@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / d848800e88
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 16, 03:15 UTC

[Misc] Move `print_*_once` from utils to logger (#11298)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk> Signed-off-by: Maxime Fournioux <55544262+mfournioux@users.noreply.github.com> Co-authored-by: Maxime Fournioux <55544262+mfournioux@users.noreply.github.com>

Cyrus Leung committed Jan 9, 2025 at 12:48 UTC d848800e884f581eeed9f154d6c2aeb38eac24de
21 files changed +129 -72
.github/workflows/lint-and-deploy.yaml
+1
@@ -64,6 +64,7 @@ jobs:
64 run: |
65 export AWS_ACCESS_KEY_ID=minioadmin
66 export AWS_SECRET_ACCESS_KEY=minioadmin
67 + sleep 30 && kubectl -n ns-vllm logs -f "$(kubectl -n ns-vllm get pods | awk '/deployment/ {print $1;exit}')" &
68 helm install --wait --wait-for-jobs --timeout 5m0s --debug --create-namespace --namespace=ns-vllm test-vllm examples/online_serving/chart-helm -f examples/online_serving/chart-helm/values.yaml --set secrets.s3endpoint=http://minio:9000 --set secrets.s3bucketname=testbucket --set secrets.s3accesskeyid=$AWS_ACCESS_KEY_ID --set secrets.s3accesskey=$AWS_SECRET_ACCESS_KEY --set resources.requests.cpu=1 --set resources.requests.memory=4Gi --set resources.limits.cpu=2 --set resources.limits.memory=5Gi --set image.env[0].name=VLLM_CPU_KVCACHE_SPACE --set image.env[1].name=VLLM_LOGGING_LEVEL --set-string image.env[0].value="1" --set-string image.env[1].value="DEBUG" --set-string extraInit.s3modelpath="opt-125m/" --set-string 'resources.limits.nvidia\.com/gpu=0' --set-string 'resources.requests.nvidia\.com/gpu=0' --set-string image.repository="vllm-cpu-env"
69
70 - name: curl test
vllm/attention/backends/torch_sdpa.py
+6 -3
@@ -13,9 +13,12 @@ from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
13 from vllm.attention.backends.utils import CommonAttentionState
14 from vllm.attention.ops.ipex_attn import PagedAttention
15 from vllm.attention.ops.paged_attn import PagedAttentionMetadata
16 -from vllm.utils import make_tensor_with_pad, print_warning_once
16 +from vllm.logger import init_logger
17 +from vllm.utils import make_tensor_with_pad
18 from vllm.worker.cpu_model_runner import ModelInputForCPUBuilder
19
20 +logger = init_logger(__name__)
21 +
22
23 class TorchSDPABackend(AttentionBackend):
24
@@ -396,8 +399,8 @@ class TorchSDPABackendImpl(AttentionImpl[TorchSDPAMetadata]):
399 raise ValueError(
400 "Torch SPDA does not support block-sparse attention.")
401 if logits_soft_cap is not None:
399 - print_warning_once("Torch SPDA does not support logits soft cap. "
400 - "Outputs may be slightly off.")
402 + logger.warning_once("Torch SPDA does not support logits soft cap. "
403 + "Outputs may be slightly off.")
404 self.num_heads = num_heads
405 self.head_size = head_size
406 self.scale = float(scale)
vllm/attention/backends/xformers.py
+5 -3
@@ -17,7 +17,9 @@ from vllm.attention.backends.utils import (
17 is_all_cross_attn_metadata_set, is_all_encoder_attn_metadata_set)
18 from vllm.attention.ops.paged_attn import (PagedAttention,
19 PagedAttentionMetadata)
20 -from vllm.utils import print_warning_once
20 +from vllm.logger import init_logger
21 +
22 +logger = init_logger(__name__)
23
24
25 class XFormersBackend(AttentionBackend):
@@ -385,8 +387,8 @@ class XFormersImpl(AttentionImpl[XFormersMetadata]):
387 raise ValueError(
388 "XFormers does not support block-sparse attention.")
389 if logits_soft_cap is not None:
388 - print_warning_once("XFormers does not support logits soft cap. "
389 - "Outputs may be slightly off.")
390 + logger.warning_once("XFormers does not support logits soft cap. "
391 + "Outputs may be slightly off.")
