[Bugfix] Add proper comparison for package versions (#22314)

Signed-off-by: Syed Muhammad Bin Asif <syedmba7@connect.hku.hk>

benchmarks/kernels/benchmark_bitblas.py

@@ -3,6 +3,8 @@
 # Copyright (c) Microsoft Corporation.
 # Licensed under the MIT License.
 
+from packaging import version
+
 from vllm.model_executor.layers.quantization.utils.bitblas_utils import (
     MINIMUM_BITBLAS_VERSION,
 )
@@ -10,7 +12,7 @@ from vllm.model_executor.layers.quantization.utils.bitblas_utils import (
 try:
     import bitblas
 
-    if bitblas.__version__ < MINIMUM_BITBLAS_VERSION:
+    if version.parse(bitblas.__version__) < version.parse(MINIMUM_BITBLAS_VERSION):
         raise ImportError(
             "bitblas version is wrong. Please "
             f"install bitblas>={MINIMUM_BITBLAS_VERSION}"

docs/design/arch_overview.md

@@ -200,7 +200,8 @@ vision-language model.
                 lora_config = vllm_config.lora_config
                 super().__init__(config, cache_config, quant_config, lora_config, prefix)
 
-        if __version__ >= "0.6.4":
+        from packaging import version
+        if version.parse(__version__) >= version.parse("0.6.4"):
             MyModel = MyNewModel
         else:
             MyModel = MyOldModel

vllm/attention/ops/triton_decode_attention.py

@@ -31,6 +31,8 @@ It supports page size >= 1.
 
 import logging
 
+from packaging import version
+
 from vllm.platforms import current_platform
 from vllm.triton_utils import tl, triton
 
@@ -40,7 +42,7 @@ logger = logging.getLogger(__name__)
 
 # Only print the following warnings when triton version < 3.2.0.
 # The issue won't affect performance or accuracy.
-if triton.__version__ < '3.2.0':
+if version.parse(triton.__version__) < version.parse('3.2.0'):
     logger.warning(
         "The following error message 'operation scheduled before its operands' "
         "can be ignored.")

vllm/model_executor/layers/quantization/bitblas.py

@@ -3,6 +3,7 @@
 from typing import Any, Optional
 
 import torch
+from packaging import version
 
 from vllm.logger import init_logger
 from vllm.model_executor.layers.linear import LinearBase, LinearMethodBase
@@ -45,7 +46,8 @@ class BitBLASConfig(QuantizationConfig):
     ) -> None:
         try:
             import bitblas
-            if bitblas.__version__ < MINIMUM_BITBLAS_VERSION:
+            if version.parse(bitblas.__version__) < version.parse(
+                    MINIMUM_BITBLAS_VERSION):
                 raise ImportError(
                     "bitblas version is wrong. Please "
                     f"install bitblas>={MINIMUM_BITBLAS_VERSION}")

vllm/model_executor/layers/quantization/bitsandbytes.py

@@ -4,6 +4,7 @@
 from typing import Any, Callable, Optional, Union
 
 import torch
+from packaging import version
 
 from vllm.model_executor.layers.fused_moe.layer import (FusedMoE,
                                                         FusedMoEMethodBase)
@@ -169,7 +170,8 @@ class BitsAndBytesLinearMethod(LinearMethodBase):
     def __init__(self, quant_config: BitsAndBytesConfig):
         try:
             import bitsandbytes
-            if bitsandbytes.__version__ < "0.46.1":
+            if version.parse(
+                    bitsandbytes.__version__) < version.parse("0.46.1"):
                 raise ImportError("bitsandbytes version is wrong. Please "
                                   "install bitsandbytes>=0.46.1.")
         except ImportError as err:
@@ -412,7 +414,8 @@ class BitsAndBytesMoEMethod(FusedMoEMethodBase):
     def __init__(self, quant_config: BitsAndBytesConfig):
         try:
             import bitsandbytes
-            if bitsandbytes.__version__ < "0.46.1":
+            if version.parse(
+                    bitsandbytes.__version__) < version.parse("0.46.1"):
                 raise ImportError("bitsandbytes version is wrong. Please "
                                   "install bitsandbytes>=0.46.1.")
         except ImportError as err:

vllm/model_executor/layers/quantization/deepspeedfp.py

@@ -6,6 +6,7 @@ from typing import Any, Optional
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+from packaging import version
 
 from vllm.model_executor.layers.linear import LinearBase, LinearMethodBase
 from vllm.model_executor.layers.quantization import QuantizationMethods
@@ -145,7 +146,7 @@ class DeepSpeedFPParameter(nn.Parameter):
                 quant_config: DeepSpeedFPConfig):
         try:
             import deepspeed
-            if deepspeed.__version__ < "0.14.2":
+            if version.parse(deepspeed.__version__) < version.parse("0.14.2"):
                 raise ImportError("deepspeed version is wrong. Please "
                                   "install deepspeed>=0.14.2.")
             from deepspeed.ops.fp_quantizer import FP_Quantize

vllm/model_executor/layers/quantization/gptq_bitblas.py

@@ -3,6 +3,7 @@
 from typing import Any, Optional
 
 import torch
+from packaging import version
 from torch.nn.parameter import Parameter
 
 from vllm.logger import init_logger
@@ -63,7 +64,8 @@ class GPTQBitBLASConfig(QuantizationConfig):
 
