import sys
 from enum import IntEnum, auto
 from typing import List, Optional, Set, Union
 
@@ -230,7 +231,7 @@ class ModelConfig:
         # Parse quantization method from the HF model config, if available.
         quant_cfg = self._parse_quant_hf_config()
 
-        if quant_cfg is not None:
+        if quant_cfg is not None and not quantization_in_turbomind(self.quantization):
             quant_method = quant_cfg.get("quant_method", "").lower()
 
             # Detect which checkpoint is it
@@ -401,3 +402,10 @@ def is_multimodal_model(model_architectures: List[str]):
 
 def is_encoder_decoder_model(model_architectures: List[str]):
     return "MllamaForConditionalGeneration" in model_architectures
+
+
+def quantization_in_turbomind(quantization: str) -> bool:
+    if quantization in ["awq_turbomind"]:
+        return True
+    else:
+        return False

python/sglang/srt/layers/linear.py

@@ -48,6 +48,7 @@ WEIGHT_LOADER_V2_SUPPORTED = [
     "GPTQLinearMethod",
     "FBGEMMFp8LinearMethod",
     "ModelOptFp8LinearMethod",
+    "AWQTurbomindLinearMethod",
     "IPEXAWQLinearMethod",
 ]
 

python/sglang/srt/layers/quantization/__init__.py

@@ -20,6 +20,7 @@ from vllm.model_executor.layers.quantization.marlin import MarlinConfig
 from vllm.model_executor.layers.quantization.qqq import QQQConfig
 from vllm.model_executor.layers.quantization.tpu_int8 import Int8TpuConfig
 
+from sglang.srt.layers.quantization.awq_turbomind import AWQTurbomindConfig
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.quantization.fp8 import Fp8Config
 from sglang.srt.layers.quantization.modelopt_quant import ModelOptFp8Config
@@ -37,6 +38,7 @@ QUANTIZATION_METHODS: Dict[str, Type[QuantizationConfig]] = {
     "gptq_marlin_24": GPTQMarlin24Config,
     "gptq_marlin": GPTQMarlinConfig,
     "awq_marlin": AWQMarlinConfig,
+    "awq_turbomind": AWQTurbomindConfig,
     "gptq": GPTQConfig,
     "compressed-tensors": CompressedTensorsConfig,
     "bitsandbytes": BitsAndBytesConfig,