392 self.num_heads = num_heads
393 self.head_size = head_size
394 self.scale = float(scale)
vllm/config.py
+4 -5
@@ -32,8 +32,7 @@ from vllm.transformers_utils.config import (
32 from vllm.transformers_utils.s3_utils import S3Model
33 from vllm.transformers_utils.utils import is_s3
34 from vllm.utils import (GiB_bytes, LayerBlockType, cuda_device_count_stateless,
35 - get_cpu_memory, print_warning_once, random_uuid,
36 - resolve_obj_by_qualname)
35 + get_cpu_memory, random_uuid, resolve_obj_by_qualname)
36
37 if TYPE_CHECKING:
38 from ray.util.placement_group import PlacementGroup
@@ -314,7 +313,7 @@ class ModelConfig:
313 sliding_window_len_min = get_min_sliding_window(
314 self.hf_text_config.sliding_window)
315
317 - print_warning_once(
316 + logger.warning_once(
317 f"{self.hf_text_config.model_type} has interleaved "
318 "attention, which is currently not supported by the "
319 "XFORMERS backend. Disabling sliding window and capping "
@@ -2758,7 +2757,7 @@ class CompilationConfig(BaseModel):
2757
2758 def model_post_init(self, __context: Any) -> None:
2759 if not self.enable_reshape and self.enable_fusion:
2761 - print_warning_once(
2760 + logger.warning_once(
2761 "Fusion enabled but reshape elimination disabled."
2762 "RMSNorm + quant (fp8) fusion might not work")
2763
@@ -3151,7 +3150,7 @@ class VllmConfig:
3150 self.scheduler_config.chunked_prefill_enabled and \
3151 self.model_config.dtype == torch.float32 and \
3152 current_platform.get_device_capability() == (7, 5):
3154 - print_warning_once(
3153 + logger.warning_once(
3154 "Turing devices tensor cores do not support float32 matmul. "
3155 "To workaround this limitation, vLLM will set 'ieee' input "
3156 "precision for chunked prefill triton kernels.")
vllm/entrypoints/chat_utils.py
+3 -4
@@ -35,7 +35,6 @@ from vllm.logger import init_logger
35 from vllm.multimodal import MultiModalDataDict
36 from vllm.multimodal.utils import MediaConnector
37 from vllm.transformers_utils.tokenizer import AnyTokenizer, MistralTokenizer
38 -from vllm.utils import print_warning_once
38
39 logger = init_logger(__name__)
40
@@ -985,14 +984,14 @@ def apply_mistral_chat_template(
984 **kwargs: Any,
985 ) -> List[int]:
986 if chat_template is not None:
988 - print_warning_once(
987 + logger.warning_once(
988 "'chat_template' cannot be overridden for mistral tokenizer.")
989 if "add_generation_prompt" in kwargs:
991 - print_warning_once(
990 + logger.warning_once(
991 "'add_generation_prompt' is not supported for mistral tokenizer, "
992 "so it will be ignored.")
993 if "continue_final_message" in kwargs:
995 - print_warning_once(
994 + logger.warning_once(
995 "'continue_final_message' is not supported for mistral tokenizer, "
996 "so it will be ignored.")
997
vllm/inputs/preprocess.py
+11 -9
@@ -10,7 +10,6 @@ from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalRegistry
10 from vllm.multimodal.inputs import MultiModalDataDict, MultiModalInputsV2
11 from vllm.prompt_adapter.request import PromptAdapterRequest
12 from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup
13 -from vllm.utils import print_info_once, print_warning_once
13
14 from .data import (DecoderOnlyInputs, EncoderDecoderInputs, ProcessorInputs,
15 PromptType, SingletonInputs, SingletonPrompt, token_inputs)
@@ -68,21 +67,24 @@ class InputPreprocessor:
67 '''
68
69 if not self.model_config.is_encoder_decoder:
71 - print_warning_once("Using None for decoder start token id because "
72 - "this is not an encoder/decoder model.")
70 + logger.warning_once(
71 + "Using None for decoder start token id because "
72 + "this is not an encoder/decoder model.")
73 return None
74
75 if (self.model_config is None or self.model_config.hf_config is None):
76 - print_warning_once("Using None for decoder start token id because "
77 - "model config is not available.")