         try:
             import bitblas
-            if bitblas.__version__ < MINIMUM_BITBLAS_VERSION:
+            if version.parse(bitblas.__version__) < version.parse(
+                    MINIMUM_BITBLAS_VERSION):
                 raise ImportError(
                     "bitblas version is wrong. Please "
                     f"install bitblas>={MINIMUM_BITBLAS_VERSION}")

vllm/model_executor/layers/quantization/ipex_quant.py

@@ -4,6 +4,7 @@
 from typing import Any, Optional
 
 import torch
+from packaging import version
 
 from vllm.model_executor.layers.linear import (LinearBase, LinearMethodBase,
                                                UnquantizedLinearMethod)
@@ -135,7 +136,8 @@ class IPEXGPTQLinearMethod(GPTQLinearMethod):
 
         try:
             import intel_extension_for_pytorch as ipex
-            if ipex.__version__ < MIN_IPEX_VERSION:
+            if version.parse(
+                    ipex.__version__) < version.parse(MIN_IPEX_VERSION):
                 raise ImportError(
                     "intel_extension_for_pytorch version is "
                     "wrong. Please install "
@@ -199,7 +201,8 @@ class IPEXAWQLinearMethod(AWQLinearMethod):
 
         try:
             import intel_extension_for_pytorch as ipex
-            if ipex.__version__ < MIN_IPEX_VERSION:
+            if version.parse(
+                    ipex.__version__) < version.parse(MIN_IPEX_VERSION):
                 raise ImportError(
                     "intel_extension_for_pytorch version is "
                     "wrong. Please install "

vllm/model_executor/layers/quantization/kernels/mixed_precision/bitblas.py

@@ -4,6 +4,7 @@
 from typing import Optional
 
 import torch
+from packaging import version
 
 from vllm.logger import init_logger
 from vllm.model_executor.layers.quantization.base_config import (
@@ -110,7 +111,8 @@ class BitBLASLinearKernel(MPLinearKernel):
 
         try:
             import bitblas
-            if bitblas.__version__ < MINIMUM_BITBLAS_VERSION:
+            if version.parse(bitblas.__version__) < version.parse(
+                    MINIMUM_BITBLAS_VERSION):
                 raise ImportError(
                     "bitblas version is wrong. Please "
                     f"install bitblas>={MINIMUM_BITBLAS_VERSION}")

vllm/model_executor/layers/quantization/utils/bitblas_utils.py

@@ -3,6 +3,7 @@
 from typing import Optional
 
 import torch
+from packaging import version
 
 from vllm.platforms import current_platform
 from vllm.scalar_type import ScalarType, scalar_types
@@ -75,7 +76,8 @@ def _check_bitblas_supported(
     # Finally, check if bitblas is installed
     try:
         import bitblas
-        if bitblas.__version__ < MINIMUM_BITBLAS_VERSION:
+        if version.parse(
+                bitblas.__version__) < version.parse(MINIMUM_BITBLAS_VERSION):
             raise ImportError("bitblas version is wrong. Please "
                               f"install bitblas>={MINIMUM_BITBLAS_VERSION}")
     except ImportError:

vllm/model_executor/layers/quantization/utils/w8a8_utils.py

@@ -4,6 +4,7 @@
 from typing import Callable, Optional, Union
 
 import torch
+from packaging import version
 
 from vllm import _custom_ops as ops
 from vllm import envs
@@ -21,8 +22,8 @@ TORCH_DEVICE_IDENTITY = None
 # torch._scaled_mm rowwise feature.
 # The condition is determined once as the operations
 # are time consuming.
-USE_ROWWISE_TORCH_SCALED_MM = (current_platform.is_rocm()
+USE_ROWWISE_TORCH_SCALED_MM = (current_platform.is_rocm() and version.parse(
-                               and torch.__version__[0:3] >= "2.7"
+    torch.__version__) >= version.parse("2.7")
                                and current_platform.has_device_capability(94))
 
 

vllm/model_executor/model_loader/bitsandbytes_loader.py

@@ -12,6 +12,7 @@ from typing import Any, Callable, Optional
 import numpy as np
 import torch
 from huggingface_hub import HfApi
+from packaging import version
 from torch import nn
 from transformers.utils import SAFE_WEIGHTS_INDEX_NAME
 
@@ -193,7 +194,8 @@ class BitsAndBytesModelLoader(BaseModelLoader):
         try:
             import bitsandbytes
 
-            if bitsandbytes.__version__ < "0.46.1":
+            if version.parse(
+                    bitsandbytes.__version__) < version.parse("0.46.1"):
                 raise ImportError("bitsandbytes version is wrong. Please "
                                   "install bitsandbytes>=0.46.1.")
         except ImportError as err:

vllm/v1/sample/ops/topk_topp_sampler.py

@@ -5,6 +5,7 @@ from typing import Optional
 
 import torch
 import torch.nn as nn
+from packaging import version
 
 from vllm import envs
 from vllm.logger import init_logger
@@ -32,7 +33,7 @@ class TopKTopPSampler(nn.Module):
         if current_platform.is_cuda():
             if is_flashinfer_available:
                 flashinfer_version = flashinfer.__version__
-                if flashinfer_version < "0.2.3":
+                if version.parse(flashinfer_version) < version.parse("0.2.3"):
                     logger.warning_once(
                         "FlashInfer version >= 0.2.3 required. "
                         "Falling back to default sampling implementation.")