python/sglang/srt/layers/quantization/awq_turbomind.py

+import logging
+import os
+import sys
+from typing import Any, Dict, List, Optional
+
+import torch
+import turbomind
+from torch.nn import Parameter
+
+turbomind_dir = os.path.split(turbomind.__file__)[0]
+sys.path.append(os.path.join(turbomind_dir, "lib"))
+import _turbomind_ext
+from vllm.model_executor.layers.linear import LinearBase
+
+from sglang.srt.layers.linear import LinearMethodBase, UnquantizedLinearMethod
+from sglang.srt.layers.parameter import GroupQuantScaleParameter, PackedvLLMParameter
+from sglang.srt.layers.quantization.base_config import (
+    QuantizationConfig,
+    QuantizeMethodBase,
+)
+from sglang.srt.layers.quantization.turbomind_utils import (
+    get_u4_slices,
+    is_layer_skipped_awq,
+    pack_u4_row,
+    unpack_awq_gemm,
+    verify_turbomind_supported,
+)
+from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
+from sglang.srt.utils import is_cuda, set_weight_attrs
+
+logger = logging.getLogger(__name__)
+
+
+class AWQTurbomindConfig(QuantizationConfig):
+    """Config class for AWQ Turbomind"""
+
+    def __init__(
+        self,
+        weight_bits: int,
+        group_size: int,
+        zero_point: bool,
+        lm_head_quantized: bool,
+        modules_to_not_convert: Optional[List[str]] = None,
+    ) -> None:
+        self.pack_factor = 32 // weight_bits  # packed into int32
+        self.group_size = group_size
+        self.zero_point = zero_point
+        self.lm_head_quantized = lm_head_quantized
+        self.weight_bits = weight_bits
+        self.modules_to_not_convert = modules_to_not_convert or []
+
+        verify_turbomind_supported(self.weight_bits, self.group_size)
+
+    def __repr__(self) -> str:
+        return (
+            f"AWQTurbomindConfig(weight_bits={self.weight_bits}, "
+            f"group_size={self.group_size}, "
+            f"zero_point={self.zero_point}, "
+            f"lm_head_quantized={self.lm_head_quantized}, "
+            f"modules_to_not_convert={self.modules_to_not_convert})"
+        )
+
+    @classmethod
+    def get_name(cls) -> str:
+        return "awq_turbomind"
+
+    @classmethod
+    def get_supported_act_dtypes(cls) -> List[torch.dtype]:
+        return [torch.half, torch.bfloat16]
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return 70
+
+    @classmethod
+    def get_config_filenames(cls) -> List[str]:
+        return ["quantize_config.json"]
+
+    @classmethod
+    def from_config(cls, config: Dict[str, Any]) -> "AWQTurbomindConfig":
+        weight_bits = cls.get_from_keys(config, ["bits"])
+        group_size = cls.get_from_keys(config, ["group_size"])
+        zero_point = cls.get_from_keys(config, ["zero_point"])
+        lm_head_quantized = cls.get_from_keys_or(config, ["lm_head"], default=False)
+        modules_to_not_convert = cls.get_from_keys_or(
+            config, ["modules_to_not_convert"], None
+        )
+        return cls(
+            weight_bits,
+            group_size,
+            zero_point,
+            lm_head_quantized,
+            modules_to_not_convert,
+        )
+
+    @classmethod
+    def override_quantization_method(cls, hf_quant_cfg, user_quant) -> Optional[str]:
+        can_convert = cls.is_awq_turbomind_compatible(hf_quant_cfg)
+        is_valid_user_quant = user_quant is None or user_quant == "awq_turbomind"
+
+        if can_convert and is_valid_user_quant:
+            msg = f"The model is convertible to {cls.get_name()} during runtime. Using {cls.get_name()} kernel."
+            logger.info(msg)
+            return cls.get_name()
+
+        if can_convert and user_quant == "awq":
+            logger.info(
+                "Detected that the model can run with awq_turbomind"
+                ", however you specified quantization=awq explicitly,"
+                " so forcing awq. Use quantization=awq_turbomind for"
+                " faster inference"
+            )
+        return None
+
+    def get_quant_method(
+        self, layer: torch.nn.Module, prefix: str
+    ) -> Optional["QuantizeMethodBase"]:
+        if isinstance(layer, LinearBase) or (
+            isinstance(layer, ParallelLMHead) and self.lm_head_quantized
+        ):
+            if is_layer_skipped_awq(prefix, self.modules_to_not_convert):
+                return UnquantizedLinearMethod()
+            return AWQTurbomindLinearMethod(self)
+
+        return None
+
+    @classmethod
+    def is_awq_turbomind_compatible(cls, quant_config: Dict[str, Any]):
+        if not is_cuda():
+            return False
+
+        # Extract data from quant config.
+        quant_method = quant_config.get("quant_method", "").lower()
+        num_bits = quant_config.get("bits")
+        group_size = quant_config.get("group_size")
+        zero_point = quant_config.get("zero_point")
+
+        if quant_method != "awq":
+            return False
+
+        # If we cannot find the info needed in the config, cannot convert.
+        if num_bits is None or group_size is None or zero_point is None:
+            return False
+
+        return verify_turbomind_supported(quant_bit=num_bits, group_size=group_size)
+
+    def get_scaled_act_names(self) -> List[str]:
+        return []
+
+
+class AWQTurbomindLinearMethod(LinearMethodBase):
+    """Linear method for AWQ Turbomind.
+
+    Args:
+        quant_config: The AWQ Turbomind quantization config.
+    """
+
+    def __init__(self, quant_config: AWQTurbomindConfig) -> None:
+        self.quant_config = quant_config
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        input_size_per_partition: int,
+        output_partition_sizes: List[int],
+        input_size: int,
+        output_size: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ) -> None:
+
+        output_size_per_partition = sum(output_partition_sizes)
+        weight_loader = extra_weight_attrs.get("weight_loader")
+
+        # Normalize group_size
+        if self.quant_config.group_size != -1:
+            group_size = self.quant_config.group_size
+        else:
+            group_size = input_size
+
+        qweight = PackedvLLMParameter(
+            data=torch.empty(
+                input_size_per_partition,
+                output_size_per_partition // self.quant_config.pack_factor,
+                dtype=torch.int32,
+            ),
+            input_dim=0,
+            output_dim=1,
+            packed_dim=1,
+            packed_factor=self.quant_config.pack_factor,
+            weight_loader=weight_loader,
+        )
+
+        num_groups = input_size_per_partition // group_size
+
+        qzeros = PackedvLLMParameter(
+            data=torch.empty(
+                num_groups,
+                output_size_per_partition // self.quant_config.pack_factor,
+                dtype=torch.int32,
+            ),
+            input_dim=0,
+            output_dim=1,
+            packed_dim=1,
+            packed_factor=self.quant_config.pack_factor,
+            weight_loader=weight_loader,
+        )
+
+        scales = GroupQuantScaleParameter(
+            data=torch.empty(
+                num_groups,
+                output_size_per_partition,
+                dtype=params_dtype,
+            ),
+            input_dim=0,
+            output_dim=1,
+            weight_loader=weight_loader,
+        )
+
+        layer.register_parameter("qweight", qweight)
+        layer.register_parameter("qzeros", qzeros)
+        layer.register_parameter("scales", scales)
+
+        layer.input_size_per_partition = input_size_per_partition
+        layer.output_size_per_partition = output_size_per_partition
+        layer.num_groups = num_groups
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+
+        qweight_turbomind = unpack_awq_gemm(layer.qweight.data)
+        qzeros_turbomind = unpack_awq_gemm(layer.qzeros.data)
+        scales_turbomind = layer.scales.data
+
+        qweight_turbomind = pack_u4_row(qweight_turbomind)
+        qzeros_turbomind = qzeros_turbomind.to(torch.half)
+
+        device_id = layer.qweight.device.index
+        properties = torch.cuda.get_device_properties(device_id)
+
+        def is_16xx_series(name):
+            import re
+
+            pattern = r"GTX 16\d\d"
+            return bool(re.search(pattern, name))
+
+        simt = is_16xx_series(properties.name)
+        qweight_turbomind = qweight_turbomind.contiguous()
+        scales_turbomind = scales_turbomind.contiguous()
+        qzeros_turbomind = qzeros_turbomind.contiguous()
+
+        self.linear = _turbomind_ext.Linear(
+            layer.input_size_per_partition,
+            layer.output_size_per_partition,
+            self.quant_config.weight_bits,
+            self.quant_config.group_size,
+        )
+
+        self.linear.post_init(
+            qweight_turbomind, scales_turbomind, qzeros_turbomind, simt
+        )
+
+        layer.qweight = Parameter(qweight_turbomind, requires_grad=False)
+        layer.scales = Parameter(scales_turbomind, requires_grad=False)
+        layer.qzeros = Parameter(qzeros_turbomind, requires_grad=False)
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        bias: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+
+        x = x.view(-1, x.shape[-1])
+        out_shape = x.shape[:-1] + (layer.output_size_per_partition,)
+        out = torch.empty(
+            (x.shape[0], layer.output_size_per_partition),
+            dtype=torch.float16,
+            device=x.device,
+        )
+        stream = torch.cuda.current_stream()
+
+        self.linear.forward(x, out, stream.cuda_stream)
+        out = torch.from_dlpack(out)
+        if bias is not None:
+            out.add_(bias)
+
+        return out.view(out_shape)