76 + logger.warning_once(
77 + "Using None for decoder start token id because "
78 + "model config is not available.")
79 return None
80
81 dec_start_token_id = getattr(self.model_config.hf_config,
82 'decoder_start_token_id', None)
83 if dec_start_token_id is None:
83 - print_warning_once("Falling back on <BOS> for decoder start token "
84 - "id because decoder start token id is not "
85 - "available.")
84 + logger.warning_once(
85 + "Falling back on <BOS> for decoder start token "
86 + "id because decoder start token id is not "
87 + "available.")
88 dec_start_token_id = self.get_bos_token_id()
89
90 return dec_start_token_id
@@ -231,7 +233,7 @@ class InputPreprocessor:
233 # updated to use the new multi-modal processor
234 can_process_multimodal = self.mm_registry.has_processor(model_config)
235 if not can_process_multimodal:
234 - print_info_once(
236 + logger.info_once(
237 "Your model uses the legacy input pipeline instead of the new "
238 "multi-modal processor. Please note that the legacy pipeline "
239 "will be removed in a future release. For more details, see: "
vllm/inputs/registry.py
+2 -2
@@ -12,7 +12,7 @@ from vllm.logger import init_logger
12 from vllm.transformers_utils.processor import cached_get_processor
13 from vllm.transformers_utils.tokenizer import AnyTokenizer
14 from vllm.utils import (ClassRegistry, get_allowed_kwarg_only_overrides,
15 - print_warning_once, resolve_mm_processor_kwargs)
15 + resolve_mm_processor_kwargs)
16
17 from .data import ProcessorInputs, SingletonInputs
18 from .parse import is_encoder_decoder_inputs
@@ -352,7 +352,7 @@ class InputRegistry:
352 num_tokens = dummy_data.seq_data.prompt_token_ids
353 if len(num_tokens) < seq_len:
354 if is_encoder_data:
355 - print_warning_once(
355 + logger.warning_once(
356 f"Expected at least {seq_len} dummy encoder tokens for "
357 f"profiling, but found {len(num_tokens)} tokens instead.")
358 else:
vllm/logger.py
+52 -5
@@ -4,11 +4,12 @@ import json
4 import logging
5 import os
6 import sys
7 -from functools import partial
7 +from functools import lru_cache, partial
8 from logging import Logger
9 from logging.config import dictConfig
10 from os import path
11 -from typing import Dict, Optional
11 +from types import MethodType
12 +from typing import Any, Optional, cast
13
14 import vllm.envs as envs
15
@@ -49,8 +50,44 @@ DEFAULT_LOGGING_CONFIG = {
50 }
51
52
53 +@lru_cache
54 +def _print_info_once(logger: Logger, msg: str) -> None:
55 + # Set the stacklevel to 2 to print the original caller's line info
56 + logger.info(msg, stacklevel=2)
57 +
58 +
59 +@lru_cache
60 +def _print_warning_once(logger: Logger, msg: str) -> None:
61 + # Set the stacklevel to 2 to print the original caller's line info
62 + logger.warning(msg, stacklevel=2)
63 +
64 +
65 +class _VllmLogger(Logger):
66 + """
67 + Note:
68 + This class is just to provide type information.
69 + We actually patch the methods directly on the :class:`logging.Logger`
70 + instance to avoid conflicting with other libraries such as
71 + `intel_extension_for_pytorch.utils._logger`.
72 + """
73 +
74 + def info_once(self, msg: str) -> None:
75 + """
76 + As :meth:`info`, but subsequent calls with the same message
77 + are silently dropped.
78 + """
79 + _print_info_once(self, msg)
80 +
81 + def warning_once(self, msg: str) -> None:
82 + """
83 + As :meth:`warning`, but subsequent calls with the same message
84 + are silently dropped.
85 + """
86 + _print_warning_once(self, msg)
87 +
88 +
89 def _configure_vllm_root_logger() -> None:
53 - logging_config: Dict = {}
90 + logging_config = dict[str, Any]()
91
92 if not VLLM_CONFIGURE_LOGGING and VLLM_LOGGING_CONFIG_PATH:
93 raise RuntimeError(
@@ -84,12 +121,22 @@ def _configure_vllm_root_logger() -> None:
121 dictConfig(logging_config)
122
123
87 -def init_logger(name: str) -> Logger:
124 +def init_logger(name: str) -> _VllmLogger:
125 """The main purpose of this function is to ensure that loggers are
126 retrieved in such a way that we can be sure the root vllm logger has
127 already been configured."""
