[Bugfix] Fix triton import with local TritonPlaceholder (#17446)

Signed-off-by: Mengqing Cao <cmq0113@163.com>

vllm/model_executor/layers/mamba/ops/ssd_combined.py

@@ -6,10 +6,11 @@
 # ruff: noqa: E501
 
 import torch
-import triton
 from einops import rearrange
 from packaging import version
 
+from vllm.triton_utils import triton
+
 from .ssd_bmm import _bmm_chunk_fwd
 from .ssd_chunk_scan import _chunk_scan_fwd
 from .ssd_chunk_state import (_chunk_cumsum_fwd, _chunk_state_fwd,

vllm/model_executor/layers/mamba/ops/ssd_state_passing.py

@@ -6,8 +6,8 @@
 # ruff: noqa: E501
 
 import torch
-import triton
-import triton.language as tl
+
+from vllm.triton_utils import tl, triton
 
 
 @triton.autotune(

vllm/model_executor/layers/quantization/awq_triton.py

 # SPDX-License-Identifier: Apache-2.0
 
 import torch
-import triton
-import triton.language as tl
+
+from vllm.triton_utils import tl, triton
 
 AWQ_TRITON_SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128]
 

vllm/model_executor/layers/quantization/compressed_tensors/triton_scaled_mm.py

@@ -3,8 +3,8 @@
 from typing import Optional, Type
 
 import torch
-import triton
-import triton.language as tl
+
+from vllm.triton_utils import tl, triton
 
 
 def is_weak_contiguous(x: torch.Tensor):

vllm/model_executor/layers/quantization/utils/fp8_utils.py

@@ -7,8 +7,6 @@ import os
 from typing import Any, Dict, List, Optional, Tuple, Union
 
 import torch
-import triton
-import triton.language as tl
 
 from vllm import _custom_ops as ops
 from vllm.logger import init_logger
@@ -17,6 +15,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
 from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
     CUTLASS_BLOCK_FP8_SUPPORTED)
 from vllm.platforms import current_platform
+from vllm.triton_utils import tl, triton
 from vllm.utils import direct_register_custom_op
 
 logger = init_logger(__name__)

vllm/model_executor/layers/quantization/utils/int8_utils.py

@@ -8,10 +8,9 @@ import os
 from typing import Any, Dict, List, Optional, Tuple
 
 import torch
-import triton
-import triton.language as tl
 
 from vllm.platforms import current_platform
+from vllm.triton_utils import tl, triton
 
 logger = logging.getLogger(__name__)
 

vllm/triton_utils/__init__.py

 # SPDX-License-Identifier: Apache-2.0
 
-from vllm.triton_utils.importing import HAS_TRITON
+from vllm.triton_utils.importing import (HAS_TRITON, TritonLanguagePlaceholder,
+                                         TritonPlaceholder)
 
-__all__ = ["HAS_TRITON"]
+if HAS_TRITON:
+    import triton
+    import triton.language as tl
+else:
+    triton = TritonPlaceholder()
+    tl = TritonLanguagePlaceholder()
+
+__all__ = ["HAS_TRITON", "triton", "tl"]

vllm/triton_utils/importing.py

@@ -16,7 +16,8 @@ if not HAS_TRITON:
     logger.info("Triton not installed or not compatible; certain GPU-related"
                 " functions will not be available.")
 
-    class TritonPlaceholder(types.ModuleType):
+
+class TritonPlaceholder(types.ModuleType):
 
     def __init__(self):
         super().__init__("triton")
@@ -27,18 +28,19 @@ if not HAS_TRITON:
         logger.warning_once(
             "Triton is not installed. Using dummy decorators. "
             "Install it via `pip install triton` to enable kernel"
-                "compilation.")
+            " compilation.")
 
     def _dummy_decorator(self, name):
 
-            def decorator(func=None, **kwargs):
-                if func is None:
+        def decorator(*args, **kwargs):
+            if args and callable(args[0]):
+                return args[0]
             return lambda f: f
-                return func
 
         return decorator
 
-    class TritonLanguagePlaceholder(types.ModuleType):
+
+class TritonLanguagePlaceholder(types.ModuleType):
 
     def __init__(self):
         super().__init__("triton.language")

vllm/v1/sample/rejection_sampler.py

@@ -3,10 +3,9 @@ from typing import Optional
 
 import torch
 import torch.nn as nn
-import triton
-import triton.language as tl
 
 from vllm.logger import init_logger
+from vllm.triton_utils import tl, triton
 from vllm.v1.sample.metadata import SamplingMetadata
 from vllm.v1.sample.ops.topk_topp_sampler import apply_top_k_top_p
 from vllm.v1.spec_decode.metadata import SpecDecodeMetadata

vllm/v1/spec_decode/eagle.py

 # SPDX-License-Identifier: Apache-2.0
 import torch
 import torch.nn as nn
-import triton
-import triton.language as tl
 
 from vllm.config import CompilationLevel, VllmConfig, set_current_vllm_config
 from vllm.forward_context import set_forward_context
@@ -11,6 +9,7 @@ from vllm.model_executor.model_loader.loader import get_model_loader
 from vllm.model_executor.model_loader.utils import set_default_torch_dtype
 from vllm.model_executor.models import ModelRegistry
 from vllm.model_executor.models.llama_eagle3 import Eagle3LlamaForCausalLM
+from vllm.triton_utils import tl, triton
 from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
 from vllm.v1.sample.metadata import SamplingMetadata