python/sglang/srt/layers/quantization/turbomind_utils.py

+# Copyright (c) OpenMMLab. All rights reserved.
+
+from typing import List
+
+import torch
+
+from sglang.srt.utils import get_device_capability
+
+
+def get_u4_slices(x: torch.Tensor, dtype: torch.dtype) -> List[torch.Tensor]:
+    assert x.dtype == torch.int32
+    xs = []
+    for _ in range(8):
+        xs.append((x & 15).to(dtype))
+        x = x >> 4
+    return xs
+
+
+def unpack_awq_gemm(x: torch.Tensor) -> torch.Tensor:
+    """
+    The int4 weights are packed into int32:
+    bit:       31-28  27-24  23-20  19-16  15-12  11-8   7-4    3-0
+    weight:    int4_1 int4_2 int4_3 int4_4 int4_5 int4_6 int4_7 int4_8
+    """
+    xs = get_u4_slices(x, torch.uint8)
+    order = [0, 4, 1, 5, 2, 6, 3, 7]
+    ys = [xs[i] for i in order]
+    return torch.stack(ys, dim=-1).view(*x.shape[:-1], -1)
+
+
+def pack_u4_row(x: torch.Tensor) -> torch.Tensor:
+    assert x.dtype == torch.uint8
+    xs = x.view(*x.shape[:-1], -1, 8).split(1, dim=-1)
+    a = torch.zeros(xs[0].shape, dtype=torch.int32, device=x.device)
+    for t in reversed(xs):
+        a = (a << 4) | t
+    return a.squeeze(dim=-1)
+
+
+def verify_turbomind_supported(quant_bit: int, group_size: int) -> bool:
+
+    if quant_bit not in [4]:
+        raise NotImplementedError(
+            f"[Tubomind] Only 4-bit is supported for now, but got {quant_bit} bit"
+        )
+    if group_size != 128:
+        raise NotImplementedError(
+            f"[Tubomind] Only group_size 128 is supported for now, "
+            f"but got group_size {group_size}"
+        )
+
+    major, minor = get_device_capability()
+    capability = major * 10 + minor
+    if capability < 70:
+        raise NotImplementedError(
+            f"[Tubomind] Only capability >= 70 is supported for now, but got {capability}"
+        )
+
+    return True
+
+
+def is_layer_skipped_awq(prefix: str, modules_to_not_convert: List[str]):
+    return any(module_name in prefix for module_name in modules_to_not_convert)

python/sglang/srt/server_args.py

@@ -375,6 +375,7 @@ class ServerArgs:
                 "marlin",
                 "gptq_marlin",
                 "awq_marlin",
+                "awq_turbomind",
                 "bitsandbytes",
                 "gguf",
                 "modelopt",

test/srt/test_turbomind_awq.py

+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.run_eval import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_MODEL_NAME_FOR_TEST,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_server,
+)
+
+
+class TestMLA(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEFAULT_MODEL_NAME_FOR_TEST
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=[
+                "--quantization",
+                "awq_turbomind",
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_mgsm_en(self):
+        args = SimpleNamespace(
+            base_url=self.base_url,
+            model=self.model,
+            eval_name="mgsm_en",
+            num_examples=None,
+            num_threads=1024,
+        )
+
+        metrics = run_eval(args)
+        assert metrics["score"] >= 0.5
+
+
+if __name__ == "__main__":
+    unittest.main()