128
92 - return logging.getLogger(name)
129 + logger = logging.getLogger(name)
130 +
131 + methods_to_patch = {
132 + "info_once": _print_info_once,
133 + "warning_once": _print_warning_once,
134 + }
135 +
136 + for method_name, method in methods_to_patch.items():
137 + setattr(logger, method_name, MethodType(method, logger))
138 +
139 + return cast(_VllmLogger, logger)
140
141
142 # The root logger is initialized when the module is imported.
vllm/lora/peft_helper.py
+4 -2
@@ -4,7 +4,9 @@ import math
4 from dataclasses import MISSING, dataclass, field, fields
5 from typing import Literal, Optional, Union
6
7 -from vllm.utils import print_info_once
7 +from vllm.logger import init_logger
8 +
9 +logger = init_logger(__name__)
10
11
12 @dataclass
@@ -42,7 +44,7 @@ class PEFTHelper:
44 def __post_init__(self):
45 self._validate_features()
46 if self.use_rslora:
45 - print_info_once("Loading LoRA weights trained with rsLoRA.")
47 + logger.info_once("Loading LoRA weights trained with rsLoRA.")
48 self.vllm_lora_scaling_factor = self.lora_alpha / math.sqrt(self.r)
49 else:
50 self.vllm_lora_scaling_factor = self.lora_alpha / self.r
vllm/lora/punica_wrapper/punica_selector.py
+5 -3
@@ -1,19 +1,21 @@
1 +from vllm.logger import init_logger
2 from vllm.platforms import current_platform
2 -from vllm.utils import print_info_once
3
4 from .punica_base import PunicaWrapperBase
5
6 +logger = init_logger(__name__)
7 +
8
9 def get_punica_wrapper(*args, **kwargs) -> PunicaWrapperBase:
10 if current_platform.is_cuda_alike():
11 # Lazy import to avoid ImportError
12 from vllm.lora.punica_wrapper.punica_gpu import PunicaWrapperGPU
11 - print_info_once("Using PunicaWrapperGPU.")
13 + logger.info_once("Using PunicaWrapperGPU.")
14 return PunicaWrapperGPU(*args, **kwargs)
15 elif current_platform.is_hpu():
16 # Lazy import to avoid ImportError
17 from vllm.lora.punica_wrapper.punica_hpu import PunicaWrapperHPU
16 - print_info_once("Using PunicaWrapperHPU.")
18 + logger.info_once("Using PunicaWrapperHPU.")
19 return PunicaWrapperHPU(*args, **kwargs)
20 else:
21 raise NotImplementedError
vllm/model_executor/custom_op.py
+1 -2
@@ -5,7 +5,6 @@ import torch.nn as nn
5 from vllm.config import get_current_vllm_config
6 from vllm.logger import init_logger
7 from vllm.platforms import current_platform
8 -from vllm.utils import print_warning_once
8
9 logger = init_logger(__name__)
10
@@ -91,7 +90,7 @@ class CustomOp(nn.Module):
90 compilation_config = get_current_vllm_config().compilation_config
91 custom_ops = compilation_config.custom_ops
92 if not hasattr(cls, "name"):
94 - print_warning_once(
93 + logger.warning_once(
94 f"Custom op {cls.__name__} was not registered, "
95 f"which means it won't appear in the op registry. "
96 f"It will be enabled/disabled based on the global settings.")
vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
+5 -3
@@ -8,6 +8,7 @@ from compressed_tensors.quantization import QuantizationStrategy
8
9 import vllm.model_executor.layers.fused_moe # noqa
10 from vllm import _custom_ops as ops
11 +from vllm.logger import init_logger
12 from vllm.model_executor.layers.fused_moe import (FusedMoE, FusedMoEMethodBase,
13 FusedMoeWeightScaleSupported)
14 from vllm.model_executor.layers.quantization.compressed_tensors.schemes import (
@@ -16,7 +17,8 @@ from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
17 all_close_1d, normalize_e4m3fn_to_e4m3fnuz, per_tensor_dequantize)
18 from vllm.model_executor.utils import set_weight_attrs
19 from vllm.platforms import current_platform
19 -from vllm.utils import print_warning_once
20 +
21 +logger = init_logger(__name__)
22
23
24 class GPTQMarlinState(Enum):
@@ -142,10 +144,10 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
144 "activation scales are None.")
145 if (not all_close_1d(layer.w13_input_scale)
146 or not all_close_1d(layer.w2_input_scale)):
145 - print_warning_once(
147 + logger.warning_once(
148 "Found input_scales that are not equal for "
149 "fp8 MoE layer. Using the maximum across experts "
148 - "for each layer. ")
150 + "for each layer.")
151 layer.w13_input_scale = torch.nn.Parameter(
152 layer.w13_input_scale.max(), requires_grad=False)
153 layer.w2_input_scale = torch.nn.Parameter(
vllm/model_executor/layers/quantization/fp8.py
+2 -3
@@ -28,7 +28,6 @@ from vllm.model_executor.parameter import (BlockQuantScaleParameter,
28 PerTensorScaleParameter)
29 from vllm.model_executor.utils import set_weight_attrs
30 from vllm.platforms import current_platform
31 -from vllm.utils import print_warning_once
31
32 ACTIVATION_SCHEMES = ["static", "dynamic"]
33
@@ -539,10 +538,10 @@ class Fp8MoEMethod(FusedMoEMethodBase):
538 "activation scales are None.")
539 if (not all_close_1d(layer.w13_input_scale)
540 or not all_close_1d(layer.w2_input_scale)):
542 - print_warning_once(
541 + logger.warning_once(
542 "Found input_scales that are not equal for "
543 "fp8 MoE layer. Using the maximum across experts "
545 - "for each layer. ")
544 + "for each layer.")
545 layer.w13_input_scale = torch.nn.Parameter(
546 layer.w13_input_scale.max(), requires_grad=False)
547 layer.w2_input_scale = torch.nn.Parameter(
vllm/model_executor/layers/quantization/kv_cache.py
+4 -2
@@ -1,8 +1,10 @@
1 import torch
2
3 +from vllm.logger import init_logger
4 from vllm.model_executor.layers.quantization.base_config import (
5 QuantizationConfig, QuantizeMethodBase)
5 -from vllm.utils import print_warning_once
6 +
7 +logger = init_logger(__name__)
8
9
10 class BaseKVCacheMethod(QuantizeMethodBase):
@@ -67,7 +69,7 @@ class BaseKVCacheMethod(QuantizeMethodBase):
69 layer._v_scale = v_scale
70 if (layer._k_scale == 1.0 and layer._v_scale == 1.0
71 and "e5m2" not in layer.kv_cache_dtype):
70 - print_warning_once(
72 + logger.warning_once(
73 "Using KV cache scaling factor 1.0 for fp8_e4m3. This "
74 "may cause accuracy issues. Please make sure k/v_scale "
75 "scaling factors are available in the fp8 checkpoint.")
vllm/model_executor/layers/quantization/utils/marlin_utils_fp8.py
+4 -2
@@ -3,11 +3,13 @@ from typing import Optional
3 import torch
4
5 import vllm._custom_ops as ops
6 +from vllm.logger import init_logger
7 from vllm.platforms import current_platform
7 -from vllm.utils import print_warning_once
8
9 from .marlin_utils import marlin_make_workspace, marlin_permute_scales
10
11 +logger = init_logger(__name__)
12 +
13
14 def is_fp8_marlin_supported():
15 return current_platform.has_device_capability(80)
@@ -47,7 +49,7 @@ def apply_fp8_marlin_linear(
49
50 def prepare_fp8_layer_for_marlin(layer: torch.nn.Module,
51 strategy: str = "tensor") -> None:
50 - print_warning_once(
52 + logger.warning_once(
53 "Your GPU does not have native support for FP8 computation but "
54 "FP8 quantization is being used. Weight-only FP8 compression will "
55 "be used leveraging the Marlin kernel. This may degrade "
vllm/model_executor/model_loader/weight_utils.py
+4 -4
@@ -25,7 +25,7 @@ from vllm.model_executor.layers.quantization import (QuantizationConfig,
25 get_quantization_config)
26 from vllm.model_executor.layers.quantization.schema import QuantParamSchema
27 from vllm.platforms import current_platform
28 -from vllm.utils import PlaceholderModule, print_warning_once
28 +from vllm.utils import PlaceholderModule
29
30 try:
31 from runai_model_streamer import SafetensorsStreamer
@@ -673,7 +673,7 @@ def maybe_remap_kv_scale_name(name: str, params_dict: dict) -> Optional[str]:
673 None: If the remapped name is not found in params_dict.
674 """
675 if name.endswith(".kv_scale"):
676 - print_warning_once(
676 + logger.warning_once(
677 "DEPRECATED. Found kv_scale in the checkpoint. "
678 "This format is deprecated in favor of separate k_scale and "
679 "v_scale tensors and will be removed in a future release. "
@@ -682,7 +682,7 @@ def maybe_remap_kv_scale_name(name: str, params_dict: dict) -> Optional[str]:
682 # NOTE: we remap the deprecated kv_scale to k_scale
683 remapped_name = name.replace(".kv_scale", ".attn.k_scale")
684 if remapped_name not in params_dict:
685 - print_warning_once(
685 + logger.warning_once(
686 f"Found kv_scale in the checkpoint (e.g. {name}), "
687 "but not found the expected name in the model "
688 f"(e.g. {remapped_name}). kv_scale is "
@@ -695,7 +695,7 @@ def maybe_remap_kv_scale_name(name: str, params_dict: dict) -> Optional[str]:
695 if name.endswith(scale_name):
696 remapped_name = name.replace(scale_name, f".attn{scale_name}")
697 if remapped_name not in params_dict:
698 - print_warning_once(
698 + logger.warning_once(
699 f"Found {scale_name} in the checkpoint (e.g. {name}), "
700 "but not found the expected name in the model "
701 f"(e.g. {remapped_name}). {scale_name} is "
vllm/model_executor/models/chameleon.py
+4 -2
@@ -11,6 +11,7 @@ from transformers import (BatchFeature, ChameleonConfig, ChameleonProcessor,
11 from vllm.attention import Attention, AttentionMetadata
12 from vllm.config import CacheConfig, VllmConfig
13 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
14 +from vllm.logger import init_logger
15 from vllm.model_executor.layers.activation import SiluAndMul
16 from vllm.model_executor.layers.layernorm import RMSNorm
17 from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
@@ -35,13 +36,14 @@ from vllm.multimodal.processing import (BaseMultiModalProcessor,
36 BaseProcessingInfo, PromptReplacement)
37 from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
38 from vllm.sequence import IntermediateTensors
38 -from vllm.utils import print_warning_once
39
40 from .interfaces import SupportsMultiModal, SupportsPP
41 from .utils import (is_pp_missing_parameter,
42 make_empty_intermediate_tensors_factory, make_layers,
43 maybe_prefix, merge_multimodal_embeddings)
44
45 +logger = init_logger(__name__)
46 +
47
48 class ChameleonImagePixelInputs(TypedDict):
49 type: Literal["pixel_values"]
@@ -1111,7 +1113,7 @@ class ChameleonForConditionalGeneration(nn.Module, SupportsMultiModal,
1113 remapped_kv_scale_name = name.replace(
1114 ".kv_scale", ".attn.kv_scale")
1115 if remapped_kv_scale_name not in params_dict:
1114 - print_warning_once(
1116 + logger.warning_once(
1117 "Found kv scale in the checkpoint (e.g. "
1118 f"{name}), but not found the expected name in "
1119 f"the model (e.g. {remapped_kv_scale_name}). "
vllm/model_executor/models/olmoe.py
+4 -2
@@ -20,6 +20,7 @@ from vllm.attention import Attention, AttentionMetadata
20 from vllm.compilation.decorators import support_torch_compile
21 from vllm.config import CacheConfig, VllmConfig
22 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
23 +from vllm.logger import init_logger
24 from vllm.model_executor.layers.fused_moe import FusedMoE
25 from vllm.model_executor.layers.layernorm import RMSNorm
26 from vllm.model_executor.layers.linear import (QKVParallelLinear,
@@ -34,13 +35,14 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
35 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
36 from vllm.model_executor.sampling_metadata import SamplingMetadata
37 from vllm.sequence import IntermediateTensors
37 -from vllm.utils import print_warning_once
38
39 from .interfaces import SupportsPP
40 from .utils import (is_pp_missing_parameter,
41 make_empty_intermediate_tensors_factory, make_layers,
42 maybe_prefix)
43
44 +logger = init_logger(__name__)
45 +
46
47 class OlmoeMoE(nn.Module):
48 """A tensor-parallel MoE implementation for Olmoe that shards each expert
@@ -446,7 +448,7 @@ class OlmoeForCausalLM(nn.Module, SupportsPP):
448 remapped_kv_scale_name = name.replace(
449 ".kv_scale", ".attn.kv_scale")
450 if remapped_kv_scale_name not in params_dict:
449 - print_warning_once(
451 + logger.warning_once(
452 "Found kv scale in the checkpoint "
453 f"(e.g. {name}), but not found the expected "
454 f"name in the model "
vllm/model_executor/models/qwen2_moe.py
+4 -2
@@ -34,6 +34,7 @@ from vllm.config import CacheConfig, VllmConfig
34 from vllm.distributed import (get_pp_group,
35 get_tensor_model_parallel_world_size,
36 tensor_model_parallel_all_reduce)
37 +from vllm.logger import init_logger
38 from vllm.model_executor.layers.activation import SiluAndMul
39 from vllm.model_executor.layers.fused_moe import FusedMoE
40 from vllm.model_executor.layers.layernorm import RMSNorm
@@ -50,13 +51,14 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
51 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
52 from vllm.model_executor.sampling_metadata import SamplingMetadata
53 from vllm.sequence import IntermediateTensors
53 -from vllm.utils import print_warning_once
54
55 from .interfaces import SupportsPP
56 from .utils import (extract_layer_index, is_pp_missing_parameter,
57 make_empty_intermediate_tensors_factory, make_layers,
58 maybe_prefix)
59
60 +logger = init_logger(__name__)
61 +
62
63 class Qwen2MoeMLP(nn.Module):
64
@@ -524,7 +526,7 @@ class Qwen2MoeForCausalLM(nn.Module, SupportsPP):
526 remapped_kv_scale_name = name.replace(
527 ".kv_scale", ".attn.kv_scale")
528 if remapped_kv_scale_name not in params_dict:
527 - print_warning_once(
529 + logger.warning_once(
530 "Found kv scale in the checkpoint "
531 f"(e.g. {name}), but not found the expected "
532 f"name in the model "
vllm/model_executor/models/vision.py
+4 -2
@@ -7,8 +7,10 @@ from transformers import PretrainedConfig
7 import vllm.envs as envs
8 from vllm.attention.selector import (backend_name_to_enum,
9 get_global_forced_attn_backend)
10 +from vllm.logger import init_logger
11 from vllm.platforms import _Backend, current_platform
11 -from vllm.utils import print_warning_once
12 +
13 +logger = init_logger(__name__)
14
15 _C = TypeVar("_C", bound=PretrainedConfig)
16
@@ -87,7 +89,7 @@ def get_vit_attn_backend(support_fa: bool = False) -> _Backend:
89 if is_flash_attn_2_available():
90 selected_backend = _Backend.FLASH_ATTN
91 else:
90 - print_warning_once(
92 + logger.warning_once(
93 "Current `vllm-flash-attn` has a bug inside vision module, "
94 "so we use xformers backend instead. You can run "
95 "`pip install flash-attn` to use flash-attention backend.")
vllm/utils.py
-12
@@ -696,18 +696,6 @@ def create_kv_caches_with_random(
696 return key_caches, value_caches
697
698
699 -@lru_cache
700 -def print_info_once(msg: str) -> None:
701 - # Set the stacklevel to 2 to print the caller's line info
702 - logger.info(msg, stacklevel=2)
703 -
704 -
705 -@lru_cache
706 -def print_warning_once(msg: str) -> None:
707 - # Set the stacklevel to 2 to print the caller's line info
708 - logger.warning(msg, stacklevel=2)
709 -
710 -
699 @lru_cache(maxsize=None)
700 def is_pin_memory_available() -> bool:
701 from vllm.platforms import current_platform